THE AXIOLOGY OF NECROLOGIES: USING NATURAL LANGUAGE
PROCESSING TO EXAMINE VALUES IN OBITUARIES
by
JACOB G. LEVERNIER
A DISSERTATION
Presented to the Department of Psychology
and the Graduate School of the University of Oregon
in partial fulfillment of the requirements
for the degree of
Doctor of Philosophy
December 2016
DISSERTATION APPROVAL PAGE
Student: Jacob G. Levernier
Title: The Axiology of Necrologies: Using Natural Language Processing to Examine
Values in Obituaries
This dissertation has been accepted and approved in partial fulfillment of the
requirements for the Doctor of Philosophy degree in the Department of Psychology
by:
Gerard Saucier Chair
Mark Alfano Core Member
Caitlin Fausey Core Member
Colin Koopman Institutional Representative
and
Scott L. Pratt Dean of the Graduate School
Original approval signatures are on file with the University of Oregon Graduate
School.
Degree awarded December 2016
ii
c© 2016 Jacob G. Levernier
This work, including text and images of this document but not including
supplemental files (for example, not including software code and data), is licensed
under a Creative Commons
Attribution 4.0 International License.
License terms for the code and other supplemental files written for this dissertation
can be found at http://doi.org/10.7264/D3WC7S.
iii
DISSERTATION ABSTRACT
Jacob G. Levernier
Doctor of Philosophy
Department of Psychology
December 2016
Title: The Axiology of Necrologies: Using Natural Language Processing to Examine
Values in Obituaries
This dissertation is centrally concerned with exploring obituaries as
repositories of values. Obituaries are a publicly-available natural language source
that are variably written for members of communities that are wide (nation-
level) and narrow (city-level, or at the level of specific groups therein). Because
they are explicitly summative, limited in size, and written for consumption by
a public audience, obituaries may be expected to express concisely the aspects
of their subjects’ lives that the authors (often family members living in the same
communities) found most salient or worthy of featuring.
140,599 obituaries nested in 832 newspapers from across the USA were
scraped with permission from Legacy.com, an obituaries publisher. Obituaries
were coded for the age at death and gender (female/male) of the deceased using
automated algorithms. For each publishing newspaper, county-level median income,
educational achievement (operationalized as percent of the population with a
Bachelor’s degree or higher), and race and ethnicity were averaged across counties,
weighting by population size.
A Neo4J graph database was constructed using WordNet and the University
of South Florida Free Association Norms datasets. Each word in each obituary in
iv
the corpus was lemmatized. The shortest path through the WordNet graph from
each lemma to 30 Schwartz value prototype words published by Bardi, Calogero,
and Mullen (2008) was then recorded. From these path lengths, a new measure,
“word-by-hop,” was calculated for each Schwartz value to reflect the relative lexical
distance between each obituary and that Schwartz value.
Of the Schwartz values, Power, Conformity, and Security were most
indicated in the corpus, while Universalism, Hedonism, and Stimulation were least
indicated. A series of seven two-level regression models suggested that, across
Schwartz values, newspaper community accounted for the greatest amount of
word-by-hop variability in the corpus. The best-fitting model indicated a small,
negative effect of female status across Schwartz values. Unexpectedly, Hedonism
and Conformity, which had conceptually opposite prototype words, were highly
correlated, possibly indicating that obituary authors “compensate” for describing
the deceased in a hedonistic way by concurrently emphasizing restraint. Future
research could usefully further expand word-by-hop and incorporate individual-level
covariates that match the newspaper-level covariates used here.
v
CURRICULUM VITAE
NAME OF AUTHOR: Jacob G. Levernier
GRADUATE AND UNDERGRADUATE SCHOOLS ATTENDED:
University of Oregon, Eugene, OR, USA
University of San Francisco, San Francisco, CA, USA
Blackfriars Hall, University of Oxford, Oxford, UK
DEGREES AWARDED:
Doctor of Philosophy, Psychology, 2016, University of Oregon
Master of Science, Psychology, 2012, University of Oregon
Bachelor of Arts, Psychology, 2010, University of San Francisco
AREAS OF SPECIAL INTEREST:
Values
Data mining to understand morals
Data ethics
PROFESSIONAL EXPERIENCE:
Instructor, Data Carpentry, 2016
Graduate Affiliate, Digital Scholarship Center, University of Oregon, 2014-
2016
Workshop Instructor (Archival Book Digitization, Scientific Programming
in R, Scientific Programming with Version Control), Digital Scholarship
Center and Department of Psychology, University of Oregon, 2015-2016
Instructor, Software Carpentry, 2014-2016
Research Data Management Assistant, Science Library, University of
Oregon, 2013-2014
Software Architect for unified participant recruitment tracking system,
Department of Psychology, University of Oregon, 2012-2015
Research Lab Manager, University of Oregon, 2012-2013
vi
Graduate Teaching Fellow (Graduate Statistics, Advanced Data Analysis,
Research Methods, Child Development, Motivation and Emotion,
Imagination, Mnemotechnics), Department of Psychology, University
of Oregon, 2011-2016
Graduate Research Fellow, University of Oregon, 2012-2013
Staff Research Associate, NCIRE - The Veterans’ Health Research Institute
/ San Francisco Veterans Affairs Medical Center, 2010-2011
Research Assistant, University of San Francisco, 2009-2010
Research Assistant, Circle of Security, Spokane, WA, 2008
Research Assistant, University of San Francisco / Numerical Reasoning Lab,
University of California, Berkeley, 2007
GRANTS, AWARDS AND HONORS:
Gregores Graduate Research Award, University of Oregon, 2013
John Templeton Foundation grant, 2012-2014, with Marjorie Taylor and
Candice Mottweiler
Graduate School Research Award, University of Oregon, 2012
Marthe E. Smith Memorial Science Fellowship, University of Oregon, 2012
Valedictorian in Arts and Social Sciences, University of San Francisco, 2010
Dean’s Medalist in Arts, University of San Francisco, 2010
Fr. Hubert Flynn, SJ, Award for highest cumulative GPA in graduating
class, University of San Francisco, 2010
Member, Alpha Sigma Nu Jesuit Honor Society, 2008
University Scholar, University of San Francisco, 2006-2010
Dean’s List, University of San Francisco, 2006-2010
PUBLICATIONS:
Taylor, M., Mottweiler, C.M., Naylor, E.R., & Levernier, J.G. (2015).
Imaginary worlds in middle childhood: A case study of two
pairs of coordinated paracosms. Creativity Research Journal.
doi:10.1080/10400419.2015.1030318
vii
ACKNOWLEDGEMENTS
My graduate experience would have been neither lasting nor fruitful had
it not been for the guidance of a wide network of individuals. I am particularly
grateful to Gerard Saucier, who, as my primary advisor, has demonstrated not only
patience as my interests developed over my course of study but also insisted that
I retain and strengthen my desire to work interdisciplinarily. I am also lastingly
grateful to Mark Alfano at Delft University of Technology in the Netherlands and
Australian Catholic University, who provided conceptual and methodological advice
throughout this project, offered advice beyond this project, and generated the
initial creative spark that eventually led to this dissertation.
I similarly appreciate of the work of the two additional members of the
dissertation committee, Caitlin Fausey and Colin Koopman, as well as the members
of other committees who have advised me throughout my time as a graduate
student, Sanjay Srivastava, Azim Shariff, Brian Westra, and Lou Moses. I am
additionally grateful to Marjorie Taylor for her supervision during the first two
years of my course of study, and for her help in naming and pursuing newly-
developing interests. With this in mind, I also acknowledge the obviously conscious
work of the entire faculty of the Department of Psychology not only to allow but
also to facilitate collaborative work across their research labs.
I am grateful to Kim Evenson and Legacy.com for their vision and open-
mindedness in allowing this project to go forward. Legacy.com graciously provided
not only permission to use obituaries from its website, but also insight through
several shared conversations about the technical and ethical considerations of a
project such as this.
viii
This project would have been infeasible without the teaching and support
of John Russell and Karen Estlund, now at Pennsylvania State University, and the
University of Oregon’s Digital Scholarship Center and its Graduate Affiliates (both
the program and the people). I am also grateful to Rose Hartman and Tyler Matta
for their generosity in offering time and advice to help me design and complete my
analyses.
The members of the Department of Psychology’s 2011 cohort have shaped
my experience incalculably, especially in my final year here. I carry forward a
growing sense of discernment that their conversation has helped to engender.
I am deeply grateful to Lori Olsen, whose logistical support and care of
students as Graduate Secretary in the Department of Psychology enabled this work.
This project has been a sobering and ultimately very special experience
in witnessing the humanity of individuals at perhaps their most fragile but also
optimistic points: making sense of the mortality of those they care about, and
of their own mortality. I am grateful to all of the authors of the obituaries that
I sampled in the creation of this dissertation for reaffirming in their writing and
anecdotes the optimism of living on, celebrating shared values and virtuous actions,
and cherishing shared memories and experiences, even following great loss. I hope
that I have responsibly curated these texts, and that this dissertation not only
provides at least some insight into the values of different communities, but also
does so in a way that is sensitive and meaningful to both obituary authors and
their subjects.
ix
To Emmit Routson, whose steps I have tried to follow throughout my education,
and who would have had many optimistic and thoughtful things to say as I
completed this project if he had had time to do so;
To Paul and Susan Levernier, whose closeness, sagacity, and advice I value
tremendously;
And to Natasha Kolosowsky, who walked with me through this process.
x
TABLE OF CONTENTS
Chapter Page
I. INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . 1
Summary Statement . . . . . . . . . . . . . . . . . . . . . . 1
Background on the Schwartz Values model . . . . . . . . . . . . . 2
Methods of Measuring the Schwartz Values . . . . . . . . . . . . . 4
Self-Report Measures . . . . . . . . . . . . . . . . . . . . . 4
Natural Language . . . . . . . . . . . . . . . . . . . . . . 6
Why Obituaries . . . . . . . . . . . . . . . . . . . . . . . . 9
A Context of Praising Examplars for Their Ideal Traits . . . . . . 9
The Logistics of Using Obituaries as Values-Containing
Data Sources . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Necrologies are in use Prospectively in Clinical Practice . . . . . . . 15
Obituaries are in Use Retrospectively in Social Research . . . . . 16
Obituaries are Biased in a Useful Way . . . . . . . . . . . . . 19
Publicly-Available, but Possibly Ethically Problematic . . . . . . . . 22
Considerations for the Deceased and Their Family
Members and Friends when Using Obituaries in Research . . . . . 22
Legal Considerations for Research that Uses Obituaries . . . . . . 26
The Goals of this Project . . . . . . . . . . . . . . . . . . . . 28
Research Questions of this Project . . . . . . . . . . . . . . . . 29
Research Question 1 . . . . . . . . . . . . . . . . . . . . . 29
Research Question 2 . . . . . . . . . . . . . . . . . . . . . 30
Individual level . . . . . . . . . . . . . . . . . . . . . . 30
xi
Chapter Page
Community level . . . . . . . . . . . . . . . . . . . . . 32
Ethnic/Racial demographics . . . . . . . . . . . . . . . . 33
II. METHODOLOGY . . . . . . . . . . . . . . . . . . . . . . . . . 34
Data Collection and Initial Processing . . . . . . . . . . . . . . . 34
Lexical Graph Database . . . . . . . . . . . . . . . . . . . 34
WordNet . . . . . . . . . . . . . . . . . . . . . . . . 34
University of South Florida Free Associations Network . . . . 38
Legacy.com Newspaper Metadata . . . . . . . . . . . . . . . 40
Newspaper Names and States Downloader . . . . . . . . . . 41
Finding Newspaper Distribution Points Using
Google Maps . . . . . . . . . . . . . . . . . . . . . . . 42
Finding United States Census County Codes
for Newspaper Locations Using the Federal
Communications Commission’s (FCC’s) Data
Conversions API . . . . . . . . . . . . . . . . . . . . . 44
Legacy.com Obituaries Downloader . . . . . . . . . . . . . . 49
Legacy.com Obituaries Data Scraper . . . . . . . . . . . . . . 50
Development of the Gender-Guessing Algorithm . . . . . . . 50
Development of the Age-Guessing Algorithm . . . . . . . . . 51
Validation of the Gender- and Age-Guessing Algorithms . . . . . 53
Initial Data Processing Steps for Calculating
Reliability Statistics . . . . . . . . . . . . . . . . . . . . 54
Gender Coding Reliability . . . . . . . . . . . . . . . . . 55
Age Coding Reliability . . . . . . . . . . . . . . . . . . 55
Filtering of Obituaries by Word Length . . . . . . . . . . . . . 56
Construction of a Matrix of Lemma - Part-of-Speech
Distances from Bardi et al.’s Dictionary . . . . . . . . . . . . 58
xii
Chapter Page
Calculation of Schwartz Value Distances from Lemma -
Bardi et al. Word Distances . . . . . . . . . . . . . . . . . . 67
III. RESULTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Further Cleaning and Pruning the Dataset . . . . . . . . . . . . . 68
Pruning Within-Newspaper Duplicate Obituaries . . . . . . . . 68
Removing Obituaries of Individuals of Unknown Age . . . . . . . 69
Removing Obituaries of Individuals of Uncertain Gender . . . . . 70
Assessing Across-Newspaper Duplicate Obituaries . . . . . . . . 70
Analyses . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Calculation of a Measure of Network Distance, “Word-
by-Hop” . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Final Calculation Formula for Word-by-Hop . . . . . . . . . 71
Initial Calculation of Word-by-Hop . . . . . . . . . . . . . 72
Correcting Word-by-Hop Values for Discrepant
Numbers of Value Lexicon Words . . . . . . . . . . . . . . 74
Correcting Word-by-Hop Values for Discrepant
Numbers of Words across Obituaries . . . . . . . . . . . . 75
Statistically Modeling Word-by-Hop Values . . . . . . . . . 78
Multiple Membership Analysis Rationale . . . . . . . . . . . . 82
Covariate Descriptive Statistics . . . . . . . . . . . . . . . . 85
Exploring the Properties of the New Word-by-Hop Measure . . . . 89
Answering Research Question 1: Which Values are
Present in Relation to One Another in Obituaries . . . . . . . . 93
Answering Research Question 2: Does Match with
the Schwartz Values Vary in Relation to Age, Gender,
Race or Ethnicity, Income, and/or Education . . . . . . . . . . 116
xiii
Chapter Page
IV. DISCUSSION . . . . . . . . . . . . . . . . . . . . . . . . . . 129
APPENDIX: R BASE AND LIBRARY VERSION NUMBERS . . . . . . . 139
REFERENCES CITED . . . . . . . . . . . . . . . . . . . . . . . . 143
xiv
LIST OF FIGURES
Figure Page
1. Bardi, Calogero, and Mullen (2008) value lexicon “Power”
words (“power,” “strength,” and “control;” lighter/green)
with their definitions (darker/blue). Words are connected
to definitions, and some definitions are connected to other
definitions. . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2. Visualization of Neo4J data model used in this project.
Node and relationship attributes are printed in italics. . . . . . . . . 39
3. Newspaper sample size by state. WY and HI had 0
newspapers that contract with Legacy.com. . . . . . . . . . . . . . 41
4. Newspaper sample size by county. WY and HI had 0
newspapers that contract with Legacy.com. . . . . . . . . . . . . . 46
5. Histogram of newspaper sample size by county. . . . . . . . . . . . 47
6. The United States’ population by county in 2014, using the
US Census Bureau Population Division’s “Annual County
Resident Population Estimates by Age, Sex, Race, and
Hispanic Origin: April 1, 2010 to July 1, 2014” dataset
(United States Bureau of the Census Population Division,
2015). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
7. Histogram of obituary word counts, cut off at 600 words
(the actual maximum word count in the corpus was 3,502
words, at the end of a long positively-skewed tail in the full
histogram’s distribution). . . . . . . . . . . . . . . . . . . . . . 59
8. Histogram of obituary word counts, cut off at 100 words
(the actual maximum word count in the corpus was 3,502
words, at the end of a long positively-skewed tail in the full
histogram’s distribution). . . . . . . . . . . . . . . . . . . . . . 60
9. Density curves for median household income and education
level (operationalized as percent of the population with a
bachelor’s degree or higher) from US census data averaged
across counties for each newspaper (after weighting by
county population size). . . . . . . . . . . . . . . . . . . . . . 87
xv
Figure Page
10. Density curves for each race or ethnicity variable from US
census data averaged across counties for each newspaper
(after weighting by county population size). Category titles
are taken from the Census dataset, and are of that category
“Alone or in Combination” with other race/ethnicity
categories (i.e., an individual included in the “White”
category could have self-described herself as both “White”
and “Hispanic”). . . . . . . . . . . . . . . . . . . . . . . . . 88
11. Schwartz “openness to change” values (Self-direction,
Stimulation, and Hedonism) plotted by gender to compare
word-by-hop values as computed using the number of Word
nodes between each obituary lemma / Part-of-Speech
combination and a value lexicon word, less one (i.e., the
number of Word node hops in the shortest path between
a given source and target word) vs. as computed using
the number of Synset-to-Synset edges in the shortest path
between a given source and target word. Pink/lighter points
represent obituaries coded as written about women. . . . . . . . . . 94
12. Schwartz “self-enhancement” values (Hedonism,
Achievement, and Power) plotted by gender to compare
word-by-hop values as computed using the number of
Word nodes between each obituary lemma / Part-of-Speech
combination and a value lexicon word, less one (i.e., the
number of Word node hops in the shortest path between
a given source and target word) vs. as computed using
the number of Synset-to-Synset edges in the shortest path
between a given source and target word. Pink/lighter points
represent obituaries coded as written about women. . . . . . . . . . 95
13. Schwartz “conservation” values (Security, Conformity, and
Tradition) plotted by gender to compare word-by-hop values
as computed using the number of Word nodes between
each obituary lemma / Part-of-Speech combination and
a value lexicon word, less one (i.e., the number of Word
node hops in the shortest path between a given source and
target word) vs. as computed using the number of Synset-
to-Synset edges in the shortest path between a given source
and target word. Pink/lighter points represent obituaries
coded as written about women. . . . . . . . . . . . . . . . . . . 96
xvi
Figure Page
14. Schwartz “self-transcendence” values (Benevolence and
Universalism) plotted by gender to compare word-by-
hop values as computed using the number of Word nodes
between each obituary lemma / Part-of-Speech combination
and a value lexicon word, less one (i.e., the number of Word
node hops in the shortest path between a given source and
target word) vs. as computed using the number of Synset-
to-Synset edges in the shortest path between a given source
and target word. Pink/lighter points represent obituaries
coded as written about women. . . . . . . . . . . . . . . . . . . 97
15. Schwartz “openness to change” values (Self-direction,
Stimulation, and Hedonism) plotted by gender and age
at death, as both coded using the automated algorithm
described in the Methods section. Pink/lighter points
represent obituaries coded as written about women. The
top row plots word-by-hop values as computed using the
number of Word nodes between each obituary lemma /
Part-of-Speech combination and a value lexicon word, less
one (i.e., the number of Word node hops in the shortest
path between a given source and target word). The bottom
row plots word-by-hop values as computed using the
number of Synset-to-Synset edges in the shortest path
between a given source and target word. . . . . . . . . . . . . . . 99
16. Schwartz “self-enhancement” values (Hedonism,
Achievement, and Power) plotted by gender and age at
death, as both coded using the automated algorithm
described in the Methods section. Pink/lighter points
represent obituaries coded as written about women. The
top row plots word-by-hop values as computed using the
number of Word nodes between each obituary lemma /
Part-of-Speech combination and a value lexicon word, less
one (i.e., the number of Word node hops in the shortest
path between a given source and target word). The bottom
row plots word-by-hop values as computed using the
number of Synset-to-Synset edges in the shortest path
between a given source and target word. . . . . . . . . . . . . . . 100
xvii
Figure Page
17. Schwartz “conservation” values (Security, Conformity, and
Tradition) plotted by gender and age at death, as both
coded using the automated algorithm described in the
Methods section. Pink/lighter points represent obituaries
coded as written about women. The top row plots word-by-
hop values as computed using the number of Word nodes
between each obituary lemma / Part-of-Speech combination
and a value lexicon word, less one (i.e., the number of Word
node hops in the shortest path between a given source and
target word). The bottom row plots word-by-hop values as
computed using the number of Synset-to-Synset edges in the
shortest path between a given source and target word. . . . . . . . . 101
18. Schwartz “self-transcendence” values (Benevolence and
Universalism) plotted by gender and age at death, as both
coded using the automated algorithm described in the
Methods section. Pink/lighter points represent obituaries
coded as written about women. The top row plots word-by-
hop values as computed using the number of Word nodes
between each obituary lemma / Part-of-Speech combination
and a value lexicon word, less one (i.e., the number of Word
node hops in the shortest path between a given source and
target word). The bottom row plots word-by-hop values as
computed using the number of Synset-to-Synset edges in the
shortest path between a given source and target word. . . . . . . . . 102
19. Schwartz “openness to change” values (Self-direction,
Stimulation, and Hedonism) plotted by gender and obituary
word count, as both coded using the automated algorithm
described in the Methods section. Pink/lighter points
represent obituaries coded as written about women. The
top row plots word-by-hop values as computed using the
number of Word nodes between each obituary lemma /
Part-of-Speech combination and a value lexicon word, less
one (i.e., the number of Word node hops in the shortest
path between a given source and target word). The bottom
row plots word-by-hop values as computed using the
number of Synset-to-Synset edges in the shortest path
between a given source and target word. . . . . . . . . . . . . . . 105
xviii
Figure Page
20. Schwartz “self-enhancement” values (Hedonism,
Achievement, and Power) plotted by gender and obituary
word count, as both coded using the automated algorithm
described in the Methods section. Pink/lighter points
represent obituaries coded as written about women. The
top row plots word-by-hop values as computed using the
number of Word nodes between each obituary lemma /
Part-of-Speech combination and a value lexicon word, less
one (i.e., the number of Word node hops in the shortest
path between a given source and target word). The bottom
row plots word-by-hop values as computed using the
number of Synset-to-Synset edges in the shortest path
between a given source and target word. . . . . . . . . . . . . . . 106
21. Schwartz “conservation” values (Security, Conformity, and
Tradition) plotted by gender and obituary word count, as
both coded using the automated algorithm described in the
Methods section. Pink/lighter points represent obituaries
coded as written about women. The top row plots word-by-
hop values as computed using the number of Word nodes
between each obituary lemma / Part-of-Speech combination
and a value lexicon word, less one (i.e., the number of Word
node hops in the shortest path between a given source and
target word). The bottom row plots word-by-hop values as
computed using the number of Synset-to-Synset edges in the
shortest path between a given source and target word. . . . . . . . . 107
22. Schwartz “self-transcendence” values (Benevolence and
Universalism) plotted by gender and obituary word count,
as both coded using the automated algorithm described
in the Methods section. Pink/lighter points represent
obituaries coded as written about women. The top row
plots word-by-hop values as computed using the number
of Word nodes between each obituary lemma / Part-of-
Speech combination and a value lexicon word, less one
(i.e., the number of Word node hops in the shortest path
between a given source and target word). The bottom row
plots word-by-hop values as computed using the number of
Synset-to-Synset edges in the shortest path between a given
source and target word. . . . . . . . . . . . . . . . . . . . . . 108
xix
Figure Page
23. Densities of word-by-hop for Schwartz “openness to change”
values (Self-direction, Stimulation, and Hedonism), plotted
by newspaper to allow visual inspection of both overall
distribution and variability across communities. The
top row plots word-by-hop values as computed using the
number of Word nodes between each obituary lemma /
Part-of-Speech combination and a value lexicon word, less
one (i.e., the number of Word node hops in the shortest
path between a given source and target word). The bottom
row plots word-by-hop values as computed using the
number of Synset-to-Synset edges in the shortest path
between a given source and target word. Darker areas
indicate higher newspaper overlap. . . . . . . . . . . . . . . . . . 109
24. Densities of word-by-hop for Schwartz “self-enhancement”
values (Hedonism, Achievement, and Power), plotted
by newspaper to allow visual inspection of both overall
distribution and variability across communities. The
top row plots word-by-hop values as computed using the
number of Word nodes between each obituary lemma /
Part-of-Speech combination and a value lexicon word, less
one (i.e., the number of Word node hops in the shortest
path between a given source and target word). The bottom
row plots word-by-hop values as computed using the
number of Synset-to-Synset edges in the shortest path
between a given source and target word. Darker areas
indicate higher newspaper overlap. . . . . . . . . . . . . . . . . . 110
25. Densities of word-by-hop for Schwartz “conservation” values
(Security, Conformity, and Tradition), plotted by newspaper
to allow visual inspection of both overall distribution and
variability across communities. The top row plots word-by-
hop values as computed using the number of Word nodes
between each obituary lemma / Part-of-Speech combination
and a value lexicon word, less one (i.e., the number of Word
node hops in the shortest path between a given source and
target word). The bottom row plots word-by-hop values as
computed using the number of Synset-to-Synset edges in
the shortest path between a given source and target word.
Darker areas indicate higher newspaper overlap. . . . . . . . . . . . 111
xx
Figure Page
26. Densities of word-by-hop for Schwartz “self-trancendence”
values (Benevolence and Universalism), plotted by
newspaper to allow visual inspection of both overall
distribution and variability across communities. The
top row plots word-by-hop values as computed using the
number of Word nodes between each obituary lemma /
Part-of-Speech combination and a value lexicon word, less
one (i.e., the number of Word node hops in the shortest
path between a given source and target word). The bottom
row plots word-by-hop values as computed using the
number of Synset-to-Synset edges in the shortest path
between a given source and target word. Darker areas
indicate higher newspaper overlap. . . . . . . . . . . . . . . . . . 112
27. Correlation matrix of word-by-hop (calculated using number
of Word nodes between a source and target word, rather
than the number of Synset-to-Synset edges) for each pair of
Schwartz values. Schwartz values in this matrix are listed
following the Schwartz circumplex model in a counter-
clockwise direction. . . . . . . . . . . . . . . . . . . . . . . . 114
28. Scatterplot of median household income and education
level (operationalized as percent of the population with a
bachelor’s degree or higher) from US census data averaged
across counties for each newspaper (after weighting by
county population size) vs. obituary word count. . . . . . . . . . . 115
29. AIC values plotted across models and Schwartz values. The
highest AIC (indicating worst fit) was found consistently
across Schwartz values in the intercept-only model (Model
#1). The lowest AIC was not in the full model (Model #7),
but rather the full model lacking level-2 predictors (Model
#3). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
30. AIC values plotted across models and Schwartz values.
The lowest AIC was not in the full model (Model #7), but
rather the full model lacking level-2 predictors (Model #3). . . . . . . 123
31. Plot showing coefficient estimates and standard errors for
all Schwartz value word-by-hop Dependent Variables and all
model predictors for Model #7, the full model. Note that
all predictors were z-scored before being entered into the
model, and that the word-by-hop DV was logit-transformed. . . . . . 126
xxi
Figure Page
32. Plot showing coefficient estimates and standard errors for
all Schwartz value word-by-hop Dependent Variables and
all model predictors for Model #7, the full model, excluding
the intercept (in order to see the other estimates in more
detail). Note that all predictors were z-scored before being
entered into the model, and that the word-by-hop DV was
logit-transformed. . . . . . . . . . . . . . . . . . . . . . . . . 126
33. Plot showing coefficient estimates and standard errors for
all Schwartz value word-by-hop Dependent Variables and all
model predictors for Model #3, the model with the lowest
AIC. Note that all predictors were z-scored before being
entered into the model, and that the word-by-hop DV was
logit-transformed. . . . . . . . . . . . . . . . . . . . . . . . . 127
34. Plot showing coefficient estimates and standard errors for
all Schwartz value word-by-hop Dependent Variables and all
model predictors for Model #3, the model with the lowest
AIC, excluding the intercept (in order to see the other
estimates in more detail). Note that all predictors were
z-scored before being entered into the model, and that the
word-by-hop DV was logit-transformed. . . . . . . . . . . . . . . 127
xxii
LIST OF TABLES
Table Page
1. The Schwartz value lexicon, reprinted from Bardi et al.
(2008). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2. Node and relationship types in the WordNet graph database. . . . . . 37
3. WordNet table names and descriptions. . . . . . . . . . . . . . . . 37
4. State Abbreviations and Names of newspapers that did
not return address results from the automated Google
Maps query procedure and thus were excluded from further
analyses. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5. TreeTagger to WordNet Part Of Speech (POS) conversions. . . . . . . 61
6. Example data showing the number of hops through WordNet
between “business” and words for three Schwartz values
from Bardi et al’s (2008) value lexicon . . . . . . . . . . . . . . . 72
7. Example implementations of the initial word-by-hop
calculation, using data from Table 6. . . . . . . . . . . . . . . . . 73
8. Descriptive statistics for newspaper-level covariates. . . . . . . . . . 86
9. Descriptive statistics for hop numbers in the shortest paths
between obituary lemma-POS combinations and value
lexicon words. . . . . . . . . . . . . . . . . . . . . . . . . . . 90
10. Percentage of final Synset (i.e., definition) node POS in
shortest paths between obituary lemma-POS combinations
and value lexicon word (excluding cases in which no path
existed). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
11. Descriptive statistics for word-by-hop calculations, ordered
by mean (descending) followed by standard deviation
(ascending). . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
12. Descriptive statistics for Word-hop-derived, logit-transformed
word-by-hop calculations, ordered by mean (descending)
followed by standard deviation (ascending). . . . . . . . . . . . . . 117
xxiii
CHAPTER I
INTRODUCTION
Summary Statement1
This dissertation is centrally concerned with exploring obituaries as
repositories of values. It seeks to lay a foundation for detecting values, defined from
the Schwartz values model (e.g., Schwartz, 2012), in obituaries, and to examine
their relation to characteristics of the individuals described in those obituaries,
and to their communities. Obituaries are a publicly-available natural language
source that are variably written for members of communities that are wide (nation-
level) and narrow (city-level, or at the level of specific groups therein). Because
they are explicitly summative, limited in size, and written for consumption by
a public audience, obituaries may be expected to express concisely the aspects
of their subjects’ lives that the authors (often family members living in the same
communities) found most salient or worthy of featuring.
This study of values using a natural language approach with obituaries
fundamentally takes a descriptive rather than normative stance. However, this
dissertation is founded in the idea that descriptive research into values can provide
insight into the normative makeup of the communities that espouse them and,
potentially, can allow useful relationships to be seen about the communities and
the individuals that they comprise (in this project, the obituaries’ subjects).
As, to some extent, rituals of praise, obituaries specifically can offer a route to
understanding the values of the communities that write them.
1The code and other supplemental files written for this dissertation can be found at http://
doi.org/10.7264/D3WC7S.
1
Background on the Schwartz Values model
The Schwartz values paradigm (e.g., Schwartz, 2012) is widely cited2
in the values literature for its wide scope and a substantial body of evidence
demonstrating its ability to describe value priorities across cultures and social
groups therein. As Schwartz (2012) summarized, the model posits that 10 basic
categories reliably encompass differences in values across cultures. The model
defines values “as desirable, transsituational goals, varying in importance, that
serve as guiding principles in people’s lives” (Schwartz et al., 2001, p. 521). Thus,
in the model, value-categories are primarily differentiated by the “[primary end-
]goal[s] or motivation[s]” toward which they are oriented (Schwartz, 2012, p. 4;
Schwartz et al., 2001)3.
Graphical depictions of the model often arrange the value categories
circularly, conceptually allowing pairs of given categories to be complementary or
in opposition to one another, and to represent that they are more continuously
than rigidly defined4. The approximate quadrants of this circle can also be named
2As of September, 2016, Google Scholar recorded that Schwartz (1992) has been cited 12,544
times.
3Schwartz’ conceptualization of values is in line with that of Haybron and Tiberius (2012,
p. 9), who defined values as “relatively robust pro-attitudes [i.e., attitudes in favor of certain
actions], or clusters of pro-attitudes, that we take to generate reasons for action and furnish
standards for evaluating how our lives are going.”
Schwartz (1992, p. 4) further defined values as “(1) concepts or beliefs [that] (2) pertain
to desirable end states or behaviors, (3) transcend specific situations, (4) guide selection or
evaluation of behavior and events, and (5) are ordered by relative importance.” This definition
agrees with Haybron and Tiberius’ (2012) at least in its second and fourth parts. To Haybron
and Tiberius, values go beyond short-term preferences in generating impulses to action, to some
extent becoming ends for action in themselves in a way that preferences do not (Haybron and
Tiberius noted that values can be called “subjectively reason-grounding” in that they are used by
individuals to justify actions even when no good or otherwise justifiable reasons for those actions
exist).
4A pictorial representation of this model is provided by Schwartz (2012), whose report is freely
available under a Creative-Commons-family license (albeit one that prohibits reuse of the image
by itself in this document).
2
(All quotes below are from Schwartz, 2012, p. 5-9; the list below moves around the
image of the circular model presented by Schwartz, 2012, in a counter-clockwise
direction):
1. Openness to change
1. Self-Direction (“Defining goal: independent thought and action–
choosing, creating, exploring.”)
2. Stimulation (“Defining goal: excitement, novelty, and challenge in life.”)
3. Hedonism (“Defining goal: pleasure or sensuous gratification for
oneself.”)
2. Self-Enhancement
1. Hedonism (Also above, in Openness to change)
2. Achievement (“Defining goal: personal success through demonstrating
competence according to social standards.”)
3. Power (“Defining goal: social status and prestige, control or dominance
over people and resources.”)
3. Conservation
1. Security (“Defining goal: safety, harmony, and stability of society, of
relationships, and of self.”)
2. Conformity (“Defining goal: restraint of actions, inclinations, and
impulses likely to upset or harm others and violate social expectations
or norms.”)
3
3. Tradition (“Defining goal: respect, commitment, and acceptance of the
customs and ideas that one’s culture or religion provides.”)
4. Self-Transcendence
1. Benevolence (“Defining goal: preserving and enhancing the welfare of
those with whom one is in frequent personal contact [the ‘in-group’].)”
2. Universalism (“Defining goal: understanding, appreciation, tolerance,
and protection for the welfare of all people and for nature.”)
While, Schwartz (2012) noted, the model does posit (and has accumulated
a large body of evidence for) 10 distinct values, the value-categories may more
usefully be thought of as pragmatically-defined boundaries along a continuum.
Thus, Schwartz argued, the value categories can be expected to blend into one
another, and can reasonably be subdivided (or superdivided) depending on the
goals of a particular project. Further, measures that assess values as defined in the
model seek to allow interpretation of “priorities” from among the value categories,
rather than of raw scores (Schwartz, 2012, p. 12; cf. Schwartz et al., 2001). With
this in mind, this dissertation project examines relative relationships among texts
in a corpus in the values that they express, and seeks to avoid taking an absolute
approach (e.g., x text is linguistically closer / expresses value y to a greater extent
than value z,” vs. “x text is about y value”).
Methods of Measuring the Schwartz Values
Self-Report Measures. The body of evidence in support of the
Schwartz model has primarily utilized self-report measures, to which it is useful
to devote attention in order to understand the history on which later attempts
to analyze values in natural language sources (such as Bardi et al., 2008,
4
and the project reported below) are founded. The model has been assessed
primarily through the use of two self-report questionnaires, the Schwartz Values
Survey (SVS), which is appropriate for use with adults, and the Portrait Values
Questionnaire (PVQ), which is appropriate for use with children (Schwartz, 2012).
As summarized by Schwartz (2012, p. 10), the SVS contains an approximate total
of 56 items in two parts that describe “potentially desirable end-states in noun
form” and “ways of acting in adjective form,” respectively. The SVS assesses
motivations as “guiding principle[s] in MY life,” and is constructed with an
asymmetric response scale to account for respondents’ tendencies to rate all values
as at least somewhat important to them. In contrast, the PVQ assesses the same 10
values as the SVS, but in a more cognitively concrete (and thus child-friendly) way,
by asking respondents to rate hypothetical others for how similar they are to the
respondent. Further decreasing cognitive load, the PVQ, unlike the SVS, does not
explicitly identify itself as a measure of values (Schwartz et al., 2001).
Schwartz et al. (2001) introduced and evaluated a preliminary version of
the PVQ after finding that SVS samples from some areas of the world (specifically,
“sub-Saharan Africa, India, Malaysia, and rural areas of less developed nations”
[p. 519], as well as with “those with minimal schooling [and] the elderly” [p. 538])
had not fit the 10-value model well. Schwartz et al. sought to examine the extent
to which this lack of fit had been a measurement issue caused by the SVS rather
than an issue with the generalizability of the theory itself. Encouragingly, Schwartz
et al. did find that fit with the values model was related to the measure used (the
SVS vs. the PVQ), interpreting this as evidence for increased generalizability of the
model (since, even if the model is not able to be called “truly universal” [p. 538],
5
the study showed that populations that previously had not been well-described by
the model could be assessed using the newer measure).
Schwartz (2012) reported on a variety of evidence that the model replicates
cross-culturally and across populations within cultures. Specifically, Schwartz
(p. 12) noted, across 82 countries, “each of the ten basic values [has been]. . .
distinguished in at least 90% of samples.” This body of evidence is not based on
sparse samples: as of their writing, Schwartz et al. (2001) stated that that the
circular values model had been supported by over 200 samples (at that time, in
60 or more countries), including several representative adult community samples
as well as samples with adolescents (Schwartz et al., 2001, e.g., reported on a
sample of 13- to 14-year-old girls from Uganda when examining the properties of
the PVQ).
Natural Language. More recently, in addition to self-report,
the Schwartz values paradigm has also been explored using natural language
corpora. Natural language, while messier to analyze than data from standardized
questionnaires such as the SVS and PVQ, opens a variety of new research
possibilities that would be more difficult using those traditional self-report
approaches, including looking across wider levels of society in an ecologically valid
way (Bardi et al., 2008). Natural Language Processing (NLP) allows not only
the use of data gathered from a variety of sources, at both individual and societal
levels (ibid.), but also allows the use of data that were not originally created as
research data. This report thus now briefly shifts to Bardi et al.’s (2008) approach
to bringing NLP concepts to the Schwartz values, and to expanding their approach
with obituaries, a type of corpus that can be expected to be particularly laden with
values.
6
Obituaries as a type of document are a subset of newspapers, which Bardi
et al. (p. 486) analyzed5, noting that “there is evidence in previous studies. . . that
popular textual media in democratic societies largely represent the salient values,
opinions, and concerns of people.” Within not just newspapers but text sources
generally, obituaries can be expected to be particularly charged with values (even
if only expressed using positive wording). Bardi et al. applied their value lexicon
to a collection of newspapers from the United States between 1900 and 2000,
searching for instances of words from the lexicon across all newspaper pages. The
approach described below for the current project expands on this approach: instead
of solely searching for word instances in a given text, the approach below includes
calculating the lexical “distance” of every word in the text from each of the words
listed in Bardi et al.’s value lexicon, allowing for greater nuance in analyses and
conclusions.
Keeping in mind the current project’s use of the value lexicon, Bardi et
al.’s (2008) findings are encouraging. Bardi et al. examined the convergent and
discriminant validity of their lexicon by comparing the relationships between the 10
Schwartz values found in their newspaper corpus with patterns found in self-reports
by contemporary participants who had completed the Schwartz Values Survey.
Bardi et al. noted that they did not expect to find a perfect correspondence
between inter-correlations among Schwartz values as expressed in newspapers
vs. through the SVS; in line with this expectation, they did not find a perfect
relationship, but did find “the same general pattern of correlations” (p. 487).
5Bardi et al. did not specifically mention exploring obituaries. However, as of this
writing (albeit eight years after Bardi et al. published their report), the resource they used,
NewspaperArchive.com, notes on its home page that its holdings include “online historical and
genealogical newspaper articles, obituaries, [and] local and international old newspapers archives”
(NewspaperArchive.com, 2016).
7
Further, and particularly interesting for the current project, Bardi et al.
examined the predictive validity of the lexicon by comparing changes in lexicon-
word usage over time with society-level behavioral indicators that could be
expected to share the same motivations as specific Schwartz values (for example,
for Power, the percentage of the population engaged in active military duty by
year; for Benevolence, the percentage of the population who had been deported by
year; and for Conformity, the number of births that occurred without the parents
being married). With the exception of Achievement (relating to successful patent
applications per capita) and Tradition (relating to churches per capita), which co-
occured regularly with words from other Schwartz values, Bardi et al. did observe
a “medium effect size” (p. 490) in the relationship between co-occurence of words
within and across Schwartz values in the newspaper corpus. Strikingly, Bardi et
al. demonstrated through a series of visualizations that usage of words from the
value lexicon over time did track major historical events in the United States (e.g.,
Security words increasing during the 1940s and early 2000s).
Bardi et al. (p. 490) concluded that, “overall, the value lexicon represents
words that can significantly discriminate between Schwartz’s values in natural
language use on the Internet.” Given their findings, the extension of their value
lexicon to allow for words that are not part of the lexicon but are nonetheless
related to the lexicon, especially using a corpus that can be expected to contain
heightened levels of value words, seems a useful path for new research. The
current project thus expands Bardi et al.’s (p. 493-495) goal of “facilitating the
measurement of values over time and in real-world settings” “without reliance on
self-report questionnaire responses,” in an ecologically-valid way (p. 495). Further,
given Bardi et al.’s conclusion that their value lexicon was sensitive enough to be
8
used at both individual and group levels, this extension in the project described
below, which includes analyses of both individual-level (age, gender) and group-
level (income, education, and ethnicity and race) predictors, seems potentially
fruitful.
Why Obituaries
As briefly mentioned above, obituaries are a particularly useful source of
language for understanding the relative value priorities of communities.
A Context of Praising Examplars for Their Ideal Traits. When
they are written to be more than simple death notices by containing biographical
information, obituaries are part of a long tradition of education in both moral
and non-moral values through the use of exemplars, especially surrounding death.
In the tradition of Western philosophy, the admonishment “not to speak evil of
the dead” has been repeated at least since ancient Greece (Diogenes Laertius, c.
350/1853, p. 33).
Pappas and Zelcer (2014, p. 100 ff.) noted that in Plato’s dialogue The
Menexenus, which primarily repeats and comments on an apocryphal oration
by Pericles, moral education was explicitly referenced as a goal of the Athenian
tradition of funerary rites. Although the dialogue expresses skepticism on the
actual ability of stories about exemplary moral agents to produce substantive and
positive moral change in an audience – Pappas and Zelcer (p. 100) paraphrased
that “high praise for the dead will make the hearers who did not know those dead
envious[, especially about] whatever exceeds an accomplishment they would be
capable of,” and that, “where envy is the most natural response, emulation is
not likely to occur” – regardless of the actual pedagogical efficacy of funerary
stories, the inclusion of valued traits in such stories is clear in this tradition.
9
Obituaries obviously share features with eulogies such as those in the Menexenus,
and exaggerate some features of eulogies: they are intended for public consumption,
but at a larger scale (whether for hundreds to thousands of people in the case of
obituaries published in local newspapers, or for millions of people in the case of
those published in national or international publications). In addition, they are
more versatile, typically running shorter in length than a transcribed eulogy, and
thus are presumably easier for audience members to digest; indeed, the publication
format of obituaries (with many together on a page) encourages reading more than
one in a sitting. The 153 obituaries published in December 2013 in The Eugene
Register-Guard, the primary newspaper in a university town of approximately
160,000 residents (and wider circulation to the surrounding county), averaged only
230 words in length. Despite their brevity, they comprised descriptions of a wide
variety of backgrounds and life experiences, with ages of the deceased ranging from
17 to 99 years. In aggregate, these smaller “snapshots” might form a much larger,
richer picture than a single eulogy of equivalent cumulative length.
Within the stream of Western civilization, the tradition of moral education
through stories has had additional intermediate steps to the present since Pericles’
oration. Reading hagiographies famously converted Ignatius Loyola, the founder
of the Jesuit order of Catholic priests (for a history, see, e.g., Decloux, 1991),
and formed a central part of his book Spiritual Exercises, several principles of
which historically prefaced many (including non-denominational) modern “spiritual
retreats.” Moral primers that include both fables and true stories about paragons
have also been used to educate children. The 1910 Ethics for Children (Cabot,
1910), as well as the more recent The Book of Virtues (Bennett, 1996), e.g.,
10
include stories about personages such as Abraham Lincoln alongside accounts of
fictional characters.
Life-summaries of the dead are, as Fowler (2005, p. 53) argued, “more than
a series of recollections about random individuals.” Rather, Fowler (p. 56) noted,
obituaries, whether written by “friends, colleagues, or even journalists,” form
a foundation of “collective memory” that, as with Alfano’s (2013) conception
of Hacking’s (1995) “looping kinds,” “feed back and shape public [history].”
Berger (1969, p. 43) called “the confrontation with death (be it through
actually witnessing the death of others or anticipating one’s own death in the
imagination). . . what is probably the most important [‘]marginal situation,[’]”
in which not only individuals but entire communities “often. . . experience. . .
‘ecstasy’ (in the literal sense of ek-stasis — standing, or stepping, outside reality
as commonly defined).” As Berger implied, mourning or otherwise facing mortality
provokes a liminal experience through which current practices and closely-held
values are reconciled, and possibly challenged. In the 15th century, the Ars
Moriendi was produced to educate Christian mourners how to address and behave
around dying individuals, and to remind those experiencing death how to do so
virtuously (specifically by avoiding vices; see Nicholson, Caxton, & de Worde,
1891, which is an English translation of the original Latin text; this work is also
noted by Doughty, 2014).
For Berger (p. 44), a “ ‘good death’ ” means one that “retain[s] to the end
a meaningful relationship with the nomos [per p. 19, “a[n]. . . order. . . of common
meaning” imposed on individuals by their society]. . . — subjectively meaningful to
oneself and objectively meaningful in the minds of others.” Doughty (2014, p. 214)
drew a similar conclusion, stating that “every culture has death values. These
11
values are transmitted in the form of stories and myths, told to children starting
before they are old enough to form memories. The beliefs children grow up with
give them a framework to make sense of and take control of their lives.” Berger
(p. 43) further asserted that the death anxiety experienced by a society needs
a ritualized, form-based method of reduction; mortality salience, Berger stated,
must be met with “legitimations of the reality of the social world,” presumably
including the moral world. Obituaries, especially through the process of authoring
them but also through that of reading them, may be seen from this perspective as
offering a chance to mourners within a community to declare the meaningfulness
and “goodness” of the lives of the deceased6, and of the shared values of those left
6Any informal survey of obituaries in a local newspaper in the USA will likely reveal that
obituaries are almost exclusively positive in disposition, apparently embracing the maxim to
“speak no evil of the dead” (this point was anecdotally substantiated in a formal [but not yet
published or fully-analyzed] survey of over 1,000 obituaries undertaken by this author and
advisors). Negative obituaries are rare enough to sometimes garner media attention specifically
for their novelty, as in the case of Marianne Theresa Johnson-Reddick, whose children wrote, in
part, “we celebrate her passing from this earth and hope she lives in the after-life reliving each
gesture of violence, cruelty, and shame that she delivered on her children. Her surviving children
will now live the rest of their lives with the peace of knowing their nightmare finally has some
form of closure.” (Mikkelson, 2013). As of this writing, a World Wide Web search for Johnson-
Reddick brings up numerous news articles reporting on her obituary and trying to make sense of
it: three of the first page of Google search results are (1) a post from Snopes.com, a website that
investigates “urban legends, folklore, myths, rumors, and misinformation” (Snopes.com, 2016),
examining the evidently difficult-to-believe claim that “[a] family member runs caustic obituary
about deceased parent” (Mikkelson, 2013); (2) a New York Daily News article titled, “Son who
helped write vicious obit for Reno mom insists ‘everything in there was completely true’ ” (The
Associated Press, 2013); and (3) a UK Daily Mail article which notes in its first summary point
that the child who authored the obituary is “completely unrepentant” (Bates, 2013). Given the
perceived newsworthiness of a negative obituary such as this, it seems reasonable to assume that
obituaries (at least in the USA, where the obituary was authored, but also possibly in the UK,
where the Daily Mail is published) carry a taboo against negative sentiments, and thus have a
function related to educating readers in shared values, typically through sharing positive traits
(and thus virtues). Indeed, an interview (McAndrew, 2014) published with Johnson-Reddick’s
son, who authored the obituary, began, “It did what obituaries don’t do. Instead of a celebration
of a life found in newspapers every day, the obituary for Marianne Theresa Johnson-Reddick was
scathing.” Notably, however, despite having violated this taboo against writing an obituary with
a negative tone, Johnson-Reddick’s son explicitly stated that he hoped to reinforce a shared value
(disgust with child abuse) in his readers: “ ‘People may see this as something we did to shame our
mother. . . . But this is to bring shame to the issue of child abuse. I want every single person to
realize this could be your obituary.’ ”
12
living; following Berger’s vocabulary, obituaries could be understood as a type of
“theodicy,” a society-level mechanism for reassuring individuals of meaning; in this
view, the writing and reading of obituaries could be an approach to quiet “anomic”
discomfort (i.e., individuals’ discomfort based on personally having to face death or
other triggers that cause the nomos to come into question). Summarizing the lives
of decedents offers an opportunity to legitimize values by publicly invoking them in
a semi-ritualized way.
More specifically within the cultural history of the USA, on which this paper
focuses because of the prevalence of obituaries from it as a data source, exemplars
have been publicly used and described as a means of values (re)affirmation
for decades if not centuries; witnessing death often requires individuals and
communities such discomfort that a reaffirmation is required. This need is
expressed variably, to the point that, as Doughty (2014, p. 225) described from
an interview with a supervising physician, medical students will sometimes choose
not to communicate terminal diagnoses rather than “ ‘face their own mortality. . .
[by] fac[ing] a dying person;’ ” cf. Rosel (1978), who also discussed at length the
difficulty faced by many contemporary Americans in acknowledging the existential
discomfort brought on by thinking about death with limited ritualized means.
Writing and reading obituaries are two processes that remain for addressing this
psychological discomfort. Given their cultural history, it makes sense that, among
all publicly-read media, obituaries would be among the best candidate bearers
of values. Because obituaries are succinct and explicitly intended to summarize
their subjects’ lives, they may be expected to include only the features that their
authors find most salient, to signal not only to their authors themselves as relatives
13
or biographers of the deceased but also to others in the community aspects of the
character of the deceased that all in common might find important.
Obituaries are public and readily-available, not only in print newspapers
but also more accessibly through newspaper websites and databases, allowing,
with the ethical and legal caveats considered below, the ability to readily conduct
text-mining analyses without having first to transcribe the texts. While obituaries
are certainly not the only bearers of values, this ready availability makes them
an attractive option for research. This attraction is particularly salient for
ethnographic research to understand the differential values of geographically or
otherwise-distinct communities across the USA (e.g., compared not only by physical
location but also by gender). In addition, it provides a foundation for research
of a psychometric mindset (e.g., using prospective obituaries written by living
individuals to index their values “match” with different communities across the
USA based on those communities’ published obituaries, and to corroborate this
index with survey data using existing measures of values such as the Schwartz
Values Survey, Schwartz, 1992).
The Logistics of Using Obituaries as Values-Containing Data Sources
Obituaries and related formats of necrologies are not only a potential data
source for community values; rather, they are currently in use, both for their
existential weight and for their accessibility as public resources. Having argued
above that obituaries may be expected to contain community values, this paper
now turns to considering the logistical realities of extracting those values from this
format of data. Specifically, instances in which obituaries have directly been used
will be considered as examples from which to build new research.
14
Necrologies are in use Prospectively in Clinical Practice
In clinical psychological practice, Hayes, Strosahl, and Wilson (2011, p. 304
ff.) defined a prospective exercise called “What do you want your life to stand
for?” for clinicians to use as part of an Acceptance and Commitment Therapy
treatment approach. In this exercise, a therapist asks a client to imagine having
died but being able to attend her own funeral. The client is encouraged to consider
the “eulogies offered by [her spouse,]. . . children,. . . friends, [and] the people
[she has]. . . worked with.” This exercise is explicitly centered on considering
what an ideal eulogy from each of those speakers would comprise, moving beyond
simply thinking about the immediate effects of one’s own death (e.g., one’s spouse
and children being distraught) and into the larger meanings that others in one’s
community might draw from one’s life. Implying the gravity of the characteristics
that eulogies (and, by extension, obituaries) typically include, Hayes et al.
encouraged clinicians to point out to clients that clients’ imagined eulogies typically
comprise different, larger goals and values than individuals are often concerned
with on a day-to-day basis — that is, these life-summaries are concerned more with
values that speak to moral, spiritual, and character accomplishments, rather than
things that a client might “berate [himself]. . . about and struggle with” on a more
local scale, such as annual income or feelings of personal inadequacy.
Hayes et al. (2011) also proposed an alternative exercise, in which a client
is encouraged to compose and then write an epitaph on an imagined or drawn
tombstone, noting that it also “often. . . reveals wide discrepancies between
the client’s values and his or her current actions.” Although Hayes et al. noted
that this exercise can be overwhelming for some individuals for the amount of
mortality salience it elicits, this writing of prospective epitaphs, eulogies, and, by
15
extension, obituaries, either about oneself or a loved one or friend, underscores the
psychological gravity associated with summarizing an individual’s life upon his or
her actual death.
Obituaries are in Use Retrospectively in Social Research.
Obituaries have also been used non-clinically to examine social processes. In
the last of a series of studies, Goodwin, Piazza, and Rozin (2014) used living
participants’ ratings of the individuals described in obituaries to examine the
importance of moral character and social warmth on individuals’ perceptions of
others. Bridging definitions of virtue, values, and personality, Goodwin et al.
(p. 148) defined “moral character” as “compris[ing] the moral7 dimensions of a
person’s personality,” implying its inclusion in modern personality research (as
Saucier, 2009, also noted). To Goodwin et al., traits exist on a continuum of
moral relevance, as well as on a continuum of general relevance in constructing
global perceptions about others. Goodwin et al. collected 250 recent obituaries
from The New York Times, which, as a major newspaper of record, typically
runs longer-format, professionally-written, more in-depth biographical accounts
than local newspapers. Obituaries published in The New York Times are also
predominantly about particularly noteworthy individuals – CEOs of major
companies, international military leaders (famous or infamous), well-known artists,
etc. Goodwin et al. noted that obituaries offer detailed views into the lives of their
subjects, and thus comprised a useful corpus for understanding person-perception
(while controlling for coders’ previous knowledge of the deceased).
Goodwin et al. (p. 162) employed (living) Research Assistants (RAs) and
participants from Amazon’s Mechanical Turk service to code the obituaries on
7Goodwin et al. did not define the word “moral.”
16
several criteria. RAs first coded the obituaries on valenced Likert scales for the
obituary subject’s described “(1) abilities or lack of abilities, (2) moral character
or immoral character,. . . [and] (3) social warmth or coldness.” Twelve or more
Mechanical Turk workers then independently rated each obituary on a wider Likert
scale for overall impression. Goodwin et al. (p. 162) concluded, in part, that
“moral character impressions are conveyed more prominently in [these] summary
accounts of people’s lives than are impressions of social warmth,” implying that
obituaries did carry significant information about moral character. Further,
Goodwin et al. found that Mechanical Turk participants’ overall impressions of
obituary subjects were able to be predicted significantly using RA’s ratings of
“morality, warmth, and ability,” indicating (in addition to a satisfactory ICC
statistic among RA coders on these dimensions) that obituaries of this type carry
information about these qualities that is consistently accessible to readers. While
shorter-format obituaries from smaller newspapers almost certainly do not carry
this information about individual subjects to the same extent as the long-format
biographies of The New York Times, they may, in aggregate, similarly provide
useful information about the community of authors and readers among whom the
deceased may have lived.
Rodler, Kirchler, and Ho¨lzl (2001, p. 829-831) used obituaries as an
“unobtrusive method” for examining longitudinal change in stereotypes about
male and female subjects who had, in life, occupied positions of organizational
power (e.g., “director[s] or head[s] of a business firm, school or other public
organization, chair[s] of a department, etc.”). Replicating and updating previous
work that sought to establish trends in the words used to describe male and female
leaders, Rodler et al. gathered 992 obituaries authored by the corporations or
17
other organizations for which the deceased had served as leaders. The sample of
obituaries allowed immediate conclusions to be drawn about the male-to-female
leadership workforce distribution over the years included in the analysis: the corpus
was heavily lopsided, with 757 obituaries about male leaders and 137 about female
leaders, even after a second round of data collection to balance the sample as
much as possible. Rodler et al. [p. 831] coded “all verbs, adjectives, and nouns”
in each obituary for representation of 58 categories that had been established in
previous work (including, e.g., “Caring,” “Consensus-oriented,” “Decision-maker,”
and “Skillful”). An iterative correspondence analysis on the frequencies at which
obituaries of men vs. women invoked each of the categories revealed two primary
dimensions: first, an authority dimension, which included “attributes like servant,
committed, professional, and humane to intelligent, efficient, skillful, experienced,
and expert” (p. 831); and second, a competence dimension, which was, “at the one
pole, described by expert, experienced, professional, intelligent, and, at the other
pole, by venerable.” Rodler et al. concluded that while descriptions of male leaders
had remained largely consistent over time (during the years 1992 and 1998, which
were included in the sample, but also building on previous research that first used
these analysis methods on obituaries during the 1970s and then 1980s), descriptions
of female leaders moved from the 1970s to 1990s from focusing on “venerable”
qualities to “professional[ism] and commit[ment].”
Rodler et al. (2001) may be seen to have been studying (even if indirectly)
community values around gender. Additionally, Rodler et al.’s analyses included
several categories that meet definitions of virtue, including “courageous,”
“ethical,” “honest,” “patriotic,” “religious,” “unselfish,” and “venerable” (see their
Appendix). From this work, as well as work like that of Goodwin et al. (2014),
18
obituaries may be seen by precedent to be an accessible, usable data source for
examining social dynamics, including the expression of values generally.
Obituaries are Biased in a Useful Way. As data sources, obituaries
likely reveal as much (and, regarding values, possibly more) about their authors
and readers as about the individuals whom they describe. In contrast to
existing measures of values and their antecedents such as the Moral Foundations
Questionnaire (Graham et al., 2011), which includes items that address disvalues
(i.e., negative values), but somewhat like the Schwartz Values Survey, which does
not, obituaries may be expected to be almost entirely positively biased in the traits
they ascribe to their subjects. Speaking conceptually on this type of sampling bias,
Fowler (2005, p. 61) noted that, “admittedly, there are limits to the obituary as
a form of witnessing, due to constraints on critical openness at the time of death
[i.e., a taboo on mentioning negative aspects of the deceased] and the conflicting
perspectives of different newspapers [(which might, e.g., each publish a death
notice of noteworthy individuals with a unique political spin)].” However, Fowler
suggested that obituaries presented together in newspapers and similar media
create a tapestry of a larger process of grieving and reflecting than any single
piece could show. In their values, two obituaries, individually read, might tell little
about a community, reflecting too strongly the idiosyncratic grief or values of the
authoring friend or family member (or, perhaps, the political or other ideological
agenda of the authoring journalist) as he or she attempts to reconcile the loss of
the deceased with some meaning in his or her own life (cf. Doughty (2014), who,
as noted above, writes at length about the meaning-making that death engenders
among the living, and the varied ways in which those left alive compensate for the
loss of and grieve the dead). In aggregate, though, obituaries “[supply] factual
19
materials which can be read in terms of a wider relational perspective, [and
thus]. . . contribute a vital resource for actively shaping and demystifying collective
memory.”
Fowler (2005, p. 61 ff.) did note that newspaper obituaries have become
“less coded [and] more subtle” since the 19th century (i.e., less adherent to
historical “rigid aristocratic formulae” that sought to shape public political and
professional consciousness by offering “verdict[s]” and “last judgment[s]” about the
deceased’s professional and personal accomplishments). While obituaries of this
“coded” form certainly still exist (especially, as noted above, in the commissioned
biography-obituaries of major political figures in international newspapers such as
The New York Times), given that obituaries of who might be called “everyday
people” have become more available in major newspapers across the USA over
the last century, Fowler implies in her argument that in modern times, obituaries
are able to be read more at face value than in the past. However, obituaries in
many newspapers continue to exhibit a heavy sampling bias toward what Fowler
called “the dominants” of a community’s social hierarchy. Fowler reported from a
content analysis of approximately 100 British obituaries across major newspapers
that the individuals summarized were disproportionately well-educated (i.e., college
educated) and of a very high social class (i.e., “Oxbridge” educated).
Fowler also suggested that obituaries often disregard explicit mention of
communalistic values in favor of individualistic ones, portraying the deceased in
personal narratives that, even if not fully divorced from the social context that
facilitated that personal journey, fail to accurately reflect the importance of that
social context. Fowler interpreted this disregard as a remnant of the historical
bias, noted above, toward aristocratic and otherwise socially “dominant” figures.
20
In line with this, Fowler (p. 63) also noted a decrease over time in the number of
prominent obituaries of individuals who served communalistic social organizations,
such as “military, clergy or elite civil servants,” with a concomitant increase in
focus on “artists, writers, musicians and actors, or academics and politicians.”
In future research, this discrepancy might be able to be controlled for or at least
understood empirically, however, using other markers of presumably communalistic
thinking on the part of obituary authors (e.g., percentage of family- and team-
related words used to describe the deceased).
In summary, the use of obituaries as markers or carriers of community values
may be expected to be beset by (at least partially) sampling bias in (a) the values
themselves (with a preference for positive terms), (b) the type of values (with a
preference for individualism over communalism), (c) the individuals described (with
a preference for higher-educated and especially influential individuals), and (d)
gender (at least in some contexts). These biases would be substantive sources of
limitation in a research project seeking to catalog community values (Goodwin et
al., 2014, also acknowledged the second form of bias in their work); however, once
acknowledged (and effectively taken into account), they would not invalidate this
research approach. One fourth type of bias that Fowler noted, but which would
be beneficial to the use of obituaries for the study of values in a given place and
time, is political motivation of authors, especially in highly-publicized obituaries.
Noting that biographical sketches in obituaries can be written either according to
or specifically to counter and raise awareness of changes in a prevailing societal
view, the ability of obituaries to change over time in their foci makes them stronger
candidates for longitudinal analyses of a given geographic or otherwise cultural
location.
21
It should be noted that Fowler’s content analysis used four major London
newspapers, plus The New York Times and Le Monde (p. 70), all of which likely
print obituaries that have been chosen on a highly selective and topical basis
and have been written by professional, commissioned authors. With access to
smaller publications (e.g., local newspapers for small- to medium-sized towns
and cities across the USA), the causes of sampling bias in the larger papers could
potentially be used as a feature rather than a bug of research, as they could allow
the differential comparison of national-level values discourse with the discourse
of smaller communities, possibly following research such as that of Fu, Plaut,
Treadway, and Markus (2014), who analyzed and contrasted the value landscapes
of quadrants of the USA. This approach could allow the comparison of local-level
newspapers with major publishers of obituaries (such as The New York Times) that
controls for geographic and ideological similarity between the publications.
Publicly-Available, but Possibly Ethically Problematic
Considerations for the Deceased and Their Family Members
and Friends when Using Obituaries in Research. Obituaries aim not only
to broadcast to readers facts about the deceased but also to convey intimate,
summative life portraits. With that context in mind, it is worth consideration
that, especially for obituaries authored by friends and family with the intent to
publish in a local newspaper for consumption by community members only, the
work of “morality mining” (Christen, Alfano, Bangerter, & Lapsley, 2013) could
be interpreted as intrusive, especially when researchers are neither members of
authors’ local communities nor manually reading and interpreting each obituary
individually as a document of a once-living person. Aggregating public records
that were likely originally intended by their authors to be read singly, while legally
22
unproblematic, could become emotionally distressing to authors or their relations
if they felt that their words were being taken out of context or used for purposes
for which they neither intended nor explicitly consented (such as generating profit
or, possibly, academic prestige, where focus might fall more on the researching
company or academic team than on the values of the authoring communities and
legacies of their decedent members).
Valeski (2012, p. 217) argued that understanding ethical implications of
the use of “new forms of communication” by third parties (including researchers)
often centers on considering what social norms have developed around those new
technologies’ predecessors. These norms, Valeski noted, often include dynamics of
control between authors and recipients (indeed, ownership and control of access are
two of the four “elements of big data ethics” proposed by Davis & Patterson, 2012,
p. 16). If obituaries are written not only to serve as official, public death notices
but also to facilitate the grieving processes of both readers and authors by shaping
reminiscence (i.e., if obituary authors are writing for themselves and their families
as much as or more so than for the readers), a sense of control over the use of the
text is likely important in that grieving process. It would be ethically problematic
for researchers, even if well-intentioned, to exploit the publicness of the medium at
the expense of authors’ (and deceased individuals’) interests.
This sentiment follows Tsvetovat and Kouznetsov (2011, p. 162), who argued
that any type of data collection “on human beings affects their lives, in hard-to-
perceive subtle ways.” This argument can be extended and applied to the use of
public datasets with previously-collected data (cf. O’Neil & Schutt, 2013, p. 354ff.,
who cite and re-print a “Hippocratic Oath of Modeling”). It also follows Scime and
Murray (2013), who posited that social scientists, especially using large, aggregated
23
datasets, are responsible to the communities in which they are performing their
research, including remaining sympathetic to changing norms and expectations of
community members around data use. Tsvetovat and Kouznetsov’s argument would
also certainly apply in research that seeks to examine prospective obituaries from
living participants (e.g., using the Acceptance and Commitment Therapy exercises
proposed by Hayes et al., 2011).
A central theme around which ethical considerations for the use of public
but emotionally-charged data such as obituary texts can center is a socially-based
concept of privacy. In line with Scime and Murray (2013), Helbing (2015, p. 135)
noted that “ethical problems. . . are related to cultural values and social norms;”
following Solove (2007), this is especially the case with potential violations of
privacy. Solove (p. 754 ff.) noted that “traditional” attempts to define privacy have
done so, both narrowly and broadly, by searching for “the essence of privacy,” a
feature that extends across all contexts and types of concerns. Solove summarized
and argued for an alternative conceptualization in which privacy is understood
not merely in terms of individual rights but rather also with reference to social
good. In this conceptual approach, privacy is about “power relationships between
people and. . . institutions” (p. 757), and is often concerned with enabling “rules
of behavior, decorum, and civility” that may relate to expressions of “people’s
autonomy and dignity” (p. 761-3).
With this conception in mind, a researcher working with this type of
data may be seen to have a greater ethical responsibility than she otherwise
would, as “misuse” of these data could be defined by the very perception of their
misuse by members of the public, and could not only damage the experience of
privacy (as control of use of the data, regardless of the data’s public nature) of the
24
obituaries’ authors and their relations, but could also diminish those and other
individuals’ future willingness to engage in this type of public ritual. Thus, for the
potential gain that could come from responsible, thoughtful use of this type of data,
there also exists a potential damage, born of perceived meddling in moments of
particular fragility. This is especially the case with obituary data, as many of the
recommendations provided by Helbing (2015, p. 137) “towards privacy-preserving
data analysis” are infeasible: the “participants” in this type of research are the
deceased individuals who are the subjects of the texts; they cannot consent to
research, and are specifically excluded from the federal definition of human research
subjects in the USA (Office for Human Research Protections, 2010, §46.102f).
Their data cannot be substantively anonymized or randomized, as the full text even
of an obituary scrubbed of names is often easily recoverable by searching the World
Wide Web for a phrase or sentence from it. As with this dissertation project, it is
possible to “coarse-grain” (Helbing, 2015, p. 141) this type of data by performing
analyses only at aggregate levels and to release only datasets that are derived from
(but do not actually include) the original texts; this, however, comes at the cost of
full reproducibility by future researchers.
To some extent, ethical questions surrounding the use of this type of
data may be empirical ones: many authors, if asked, may be happy to have the
obituaries they authored be included in aggregative inquiry, especially if it is
performed with the goal of understanding community values. Regardless, values
research conducted with obituaries should take place with a mindset of sensitivity
and respect to the obituaries’ authors and their subjects, and should proceed
with as much transparency as possible, perhaps also including explicit licenses on
published datasets, clarifying expectations around data reuse. Ideally, these issues
25
should be explicitly considered in a Data Management Plan before beginning a
project.
In summary, the potential for negative emotional reactions among authors
reveals an ethical weight associated with this class of data, and also points to the
value of intent in interacting with these records. In addition, this ethical weight
indicates the richness that this type of corpus may hold for research that seeks
to understand community values in order to promote the social good. Although
this type of research could be undertaken to manipulate a ritual of public grieving
into an abstract data-generating mechanism divorced from its original context,
those wishing to understand the moral language of local communities responsibly,
especially to perform value-relevant work within or translating across those
author communities, may find valuable insight in this approach. To this end, it
is worth reflecting on Schwartz’ (2012) argument that “the critical focus of value
transmission is to develop commitment to positive relations, identification with
the group, and loyalty to its members.” The potential importance of obituaries as
values-transmitting artifacts, at least within the United States, may be expected to
relate to the enormity of betrayal and upset that members of a community could
experience if they perceived a violation of propriety surrounding their use.
Legal Considerations for Research that Uses Obituaries. The
American Psychological Association’s (APA) Ethical Principles of Psychologists
and Code of Conduct guidelines (American Psychological Association, 2010, §8.14)
advise that,
After research results are published, psychologists do not withhold
the data on which their conclusions are based from other competent
professionals who seek to verify the substantive claims through
26
reanalysis and who intend to use such data only for that purpose,
provided that the confidentiality of the participants can be protected
and unless legal rights concerning proprietary data preclude their
release” [emphasis added].
Whether considered for allowing publication of a dataset or for using the
dataset to begin with, the legal status of obituaries is unclear. Obituaries are to
some extent public records (and thus especially useful as research data), since
they are published in newspapers to serve as death notices (source: personal
correspondence with Legacy.com, one of the USA’s largest obituary publishers,
which contracts with newspapers nationally), but their actual legal status is
often ill-defined; e.g., many states in the USA do not require the publication of
death notices, meaning that obituaries’ publication is not necessarily protected
under public notice laws in those states. Russell (2012)8 noted that copyright of
obituaries typically remains with their original authors (unless the authors have
atypically signed over rights to the publishing newspaper), but that the Fair
Use doctrine in the USA also protects their inclusion in research under certain
circumstances. These points highlight that while obituary data begins “in the
open” (i.e., published for public consumption), their availability for re-use remains
uncertain, especially around issues of republication of datasets. Especially with
obituaries, if data are to be made available for further use by other researchers,
the admonitions from the APA’s guidelines above may be seen to require the
8The analysis of Russell (2012) is not a source of official legal advice, but is a useful summary
of legal issues specifically surrounding the use of obituaries.
27
researchers who initially gather the data to serve as ethical “gatekeepers” for re-
use of the data9.
The Goals of this Project
This dissertation project has two purposes. First, the project examines the
Schwartz values paradigm in a naturalistic context (that of obituaries). To this end,
the project seeks to describe which values are discussed most across contemporary
newspaper-based “communities” across the United States, taking into account
potential individual- and group-level predictors of values-differentiation: at the
individual level, the age and gender of the deceased; and at the community level,
median income, education level, and ethnic and racial demographic composition in
the counties in which each obituary’s publishing newspaper(s) are based.
Second, this project seeks to develop tools for a novel method of assessing
the “fit” of obituary texts with the Schwartz values for future researchers to use.
The creation and documentation of a graph database from Princeton University’s
(2010) WordNet thesaurus10, following but expanding on the work of Nagi (2013),
as well as readily-usable algorithms for employing it in the analysis of this obituary
9Indeed, when I negotiated a data use agreement with Legacy.com, one of the largest obituary
warehousing websites in the USA, the agreement centered around avoiding breaches of public trust
not only in my own research (specifically by conducting analyses unlikely to cause resentment
or visceral unease on the part of obituary authors or other family members of the deceased) but
also in future research that others might perform from data gathered and processed through this
project (minimizing these through, e.g., data cleaning and partial anonymization techniques).
While, as discussed above, full anonymization of the obituaries would likely not be possible (as
a simple internet search of any of the obituary’s text would allow finding the original source),
names (e.g., surnames) could be stripped out of published versions of derivative datasets,
encouraging future researchers to perform analyses of the texts over potentially more problematic
analyses such as the race of obituary subjects.
10During the course of writing this dissertation, the official WordNet project released a
canonical Neo4J graph database, currently available at https://github.com/wordnet/wordnet. The
inclusion of the University of South Florida Free Association Norms dataset (Nelson, McEvoy, &
Schreiber, 1998) in this dissertation project’s database, as well as the documentation provided
as part of this project, will hopefully help to supplement this new, official Neo4J WordNet
implementation.
28
corpus, can hopefully be utilized in future Natural Language Processing research in
psychology surrounding “linguistic distance.”
Although not its primary motivation, this project can also enable future
research that could make clear areas in which the Schwartz values paradigm may
not be comprehensive. By subtracting out of the corpus words that have low
“word-by-hop” values (defined below) for all Schwartz categories, future research
could use the graph database approach used in this project to examine new areas
of consistent speech across or within communities in the USA that could be
considered values without fitting well in the current Schwartz values paradigm
(e.g., by topic modeling the words that remain after removing those that are highly
connected to Bardi et al’s [2008] value lexicon words). The project could also, using
the same approach, potentially demonstrate wide comprehensiveness of the current
Schwartz values model, contributing to its wide body of existing evidence.
Research Questions of this Project
The current project seeks to answer two primary research questions, each of
which comprises additional sub-questions.
Research Question 1. First, the current project asks whether the
Schwartz values can be detected in obituaries, using a thesaurus-based expansion
of Bardi et al.’s (2008) value lexicon approach (described below). To this end, it
additionally asks the following questions:
1. Are all of the Schwartz values indicated in the corpus?
Given the nature of values in the Schwartz model as differentially prioritized
across communities, it is reasonable to expect that not all Schwartz values will be
equally represented in different communities; further, it is also reasonable to expect
that even in aggregate, not all Schwartz values will be equally represented in the
29
obituaries corpus overall. Given that obituaries are subject to cultural taboos
around topic-appropriateness, as discussed above, I expect that Hedonism and
Power specifically will be represented less than other values in aggregate, if at all11.
2. In what relative proportions are the Schwartz values talked about in the
corpus?
3. For each of the 10 Schwartz values, what is the mean word-by-hop numerical
value across the obituary corpus?12
Research Question 2. Second, the current project asks the following
question:
To what extent is discussion of particular Schwartz values related to
characteristics of obituaries, at the levels of individuals and their communities?
The current project seeks to answer this question by expanding Bardi et al.’s
(2008) approach with word-by-hop calculations as a dependent measure.
Individual level.
11I do not mean to imply that Hedonism and Power, especially as defined in the Schwartz values
model, are “bad” values, but rather that they specifically may be indicated by a larger number
of potential anecdotes that would be less likely to be printed in an obituary than anecdotes that
reference the other values.
12In the originally-proposed version of this dissertation, Research Question 1 contained two
additional components, both of which “fell away” as nonsensical as the word-by-hop calculation
described below developed. First, sub-question 3 asked what the mean word-by-hop value
across all Schwartz values was. This became less relevant when word-by-hop became a relative
measure (i.e., one that is not meaningful unless compared with another word-by-hop value).
Second, a fourth sub-question asked what percentage of words found in the corpus are linked
with Schwartz values. This question became less relevant as it became clear that the WordNet
graph database described below was much more robust than expected: almost all words (> 90%)
in the corpus had a connection to at least one of the value lexicon words described in the Methods
section. Thus, to clarify, these additional questions have not been ignored in this complete draft;
rather, they have fallen away as the attention that was originally directed to them has turned to
exploring the properties of word-by-hop, in consultation with members of the Committee.
30
Approximate age of the deceased Schwartz et al. (2001, p. 533) stated
that age is positively related to Tradition, Conformity and Security, and that it
is negatively related with Self-direction, Stimulation, and Hedonism. In line with
these past findings, Schwartz et al. found in an international pair of samples that
age correlated significantly with “self-transcendence (Benevolence, Universalism)
and negatively with self-enhancement (Power, Achievement).” It is reasonable to
expect that age should show similar main effects in the obituary corpus.
Gender Schwartz et al. (2001) found small correlations (primarily less than
or equal to .11) between several of the 10 Schwartz values and gender (defined
as male vs. female)13. The largest relationships reported by Schwartz et al. (r
> .11) included women over men in Benevolence and Tradition values, and men
over women in Stimulation. However, these were localized to specific cultures
from those that were sampled (Israel, South Africa, and Italy, respectively).
Similarly, Schwartz and Rubel (2005) found increased importance of Tradition
in women over men, and of Hedonism in men over women, but inconsistently in
both cases over four studies. Schwartz and Rubel presented evidence, however, that
this inconsistency may have been a result of a sample (student vs. community)
x measure (SVS vs. PVQ) interaction. When controlling for sample type and
measure, Schwartz and Rubel (p. 1019) found that men consistently rated
Stimulation, Hedonism, Achievement, and Self-Direction higher than women, while
13Schwartz and Rubel (2005) called this variable “sex” rather than “gender.” The current
project uses “gender;” thus, I have re-worded Schwartz and Rubel’s report in this summary,
despite the differences between the two concepts, given the rationale for using “gender” as a
binary variable in this project. Since obituaries are written about individuals as they presented
themselves to others, but typically lack the level of detail that would be required for more
nuanced classification than a simplified “man” vs. “woman” distinction, the current project refers
to “gender” but still uses a “sex”-like binary classification.
31
women consistently rated Universalism and (to a lesser extent) Security higher
than men, especially in older (non-student) samples. As in Schwartz et al.’s (2001)
report, however, the effect sizes for these findings were small, and were not always
consistent across countries (for example, Schwartz and Rubel, p. 1022, noted, for
Self-Direction, “The sex difference (men higher) is smaller the more autonomous
versus embedded. . . and the more individualist. . . the culture of the country,” and
is “smaller. . . the richer the country”).
Thus, I expected gender to have a minimal main effect in the obituary
corpus. However, for communities that value Tradition highly overall, given the
ritual nature of obituaries as values-affirming documents intended to be read by
members of the community, it is possible that this particular corpus would show
higher gender effects than have been previously observed. Perhaps more usefully,
age and gender were allowed to interact in the models explored below, as well
as gender with community education level. These follow Schwartz and Rubel’s
(2005) call for future research to examine interactions between “[gender] and such
demographic variables as age, education,. . . and social class.”
Community level.
Median income and median education level Although education and
income14 could be aggregated into a single measure conceptually approaching
“wealth,” they were instead both included individually in the models below and
allowed to interact.
14The Gini coefficient of income inequality could be useful in these analyses, but was not
included. Instead, I chose to use median income level because obituaries are almost certain not
to reflect the full range of individuals on the wealth spectrum. Assuming that obituaries across
locations tend to be of those who are at a certain level of socio-economic status and above, it is
reasonable to expect that variability that would otherwise be related to the Gini coefficient would
not be available to a statistical model for explanation when using this corpus.
32
Schwartz et al. (2001, p. 534) found positive, significant relationships
between education (from none through college [“beyond high school”]) and “self-
direction and stimulation[,] and negative correlations with conformity and tradition
values.” I therefore expected similar effects in the corpus of obituaries.
Valuing Power, by its definition above, includes a motivation toward
“control[ling]. . . resources” (Schwartz et al., 2001, p. 5), which conceptually relates
to income. Thus, higher income was expected to be positively related to Power,
as well as possibly to Achievement (which, as above, includes a motivation to
“[demonstrate] competence according to social standards,” Schwartz, 2012, p. 5).
Ethnic/Racial demographics. Additionally, a coarse indicator of
the ethnic makeup of geographic communities will be included, in five covariates,
respectively reflecting the percentage of individuals classified as “American Indian
or Native Alaskan,” “Black or African American,” “Hispanic,” “Native Hawaiian
and other Pacific Islander,” and “White” (each in isolation or in combination with
other categories) by the United States Census Bureau in that area (at the county
level)15.
15The US Census dataset from which these data were gathered (United States Bureau of the
Census Population Division, 2015) uses variable names that sometimes conflate race and ethnicity
(e.g., “Black or African-American,” “White,” etc.). Thus, the admittedly vague phrase “racial and
ethnic groups/categories” is used throughout this report to accurately reflect the data used.
33
CHAPTER II
METHODOLOGY
Data Collection and Initial Processing
Lexical Graph Database.
WordNet. Bardi et al. (2008) composed a “value lexicon” of three
“lexical indicators” (i.e., prototype words) for each of the 10 Schwartz values
categories, for a total of 36 words (1 category title + 3 prototype words for each
of the 10 categories, with some titles overlapping with prototype words). This
Schwartz values dictionary is reprinted in Table 1.
Table 1. The Schwartz value lexicon, reprinted from Bardi et al. (2008).
Value Lexical indicators for each value
Power power, strength, control
Achievement achievement, ambition, success
Hedonism luxury, pleasure, delight
Stimulation excitement, novelty, thrill
Self-direction independence, freedom, liberty
Universalism unity, justice, equality
Benevolence kindness, charity, mercy
Tradition tradition, custom, respect
Conformity restraint, regard, consideration
Security security, safety, protection
In order to quantify the extent to which obituaries in the current project’s
corpus were concerned with each of the 10 Schwartz values, I paired Bardi et al.’s
34
value lexicon with a network-graph-based dictionary/thesaurus of the full English
language. In doing so, I enabled analyses of “lexical distance,” as explained below
(see the section defining “word-by-hop”), between words in any given obituary and
the words from Bardi et al.’s value lexicon.
Beginning in 2005, and with updates periodically through 2010 (when it
released its current version as of this writing, 3.1), Princeton University (2010)
published WordNet, a computer-readable dictionary/thesaurus of 147,478 words
grouped in 117,791 definitions. WordNet version 3.1 was downloaded in SQLite1
format from the WordNet SQL (WNSQL) project (Bou & Princeton University,
2014), which repackaged the WordNet database for ready use. The database tables
were converted into separate CSV files, which were then imported into Neo4J v2.3.3
Community Edition, an open-source graph database platform.
A traditional “relational database” (e.g., WordNet in its original form),
which models data in a collection of linked “tables” (each conceptually equivalent
to a spreadsheet, with ID numbers from one table recorded in columns in linked
subsequent tables allowing related data to be joined during queries). In contrast,
a graph database models data directly as nodes and relationships between them.
Graph databases can thus be used more straightforwardly than traditional
relational databases to find paths between nodes of different types (for example,
from a word x through a series of definitions and words to a final word, y). This
graph approach is visualized in Figure 1, which shows Bardi et al.‘s “Power” words
in green, with the words’ definitions in blue.
The import of the WordNet tables into Neo4J expanded on work by Nagi
(2013), who published a schema for representing the primary tables of WordNet
1SQLite is an SQL database format that stores the entire database in a single file, and can thus
be more straightforward than other SQL database formats to manage and inspect.
35
Figure 1. Bardi et al. (2008) value lexicon “Power” words (“power,” “strength,”
and “control;” lighter/green) with their definitions (darker/blue). Words are
connected to definitions, and some definitions are connected to other definitions.
36
in Neo4J. Nagi’s schema specifically encompassed the Words, Senses, Synsets,
Semlinks, and Linktypes tables of WordNet; the graph database constructed for
this project additionally included WordNet’s Morphs and Morphmaps tables.
Lacking consistent documentation as of this writing, WordNet’s tables are
summarized here in Table 3. The graph database constructed for this project
used three WordNet-derived relationship types (“Is a form of,” “Is related to,”
and “Is defined as”), as well as one relationship type (not used in the analyses
reported below) derived from the University of South Florida Free Association
Norms dataset (Nelson et al., 1998, see below), “Is freely associated with.” These
relationships are summarized in Table 2. The full data schema is summarized
visually in Figure 2.
Table 2. Node and relationship types in the WordNet graph database.
From Node Relationship Name To Node
Word IS FREELY ASSOCIATED WITH Word
Morph IS A FORM OF Word
Synset IS RELATED TO Synset
Word IS DEFINED AS Synset
Table 3. WordNet table names and descriptions.
Table
Name Table Description
Words Lemmas (the basest forms of words – e.g., for “is,” “be”; for
“shouted,” “shout”)
Synsets Part of speech, lexdomain reference, definition
37
Table
Name Table Description
Senses Linking table for words and synsets
Postypes Defines Part of Speech abbreviations
Semlinks Links synset meanings with each other using a list of types of
relationships
Linktypes Definitions for the types of relationships between semlinks (e.g.,
“similar,” “antonym,”2 “hypernym,” etc.)
Morphs Gives different/alternative forms (e.g., noun, verb) of words in the
Words table
MorphmapsRecords words and their different forms as written/recorded in the
Morphs table
The “Postypes” table was not imported into the graph database, but was
used when translating Part of Speech (POS) abbreviations between TreeTagger
(described below) and WordNet.
University of South Florida Free Associations Network.
In addition to WordNet, the graph database utilized Nelson, McEvoy, and
Schreiber’s (1998) University of South Florida word association, rhyme, and
word fragment norms dataset. As noted in the dataset’s introductory webpage
(http://w3.usf.edu/FreeAssociation/Intro.html), beginning in 1975, Nelson et
al. (1998) asked approximately 6,000 participants each to list one word that the
2All linktypes were included when calculating hop numbers between obituary lemmas and
value lexicon words, based on the understanding that (e.g., with antonyms), obituaries that
include words that are the opposite of a given value are nonetheless invoking that value (even
if with a negative disposition).
38
Figure 2. Visualization of Neo4J data model used in this project. Node and
relationship attributes are printed in italics.
39
participant freely associated with each of a sample of 5,019 target words. These
responses were then normed.
Because of its likely heavy dependence on the temporal and geographic
context of its participants, Nelson et al.’s dataset was not used in the analyses
described below; specifically, it was not used to generate “hop” numbers between
lemmas from the obituary corpus and each of Bardi et al.’s value lexicon lemmas.
However, because of the effort required to import the free association norms
dataset into the graph database, as well as Dr. Nelson’s generous permission to me
(personal correspondence, May 26, 2016) to republish the dataset under a WordNet
3.1 license, this dataset is included in the supplemental material made available as
part of this dissertation, and so is described here.
Legacy.com Newspaper Metadata. Legacy.com is an international
obituaries warehouser that contracts with 1,041 newspapers across 48 of the
50 United States, as well as Guam, to display newspapers’ obituaries online3.
Although, as discussed above, the copyright for a given obituary typically
remains the property of the obituary’s author, Legacy.com retains a transferable
license to republish and use the content that is displayed on its site. In February
2015, I negotiated and signed a data use agreement with Legacy.com’s Chief
Marketing Officer, who graciously allowed me to scrape and analyze obituaries from
Legacy.com’s website for this and related projects. Thus, a series of downloading
scripts was developed, first to scrape and process metadata about Legacy.com’s
3Legacy.com notes on its “About Legacy.com, Inc.” page (Legacy.com, 2016a; cf. Legacy.com’s
“Our International Newspaper Partners” page, Legacy.com, 2016c) that it partners with over
1,500 newspapers total, including newspapers in Canada, Australia, New Zealand, the United
Kingdom, and Bermuda. Legacy.com claims in the “About Us” page of their Memorial Websites
subdomain (Legacy.com, 2016b) that Legacy.com “partners with 76 of the 100 largest newspapers
in the U.S. and features obituaries and Guest Books for more than 60 percent of people who die in
the United States.”
40
domestic affiliate newspapers, and then to locally download copies of obituaries
displayed for each of those newspapers for further processing and analysis.
Newspaper Names and States Downloader. First, a downloader
was written in Bash and used to scrape the Legacy.com map of affiliate
newspapers4 in order to record the website’s internal state/location ID values.
These ID values were then used to download each of a series of state- and territory-
specific pages listing affiliate newspapers. The locally-downloaded files were
subsequently scraped for each newspaper’s name, unique Legacy.com internal ID
shortcode, Legacy.com newspaper-specific website URL, and geographic state.
The geographic distribution of Legacy.com domestic newspaper affiliates can
be seen in Figure 3.
Figure 3. Newspaper sample size by state. WY and HI had 0 newspapers that
contract with Legacy.com.
4As of this writing, the map is located at http://www.legacy.com/ns/about/newspapers/.
41
Finding Newspaper Distribution Points Using Google Maps. A
script was then written in R5 to query the Google Maps Application Programming
Interface (API) in order to retrieve all addresses available for each newspaper’s
name and state abbreviation (e.g., in response to the search term “The Register
Guard newspaper OR” for The Register Guard newspaper, which primarily
operates in Lane County, which encompasses Eugene, Oregon). This approach
was founded in the understanding that many newspapers have multiple counties of
primary coverage (for example, The Spokesman Review newspaper has main offices
in both Spokane county in Washington and Kootenai County in Idaho). Thus, this
approach allowed incorporating all addresses returned by Google Maps for each
newspaper; ideally, each of these addresses referred to a field or administrative
office for a given newspaper.
For newspapers for which Google Maps returned more than six results
(a threshold set simply by manually searching for several newspapers using the
query format given above and noting at what approximate level search result
sets became consistently filled with false positives), R’s agrep function was used
to selectively remove results. The agrep function allowed running a fuzzy (i.e.,
approximate) search among the list of returned business and place names for the
“canonical” name taken from the Legacy.com pages that listed newspaper affiliates.
maximum.distance, the argument that controlled the precision with which names
were treated as matching the canonical name, was set to 0.3 after iterative trial-
and-error tests with several high-search-result Google Maps queries (specifically,
“The Arizona Republic,” which yielded nine search results; the “Baltimore Sun,”
which yielded six search results; the “Times-News,” which yielded 20 search results;
5The version numbers of all R packages used in this project that are not presented in-text can
be found in the Appendix of this report.
42
and “The Birmingham News,” which yielded 14 search results). This value was
chosen with the intent to run searches that were conservative but not overly
restrictive.
In these cases of high search results, a three-step process was used:
1. Always accept the first result (this was based on the assumption that Google
Maps returned results ordered by relevance).
2. Search among the remaining returned business/place names for the canonical
name of the newspaper, as defined using Legacy.com’s affiliate newspaper list.
3. Filter for unique place names in the remaining Google Maps search set.
Each result’s address, Google Maps place name, latitude, longitude,
Legacy.com newspaper code, and date of download were recorded.
Of the original sample of 1,041 newspapers, 1.441% (15 newspapers)
returned no search results, and thus were excluded from further analyses. Thus,
1,026 newspapers were included in the subsequent analyses. Excluded newspapers
were not isolated to a single geographic region, as is shown in Table 4.
Table 4. State Abbreviations and Names of newspapers that did not return address
results from the automated Google Maps query procedure and thus were excluded
from further analyses.
State Abbreviation Newspaper Name Shortcode
CA Pomerado News pomeradonews
CA Inside Bay Area insidebayarea
IL Chicago Sun-Times chicagosuntimes
IL Daily Southtown daily-southtown
MA The Country Gazette - Plainville wickedlocal-plainville
43
State Abbreviation Newspaper Name Shortcode
MA Falmouth Bulletin wickedlocal-falmouth
MA Waltham News Tribune WalthamNewsTribune
MI Muskegon Chronicle muskegon
NJ South Jersey Times southjerseytimes
OH The Marion Star marionstar
OR Hillsboro Argus/ Forest Grove Leader /
Beaverton Leader
argus
PA Susquehanna County Independent independentweekender
SC The Cheraw Chronicle thecherawchronicle
VA RichmondObitNews.com richmond-VA
WI Kenosha News kenoshanews
Finding United States Census County Codes for Newspaper
Locations Using the Federal Communications Commission’s (FCC’s)
Data Conversions API. I then submitted the latitude and longitude of each
unique recorded Google Maps search result to the Federal Communications
Commission’s (FCC’s) Data Conversions API6, a service provided by the FCC to
translate location data such as latitude and longitude into United States Census
Federal Information Processing Standard (FIPS) codes, unique numeric identifiers
used by the United States Census Bureau for places down to the county level.
FIPS codes are embedded in many US Census Bureau datasets and, further, are
standardized geographically and temporally, unlike ZIP codes, the boundaries
6As of this writing, the API can be found at https://www.fcc.gov/general/census-block-
conversions-api.
44
of which can be amended by the US Postal Service as postal requirements of a
community change (put differently, FIPS codes seemed more appropriate than
ZIP codes for use in this project given the relative stability of the former). The
FCC’s Data Conversions API provides a 15-digit FIPS code for a given latitude and
longitude, which includes the following geographic identifiers:
– State (2 digits)
– County (3 digits)
– Census tract (6 digits)
– Block group (1 digit)
– Block (3 digits)
Thus, with an interest in the county level, the first five digits (State +
County) of each location’s FIPS code were recorded for further analyses.
Figure 3, a map displaying newspaper sample size by state, could be re-
visualized with these county-level results as in Figure 4. Given the range of
newspapers by county, Figure 4 is accompanied by Figure 5, which provides a
histogram of newspaper frequency by county. Figure 3 is also accompanied by
Figure 6, a visualization of US population by county in 2014, the latest year for
which data was available at the time of writing (data is from the US Census
Bureau Population Division’s “Annual County Resident Population Estimates by
Age, Sex, Race, and Hispanic Origin: April 1, 2010 to July 1, 2014” dataset, United
States Bureau of the Census Population Division, 2015).
45
Figure 4. Newspaper sample size by county. WY and HI had 0 newspapers that
contract with Legacy.com.
46
Figure 5. Histogram of newspaper sample size by county.
47
Figure 6. The United States’ population by county in 2014, using the US Census
Bureau Population Division’s “Annual County Resident Population Estimates
by Age, Sex, Race, and Hispanic Origin: April 1, 2010 to July 1, 2014” dataset
(United States Bureau of the Census Population Division, 2015).
48
Legacy.com Obituaries Downloader. A Bash script was written to
use the Legacy.com newspaper-affiliate-specific URLs gathered as described above,
appending to them query arguments to request the newspaper’s “Last 7 Days”
of results. Results from these requests were returned from Legacy.com’s servers
in pages of up to 10 results, and were wrapped in HTML <entryContainer>
elements. Page numbers in the URLs being queried were incremented by 1 (e.g.,
“page=1”, “page=2”, etc.) until the most recently downloaded page contained
no results (defined as a lack of <entryContainer> elements). The final (empty)
downloaded page was then discarded, and the process re-started for the next
newspaper. Downloads were rate-limited (all queries spaced two seconds apart)
to minimize the intrusiveness of the process on Legacy.com’s servers, and were
launched during weekend nights. Downloads were run over seven consecutive weeks,
on the following dates:
– Saturday, April 16, 2016
– Sunday, April 24, 2016
– Saturday, April 30, 2016
– Saturday, May 07, 2016
– Friday, May 13, 2016, to Saturday, May 14th, 2016
– Friday, May 20, 2016, to Saturday, May 21, 2016
– Friday, May 27, 2016, to Saturday, May 28, 2016
49
Combined, these downloads yielded 34,640 search result pages, totaling
9.1 GB of HTML data (i.e., the not-yet-scraped obituaries, as well as all of their
HTML markup), including 313,607 obituaries7.
Legacy.com Obituaries Data Scraper. Using R, a scraper was
developed to take the raw HTML content of each locally-downloaded search results
pages and, for each result, extract the name of the deceased, the obituary text,
the date of download, the URL of the individual obituary, and the name and
state of the publishing newspaper. In addition, the scraper included algorithms for
automatically coding the gender and age of the deceased (as necessary, calculating
the latter by guessing date of birth from the information included in the obituary).
Reliability statistics for these algorithms are included below.
Development of the Gender-Guessing Algorithm. The gender-
coding algorithm utilized a two-pass approach to infer the gender of the deceased
from the obituary text. Given the limited information available in obituaries
and the fact that obituaries are very rarely written by the deceased and do not
typically contain direct statements of gender identity, gender was coded using three
categories: “female,” “male,” and “uncertain.”
In the first pass, the gender algorithm counted the number of masculine
(“he,” “his,” “him”) and feminine (“she,” “her”) pronouns in the obituary text. If
the obituary text contained more masculine than feminine pronouns, the gender of
the deceased was coded as “male;” if more feminine than masculine pronouns were
counted, the gender of the deceased was coded as “female.”
In cases in which there were no gendered pronouns, or an equal number
of masculine and feminine pronouns, the gender algorithm undertook a second
7It should be noted that, given date overlaps, this number includes some obituaries that were
downloaded more than once. Pruning of obituaries is described in the Results section below.
50
pass, in which it attempted to use historical Census data to guess gender based
on the forename of the deceased. The gender package for R was used for this
purpose. While the package’s gender function can tailor its guesses based on the
approximate date of birth of the deceased, in order to keep the gender- and age-
guessing algorithms separate from one another (such that imperfections in one
would not unduly affect the output of the other), no date information was passed
to the gender function, causing it to make a guess using Census data from across
the timespan encompassed in its dataset.
In all cases, the gender-guessing algorithm recorded the source of its final
output (pronouns vs. pronouns followed by forename), to facilitate debugging
during the development process. Additionally, the gender algorithm was
constructed to find and strip leading initials (e.g., “J. Henry Doe”), which were
not usable by the gender Census-lookup function, from all names before sending
them to the function.
Development of the Age-Guessing Algorithm. In order to guess
age at death (or, as necessary, to calculate it by guessing year of birth), the age-
guessing algorithm completed four separate searches through the text of each
obituary and combined the information gathered from each to code age:
1. The lowest four-digit number (as a possible year of birth) in the obituary text
2. The first instance (if available) of the phrase “born. . . x,” where x was a
four-digit number.
3. A one- to three-digit number (as possible age at death) after a space and
before a comma or period (for phrases such as “John Doe, 82, passed
away. . . ”). Candidate numbers needed to meet the following criteria:
51
1. Not be immediately preceeded by any long- (“January”) or short-form
(“Jan”) month name or a one- to four-digit number followed by a space
or forward-slash (e.g., “10/12,” which likely signified a date rather than
an age).
2. Be immediately preceeded by a space.
3. Be immediately followed by either one or more spaces or a comma.
4. Not be immediately followed by zero or more spaces followed by “AM,”
“am,” “PM,” “pm,” or any combination of these abbreviations using
uppercase and lowercase letters. This search also checked for versions
of these abbreviations that included periods (e.g., “p.m.”). This search
avoided recognizing numbers in phrases such as “5 pm”) as candidate
ages.
5. Not be immediately followed by whitespace or a comma and then the
year of the obituary’s download (2016). This search avoided recognizing,
e.g., the “9” in “April 9, 2016,” which was assumed to be more likely
a date of death or of scheduled funerary services. This search had the
notable downside of systematically missing children who died as infants
at less than one year. This decision was made after reading several
hundred obituaries manually and seeing few infant obituaries, especially
with a date of birth printed explicitly.
4. A two- to four-digit number that might be a year in the US-common date
form of MM/DD/YY[YY] or MM-DD-YY[YY].
The age algorithm then removed any candidate year of birth numbers (not
candidate age at death numbers) higher than the year of download (e.g., 2016) or
52
lower than the year of download minus 110, which was expected to be a liberal
estimate of the highest age at death to be found in the corpus. Since no individuals
were expected to be older than 1108, potential age at death numbers higher than
110 were thus assumed to be misidentified, and were discarded.
The algorithm then considered all of the gathered candidate years for each
obituary, and for each obituary, took the lowest candidate year as the likeliest year
of birth. This choice was based on the assumption that, since obituaries typically
craft a narrative around the life of the deceased, the earliest date was likely to be
the correct one to use.
Following the rationale described above for systematically excluding guesses
for infant children less than one year old at death, the age algorithm declined to
make an age guess if the only guess to make was the year of download (e.g., 2016).
In cases in which the algorithm made a guess for year of birth but not for
age, approximate age was calculated using the equation
year of publication− year of birth = approximate age
Validation of the Gender- and Age-Guessing Algorithms. The
algorithms for categorizing gender and age did not need to be perfect; they did,
however, need to be approximately as accurate as would be expected from a human
coder working as part of a team (for example, a Research Assistant).
8Following a sample in Spain, Go´mez-Redondo and Garc´ıa Gonza´lez (2010, p. 164-165)
concluded that “there is no relationship between the population size of the regions or provinces,
and the number of supercentenarians[, individuals who are 110 years or older].” With that in
mind, it can be useful to consider historical rates of supercentenarianship. Kestenbaum and
Ferguson (2010) found 325 individuals who had died at age 110 or greater between 1980 and
2003, a span of over two decades in a country of a minimum of 226.5 million (at the beginning
of that span, in 1980, as reported by Forstall & United States Bureau of the Census Population
Division, 1996).
53
Using a “development” corpus of 27,923 additional obituaries that were
downloaded and scraped separately by approximately two weeks from the larger
“production” corpus (on June 11th, 2016; this “development” corpus was not
included in the final analyses reported below; rather, it was downloaded and
scraped specifically for use in the code development required to enable those
analyses), I randomly selected and manually coded 50 obituaries for gender and
age or date of death (whichever was easiest given the information presented in each
obituary) before looking at the guesses made by the automated algorithms. This
allowed me to see initially whether the algorithms needed major adjustment. Seeing
that inter-rater reliability estimates were high with this sample of 50 (following
calculations described below), I took a new random sample (with replacement) of
100 obituaries. This new sample was coded separately (and, as before, without
knowledge of the automated algorithms’ output) by two coders: a member of the
Dissertation Committee (Mark Alfano) and me.
Initial Data Processing Steps for Calculating Reliability
Statistics. In order to increase the meaningfulness of the reliability statistics to
be computed, the matrix of automatically- and manually-coded genders and ages
was first processed using the following steps. These steps were performed separately
between each of the two humans’ codes and the automated algorithm’s codes.
1. I replaced all blank values with 0s, in order that rows in which the computer
or the human coder did not have enough information to make a guess would
not be skipped in the calculations. This had the effect of penalizing cases in
which the human coder made a guess and the automated algorithm did not,
54
and rewarding cases in which both the computer and the human coder agreed
that there was no basis for a guess.9
2. For rows in which the automated algorithm and human coder’s age guesses
were within one year of each other, the two guesses were considered the same
(by setting them equal to one another). This step was warranted because for
both the human and computer coders, age guesses were sometimes calculated
from a guess of year of birth. In those cases, age was calculated as “2016
[the current year] - year of birth”. This equation yields a result one year off
in cases in which the birthday of the deceased has not yet occurred in the
current year.10
Gender Coding Reliability. Cohen’s κ, computed between the
automated algorithm and me was 0.886, and was 0.881 between the automated
algorithm and Mark Alfano for gender codes. Percent agreement was 94% between
the automated algorithm and each of the human coders.
Age Coding Reliability. Cohen’s κ for age codes between the
automated algorithm and Mark Alfano was 0.803, and was 0.865 between the
automated algorithm and me. Respectively, percent agreement was 81% and 87%
between the automated algorithm and the two coders. Respectively, the correlation
of age guesses between the automated algorithm and the two coders was 0.760
9This processing step had the downside of making Pearson’s correlation statistics less
meaningful in this context (nevertheless, Pearson’s correlations are still reported below, for
completeness), in cases in which the computer made a guess and the human coder did not, or
vice versa.
10For example, if an individual’s date of birth were February 20th, 1987, and the person were 28
years old at the time of death, this equation would yield a date of birth one year off (2016 - 1987
= 29), until after February 20th (when the individual would have turned 29). I considered the
potential one-year difference acceptable for the analyses described later in this report; for assessing
reliability of the coding algorithm, however, it seemed necessary to allow a one-year tolerance
when evaluating computer and human age guesses as matching.
55
and .830. While this statistic is reported for completeness, it is not as useful as
the above statistics in this context, given that the data processing steps described
above could cause disproportionate decreases in any case in which the automated
algorithm guessed the age of the deceased in an obituary that a human coder
decided did not contain sufficient information to support a guess, or vice versa.
Filtering of Obituaries by Word Length. Obituaries published
in the USA follow two primary formats (defined as endpoints on a continuum).
First, some obituaries comprise short-form death notices, which often contain
the deceased’s name and age or date of birth, but no or almost no biographical
content. These short-form obituaries seem primarily motivated to serve as public
notices and to provide logistical information about funerary rites for the deceased
(e.g., the date and time of a wake or funeral, locations to which to send flowers,
and the name of the managing funeral home). Short-form obituaries of this type
might also contain limited familial information (using phrases such as “Survived
by. . . ”). Long-form obituaries, which might be several hundred words in length, by
contrast, contain all of the information expected in a short-form obituary, but also
include a richer biographical sketch of the deceased. In some newspapers, such as
The New York Times, long-form obituaries can be commissioned and professionally-
authored, reading much more like biographies possibly to the exclusion of more
detailed logistical or familial information. Many obituaries, however, are written
by friends or family members of the deceased (and occasionally by the deceased
him- or herself in advance of death, perhaps during a long illness), either following
templates provided by the publishing newspaper or an assisting funeral home
(supplying unique biographical information but in a consistent format), or
composed wholly by the author.
56
I expected that short-form obituaries, as death notices, would likely
not contain information useful for understanding values that might differ from
community to community, or from author to author; short-form death notices
can be thought of as demonstrating shared values about notifying community
members about deaths, sharing with community members in funerary rites, and
enumerating ties to still-living relations, but these values are, by the very fact
that obituaries are published across the USA, not likely to vary or be unique to
particular communities nearly as much as richer biographical information. For this
reason, I filtered obituaries from the corpus gathered over the seven-week recording
period to contain as few short-form death notices as possible. Since short-form
necrologies are defined in part by their length, obituaries were filtered based on
word count. It should be noted that this filtering decision was quite granular: it is
possible, for example, that an obituary could comprise a brief biographical sketch
and eschew any logistical funeral information in such a way that it would not solely
be a death notice but would still be excluded from the final corpus based on the
word-count filter. I made the decision to use word count as a useful indicator
simply from having read many obituaries manually and not remembering having
seen any short-form obituaries that did not contain only logistical and basic familial
information. While constructing a more advanced algorithm for pruning obituaries
could be useful fodder for future research, the decision can be seen as valid for
this project in its pragmatic, bottom-up approach (from manual reading), which
enabled progression towards the project’s larger questions and goals.
In order to set a word-count threshold below which to exclude obituaries
from further analyses, I qualitatively examined a series of obituary word-count
histograms of successively finer granulation (two of these are reproduced here as
57
Figures 7 and 8). An initial histogram of word-counts across the corpus showed
a notable drop-off in obituaries just below the 100-word mark. Creating further
histograms with smaller x-axis ranges and finer-grained bins suggested that a
qualitative change might take place in obituaries (from the highly-templated and
thus consistent-length short-form death notices to longer-form obituaries in which
authors wrote about the deceased in more detail and with more idiosyncrasy) at
lengths somewhere between 40 words and 120 words.
I thus took random samples (for each sample, n = 30) of obituaries within
ranges of 10 words, starting with the range of 21 to 30 words, and manually read
every obituary in the sample to look for non-templated biographical content. The
intent of this approach was to stop sampling in this way and set a filter cutoff as
soon as I began to see biographical information (even in just one or two obituaries
in these small samples) in order to set the cutoff as conservatively as possible. This
approach relied on the assumption that a random sample of 30 obituaries (from
a larger corpus of several thousand for that 10-word range) would be sufficient to
manually detect this qualitative change in the corpus from all or almost all short-
form death notices to longer-form biographical obituaries. While this approach was
admittedly imprecise, it was also highly functional. After examining samples of
obituaries in the ranges of 21 to 30 words, 31 to 40 words, 41 to 50 words, 51 to 60
words, and greater than 100 words, I set the filter cutoff at 60 words.
Construction of a Matrix of Lemma - Part-of-Speech Distances
from Bardi et al.’s Dictionary. Of the original 313,607 obituaries in the corpus
gathered during the five-week collection period, 211,963 obituaries were above the
58
Figure 7. Histogram of obituary word counts, cut off at 600 words (the actual
maximum word count in the corpus was 3,502 words, at the end of a long
positively-skewed tail in the full histogram’s distribution).
59
Figure 8. Histogram of obituary word counts, cut off at 100 words (the actual
maximum word count in the corpus was 3,502 words, at the end of a long
positively-skewed tail in the full histogram’s distribution).
60
minimum word count cutoff, and were thus retained for further analyses11. For each
of these remaining obituaries, the following steps were taken:
1. The text of the obituary was lower-cased (e.g., “John Doe was very brave”
became “john doe was very brave”) in order to normalize text across
obituaries to correct for possible inconsistencies in capitalization, whether
from typos or noun usage (e.g., “. . . was a member of the Methodist church”
[a particular church building in an area] vs. “. . . was a member of the
Methodist Church” [an overall religious organization]).
2. All words in the obituary were tagged for Part of Speech (POS) using
TreeTagger (Schmid, 2013), an open-source, cross-platform language tagging
and processing tool.
3. TreeTagger’s POS dictionary (i.e., the list of Parts of Speech it had available
for use) was substantially more detailed than WordNet’s. Thus, in order to
enable compatibility with the WordNet graph database (the usage of which is
described below), TreeTagger POS tags were “translated” into WordNet POS
tags, using a conversion table (Table 5).
Table 5. TreeTagger to WordNet Part Of Speech (POS) conversions.
TreeTagger POS WordNet POS
verb v
adjective a, s
adverb r
11The “development” dataset was similarly cleaned: of its initial 23,946 rows, 33.241% were
removed for being under the minimum word length, leaving 15,986 obituaries.
61
TreeTagger POS WordNet POS
noun n
The TreeTagger “adjective” tags were given two matching WordNet tags,
signifying WordNet’s “adjective” and “adjective satellite” categories. In the
WordNet query used in the steps below, these two tags were combined in an “OR”
statement (e.g., “match any word in the WordNet database that is tagged as an
adjective OR as an adjective satellite.”).
4. TreeTagger was used to “lemmatize” all of the words in the obituary. Unlike
coarser “stemming” methods (e.g., using the Porter stemming algorithm12
– see, e.g., Porter, 1980), which normalizes words in a corpus by removing
endings (changing “speaking” and “speaker” into “speak,” and “happy”
and “happiness” to “happ”), “lemmatizing” seeks to change a word into its
most basic form, given its identified POS (e.g., for “saw” as a noun, “saw;”
for “saw” as a verb, “see;” for “was” as a verb, “is”). Because it requires
knowledge about POS, lemmatizing is more difficult than stemming, but also
potentially more useful for its increased nuance (assuming the accuracy of the
POS tags that it uses).
5. Words of the following Parts of Speech were removed from the list of
lemmatized words, on the assumption that they would not be relevant for
calculating distance from Bardi et al. words:
12As of this writing, the official website for the Porter stemming algorithm is at http://
tartarus.org/~martin/PorterStemmer/. The citation provided here is for one of the algorithm’s
canonical publications.
62
1. “Cardinal numbers” (e.g., “76”) were excluded, since the “lemma” value
of each was replaced by TreeTagger with the string “@card@”, and so
was not expected to contribute to this project’s analyses.
2. “Modal” lemmas (e.g., “will” in “will follow”) were also excluded on
the expectation that they would also not be relevant (especially as
compared to the verbs they were paired with) for this project’s analyses
of distances from value lexicon words.
6. Stopwords were removed, following the R tm (“TextMining,” v. 0.6.2)
package’s standard English stopword list: “i”, “me”, “my”, “myself”, “we”,
“our”, “ours”, “ourselves”, “you”, “your”, “yours”, “yourself”, “yourselves”,
“he”, “him”, “his”, “himself”, “she”, “her”, “hers”, “herself”, “it”, “its”,
“itself”, “they”, “them”, “their”, “theirs”, “themselves”, “what”, “which”,
“who”, “whom”, “this”, “that”, “these”, “those”, “am”, “is”, “are”, “was”,
“were”, “be”, “been”, “being”, “have”, “has”, “had”, “having”, “do”, “does”,
“did”, “doing”, “would”, “should”, “could”, “ought”, “i’m”, “you’re”, “he’s”,
“she’s”, “it’s”, “we’re”, “they’re”, “i’ve”, “you’ve”, “we’ve”, “they’ve”, “i’d”,
“you’d”, “he’d”, “she’d”, “we’d”, “they’d”, “i’ll”, “you’ll”, “he’ll”, “she’ll”,
“we’ll”, “they’ll”, “isn’t”, “aren’t”, “wasn’t”, “weren’t”, “hasn’t”, “haven’t”,
“hadn’t”, “doesn’t”, “don’t”, “didn’t”, “won’t”, “wouldn’t”, “shan’t”,
“shouldn’t”, “can’t”, “cannot”, “couldn’t”, “mustn’t”, “let’s”, “that’s”,
“who’s”, “what’s”, “here’s”, “there’s”, “when’s”, “where’s”, “why’s”, “how’s”,
“a”, “an”, “the”, “and”, “but”, “if”, “or”, “because”, “as”, “until”, “while”,
“of”, “at”, “by”, “for”, “with”, “about”, “against”, “between”, “into”,
“through”, “during”, “before”, “after”, “above”, “below”, “to”, “from”, “up”,
“down”, “in”, “out”, “on”, “off”, “over”, “under”, “again”, “further”, “then”,
63
“once”, “here”, “there”, “when”, “where”, “why”, “how”, “all”, “any”,
“both”, “each”, “few”, “more”, “most”, “other”, “some”, “such”, “no”, “nor”,
“not”, “only”, “own”, “same”, “so”, “than”, “too”, “very”
7. Punctuation was removed using the following list: “.”, “,”, [single quote],
[double quote], “;”, “:”, “!”, “[”, “]”, “{”, “}”, “/”, “ ”, “&”, “$”, “#”, “!”,
“(”, “)”.
8. For each lemma-POS combination (e.g., “see” used as a verb), the WordNet
graph database was queried to find the shortest distance through the graph
between the lemma and each of the 36 words in Bardi et al.’s dictionary.
Candidate paths were capped at 15 total hops (counting all node types).
Within this 15-hop cap, separate counts were performed for the following
(all from the same shortest path):
1. Total number of hops
2. Number of hops only counting Word nodes
3. Number of hops only counting Morph nodes
4. Number of hops only counting Synset nodes
5. Number of Synset-to-Synset relationships
6. Number of Word-to-Synset relationships
7. Number of Morph-to-Word relationships
8. Number of Free Association relationships (this was expected to be zero
in every case, given the decision to exclude IS FREELY ASSOCIATED WITH
relationships from the current project’s analyses, and was included
64
to both check that expectation, and to provide ready code for future
projects that release this constraint)
Put differently, in the path x:Word --> y:Synset --> z:Word -->
a:Synset --> b:Word (where “x:Word” means “a node, x, which is a Word”),
the count of Word nodes between x and b would be 2, and the count of total hops
would be 4. The cap of 15 total hops was based on an arbitrary but arguably
common-sense decision that words more than 15 hops distant from each other
would not be reasonably defined as “related” in a meaningful way in the context
of this project13. The lemma’s POS tag was used to filter the first Synset (i.e.,
definition) encountered in the shortest path (in plain language, this aspect of the
query would have read, “Find the shortest path between ‘saw’ and ‘benevolence,’
where the first Synset node encountered is tagged as a ‘verb.’ ”14 Lemmas that were
not successfully POS-tagged by TreeTagger were also queried, without the first-
Synset filter (allowing any POS/definitions in the first hop). Since the words in
Bardi et al.’s dictionary were not POS tagged by Bardi et al., their definitions in
the graph were not filtered during these queries. The Bardi et al. words were not
13Although in this case doing so would be preferable, as of this writing, Neo4J is not capable
of filtering to, e.g., 15 Word hops specifically, without setting a maximum constraint on
total number of hops at an arbitrarily larger threshold. Setting such a threshold would be
computationally expensive without necessarily succeeding in allowing the desired maximum
number of Word hops, e.g., if a network path between a lemma and a value lexicon word
contained enough Synset-to-Synset relationships.
14Unlike the constraint imposed on the first Synset encountered in the shortest path between a
given obituary lemma and a target value lexicon word, I decided not to constrain the final Synset
encountered before a value lexicon word to be a noun (i.e., the POS type of all of the value lexicon
words: per Bardi et al., 2008, p. 485, “In order to hold part of speech constant across different
values, we considered only nouns. For example,for the value Hedonism, the noun pleasure was
used rather than the adjective pleasing.”). While Bardi et al’s value lexicon words are all nouns,
given the exploratory aspect of this project, I decided not to impose this constraint in order that
traversal through the network initially be allowed to be liberal rather than conservative (put
differently, to allow more rather than fewer connections). However, POS of the final Synset was
recorded when computing shortest paths, in order to explore whether the final Synset POS was
notably variable without this constraint.
65
lemmatized during this graph query process, as they were understood to have been
published as lemmas.
The graph database query was parallelized using an Amazon Web Services
Virtual Machine running Ubuntu Linux 14.04 with 40 virtual CPUs and 160GB of
memory. Query results were appended to a 12-column matrix:
1. From (the lemma in the corpus)
2. To (the target value lexicon word)
3. Part of Speech of first hop Synset (the POS of the lemma as it was used in
the corpus)
4. POS of final Synset encountered (used to check whether not constraining the
final Synset to be of a noun POS was likely to affect analysis outcomes)
5. Total number of hops in the shortest path
6. Shortest number of hops, only counting Word nodes
7. Shortest number of hops, only counting Morph nodes
8. Shortest number of hops, only counting Synset nodes
9. Number of Synset-to-Synset relationships in the shortest path
10. Number of Word-to-Synset relationships in the shortest path
11. Number of Morph-to-Word relationships in the shortest path
12. Number of free association relationships in the shortest path (this was used
to check the constraint that free association relationships not be allowed in
queries for this project, as explained above)
66
Lemma-POS combinations that did not connect to any Bardi et al. words
were appended with a null value in all count columns, in order that they not
be looked up again from subsequent obituaries.
Calculation of Schwartz Value Distances from Lemma - Bardi et
al. Word Distances. The final matrix of terms showed that the filtered obituary
corpus comprised 98,499 unique lemma-POS combinations, with 1,148,983 total
rows (each for a unique lemma-POS-Bardi et al. word combination)15. The matrix
of these combinations was grouped by the Schwartz values of the value lexicon
words and aggregated using a new calculation, “word-by-hop,” which is defined
below.
15The “development” dataset contained 42,784 unique lemma-POS combinations, with 635,124
total rows (each for a unique lemma-POS-Bardi et al. word combination).
67
CHAPTER III
RESULTS
Further Cleaning and Pruning the Dataset
Beyond the initial data “cleaning” steps described above in the Methods
section, I took several steps to better organize the dataset to make it more ready
for use in analyses.
Pruning Within-Newspaper Duplicate Obituaries. As noted
above, overlap in obituary download dates caused some obituaries to be
downloaded twice. In addition, several obituaries were downloaded repeatedly
from the same newspapers because they appeared over time in multiple forms
in Legacy.com’s search results. Several of these forms are non-exhaustively listed
below:
– “Teaser” obituaries, which seemed to serve as placeholders for full,
biographical texts, but presented only logistical death notice information
(e.g., time and location of funerary services). These obituaries also sometimes
contained a note that the full obituary would be published on a given date.
– Variants of the same text, but including a nickname of the deceased (e.g.,
“John Doe. . . ” vs. “John ‘Johnny’ Doe. . . ”)
– Variants of the same text, but including spelling corrections (especially for
place names).
To remove these duplicates, I grouped obituaries by newspaper and
obituary text (i.e., condensing by exactly-matching obituary text strings)1. This
1I compared this method to grouping by newspaper, obituary text, and name of the deceased.
The two methods produced a difference of 218 rows. I then compared the names of the deceased
68
removed approximately 16% of the original dataset’s rows, leaving 177,272 rows.
I then further condensed the dataset using the URL of each obituary (as each
obituary listed in a Legacy.com search result page contains a link to a unique
page comprising only that obituary). 1.66% of the remaining dataset contained
a duplicate URL (e.g., from the “teaser” obituaries mentioned above). After
confirming that no URLs were duplicated across newspapers (cross-newspaper
duplicates are discussed below, as a separate area of consideration), for each URL
in the dataset, I retained the obituary with the longest word count, assuming it to
be the most updated version of the obituary text. This removed an additional 1,482
rows (0.836% of the remaining dataset) from the dataset (a portion of the 1.66%
mentioned above), leaving 175,790 rows2.
Removing Obituaries of Individuals of Unknown Age. Of the
remaining dataset, 6.256% did not contain sufficient information for the automated
age-coding algorithm to make a guess. Given the relatively low percentage of the
dataset lacking an age guess, and the intention to use age as a predictor in the
regression models below, I excluded these data from further analyses, rather than
attempt (likely problematically) to impute age values. This left 164,792 rows in the
dataset3.
for each pair of rows that contained matching obituary text but non-matching names (in order to
confirm that obituaries of two separate people had not both been written from similar templates,
e.g.), and found that in all cases, the paired obituaries qualitatively obviously referred to the same
individuals, differing only in punctuation, capitalization, or other minor differences in the printed
name of the deceased.
2The “development” dataset was similarly cleaned: of its remaining rows, 0.181% were
removed, leaving 15,957 rows (zero rows were removed from the development dataset for
containing duplicate URLs).
3The “development” dataset was similarly cleaned: of its remaining rows, 5.634% were removed
for being of uncertain age, leaving 15,058 rows.
69
Removing Obituaries of Individuals of Uncertain Gender. Of
the remaining dataset, 0.439% (724 rows) were assigned a gender category of
“uncertain” by the automated coding algorithm. Given the low percentage of data
in this third category (vs. “female,” which 48.867% of the remaining dataset was
coded as; and “male,” which 50.694% of the remaining dataset was coded as), I
excluded those rows from further analyses. This left 164,068 rows in the dataset4.
Assessing Across-Newspaper Duplicate Obituaries. A frequency
table of obituary text strings (i.e., collapsing exactly-matching obituary texts and
counting the frequency of each) revealed that 20.294% of the remaining dataset
were duplicate obituaries across newspapers (within-newspaper duplicates having
been removed following the description above; a check was also performed at this
step to confirm that no within-newspaper duplicate obituary texts remained in the
dataset).
These duplicate obituaries were not isolated to only a small number of
newspapers: 393 papers in the sample printed at least one duplicate obituary. This
is not conceptually problematic: it is understandable that an obituary be published
in more than one newspaper, depending on the notoriety or width of the social
circle of the deceased, and/or the number of newspapers operating simultaneously
in the same geographic region. However, from an analysis perspective, this overlap
required consideration. Steps taken to accommodate these duplicate obituaries are
described further below.
4The “development” dataset was similarly cleaned: of its remaining rows, 0.578% were removed
for being of uncertain gender, leaving 14,971 rows (before removing these unknown rows, the
percentages of obituaries coded as female vs. male were 49.416% and 50.001%, respectively).
70
Analyses
Calculation of a Measure of Network Distance, “Word-by-
Hop”. Because not all lemmas contained in the obituary corpus were related
to all Schwartz values (or, more specifically, to all of the words from Bardi et
al.’s [2008] value lexicon), it was necessary to calculate a numeric value for each
word’s relationship with each of the value lexicon words that was able to represent
“no relationship.” Number of hops through the WordNet graph, while intuitive
to use when answering Research Question 1, was inapposite for use especially in
the regression context of Research Question 2, since it was only able to represent
a lack of relationship (i.e., no path within 15 hops between two lemmas) with
missing values; “0” hops, rather than signifying no relationship, signified a perfect
relationship, that the lemmas being compared were identical. Thus, in a regression
context, lemmas that were unrelated to given Bardi words would need to be
either a) excluded from analyses in either a listwise or pairwise fashion, or b)
imputed with an arbitrary value (such as 100). Neither of these solutions seemed
satisfactory; therefore, a new calculation, called “word-by-hop5,” was used.
Word-by-hop values are not meaningful in isolation, but can be compared to
one another to determine the relative fit between different pairs of lemmas. Larger
word-by-hop values represent better “fit” with a Schwartz value, where fit is defined
as “greater connection,” rather than solely by “fewer hops.”
Final Calculation Formula for Word-by-Hop. This report first
presents the final calculation for “word-by-hop,” in order that it be presented as
early as possible. Following that, it explains the step-by-step development of the
calculation.
5“Word-by-hop” is a shorthand for “Words divided by median hops.”
71
The equation for word-by-hop is as follows:
word-by-hop =
Number of words that have a path to Value v
Total number of words in the obituary
1 + Median number of hops of the connected words to Value v
Word-by-hop has boundaries at [0, 1]6.
Initial Calculation of Word-by-Hop. In its first stage of
development, word-by-hop was initially calculated by counting the total number of
lemmas (including repeats) in an obituary that have any path to a given word from
Bardi et al.’s value lexicon (i.e., in this project, from 0 to 15 hops) and dividing
that count by one plus the median number of hops in each lemma’s shortest path
to the given word. In cases in which there are no connections, word-by-hop was set
to 0.
The addition of one to the median number of hops corrects for cases in
which there is complete overlap between obituary and value lexicon words (i.e.,
a median hop number of 0) – rather than dividing by 0 or a fraction less than 1
(which would multiply the count of words unlike for non-zero median hop values),
an ideal word-by-hop value is simply not penalized, by dividing by 1.
An example word-by-hop calculation between a hypothetical obituary that
comprises only a single lemma and three Schwartz values is illustrated in Tables 6
and 7.
Table 6. Example data showing the number of hops through WordNet between
“business” and words for three Schwartz values from Bardi et al’s (2008) value
lexicon
Lemma in Obituary Value Lexicon Word Number of Word Hops
business novelty 6
6The notation is meant to denote a “closed” interval, i.e., one that includes its endpoints.
72
Lemma in Obituary Value Lexicon Word Number of Word Hops
business excitement 5
business thrill 4
business charity (No path)
business kindness (No path)
business mercy (No path)
business tradition (No path)
business custom 7
business respect 4
Table 7. Example implementations of the initial word-by-hop calculation, using
data from Table 6.
Schwartz Value word-by-hop calculation word-by-hop value
Stimulation 3 words ÷ (1 + median[6, 5, 4]) hops 0.500
Benevolence 0 words 0
Tradition 2 words ÷ (1 + median[7, 4]) hops 0.308
Table 7 illustrates that the hypothetical obituary (which, in this example,
contains only one word, “business”), is more connected to Stimulation (word-by-
hop = 0.500) than to Tradition (word-by-hop = 0.308), and that it is not connected
to benevolence (word-by-hop = 0), using this initial word-by-hop calculation.
73
This process can be aggregated over as many lemmas in the obituary that
have a path to words from the value lexicon. Other, more informal, example
calculations illustrate the properties of this initial version of word-by-hop:
– For a given Schwartz value, 1 connected word ÷ (1 + median 3 hops) =
.250, whereas 5 connected words ÷ (1 + median 3 hops) = 1.250 (a higher
value). Thus, the calculation considers obituaries that contain larger numbers
of connected words “more connected” to a Schwartz value than those that
contain smaller numbers of similarly-connected words.
– For a given Schwartz value, 7 connected words ÷ (1 + median 3 hops) =
1.75, whereas 4 connected words ÷ (1 + median 2 hops) = 1.333. Thus, the
calculation considers obituaries that contain a larger number of connected
words at a slightly larger median graph distance “more connected” to a
Schwartz value than those that contain a smaller number of connected words
at a slightly smaller median graph distance. This is unlike using median hop
count only, as hop count would prefer the latter case over the former.
This word-by-hop approach is partially conceptually similar to Bardi et al.’s
(2008), which included a preference for more vs. fewer indicator words (Bardi et
al. only counted pages in their newspaper corpus that included all three indicator
words for a given value).
The initial word-by-hop calculation defined above required several
corrections for use in the current project, however, as described below.
Correcting Word-by-Hop Values for Discrepant Numbers
of Value Lexicon Words. Bardi et al.’s (2008) value lexicon includes three
indicators for each Schwartz value. For some Schwartz values, but not all, the name
74
of the Schwartz value could be included as an additional (fourth) indicator: for
some Schwartz values, Bardi et al.’s lexicon uses the value’s name as an indicator
(e.g., for Power, “power,” “strength,” and “control”), while for others, the value’s
name is not included (e.g., for Hedonism, “luxury,” “pleasure,” and “delight”).
Thus, in total, Bardi et al.’s value lexicon contains 36 words, comprising four
Schwartz values of three indicators, and six Schwartz values of four indicators
(when the value’s name is included)7.
Word-by-hop calculations could be affected by the inclusion of a fourth
indicator for a given value by providing additional opportunities for connections
to a value lexicon word. Thus, the analyses described below were run with only the
three primary indicator words for each Schwartz value8.
Correcting Word-by-Hop Values for Discrepant Numbers of
Words across Obituaries. Calculating word-by-hop values using the count of
words that have a connection to a given Schwartz value through Bardi et al.’s value
lexicon potentially preferences longer obituaries over shorter obituaries.
This preference could be conceptually problematic, given that newspapers
may charge fees to publish obituaries based on word count (potentially confounding
wealth with word-by-hop calculations). This was remedied by replacing count of
words that have a connection to a Schwartz value with that count divided by the
total number of words in the obituary (i.e., using percentage of total words in
7Bardi et al. (2008, p. 485) explained this discrepancy in their report, stating, “A preference
was given to use the value label itself (e.g., power, security) as one of the words to represent
each value; however, this was not always possible because some of these value labels yielded
prohibitively low word frequencies (e.g., universalism, hedonism).”
8Future follow-up work could repeat the analyses, including the fourth indicator words to
examine the results’ robustness. Given Bardi et al.‘s (2008, p. 485) note that the value titles that
were not used as indicators were not included because they were seen infrequently in their corpus,
however, those value titles’ addition can likely be expected not to change the results dramatically.
75
the obituary, and subsequently dividing that percentages by one plus the median
number of hops of those words)9.
This adjustment has the secondary effect of causing the word-by-hop
calculation to be bounded at [0, 1], as can be seen in these examples:
– An obituary with “perfect” connection with a Schwartz value would be one
that contained only words that exactly matched those from the value lexicon
for that Schwartz value: for example, 10 words (including repeated words)
that exactly correspond with value lexicon words for a given value ÷ 10 total
words in the obituary ÷ (1 + 0 median hops) = 1.000
– 15 words (including repeated words) that have a connection to a given value
÷ 120 total words in the obituary ÷ (1 + 4 median hops) = 0.025, whereas
10 words (including repeated words) ÷ 120 total words in the obituary ÷
(1 + 4 median hops) = 0.0167, representing that the first case is “better
connected” to the value than the second.
– 15 words (including repeated words) that have a connection to a given value
÷ 120 total words in the obituary ÷ (1 + 4 median hops) = 0.025, whereas
10 words (including repeated words) ÷ 120 total words in the obituary ÷ (1
+ 3 median hops) = 0.021, representing that, as above, the calculation still
considers obituaries that contain a larger number of connected words at a
slightly larger median distance “more connected” to a Schwartz value than
9Alternatively, I could have replaced count with count divided by the number of words in the
obituary that have a connection to any Schwartz value (i.e., using percentage of values words in
the obituary, and subsequently dividing that by one plus the median number of hops of those
words. However, this could build into the calculation the possibly hazardous assumption that that
all obituaries are equally laden with Schwartz values overall.
76
those that contain a smaller number of connected words at a slightly smaller
median distance.
Word-by-hop is conceptually related to several other measures of
graph connectedness. Unlike in-degree and out-degree, which respectively
measure the number of incoming and outgoing edges from a given node
in a directed graph, word-by-hop ignores the direction of relationships
in the graph, allowing bidirectional traversal between source and target
lemmas. Word-by-hop does not consider the centrality of a given node, since
it is primarily concerned with the existence and distance of relationships
between a source node and a given set of target nodes, rather than
whether a given node is a hub for other nodes; however, it does incorporate
some elements that are similar to those used to calculate betweenness
centrality, the equation for which is how many of the paths include node n
number of shortest paths between node a and node b
(assuming that nodes a, b, and n are not the same) (see, e.g., O’Neil
& Schutt, 2013, p. 258). Word-by-hop’s equation could be re-written
number of shortest paths between obituary word a and the value lexicon words b, c, and d for a given value
the median distance of those shortest paths
; like
betweenness centrality, word-by-hop incorporates information about the number of
paths between a set of nodes. Although word-by-hop is not the same as measures of
cross-clique centrality, it does share the approach of assessing the extent to which
a given node connects to “cliques” of nodes (here, defined using the value lexicon,
which is divided following the Schwartz values paradigm; it is not necessarily the
case, however, that all words in the value lexicon for a given value connect to one
another).
Two potential limitations of word-by-hop are that it does not take into
account the clustering coefficient of nodes (i.e., how prone to clustering an
77
obituary’s words are) nor the baseline frequency of words as they are used in
written English. These are both additions that would be useful to consider in
future projects.
Statistically Modeling Word-by-Hop Values. Because it is a
proportion, word-by-hop is theoretically bounded at [0, 1]. Although word-by-hop
values of 1.0 are almost certain never to be seen in this type of corpus (as it would
require that every word in an obituary exactly matched a word from the value
lexicon for a particular Schwartz value), word-by-hop values of 0 are realistically
possible; these values would indicate that a given obituary has no words that have
any connection (within 15 hops through the lexical graph database) to any of the
value lexicon words for a given Schwartz value. Although no word-by-hop values
of 0 were observed in the dataset used for this project (see below, in response to
Research Question 1), it is worth pausing to consider the analytic approaches that
would be best suited to modeling this new measure, in case zeros are observed in
future datasets to which it is applied.
The [0, 1] boundaries of the corrected word-by-hop calculation defined above
could be problematic to use in a typical multiple regression (including multi-level
multiple regression) context, for two reasons. First, multiple regression could poorly
predict word-by-hop as a dependent measure if, as is reasonable to expect, much of
the dataset pools at the bottom of its range. Put differently, the expectation that
there will be a substantial number of values at and near 0 could cause difficulty for
a regression model, which would otherwise mathematically assume that the model’s
Dependent Variable (DV) has a theoretical range from −inf to inf and thus might
present predicted values that are beyond the actual boundaries of the DV (Ferrari
& Cribari-Neto, 2004). Second, beyond pooling near 0, word-by-hop values could
78
be considered more aptly described by a beta distribution than by the normal
distribution assumed in a typical multiple regression context. Beta distributions
can take a variety of both normal and non-normal shapes, and are bounded at (0,
1) (i.e., between 0 and 1, without including values of exactly 0 or 1).
For datasets that include word-by-hop as a dependent measure but do not
include values of 0 or 1, several analytic possibilities exist. Using a Generalized
Linear Model (GLM), word-by-hop values may be modeled directly as beta-
distributed, using, for example, the betareg, glmmADMB, and zoib packages for R,
each of which allows beta-distributed dependent variables as well as the inclusion
of random effects (here, e.g., the newspaper that published a given obituary). This
approach is described in the literature as a “beta regression,” and was introduced
by Ferrari and Cribari-Neto (2004). Alternatively, the data may be modeled using
a wider variety of tools (such as R’s glmer package) using a logit link function and
a binomial distribution (University of California, Los Angeles Institute for Digital
Research and Education, 2016). A test model had trouble converging using this
approach, however, with the full obituaries dataset, possibly due to a large number
of small (non-zero, but near zero) word-by-hop values10.
As a third alternative, the data may be logit-transformed, using the formula
log word-by-hop
1−word-by-hop . The logit-transformation will re-scale values to be bounded at
(−inf , inf), allowing them to be used in a normal multi-level regression.
For future datasets that do include word-by-hop values of 0, the DV
will not be able to be directly modeled by either a beta regression as described
above or a normal regression using logit-transformed data, as in the former case,
10I include this speculative explanation because applying a test transformation to word-by-hop
values such that all values were moved up by a constant until the maximum value was just below
1 caused the warning produced when modeling the raw word-by-hop values with this approach to
cease.
79
the beta distribution cannot include zeros, and in the latter case, the normal
distribution cannot include −inf (which is the result of logit-transforming 0).
Ospina and Ferrari (2010; cf. Ospina & Ferrari, 2012) proposed a “zero-inflated”
beta regression for this case, which simultaneously models non-zero values of the
DV as beta-distributed and DV values of 0 as binomial-distributed. The zoib
package for R allows for zero-inflation (Liu & Kong, 2015); however, it is built on
JAGS and rjags, which perform Markov-Chain Monte Carlo (MCMC) estimation,
and thus performed very slowly in tests even for simple models when adding a
random effect for newspaper with the development dataset of ˜15,000 obituary
rows. As of this writing, R’s glmmADMB package also allows for zero-inflation, but,
per its documentation, only when modeling outcome variables using Poisson or
binomial distributions.
For future datasets that do include word-by-hop values of 0, in contexts
in which performing a zero-inflated beta regression is either not desirable or (as
with this project) logistically untenable because of dataset size causing large model
convergence times, Smithson and Verkuilen (2006, p. 55) proposed transforming the
data using the formula
y×(n−1)+ 1
2
n
, “where n is the sample size,” in order to slightly
inflate values of 0 and slightly deflate values of 1 (cf. the R betareg package
vignettes documentation, Cribari-Neto & Zeileis, n.d., p. 3, which quotes Smithson
and Verkuilen).
The current project featured 10 DVs, each comprising word-by-hop values
for a different Schwartz value. Thus, each regression model that I sought to test
in the current project needed to be run 10 times, once for each DV (e.g., in an
iterative model-building exercise in which seven models were compared to one
another, 70 total regressions would need to be run). Numerous tests indicated
80
that, even under conditions of parallelization, all beta regression options for R
described above performed too slowly on the full dataset to be usable, particularly
when adding a random intercept term for newspaper (even without random slope
components for age of the deceased, gender of the deceased, and their interaction).
Thus, for the regression analyses described below, word-by-hop values were logit-
transformed and then passed into R’s lme4 package’s lmer function, which is
designed for more traditional multi-level modeling.
Logit-transforming the data in this way presents several potential issues.
First, any deviation from modeling the data directly (i.e., as beta-distributed)
is open to a rightful conceptual criticism that interpretation of the resulting
models’ output may be more difficult. In this case, however, it is useful to consider
that word-by-hop is a new measure, which could be made to include the logit
transformation in its definition. More strongly, word-by-hop values are not meant
to be interpreted in isolation, but rather relative to one another (as higher values
indicate greater “connectedness” to the target words). As higher word-by-hop
values correspond to higher logit-transformed word-by-hop values, the potential
issue of interpretation is minimized in this case. Second, logit-transforming data
does not cause them to become normally-distributed. Thus, while word-by-hop
values for several Schwartz values do appear to be normally distributed, several
appear binomially-distributed at least in some newspapers (See Figures 23, 24, 25,
and 26, which are further explored below); output from models that utilized those
non-normal word-by-hop variables should be interpreted with this in mind. Future
work could also usefully compare the results of using this analysis approach vs. a
beta regression approach on a subset of word-by-hop data.
81
Multiple Membership Analysis Rationale. As noted above,
20.294% of the remaining dataset comprised obituaries that had been published
across more than one newspaper. Because “communities” in this project
were defined by newspaper, these data would ideally be modeled as “multiple
members11” of the newspaper-communities (henceforward called simply
“newspapers”) that published them, following Leckie’s (2013) terminology.
Incorporating the multiple membership of obituaries in newspapers would
include explicitly modeling obituaries (at level 1 of the model), collapsed by their
text strings such that duplicate obituaries were all counted as a single row, as
potentially having been published in more than one newspaper (at level 2 of the
model), rather than being strictly nested each within one newspaper. Following the
approach advocated by Leckie (2013), this process could also include weights to
show each obituary’s percentage of “membership” in each of the newspapers that
published it (e.g., weighting could be done by the newspapers’ counties’ relative
population sizes). In this case, a random-intercepts regression model with age and
gender as obituary-level predictors and newspapers’ counties’ median income as a
newspaper-level predictor would be of the following form:
11Leckie (2013) and the Center for Multilevel Modelling at the University of Bristol use the
term “multiple membership” to describe this data structure. In other literatures, the same
structure is alternatively called “mixed membership” and (in some cases) a “non-nested” design.
82
word-by-hopSchwartz valuev =
β0
+ β1ageo
+ β2maleo
+ β3newspaper median incomeo
+
∑
n∈newspapers(o)
w(2)n,ou
(2)
0n
+ eo
In this equation, each β functions as in a typical regression, as a measure
of the expected logit-transformed word-by-hop change for a one-unit increase in
the associated predictor, holding all other predictors at a constant level (e.g., 0).
“(2)” superscripts denote terms at level 2 of the model (the newspaper level).
newspaper median incomeo denotes the median income of the counties in which a
newspaper is published. Following Leckie’s approach and terminology, this term
would be defined for each obituary as the weighted sum of the income values of
each of the newspapers that published the obituary:
newspaper median incomeo =∑
n∈newspapers(o)︸ ︷︷ ︸
Sum over obituary o’s newspapers...
w(2)n,onewspaper median incomeo︸ ︷︷ ︸
For each newspaper, weight the median income (e.g., by population)
u
(2)
on would denote a random effect at the newspaper level (i.e., the variance
attributable to newspaper membership).
The term
∑
c∈counties(o)w
(3)
c,ou
(3)
0c could also be added to the first equation
above to explicitly model newspapers as being multiple members of counties (rather
83
than manually computing a weighted average by population size for each newspaper
across the counties in which it operates, as was done for this project’s analyses).
Unfortunately, as of this writing, software available for conducting analyses
was not readily able to combine this type of multiple membership approach with
the beta regression approach described above, or, in the case of R, to readily
incorporate multiple membership data structures generally. Given that it was
not logistically feasible (although it would be ideal) to incorporate multiple
membership in either a weighted or non-weighted way, two primary options for
analyses were considered. First, the multiple membership structure of the data
could be ignored. In this approach, obituaries that were in fact duplicates across
newspapers would be treated in each of their instances as unique. Previous research
has demonstrated that ignoring multiple membership structure underestimates
higher-level variance while overestimating level-1 variance (see Leckie, 2013),
which would be especially problematic in this case, since the newspaper level was
of primary interest. Alternatively choosing to consider only one instance of each
obituary would be, according to Leckie, a “naive” approach (p. 3); nonetheless,
this alternative approach seemed preferable to proceeding while ignoring the
data’s multiple membership structure. Systematically selecting a single obituary
for each set of duplicates (e.g., by the newspaper serving the counties with
largest populations) would be conceptually problematic, as it would build into all
downstream models the assumption that obituaries are most intended for as wide
an audience as possible (vs., e.g., being intended first for a smaller community,
with which the deceased may have had particularly close ties, and secondly for
a larger community). Thus, each set of across-newspaper duplicate obituaries
was randomly downsampled to one instance of the obituary. This approach did
84
essentially discard data; however, it retained as much data as possible while not
ignoring the data’s multiple membership structure. Conceptually, this approach
is equivalent in Leckie’s schema to randomly weighting newspaper membership
to 0, 0, and 1, vs. an (e.g., arbitrary) weighting scheme of 1
3
, 1
3
, 1
3
, for an obituary
published in three newspapers.
As mentioned above, county-level predictors for the analyses below
were manually aggregated to the newspaper level by creating averages for each
newspaper weighted by county population size. The newspaper-in-county data
structure could itself be modeled using a multiple-membership approach. Failing
to do this likely resulted in shunting county-level variance elsewhere in the model
(hopefully, but not necessarily, to the newspaper level). However, since counties
were not the focus of this project (and were originally incorporated with the
erroneous understanding that newspapers primarily would operate in only one
county), this decision, which also simplified the regression models used below,
seemed appropriate. With this in mind, future work focused more directly on
these decisions’ analytic implications could determine the extent to which choosing
not to include a random county-level variance component substantially alters
conclusions. Future work could also re-downsample the duplicate obituaries and
re-run the regression models below in order to explore the analyses’ robustness to
this approach.
Covariate Descriptive Statistics. After randomly down-sampling
obituaries published in multiple newspapers to single instances (as described
above), the dataset contained 140,599 obituaries, nested in 832 newspapers. The
remaining newspapers and their associated obituaries were eliminated for the
analyses presented below because county-level information for them was unavailable
85
(i.e., for the remaining newspapers, an SQL JOIN of the newspaper data onto the
county-level data using the Google Maps-derived county FIPS codes as shared
identifiers returned no matches).
Age at death (as coded by the automated algorithm described above)
ranged from 1 to 110 (M = 73.5, SD = 21). The percentages of obituaries coded as
having been written about women vs. men in the final sample were approximately
even, with 49.1% female and 50.9% male. 71.3% of gender codes (which were all
assigned using the automated algorithm described above) were made solely using
counts of gendered pronouns in the obituary text, while 27.9% were made using
the gender package for R. 0.892% of codes were assigned using a combination of
pronoun-counting and the gender package. Descriptive statistics for all newspaper-
level predictors (mixed ethnic/racial demographic variables; county population
size; median household income; and education, operationalized as percent of the
population with a Bachelor’s degree or higher) are presented in Table 8.
Table 8. Descriptive statistics for newspaper-level covariates.
Variable min max mean sd
Race/Ethnicity: “White” 20.210 99.08 79.634 13.687
Race/Ethnicity: “Black & African American” 0.569 79.14 16.157 13.251
Race/Ethnicity: “American Indian & Native Alaskan” 0.400 41.31 1.639 2.302
Race/Ethnicity: “Native Hawaiian & Pac. Islander” 0.020 3.14 0.273 0.334
Race/Ethnicity: “Hispanic” 0.496 88.69 12.639 12.852
Income (divided by 1,000 for this table) 26.351 108.48 54.245 12.115
Percentage with Bachelor’s degree or higher 7.300 64.00 30.046 8.886
86
Figure 9 visualizes the densities of median household income (United States
Bureau of the Census, 2015b) and education level (United States Bureau of the
Census, 2015a). Education was operationalized as percentage of the population
with a Bachelor’s degree or higher after comparing its distribution across counties
to the percentage of the population with a high school degree or higher, and finding
the variability of the former to be greater than that of the latter, and thus of
greater use in a regression context.
Figure 9. Density curves for median household income and education level
(operationalized as percent of the population with a bachelor’s degree or higher)
from US census data averaged across counties for each newspaper (after weighting
by county population size).
Figure 10 visualizes the densities of racial and ethnic groups across
newspapers (United States Bureau of the Census Population Division, 2015).
87
Figure 10. Density curves for each race or ethnicity variable from US census data
averaged across counties for each newspaper (after weighting by county population
size). Category titles are taken from the Census dataset, and are of that category
“Alone or in Combination” with other race/ethnicity categories (i.e., an individual
included in the “White” category could have self-described herself as both “White”
and “Hispanic”).
88
Exploring the Properties of the New Word-by-Hop Measure.
Because word-by-hop is a new measure, it is worthwhile to pause here to explore
its properties before reporting statistics directly relevant to the project’s two major
research questions.
As noted in the Methods section above, for each shortest path between
an obituary lemma-POS combination and a target value lexicon word, eight hop
counts were recorded, each following a different definition of “shortest path” that
could be relevant as part of this project:
1. Total number of hops
2. Number of hops only counting Word nodes
3. Number of hops only counting Morph nodes
4. Number of hops only counting Synset (i.e., Definition) nodes
5. Number of Synset-to-Synset relationships
6. Number of Word-to-Synset relationships
7. Number of Morph-to-Word relationships
8. Number of Free Association relationships
In order to facilitate understanding of word-by-hop computations, raw hop
numbers for each of these definitions of “shortest path” are summarized in Table
9. The number of free association relationships was 0 across shortest paths, as
expected; as noted above, because free association edges were excluded from the
graph queries for this project, this descriptive statistic was included as a check
that the query was working properly. Morph nodes were seen infrequently in the
89
shortest paths, as also expected, given that Morphs are uncommon variants of
Words, and only connect to Words.
Table 9. Descriptive statistics for hop numbers in the shortest paths between
obituary lemma-POS combinations and value lexicon words.
min max mean sd
Total Number of Hops 0 15 8.305 1.516
Shortest Number of Hops Only Counting Words 0 7 1.723 0.710
Shortest Number of Hops Only Counting Morphs 0 1 0.000 0.017
Shortest Number of Hops Only Counting Synsets 0 14 6.581 1.468
Number of Synset to Synset Relationships in Shortest Path 0 13 4.859 1.737
Number of Word to Synset Relationships in Shortest Path 0 14 3.445 1.420
Number of Morph to Word Relationships in Shortest Path 0 2 0.001 0.034
Number of Free Association Relationships in Shortest Path 0 0 0.000 0.000
The liberal decision not to constrain the final Synset node encountered
in each shortest path when calculating hop numbers through WordNet can be
evaluated with reference to Table 10, which indicates that although in the majority
(˜95%) of cases, the final Synset encountered was of POS noun, some paths did
finish on verb or adjective Synsets. This table suggests that allowing the final
Synset POS to be unconstrained was likely useful given the exploratory nature of
this work, but that future work that chooses to add a noun constraint to the final
Synset POS encountered in each shortest path would likely not affect hop counts
(and thus derived word-by-hop values) substantively.
90
Table 10. Percentage of final Synset (i.e., definition) node POS in shortest paths
between obituary lemma-POS combinations and value lexicon word (excluding
cases in which no path existed).
POS Frequency as Percentage
a 0.105
n 94.501
v 5.393
Of the eight measurements above for shortest path length, the two that
I considered most relevant for this project were number of hops only counting
Word nodes, and number of hops counting only Synset-to-Synset relationships,
as they seem especially to follow lay definitions of word closeness12. Thus, for
each obituary, the word-by-hop value for each Schwartz value was calculated and
recorded twice, first using Word node counts, and then using Synset-to-Synset
relationship counts. These two word-by-hop calculations for each Schwartz value
are compared below, both in response to Research Question 1 and because these
comparisons created a foundation for choosing one computation over the other
for use in the regressions reported below in response to Research Question 2.
Table 11 presents the minimum, maximum, mean, and standard deviation for each
computation of word-by-hop.
12Free association also seems to me to follow a lay definition of word closeness; for this project’s
analyses, however, free associations were not included in the graph, following the rationale above.
91
Table 11. Descriptive statistics for word-by-hop calculations, ordered by mean
(descending) followed by standard deviation (ascending).
Schwartz Value Calculated Using min max mean sd
power word hop 0.005 0.320 0.207 0.029
conformity word hop 0.005 0.320 0.202 0.033
security word hop 0.005 0.320 0.201 0.034
self direction word hop 0.005 0.320 0.190 0.039
benevolence word hop 0.005 0.320 0.176 0.042
achievement word hop 0.005 0.320 0.170 0.041
tradition word hop 0.004 0.320 0.156 0.037
universalism word hop 0.004 0.317 0.142 0.025
hedonism word hop 0.004 0.269 0.139 0.018
stimulation word hop 0.004 0.213 0.139 0.018
power synset to synset relationship 0.002 0.148 0.086 0.013
universalism synset to synset relationship 0.002 0.143 0.084 0.011
stimulation synset to synset relationship 0.002 0.148 0.083 0.011
achievement synset to synset relationship 0.002 0.130 0.081 0.012
security synset to synset relationship 0.002 0.138 0.081 0.012
conformity synset to synset relationship 0.002 0.136 0.081 0.012
tradition synset to synset relationship 0.002 0.143 0.077 0.012
self direction synset to synset relationship 0.002 0.123 0.076 0.013
hedonism synset to synset relationship 0.002 0.122 0.075 0.012
benevolence synset to synset relationship 0.002 0.123 0.071 0.010
92
Table 9 demonstrates that counts of Word nodes were smaller on average
and with a smaller range than counts of Synset-to-Synset relationships. This had
the effect of producing larger word-by-hop values with larger ranges, as can be
seen in Table 11 and many of the figures below. Because the Word node count
calculation provided higher levels of variance for to account for, word-by-hop values
derived from Word node counts were chosen over those derived from Synset-to-
Synset relationship counts for use in the regression models reported below.
The relationship between the two calculations of word-by-hop is visualized
in Figures 11, 12, 13, and 14. These figures indicate a clear positive relationship
between the two calculations, such that word-by-hop values calculated from
Word node counts do tend to correspond to higher word-by-hop values calculated
from Synset-to-Synset relationship counts. However, these figures also show a
consistently multi-linear relationship between the two calculations, resulting in a
low linear correlation value between the means of each calculation across Schwartz
values (taken from Table 9), r = 0.122.
Answering Research Question 1: Which Values are Present in
Relation to One Another in Obituaries. The distributions of male vs. female
word-by-hop values plotted by age at death are visualized in Figures 15, 16, 17,
and 18. These figures illustrate the differential variances of the two word-by-hop
calculation approaches, and also, surprisingly, suggest that if an age effect exists
in the data at a statistically noteworthy level, it is small enough perhaps not to
pass the threshold for conceptual noteworthiness. Overall, the figures suggest that
obituaries coded as written about women also tended to be coded as written about
older individuals; it is worth noting, however, that the range of ages displayed
in the figures (from 1, the minimum age allowed by the automated age-guessing
93
Figure 11. Schwartz “openness to change” values (Self-direction, Stimulation,
and Hedonism) plotted by gender to compare word-by-hop values as computed
using the number of Word nodes between each obituary lemma / Part-of-Speech
combination and a value lexicon word, less one (i.e., the number of Word node hops
in the shortest path between a given source and target word) vs. as computed using
the number of Synset-to-Synset edges in the shortest path between a given source
and target word. Pink/lighter points represent obituaries coded as written about
women.
94
Figure 12. Schwartz “self-enhancement” values (Hedonism, Achievement, and
Power) plotted by gender to compare word-by-hop values as computed using the
number of Word nodes between each obituary lemma / Part-of-Speech combination
and a value lexicon word, less one (i.e., the number of Word node hops in the
shortest path between a given source and target word) vs. as computed using the
number of Synset-to-Synset edges in the shortest path between a given source
and target word. Pink/lighter points represent obituaries coded as written about
women.
95
Figure 13. Schwartz “conservation” values (Security, Conformity, and Tradition)
plotted by gender to compare word-by-hop values as computed using the number
of Word nodes between each obituary lemma / Part-of-Speech combination and a
value lexicon word, less one (i.e., the number of Word node hops in the shortest
path between a given source and target word) vs. as computed using the number
of Synset-to-Synset edges in the shortest path between a given source and target
word. Pink/lighter points represent obituaries coded as written about women.
96
Figure 14. Schwartz “self-transcendence” values (Benevolence and Universalism)
plotted by gender to compare word-by-hop values as computed using the number
of Word nodes between each obituary lemma / Part-of-Speech combination and a
value lexicon word, less one (i.e., the number of Word node hops in the shortest
path between a given source and target word) vs. as computed using the number
of Synset-to-Synset edges in the shortest path between a given source and target
word. Pink/lighter points represent obituaries coded as written about women.
97
algorithm, to 110, the maximum age allowed) suggest that a sizable number of
obituaries may have been misclassified (as it is unlikely that so many 110-year-
olds actually died during the data collection period, which covered approximately
one month). This is to be expected, given that the validation exercise for the
age-coding algorithm reported above in the Methods section indicated that the
algorithm was in high but not perfect agreement with the human coders, and
because the full dataset is large (and thus provided ample opportunities for
misclassification, even at a low overall level); however, it does suggest that age-
related results should be interpreted with caution.
Consistent “banding” can be observed at the bottom of several of Figures
15, 16, 17, and 18. The obituaries that these bands comprise are low in Schwartz
value connectedness but tend to be written for individuals who are of higher ages at
death. It is possible that these indicate a particular template of obituary, which
deflates its word-by-hop values while inflating its word count by naming large
numbers of family members (cf. Figures 19, 20, 21, and 22, which are discussed
below, and which also show banding at consistent ranges of word counts).
The distributions of male vs. female word-by-hop values plotted by obituary
word count are visualized in Figures 19, 20, 21, and 22. The consistent conical
distribution shapes in these figures may indicate several properties of word-by-
hop. First, it is possible that the cones show that obituaries at lower word counts
contain more “noise” in relation to clearly understanding their relationship to
a given Schwartz value. This is reasonable, given that obituaries are generally
expected to perform several roles simultaneously, including providing biographical
information about the deceased, naming surviving family members, and providing
logistical information about funerary services. Thus, obituaries with lower word
98
Figure 15. Schwartz “openness to change” values (Self-direction, Stimulation,
and Hedonism) plotted by gender and age at death, as both coded using the
automated algorithm described in the Methods section. Pink/lighter points
represent obituaries coded as written about women. The top row plots word-by-
hop values as computed using the number of Word nodes between each obituary
lemma / Part-of-Speech combination and a value lexicon word, less one (i.e., the
number of Word node hops in the shortest path between a given source and target
word). The bottom row plots word-by-hop values as computed using the number
of Synset-to-Synset edges in the shortest path between a given source and target
word.
99
Figure 16. Schwartz “self-enhancement” values (Hedonism, Achievement, and
Power) plotted by gender and age at death, as both coded using the automated
algorithm described in the Methods section. Pink/lighter points represent
obituaries coded as written about women. The top row plots word-by-hop values
as computed using the number of Word nodes between each obituary lemma /
Part-of-Speech combination and a value lexicon word, less one (i.e., the number of
Word node hops in the shortest path between a given source and target word). The
bottom row plots word-by-hop values as computed using the number of Synset-to-
Synset edges in the shortest path between a given source and target word.
100
Figure 17. Schwartz “conservation” values (Security, Conformity, and Tradition)
plotted by gender and age at death, as both coded using the automated algorithm
described in the Methods section. Pink/lighter points represent obituaries coded
as written about women. The top row plots word-by-hop values as computed
using the number of Word nodes between each obituary lemma / Part-of-Speech
combination and a value lexicon word, less one (i.e., the number of Word node hops
in the shortest path between a given source and target word). The bottom row
plots word-by-hop values as computed using the number of Synset-to-Synset edges
in the shortest path between a given source and target word.
101
Figure 18. Schwartz “self-transcendence” values (Benevolence and Universalism)
plotted by gender and age at death, as both coded using the automated algorithm
described in the Methods section. Pink/lighter points represent obituaries coded
as written about women. The top row plots word-by-hop values as computed
using the number of Word nodes between each obituary lemma / Part-of-Speech
combination and a value lexicon word, less one (i.e., the number of Word node hops
in the shortest path between a given source and target word). The bottom row
plots word-by-hop values as computed using the number of Synset-to-Synset edges
in the shortest path between a given source and target word.
102
counts may be more variable in their word-by-hop values because they include
a higher percentage of words that may be only incidentally value-relevant while
serving one of the more logistical functions of obituaries as death notices. Following
this, it is possible that longer obituaries converge on a “true” level of connectedness
with a Schwartz value, as expressed in their word-by-hop values. Less strongly,
this may indicate that there exists a minimum word count threshold beyond which
value words become more likely to appear, and that the 60-word threshold applied
in this project for biographical content is below that threshold for values. If such an
additional threshold exists, however, it is likely to be more complicated than based
simply on word count (anecdotally, many obituaries of only a few hundred words
can be seen readily to invoke values).
Second, it is possible that the cone-shaped distributions indicate a property
of the word-by-hop formula itself, such that longer obituaries are less able to fit
a larger number of word-by-hop values. This possibility is one that would be
particularly useful to explore in follow-up research. The bimodal distributions of
word-by-hop values in these figures for Self-direction, Achievement, and all of the
Schwartz “conservation” values (Security, Conformity, and Tradition) and “self-
transcendence” values (Benevolence and Universalism) are also noteworthy, and
potentially point to differences across communities particularly in those Schwartz
values (as explored below in the random-intercepts regression models, in which
word-by-hop values are estimated for each newspaper separately). Looking from a
different perspective, it is worth noting that only two Schwartz values (Stimulation,
Hedonism, as well as, to a lesser extent, Power) lack a bimodal distribution when
plotted against word count. This widespread bimodality could be an artifact of an
interaction between the use of templates by obituary authors and the connections
103
between words in the WordNet graph. If a small number of highly-connected words
tended to be used by obituary authors to describe the deceased (perhaps as part of
a templated phrase), the inclusion of a small number of additional highly-connected
words could presumably “bump up” the word-by-hop value for that obituary by
a consistent level. If this were true, Hedonism and Stimulation, as values that
occur less frequently, might have fewer templates available and thus show greater
normality. This explanation is bolstered by the fact that bimodality is not seen
in these figures where Synset-to-Synset-derived word-by-hop values are used as
DVs; this difference in plots between the two calculations implies that distribution
is linked to Word connectedness (as well as, possibly, an inherent property of the
word-by-hop formula that manifests more when the median number of hops is
lower, as when using Word node counts).
Figures 23, 24, 25, and 26 illustrate the densities of word-by-hop values
by Schwartz value, showing distributions without reference to any covariates.
These figures are plotted by newspaper (with darker areas indicating greater
overlap among separate newspapers), and also show distributional variability
(e.g., bimodality vs. normality) across communities not only in Self-direction,
Achievement, Tradition, and Benevolence, but in every value when looking across
newspapers.
The correlation matrix of word-by-hop values is presented in Figure 27.
The minimum correlation was 0.462. Schwartz values in this matrix are listed
following the Schwartz circumplex model in a counter-clockwise direction. A
visual inspection of this matrix does indicate that, generally, Schwartz values as
reconstructed using word-by-hop values do tend to correlate more highly with
values that are closer to them in the circle (i.e., the beginning and end of each
104
Figure 19. Schwartz “openness to change” values (Self-direction, Stimulation,
and Hedonism) plotted by gender and obituary word count, as both coded using
the automated algorithm described in the Methods section. Pink/lighter points
represent obituaries coded as written about women. The top row plots word-by-
hop values as computed using the number of Word nodes between each obituary
lemma / Part-of-Speech combination and a value lexicon word, less one (i.e., the
number of Word node hops in the shortest path between a given source and target
word). The bottom row plots word-by-hop values as computed using the number
of Synset-to-Synset edges in the shortest path between a given source and target
word.
105
Figure 20. Schwartz “self-enhancement” values (Hedonism, Achievement, and
Power) plotted by gender and obituary word count, as both coded using the
automated algorithm described in the Methods section. Pink/lighter points
represent obituaries coded as written about women. The top row plots word-by-
hop values as computed using the number of Word nodes between each obituary
lemma / Part-of-Speech combination and a value lexicon word, less one (i.e., the
number of Word node hops in the shortest path between a given source and target
word). The bottom row plots word-by-hop values as computed using the number
of Synset-to-Synset edges in the shortest path between a given source and target
word.
106
Figure 21. Schwartz “conservation” values (Security, Conformity, and Tradition)
plotted by gender and obituary word count, as both coded using the automated
algorithm described in the Methods section. Pink/lighter points represent
obituaries coded as written about women. The top row plots word-by-hop values
as computed using the number of Word nodes between each obituary lemma /
Part-of-Speech combination and a value lexicon word, less one (i.e., the number of
Word node hops in the shortest path between a given source and target word). The
bottom row plots word-by-hop values as computed using the number of Synset-to-
Synset edges in the shortest path between a given source and target word.
107
Figure 22. Schwartz “self-transcendence” values (Benevolence and Universalism)
plotted by gender and obituary word count, as both coded using the automated
algorithm described in the Methods section. Pink/lighter points represent
obituaries coded as written about women. The top row plots word-by-hop values
as computed using the number of Word nodes between each obituary lemma /
Part-of-Speech combination and a value lexicon word, less one (i.e., the number of
Word node hops in the shortest path between a given source and target word). The
bottom row plots word-by-hop values as computed using the number of Synset-to-
Synset edges in the shortest path between a given source and target word.
108
Figure 23. Densities of word-by-hop for Schwartz “openness to change” values
(Self-direction, Stimulation, and Hedonism), plotted by newspaper to allow visual
inspection of both overall distribution and variability across communities. The
top row plots word-by-hop values as computed using the number of Word nodes
between each obituary lemma / Part-of-Speech combination and a value lexicon
word, less one (i.e., the number of Word node hops in the shortest path between
a given source and target word). The bottom row plots word-by-hop values as
computed using the number of Synset-to-Synset edges in the shortest path between
a given source and target word. Darker areas indicate higher newspaper overlap.
109
Figure 24. Densities of word-by-hop for Schwartz “self-enhancement” values
(Hedonism, Achievement, and Power), plotted by newspaper to allow visual
inspection of both overall distribution and variability across communities. The
top row plots word-by-hop values as computed using the number of Word nodes
between each obituary lemma / Part-of-Speech combination and a value lexicon
word, less one (i.e., the number of Word node hops in the shortest path between
a given source and target word). The bottom row plots word-by-hop values as
computed using the number of Synset-to-Synset edges in the shortest path between
a given source and target word. Darker areas indicate higher newspaper overlap.
110
Figure 25. Densities of word-by-hop for Schwartz “conservation” values (Security,
Conformity, and Tradition), plotted by newspaper to allow visual inspection of both
overall distribution and variability across communities. The top row plots word-by-
hop values as computed using the number of Word nodes between each obituary
lemma / Part-of-Speech combination and a value lexicon word, less one (i.e., the
number of Word node hops in the shortest path between a given source and target
word). The bottom row plots word-by-hop values as computed using the number
of Synset-to-Synset edges in the shortest path between a given source and target
word. Darker areas indicate higher newspaper overlap.
111
Figure 26. Densities of word-by-hop for Schwartz “self-trancendence” values
(Benevolence and Universalism), plotted by newspaper to allow visual inspection
of both overall distribution and variability across communities. The top row plots
word-by-hop values as computed using the number of Word nodes between each
obituary lemma / Part-of-Speech combination and a value lexicon word, less one
(i.e., the number of Word node hops in the shortest path between a given source
and target word). The bottom row plots word-by-hop values as computed using the
number of Synset-to-Synset edges in the shortest path between a given source and
target word. Darker areas indicate higher newspaper overlap.
112
row of the matrix) than with those that are further away. The matrix does not
always follow this pattern, however: Benevolence and Universalism, which exist
side-by-side in the circumplex model, had a low correlation with each other. It may
be that the value lexicon words for Benevolence and Universalism (Table 1), while
related, would be invoked in different contexts, especially based on the occupation
of the deceased. “Kindness,” “charity,” and “mercy,” the prototype words for
Benevolence, seem able to be applied to a variety of individuals. The prototype
words for Universalism, however, “unity,” “justice,” and “equality,” might be
invoked more for individuals of particular professions (e.g., lawyers, judges, and
activists), diminishing the overall correlation between the two values.
Similarly, Hedonism was highly correlated with Conformity, its polar
opposite in the circumplex model. Hedonism’s relationship to Conformity does
make sense with reference to Bardi et al.’s lexicon words for these two values. The
lexicon words for Hedonism, “luxury,” “pleasure,” and “delight,” are conceptually
opposite of those for Conformity, “restraint,” “regard,” and “consideration.” Given
the expectation in this project that Hedonism is a taboo value to invoke in an
obituary, this correlation may indicate that obituary authors “compensate” for
describing the deceased in a hedonistic way by concurrently emphasizing restraint
(and therefore Conformity).
Because it is possible that income and education might be associated with
obituary word counts, which are used when calculating word-by-hop values, Figure
28 visualizes the relevant relationships between these variables. Visual inspection
of this figure suggests that no noteworthy relationship exists (put differently, it
does not appear to be the case that wealthier or more highly-educated communities
systematically produced longer obituaries).
113
Figure 27. Correlation matrix of word-by-hop (calculated using number of Word
nodes between a source and target word, rather than the number of Synset-to-
Synset edges) for each pair of Schwartz values. Schwartz values in this matrix are
listed following the Schwartz circumplex model in a counter-clockwise direction.
114
Figure 28. Scatterplot of median household income and education level
(operationalized as percent of the population with a bachelor’s degree or higher)
from US census data averaged across counties for each newspaper (after weighting
by county population size) vs. obituary word count.
115
Overall, these results demonstrate that all 10 Schwartz values are present in
this obituary corpus, but that some Schwartz values are more lexically indicated
than others13. Following Table 11, as expected, Hedonism was among the least-
indicated Schwartz values in the corpus. However, counter to predictions, Power,
rather than also being low in its mean word-by-hop value, was the single most
indicated Schwartz value in the corpus. Schwartz “conservation” values (Conformity
and Security) were also particularly highly-indicated in the corpus overall. With
this in mind, it is useful to note that a substantial amount of variability exists
across newspapers. This variability is addressed below in response to Research
Question 2.
Answering Research Question 2: Does Match with the Schwartz
Values Vary in Relation to Age, Gender, Race or Ethnicity, Income,
and/or Education. Following the rationale discussed above, word-by-hop values
were logit-transformed and entered into a series of two-level linear models, with
obituaries as level 1 and newspapers as level 2. Also as discussed above, across-
newspaper duplicate obituaries were randomly downsampled until one per duplicate
set remained in the dataset. For each model, 10 regression analyses were performed,
each using a word-by-hop DV aligned with a different Schwartz value. To be
interpreted alongside Table 11, Table 12 presents the minimum, maximum, mean,
and standard deviation for each Word-hop-derived logit-transformed word-by-hop
DV.
13All results discussed from this point forward employed word-by-hop values calculated using
the number of Word nodes between a source and target word, rather than the number of Synset-
to-Synset edges.
116
Table 12. Descriptive statistics for Word-hop-derived, logit-transformed word-by-
hop calculations, ordered by mean (descending) followed by standard deviation
(ascending).
Schwartz Value min max mean sd
power -5.21 -0.753 -1.36 0.230
conformity -5.21 -0.753 -1.39 0.252
security -5.21 -0.753 -1.40 0.259
self direction -5.21 -0.753 -1.48 0.296
benevolence -5.21 -0.753 -1.58 0.326
achievement -5.21 -0.753 -1.61 0.318
tradition -5.62 -0.753 -1.72 0.303
universalism -5.62 -0.765 -1.81 0.234
hedonism -5.62 -1.001 -1.84 0.205
stimulation -5.62 -1.304 -1.84 0.205
Five predictors were prepared for use in the regression models:
– Level 1 (Obituaries):
∗ Age of the deceased (automatically coded using the algorithm described
above)
∗ Gender (female/male) of the deceased (automatically coded using the
algorithm described above). This variable was dummy-coded (0 = male,
1 = female)
117
– Level 2 (Newspaper) (as described above, computed from data from the
counties in which each newspaper operated, using a weighted average by
county population size):
∗ Median household income (United States Bureau of the Census, 2015b)
∗ Education, operationalized as percentage of the population with a
Bachelor’s degree or higher (United States Bureau of the Census, 2015a)
∗ Percentage of the population that identified (in isolation or in
combination) as one of five racial and/or ethnic identities (each category
was represented with its own predictor variable): “American Indian or
Native Alaskan,” “Black or African American,” “Hispanic,” “Native
Hawaiian and other Pacific Islander,” and “White” (United States
Bureau of the Census Population Division, 2015)
This project’s analyses were focused not only on the fixed effects of these
predictors, but also centrally on the effects of adding random intercepts by
newspaper, and random slopes by gender, age, and the interaction between gender
and age. Thus, a series of seven nested or semi-nested regression models were run
(for a total of 70 regression analyses, one for each word-by-hop Schwartz value DV).
These models were written and run before the results of each were analyzed (put
differently, these models were written in an a priori fashion), with the exception of
one final model (listed as model 3 below, given that it conceptually builds on model
2), which was developed after viewing the results of the others. Given the number
of models run, p-values are eschewed in the report below in favor of considering
estimates and their standard errors directly, with a particular additional focus
on comparing Akaike Information Criterion (AIC) values across models to aid
118
in determining fit. All models were run using the lmer function from R’s lme4
package, with the exception of the first, most basic model, which did not include
any random effects and was thus run with R’s related lm (“linear model”) function.
Because of the vast difference in scales among the predictor variables (e.g., income,
which had a mean in the tens of thousands, vs. age at death, which had a mean
in the tens), all continuous predictor variables were z-scored (i.e., made to have a
mean of 0 and a Standard Deviation of 1).
The seven models are described below. Although the model descriptions do
not mention it, all models contained an individual-level error term, as is standard
in multiple regression. All formulae are in R’s lm/lmer format, where a tilde “˜”
separates a DV from the rest of the formula, and a pipe symbol (“|”) indicates a
random factor grouped by the term on the right side of the pipe (in this notation,
“(1 | group)” would indicate a random intercept, and “(predictor | group)” would
indicate a random slope for the predictor).
1. A basic model, consisting of only a single fixed intercept (i.e., for each DV,
modeling word-by-hop values using only the mean word-by-hop value).
Labeled in tables and plots below as “Model with only intercept.”
2. A model consisting of a random intercept term (i.e., for each DV, modeling
word-by-hop values for each newspaper separately using the mean word-by-
hop value of that newspaper). Labeled in tables and plots below as “Model
with random intercept.” The formula for this model was (for an example
Schwartz value word-by-hop DV) “Achievement ˜ (1 | newspaper shortcode)”.
3. Model #2 expanded to include both level-1 predictors (age and gender), as
well as random slopes for age, gender, and the interaction of age and gender
119
(i.e., allowing the individual and interactive effects of age and gender to
differ across newspapers in the same way that the intercept was allowed
to vary across newspapers). Labeled in tables and plots below as “Full
model with random interaction effects without level-2 predictors.” The
formula for this model was (for an example Schwartz value word-by-hop
DV) “Achievement ˜ Female + Age at death + (1 + Female * Age at death |
newspaper shortcode)”.
4. Model #3 without the random effects for age, gender, and age-by-gender
interaction (i.e., a model consisting of a random intercept by newspaper and
fixed effects for age and gender, without their interaction; this modeled the
effect of age and gender as being the same for every newspaper). Labeled
in tables and plots below as “Full model except for random interaction and
level-2 predictors.” The formula for this model was (for an example Schwartz
value word-by-hop DV) “Achievement ˜ Female + Age at death + (1 |
newspaper shortcode)”.
5. Model #2 with fixed effects for gender and age, and fixed effects for income,
education, the interaction between income and education, and the interaction
between gender and education. Labeled in tables and plots below as “Full
model except for random interaction and race/ethnicity.” The formula for this
model was (for an example Schwartz value word-by-hop DV) “Achievement
˜ Female + Age at death + Income + Education + Female * Education +
Income * Education + (1 | newspaper shortcode)”.
6. Model #5 with all race/ethnicity predictors added as fixed effects. Adding
these as a block allowed examining the overall effect of race/ethnicity.
120
Labeled in tables and plots below as “Full model except for random
interactions.” The formula for this model was (for an example Schwartz
value word-by-hop DV) “Achievement ˜ Female + Age at death +
Income + Education + Female * Education + Income * Education +
Race/Ethnicity:”White” + Race/Ethnicity: “Black or African American”
+ Race/Ethnicity: “American Indian or Native Alaskan” + Race/Ethnicity:
“Native Hawaiian and Other Pac. Islander” + Race/Ethnicity: “Hispanic” +
(1 | newspaper shortcode)“.
7. Model #6 combined with Model #3. Labeled in tables and plots below as
“Full model.” The formula for this model was (for an example Schwartz
value word-by-hop DV) “Achievement ˜ Female + Age at death +
Income + Education + Female * Education + Income * Education +
Race/Ethnicity:”White” + Race/Ethnicity: “Black or African American”
+ Race/Ethnicity: “American Indian or Native Alaskan” + Race/Ethnicity:
“Native Hawaiian and Other Pac. Islander” + Race/Ethnicity: “Hispanic” +
(1 + Female * Age at death | newspaper shortcode)“.
Given the large number of AIC values to present (as 70 regression analyses
were run in total), rather than presenting these results in a table, Figure 29
graphically shows AIC values across models and Schwartz values. Figure 30
removes the worst-fitting model (Model #1, which included only a fixed intercept)
from Figure 29 in order to show with more detail the differences among the
remaining models for each Schwartz value.
Lower AIC values indicate better model fit. AIC incorporates both the
number of parameters and the model’s deviance; thus, given two models with
similar deviance, AIC will prefer the more parsimonious model (i.e., the model with
121
Figure 29. AIC values plotted across models and Schwartz values. The highest AIC
(indicating worst fit) was found consistently across Schwartz values in the intercept-
only model (Model #1). The lowest AIC was not in the full model (Model #7), but
rather the full model lacking level-2 predictors (Model #3).
122
Figure 30. AIC values plotted across models and Schwartz values. The lowest AIC
was not in the full model (Model #7), but rather the full model lacking level-2
predictors (Model #3).
123
fewer parameters). Looking at Figure 29, it is striking that, by far, the greatest
amount of variability across Schwartz values, as measured using word-by-hop, was
accounted for simply by including a random intercept by newspaper (moving from
Model #1 to Model #2). As expected, this indicates, either or together, that,
as expected, obituaries within a given newspaper do cluster around particular
value levels, and that word-by-hop as a measure is sensitive to these differences.
Figure 30 shows more clearly that, surprisingly, AIC values were consistently lowest
across Schwartz values not for the full model (Model #7), but for Model #3, which
allowed random slopes for age, gender, and their interaction, but excluded level-2
covariates. The full model consistently showed the next-lowest AIC values. While
it is to be expected that adding parameters (even poorly predictive ones) to a
model will increase its fit, AIC’s penalization of larger numbers of parameters likely
resulted in Model #3, which is more parsimonious, being given the smaller AIC
values. Models #4 and #6 shows that it was not age and gender in themselves as
fixed effects that resulted in better model fit, but specifically their being modeled
as random effects.
Models #5 and #6, which lacked those random effects but differed from
each other only in the inclusion of the race/ethnicity variables, indicate that the
race/ethnicity covariates were not useful additions to this model. Taken together
with Model #7, which had a higher AIC value than the model that lacked those
covariates as well as income and education (Model #3), this could indicate that
value levels in obituaries do not vary systematically by income, education, or
race/ethnicity. However, this could also indicate that the measurement of these
covariates was problematic in some way, perhaps either in that counties were
not explicitly modeled and were instead aggregated by newspaper, or in that the
124
automated newspaper-location lookup algorithm included a large number of false
positives and/or false negatives.
It is also worth remarking on the fact that AIC values were different
across Schwartz values. This is likely the result of the word-by-hop DVs having
different levels of variability for the models to account for. Word-by-hop values for
Hedonism, for example, had higher uniformity (i.e., a lower Standard Deviation)
than other values (as shown in Tables 11 and 12), perhaps because of norms
against invoking Hedonism in obituaries (see the Discussion section below). For
Benevolence, by contrast, the regression models had larger amounts of variability
(indicated by a larger Standard Deviation) to account for: more so than Hedonism,
it seems reasonable to expect that descriptions of the deceased in some obituaries
would be much more indicative of Benevolence than in others.
As above, given the large number of analyses, Figures 31 and 33 graphically
present the estimates and their standard errors for Models #3 and #7, respectively
(i.e., the best-fitting model and the next-best fitting, full, model). Figures 32
and 34 feature the same respective information, but lacking the models’ intercept
estimates, allowing more detailed inspection.
Figure 34 indicates that in Model #3, of gender and age, gender was
estimated to have had a lower but consistently non-zero effect across Schwartz
values. This effect was most pronounced in Self-direction (with females having
˜.04 lower logit-transformed word-by-hop values than men, holding age constant),
while the effect was least pronounced in Achievement. The estimates of age’s effect
encompassed zero for several values, and were not consistently positive or negative
for the rest.
125
Figure 31. Plot showing coefficient estimates and standard errors for all Schwartz
value word-by-hop Dependent Variables and all model predictors for Model #7,
the full model. Note that all predictors were z-scored before being entered into the
model, and that the word-by-hop DV was logit-transformed.
Figure 32. Plot showing coefficient estimates and standard errors for all Schwartz
value word-by-hop Dependent Variables and all model predictors for Model #7,
the full model, excluding the intercept (in order to see the other estimates in more
detail). Note that all predictors were z-scored before being entered into the model,
and that the word-by-hop DV was logit-transformed.
126
Figure 33. Plot showing coefficient estimates and standard errors for all Schwartz
value word-by-hop Dependent Variables and all model predictors for Model #3,
the model with the lowest AIC. Note that all predictors were z-scored before being
entered into the model, and that the word-by-hop DV was logit-transformed.
Figure 34. Plot showing coefficient estimates and standard errors for all Schwartz
value word-by-hop Dependent Variables and all model predictors for Model #3,
the model with the lowest AIC, excluding the intercept (in order to see the other
estimates in more detail). Note that all predictors were z-scored before being
entered into the model, and that the word-by-hop DV was logit-transformed.
127
Figure 32 indicates that, for the full model (Model #7), some race/ethnicity
percentages did vary consistently with logit-transformed word-by-hop across
Schwartz values. An increase of one SD in the county-level percentage of the
population categorizing itself as “Black or African American” was associated
with an estimated increase of over .10 logit-transformed word-by-hop values for
Achievement, holding all other predictors constant. Percentage of the population
that categorized itself as “White” consistently had the next-highest effect estimates
across Schwartz values. Having said that, it is worth keeping in mind that race
and ethnicity as a whole (i.e., as a set of predictors) did not substantially increase
model fit.
128
CHAPTER IV
DISCUSSION
The current project explored the extent to which a sample of contemporary
obituaries from across the USA used words related to values from Schwartz’ (e.g.,
2012) model. To do this, it employed word-by-hop, a new measure of lexical
distance from a given set of target words and the Schwartz values they indicate.
Obituaries were expected to be a particularly useful potential repository of values
when viewed in aggregate, as, given their place in death-related ritual in the USA,
they could be expected to be particularly charged with value-expressions (even if
only of a positive valence), more so than general newspaper texts (such as those
used by Bardi et al., 2008) or other public corpora.
The approach used in this project did reveal differences across Schwartz
values as operationalized using word-by-hop. Power, the most-indicated Schwartz
value, had a mean word-by-hop value across the corpus that was substantially
(more than 2.25 of its Standard Deviations) above Hedonism and Stimulation,
the least-indicated Schwartz values. Word-by-hop values were consistently small,
as expected (as a value of 1 would indicate perfect overlap between lemmas from
an obituary and target value lexicon words); despite this, it is encouraging and
reasonable to interpret these findings as evidence that the word-by-hop measure is
sensitive to relative variation, at least in aggregate, in the values (or at least the
template-expressions of values) that a corpus indicates.
Stimulation and Hedonism were the least-indicated Schwartz values in the
corpus. This value ordering makes sense with reference to Berger (1969, p. 43),
who, as discussed above in this dissertation’s introduction, argued that responses
to death rely on “legitimations of the reality of the social world,” presumably
129
including the moral world. Many systems of morality have to do with regulating
Hedonism and Stimulation specifically, more so than the other Schwartz values.
Further, compared with the other Schwartz values (including Power, Conformity,
and Security, which were most indicated in the corpus), Hedonism and Stimulation
are perhaps the least-social (even Self-direction, which, according to Schwartz,
2012, p. 5, is motivated by “independent thought and action–choosing, creating,
exploring,” is defined as belonging to a social context, even if eschewing it). These
moral and less-social aspects of Hedonism and Stimulation may also interact with
the ritual nature of obituary-writing in the USA, which places taboos on including
statements that could be perceived by readers as immoral.
Power was the most-indicated value in the corpus. This result was
unexpected, but does make sense following Berger, as above, as well as Fowler
(2005), who noted that obituaries tend to emphasize individualism over
communalism. Within Berger’s framework of responding to death by emphasizing
the social over the anti-social, Power is the Schwartz value that could be seen as
the most individualistic while also social. While Achievement, e.g., is motivated by
“personal success” (also emphasizing individualism; Schwartz, 2012, p. 5-9), Power
is motivated by “social status and prestige, control or dominance over people and
resources,” emphasizing a distinct, more direct social concern.
One central question to address in follow-up work is what causes the
bimodality seen when plotting word-by-hop values against obituary word-count
(as in Figures 19, 20, 21, and 22). As noted above in the Results section, bimodal
distributions were seen at least somewhat in all Schwartz values except for
Hedonism and Stimulation, the two least-indicated values in the corpus. This
could suggest that writing obituaries from templates is more common in longer
130
obituaries. If this is the case, it presents an interesting additional impetus for
future research: following a template, whether explicit (e.g., “Fill in the names
of the spouse, children, etc. of the deceased”) or implicit (e.g., “Many people use
the phrase ‘Went to her eternal home,’ so I will, as well”), is inherently a cultural
act: it is based in imitation of others in one’s community. Future work could
thus not only explore obituaries from this corpus for markers of greater linguistic
similarity (which could indicate that authors tended to follow an implicit or explicit
template), but also the traits of authors of prospective obituaries, whose writing
process could be observed in a more controlled setting. Incorporating the clustering
coefficient of nodes (i.e., how prone to clustering an obituary’s words are) into
future analyses and even the word-by-hop calculation itself would be one way to
assess templating tendencies.
In the best-fitting regression model, gender (male/female) had a consistently
non-zero estimated effect on predicted logit-transformed word-by-hop values.
This finding is in line with those of Rodler et al. (2001), who observed different
changes over time in the attributes used in obituaries to describe male vs. female
leaders. Given that the effect was small even in the most extreme case (for Self-
direction), however, they also follow Schwartz et al. (2001), who observed that
gender tended to be useful for explaining variance in Schwartz values data, but
not to a large extent (and not in all cultures). Interestingly, disregarding the
idiosyncrasy of individual newspapers, female status resulted in lower predictions
across all Schwartz values. Acknowledging that these differences were small
(the effect estimates are on the scale of logit-transformed word-by-hop values),
this finding is in contrast with Schwartz and Rubel (2005), who found (albeit
possibly as the result of a sample/measurement interaction) that women did tend
131
to place higher emphasis on some values (such as Universalism) than men. If
this result in the current project is both accurate and large enough to warrant
conceptual attention (vs. solely statistical attention), however, it is useful to note
that obituaries as a data source are unlike typical self-report measures of Schwartz
values: most relevantly, they are written by informants (family members, friends, or
colleagues, commissioned biographers), and are possibly written not actually about
the deceased individual herself, but rather about an idealized (or at least curated)
version of the individual, reflecting perhaps as much in this context about the
author as about the individual. Thus, this finding prompts the question, Regardless
of whether women report prioritizing certain values at a higher rate than men
at the individual level (as in Schwartz and Rubel’s report), do authors (whether
women or men) express these same value preferences when publicly signalling
them about a gendered other (such as a decedent)? If not, obituaries could be
expected to show higher gender differences for communities that have higher levels
of Tradition, and could possibly indicate that some communities culturally see men
more than women as containers of values when described by others.
The best-fitting statistical model (considering AIC) for predicting word-by-
hop values was one that excluded all newspaper/county-level covariates, including
race/ethnicity, income, and education. Given findings reported in previous
literature, this result was unexpected. Specifically, it is counter to Schwartz et
al. (2001, p. 534), who found positive, significant relationships between education
and Self-direction and Stimulation: in the full model that included it, education
had a predicted near-zero or negative effect for both of these values as measured
through word-by-hop. It is also not in line with Schwartz et al.‘s finding of negative
correlations between education and Conformity and Tradition: while education’s
132
predicted relationship in the full model was negative for both of these Schwartz
values, the models that included education were not the best-fitting. It is worth
considering, however, that education was operationalized in this project quite
differently than in Schwartz et al.’s report. First, this project considered the
percentage of the population with a Bachelor’s degree or higher, rather than, as in
Schwartz et al.’s report, a scale between no educational experience and experience
“beyond high school” (p. 534). Second, in the current project, education was
measured not at the individual level but at the county level. The inclusion of
this covariate, as well as those for income and race/ethnicity, was based on the
assumption that, in aggregate, community-level demographic indicators would
provide a “signal” detectable even through the “noise” introduced by not modeling
these characteristics at the individual level, which would have been infeasible
or impossible to do with this dataset. The current project’s findings cast doubt
on this assumption and suggest that future research with a smaller number of
obituaries that can be manually coded for at least some of these and related
covariates (e.g., education, even if not income or self-identified race/ethnicity) could
be fruitful for re-examining those covariates’ statistical utility. As it was modeled
in this project, even income, which, as discussed above in this dissertation’s
introduction, seemed definitionally related to Power, had among the lowest fixed
effect estimates for Power in the full model.
Having acknowledged that the race/ethnicity covariates did not contribute
to the model overall, it is worthwhile to tentatively interpret the full model (Model
#7; Figure 32) as a starting point for future work. Even with a small effect, the
model indicated most notably that for every increase of 1 SD in the community-
level percentage of individuals who self-identified as “Black or African American”
133
and, to a lesser extent, of those who self-identified as “White,” estimated logit-
transformed word-by-hop values were expected to increase by a non-zero amount
for all Schwartz values. This may indicate that areas with higher percentages
of individuals who identify as part of those groups may have norms for less
formalization or greater vividness in their language. However, they may also
indicate an effect that may be true for other racial/ethnic groups as well but is
masked by the data. Only English-language newspapers were considered in this
project, possibly heavily misestimating the “true” word-by-hop values that would
result from incorporating obituaries from Spanish-speaking media in areas with
high populations that identify as “Hispanic.” Somewhat similarly, Figure 10
indicates that most communities included in the sample contained less than 0.5%
individuals who self-identified as “Native Hawaiian and Other Pacific Islander.” It
is an open question whether the effect estimate for this group would be similar if a
newspaper from Hawaii had been included in the sample (as the caption for Figure
3 notes, Legacy.com does not currently contract with any newspapers in Hawaii).
This project expands Bardi et al.’s (2008) approach of understanding values
through ecologically-valid text analysis in two primary aspects. First, the obituaries
used in this project may be expected to comprise a more values-focused corpus
than newspaper texts in general, potentially amplifying the ability to detect values
in the text. Second, this project expands the approach of values-detection from
counting instances of specific words to analyzing the lexical distance for every
word in a text to those words, even if they do not appear explicitly in a given
text. The development of word-by-hop will, I hope, be of use for future research
not only in its concept but also in its implementation: the code that was written
and documented as part of this project will be released alongside this dissertation
134
under an open license, ideally enabling future analyses of other sources of natural
language, possibly with expanded or alternative target lexica. Additionally, the
codebase includes automated algorithms for guessing gender and age; a WordNet
specification (and documentation) for the Neo4J graph database platform; the
University of South Florida Free Associations dataset released (to this author’s
knowledge) for the first time publicly under an explicit and open license, thanks to
the generosity of Nelson, McEvoy, and Schreiber (2004); and the hop count dataset
derived from the obituaries corpus1.
The approach taken by this project included several notable limitations,
several of which may serve as fodder for follow-up research. Most sweepingly,
this project necessarily incorporated several conceptually “fuzzy” methodological
assumptions and approaches, as is common in analyses of free-text data. As
noted above, Figures 15, 16, 17, and 18 suggest that some obituaries were
misclassified by the automated age-guessing algorithm (and thus likely also by
the automated gender-coding algorithm). The use of some non-noun final Synset
nodes when calculating word-by-hop values (as shown in Table 10) reflects the
open, exploratory conceptual approach of this project, which had to release some
definitional rigidity in order to utilize the free-text dataset. The lack of readily-
available baseline English-language word frequency data (such as from the Corpus
of Contemporary American English) prevented word-by-hop values from being
normalized using lemmas’ baseline frequencies; it is possible that future inclusion of
this type of auxiliary data could substantively improve the word-by-hop calculation.
These aspects require that this project be seen accurately as a first step toward
expanding values research that uses this type of corpus. Moving forward from the
1With Legacy.com’s permission, raw obituary data may also be released in the future.
135
foundation provided by this project, including not only its findings but also its
associated code, algorithmic refinements would be a fruitful next area of focus.
This project’s findings point toward several additional exciting areas of
future research. Having demonstrated in a fully free-text corpus that word-by-hop
values do show sensitivity to relative differences in value-expressions, future work
could further explore both the properties of word-by-hop (seeking to explain, e.g.,
the cone-shaped distributions seen in Figures 19, 20, 21, and 22), and new ways
to model it (e.g., comparing the logit-transformation approach used in the models
for this project with beta- and/or multiple-membership regression approaches).
Future analyses on this obituary corpus could explore additional individual-level
characteristics, such as veteran status (e.g., by searching for phrases such as
“Army,” “Navy,” etc.), family membership (by attempting to create a graph of
types of familial relations from the descriptions of surviving family members),
and the overall sentiment of words in each obituary (using, e.g., SentiWordNet,
a fork of the WordNet database used in this project with an additional positive-
vs. negative-sentiment layer; see Baccianella, Esuli, & Sebastiani, 2010). While
obituaries may be expected to be generally positive in tone, i.e., not disparaging
the deceased, they may vary in positive vs. negative word choice.
Given that newspaper/county-level predictors were not statistically useful
(comparing models that included them with a model that excluded them), future
work could usefully explore the effects of aggregating covariates either by larger
geographic regions or, perhaps most likely to alter results, by explicitly modeling
newspapers as multiple members of counties (and, relatedly, obituaries as multiple
members of newspapers, and therefore counties; see Leckie, 2013). Alternatively,
and given the gender-related question posed above, a future project could compare
136
the values invoked in free-text prospective obituaries (either about self or a close
other) written by living participants with their responses to the Schwartz Values
Survey (SVS).
As measured by word-by-hop, Hedonism was highly correlated with
Conformity, its opposite in the Schwartz circumplex model, in the obituaries
corpus. The reasons for this are unclear. However, they do suggest that the
correlation matrix presented in Figure 27 may not reproduce the structure of
Schwartz values that is expected based on previous analyses of individual-level
responses to measures such as the SVS (Schwartz, 2012). This may be the result
of obituary texts being able to express values that are “opposed” to one another
more consistently in aggregate than value rankings made by living people (e.g., in
response to the SVS). The structure of the correlation matrix from the obituaries
corpus could, therefore, usefully be explored (e.g., through factor analysis) in future
work in order to examine whether obituaries not only show differences in values
but also reproduce the expected mappings of those values in relation to each other.
Similarly, using the word-by-hop approach, all Schwartz-value-relevant words could
be removed from the corpus (by removing all words that have a word-by-hop value
above a given threshold for a Schwartz value). The occurrence of the remaining
words in the corpus could then be factor-analyzed to look for clusters of words
that fit Schwartz’ (1992, p. 4) general definition of values, but do not fit any of
the Schwartz values themselves.
This dissertation project had two goals at its outset: first, to describe which
Schwartz values are discussed most across newspaper-communities in the USA, and
second, to provide the initial development of new tools for future obituary-based
“morality mining” work (Christen et al., 2013). Addressing the first goal, results
137
revealed that Power was the most-indicated value across the corpus, followed by
Conformity, Security, Self-direction, and Benevolence. Stimulation and Hedonism
were least-indicated, followed by Universalism, Tradition, and Achievement.
Addressing the second goal, this dissertation is supplemented with algorithm
definitions and code as well as datasets that I hope will facilitate responsible future
research of this type. Results of this project indicated that the answer to Research
Question 1, which asked whether the Schwartz values can be detected in obituaries,
is “Yes,” with the qualification that part of the “signal” being detected may be
an artifact of authors following explicit or implicit templates (although template-
following itself can indicate an adherence to community norms). In response to
Research Question 2, whether Schwartz value expression is related to characteristics
of obituaries at the individual and community levels, the answer is a more qualified
“Yes.” Results indicated that obituaries certainly vary by community in their
distance from the Schwartz values; however, the effect sizes for individual and
community-level covariates were small.
Obituaries are written in response to individual events, but can be
viewed in aggregate as the output of a shared cultural ritual of externalizing
grief and presenting it to one’s wider community. This project explored an
expanded methodology for traversing the constellation of obituaries’ words’
meanings, indicating that obituaries, whether read singly or together, can inform
understanding of community value priorities.
138
APPENDIX
R BASE AND LIBRARY VERSION NUMBERS
Version numbers of R base and R packages used in this project. This table was
generated automatically from package documentation within R; author names are
therefore as the authors wished them to be printed.
Package
Name
Version Number (NA if not installed
through CRAN) Package Citation
base 3.3.1 R Core Team (2016)
bbmle 1.0.18 Bolker and Team (2016)
betareg 3.1.0 Zeileis, Cribari-Neto,
Gruen, and Kosmidis
(2016)
boot 1.3.18 Canty and Ripley (2016)
choroplethr 3.5.2 Lamstein and Johnson
(2016)
choroplethrMaps 1.0 Lamstein (2014)
choroplethrZip (NA)
coda 0.18.1 Plummer et al. (2015)
coefplot 1.2.4 Lander (2016)
coefplot2 (NA)
corrplot 0.77 Wei and Simko (2016)
cowplot 0.6.3 Wilke (2016)
doParallel 1.0.10 Analytics and Weston
(2015)
139
Package
Name
Version Number (NA if not installed
through CRAN) Package Citation
dplyr 0.5.0 Wickham and Francois
(2016)
fmsb (NA)
foreach 1.4.3 Revolution Analytics
and Weston (n.d.)
foreign 0.8.66 R Core Team (2015)
gender 0.5.1 Mullen (2015)
ggmaps (NA)
ggmcmc 1.1 i Marn (2016)
ggplot2 2.1.0 Wickham and Chang
(2016)
glmmadmb (NA)
glmmADMB 0.8.3.3 Skaug, Fournier, Nielsen,
Magnusson, and Bolker
(2016)
grid 3.3.1 ?
gridExtra 2.2.1 Auguie (2016)
gtable 0.2.0 Wickham (2016a)
Hmisc 3.17.4 Harrell (2016)
irr 0.84 Gamer, Lemon, and
¡puspendra.pusp22@gmail.com¿
(2012)
140
Package
Name
Version Number (NA if not installed
through CRAN) Package Citation
jsonlite 1.0 Ooms, Temple Lang, and
Hilaiel (2016)
knitr 1.13 Xie (2016)
koRpus 0.6.5 m.eik michalke (2016)
lme4 1.1.12 Bates, Maechler, Bolker,
and Walker (2016)
lme4a (NA)
magrittr 1.5 Bache and Wickham
(2014)
methods 3.3.1 (Not provided by
package author)
optparse 1.3.2 documentation et al.
(2015)
parallel 3.3.1 (Not provided by
package author)
plotrix 3.6.3 Lemon et al. (2016)
plyr 1.8.4 Wickham (2016b)
RColorBrewer 1.1.2 Neuwirth (2014)
readODS 1.6.2 Schutten and hong Chan
(2016)
reshape 0.8.5 Wickham (2014b)
reshape2 1.4.1 Wickham (2014a)
141
Package
Name
Version Number (NA if not installed
through CRAN) Package Citation
rjags 4.6 Plummer (2016)
rJava 0.9.8 Urbanek (2016)
RNeo4j 1.6.4 White (2016)
rstan 2.12.1 Guo, Gabry, and
Goodrich (2016)
rvest 0.3.2 Wickham (2016c)
shiny 0.13.2 Chang, Cheng, Allaire,
Xie, and McPherson
(2016)
shinystan 2.2.1 Gabry (2016)
sqldf 0.4.10 Grothendieck (2014)
stringdist 0.9.4.1 van der Loo (2016)
stringr 1.0.0 Wickham (2015)
stringr (NA)
testthat (NA)
visNetwork 1.0.1 Almende B.V. and
Thieurmel (2016)
wordnet 0.1.11 Feinerer and Hornik
(2016)
zoib 1.4.1 with contributions from
Yunchuan Kong (2016)
142
REFERENCES CITED
Alfano, M. (2013). Character as moral fiction. Cambridge: Cambridge University
Press.
Almende B.V., & Thieurmel, B. (2016). visnetwork: Network visualization using
’vis.js’ library [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=visNetwork (R package version
1.0.1)
American Psychological Association. (2010). Ethical principles of psychologists and
code of conduct. http://www.apa.org/ethics/code/index.aspx.
Analytics, R., & Weston, S. (2015). doparallel: Foreach parallel adaptor for the
’parallel’ package [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=doParallel (R package version
1.0.10)
Auguie, B. (2016). gridextra: Miscellaneous functions for ”grid” graphics
[Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=gridExtra (R package version
2.2.1)
Baccianella, S., Esuli, A., & Sebastiani, F. (2010). SentiWordNet 3.0: An Enhanced
Lexical Resource for Sentiment Analysis and Opinion Mining. In LREC
(Vol. 10, pp. 2200–2204).
Bache, S. M., & Wickham, H. (2014). magrittr: A forward-pipe operator for r
[Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=magrittr (R package version 1.5)
Bardi, A., Calogero, R. M., & Mullen, B. (2008, May). A new archival approach to
the study of values and value–behavior relations: Validation of the value
lexicon. The Journal of Applied Psychology , 93 (3), 483–497. doi:
10.1037/0021-9010.93.3.483
Bates, D. (2013, September). Son behind vicious obit to mother that went viral
speaks. http://www.dailymail.co.uk/news/article-2419429/Patrick
-Reddick-wrote-vicious-obituary-abusive-mother-unrepentant.html.
Bates, D., Maechler, M., Bolker, B., & Walker, S. (2016). lme4: Linear
mixed-effects models using ’eigen’ and s4 [Computer software manual].
Retrieved from https://CRAN.R-project.org/package=lme4 (R package
version 1.1-12)
143
Bennett, W. J. (1996). The book of virtues (1st ed.). New York: Simon & Schuster.
Berger, P. L. (1969). The sacred canopy: Elements of a sociological theory of
religion. Garden City, N.Y.: Doubleday.
Bolker, B., & Team, R. D. C. (2016). bbmle: Tools for general maximum likelihood
estimation [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=bbmle (R package version 1.0.18)
Bou, B., & Princeton University. (2014). WordNet SQL (WNSQL).
http://wnsql.sourceforge.net/.
Cabot, E. L. (1910). Ethics for children: A guide for teachers and parents. Boston:
Houghton Mifflin.
Canty, A., & Ripley, B. (2016). boot: Bootstrap functions (originally by angelo
canty for s) [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=boot (R package version 1.3-18)
Chang, W., Cheng, J., Allaire, J., Xie, Y., & McPherson, J. (2016). shiny: Web
application framework for r [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=shiny (R package version 0.13.2)
Christen, M., Alfano, M., Bangerter, E., & Lapsley, D. (2013, May). Ethical issues
of ’morality mining:’ Moral identity as a focus of data mining. In
H. Rahman & I. Ramos (Eds.), Ethical data mining applications for
socio-economic development (pp. 1–21). Idea Group Inc (IGI).
Cribari-Neto, F., & Zeileis, A. (n.d.). Beta regression in R.
Davis, K., & Patterson, D. (2012). Ethics of big data (1st ed.). Cambridge:
O’Reilly Media.
Decloux, S. (1991). The Ignatian Way (C. M. Buckley, Trans.). Chicago: Loyola
University Press.
Diogenes Laertius. (c. 350/1853). The lives and opinions of eminent philosophers
(C. D. Yonge, Trans.). London: H. G. Bohn.
documentation, T. L. D. S., examples ported from Allen Day’s getopt package.
Some documentation from the optparse Python module by the Python
Software Foundation. Contributions from Steve Lianoglou, Nikelski, J.,
Mller, K., Humburg, P., & FitzJohn., R. (2015). optparse: Command line
option parser [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=optparse (R package version
1.3.2)
144
Doughty, C. (2014). Smoke gets in your eyes: And other lessons from the
crematory (1st ed.). New York: W. W. Norton & Company.
Feinerer, I., & Hornik, K. (2016). wordnet: Wordnet interface [Computer software
manual]. Retrieved from https://CRAN.R-project.org/package=wordnet
(R package version 0.1-11)
Ferrari, S., & Cribari-Neto, F. (2004, August). Beta Regression for Modelling Rates
and Proportions. Journal of Applied Statistics , 31 (7), 799–815. doi:
10.1080/0266476042000214501
Forstall, R. L., & United States Bureau of the Census Population Division. (1996).
Population of States and Counties of the United States: 1790 - 1990.
Fowler, B. (2005, January). Collective memory and forgetting components for a
study of obituaries. Theory, Culture & Society , 22 (6), 53–72. doi:
10.1177/0263276405059414
Fu, A. S., Plaut, V. C., Treadway, J. R., & Markus, H. R. (2014). Places, products,
and people ”make each other up:” Culture cycles of self and well-being. In
Geographical psychology: Exploring the interaction of environment and
behavior (pp. 275–300). Washington, DC, US: American Psychological
Association.
Gabry, J. (2016). shinystan: Interactive visual and numerical diagnostics and
posterior analysis for bayesian models [Computer software manual].
Retrieved from https://CRAN.R-project.org/package=shinystan (R
package version 2.2.1)
Gamer, M., Lemon, J., & ¡puspendra.pusp22@gmail.com¿, I. F. P. S. (2012). irr:
Various coefficients of interrater reliability and agreement [Computer
software manual]. Retrieved from
https://CRAN.R-project.org/package=irr (R package version 0.84)
Go´mez-Redondo, R., & Garc´ıa Gonza´lez, J. M. (2010). Emergence and verification
of supercentenarians in Spain. In H. Maier, J. Gampe, B. Jeune,
J.-M. Robine, & J. W. Vaupel (Eds.), Supercentenarians (pp. 151–172).
Springer.
Goodwin, G. P., Piazza, J., & Rozin, P. (2014, January). Moral character
predominates in person perception and evaluation. Journal of Personality
and Social Psychology , 106 (1), 148–168. doi: 10.1037/a0034726
Graham, J., Nosek, B. A., Haidt, J., Iyer, R., Koleva, S., & Ditto, P. H. (2011).
Mapping the moral domain. Journal of Personality and Social Psychology ,
101 (2), 366–385. doi: 10.1037/a0021847
145
Grothendieck, G. (2014). sqldf: Perform sql selects on r data frames [Computer
software manual]. Retrieved from
https://CRAN.R-project.org/package=sqldf (R package version 0.4-10)
Guo, J., Gabry, J., & Goodrich, B. (2016). rstan: R interface to stan [Computer
software manual]. Retrieved from
https://CRAN.R-project.org/package=rstan (R package version 2.12.1)
Hacking, I. (1995). The looping effects of human kinds. In D. Sperber, D. Premack,
& A. J. Premack (Eds.), Causal cognition: A multidisciplinary debate (pp.
351–394). Clarendon Press.
Harrell, F. E., Jr. (2016). Hmisc: Harrell miscellaneous [Computer software
manual]. Retrieved from https://CRAN.R-project.org/package=Hmisc
(R package version 3.17-4)
Haybron, D. M., & Tiberius, V. (2012, February). Normative foundations for
well-being policy (Papers on Economics and Evolution). Philipps University
Marburg, Department of Geography.
Hayes, S. C., Strosahl, K. D., & Wilson, K. G. (2011). Acceptance and commitment
therapy, second edition: The process and practice of mindful change (2nd
ed.). New York: The Guilford Press.
Helbing, D. (2015). Thinking Ahead - Essays on Big Data, Digital Revolution, and
Participatory Market Society. Cham: Springer International Publishing.
i Marn, X. F. (2016). ggmcmc: Tools for analyzing mcmc simulations from
bayesian inference [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=ggmcmc (R package version 1.1)
Kestenbaum, B., & Ferguson, B. R. (2010). Supercentenarians in the United
States. In H. Maier, J. Gampe, B. Jeune, J.-M. Robine, & J. W. Vaupel
(Eds.), Supercentenarians (pp. 219–230). Springer.
Lamstein, A. (2014). choroplethrmaps: Contains maps used by the choroplethr
package [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=choroplethrMaps (R package
version 1.0)
Lamstein, A., & Johnson, B. P. (2016). choroplethr: Simplify the creation of
choropleth maps in r [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=choroplethr (R package version
3.5.2)
146
Lander, J. P. (2016). coefplot: Plots coefficients from fitted models [Computer
software manual]. Retrieved from
https://CRAN.R-project.org/package=coefplot (R package version
1.2.4)
Leckie, G. (2013). Module 13: Multiple Membership Multilevel Models. In (pp.
1–61). Bristol: Center for Multilevel Modelling.
Legacy.com. (2016a). About Legacy.com - Life Stories [Text].
http://www.legacy.com/about/about-us.
Legacy.com. (2016b, October). About us: Our history and our dedication to
obituaries and memorials.
http://memorialwebsites.legacy.com/aboutus.aspx.
Legacy.com. (2016c, October). Our international newspaper partners.
http://www.legacy.com/ns/about/newspapers/international.aspx.
Lemon, J., Bolker, B., Oom, S., Klein, E., Rowlingson, B., Wickham, H., . . . Ogle,
D. (2016). plotrix: Various plotting functions [Computer software manual].
Retrieved from https://CRAN.R-project.org/package=plotrix (R
package version 3.6-3)
Liu, F., & Kong, Y. (2015). Zoib: An R Package for Bayesian Inference for Beta
Regression and Zero/One Inflated Beta Regression. R Journal , 7 (2), 34–51.
McAndrew, S. (2014, June). Brutal obituary reveals lives of abuse, neglect.
http://www.rgj.com/story/news/2014/06/25/
brutal-obituary-reveals-lives-of-abuse-neglect/11322269/.
m.eik michalke. (2016). korpus: An r package for text analysis [Computer software
manual]. Retrieved from https://CRAN.R-project.org/package=koRpus
(R package version 0.06-5)
Mikkelson, B. (2013, September). Death penalty.
http://www.snopes.com/media/iftrue/obituary.asp.
Mullen, L. (2015). gender: Predict gender from names using historical data
[Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=gender (R package version 0.5.1)
Nagi, K. (2013). A new representation of WordNet using graph databases. In
DBKDA 2013: The Fifth International Conference on Advances in
Databases, Knowledge, and Data Applications (pp. 1–8).
147
Nelson, D. L., McEvoy, C. L., & Schreiber, T. A. (1998). The University of South
Florida word association, rhyme, and word fragment norms.
http://w3.usf.edu/FreeAssociation/.
Nelson, D. L., McEvoy, C. L., & Schreiber, T. A. (2004, August). The University of
South Florida free association, rhyme, and word fragment norms. Behavior
Research Methods, Instruments, & Computers , 36 (3), 402–407. doi:
10.3758/BF03195588
Neuwirth, E. (2014). Rcolorbrewer: Colorbrewer palettes [Computer software
manual]. Retrieved from
https://CRAN.R-project.org/package=RColorBrewer (R package version
1.1-2)
NewspaperArchive.com. (2016). NewspaperARCHIVE.com: Find Your Family in
4.6 Billion Historic Names. http://newspaperarchive.com/.
Nicholson, E. W. B., Caxton, W., & de Worde, W. (1891). Ars moriendi: That is
to saye the craft for to deye for the helthe of mannes sowle. London:
Bernard Quaritch.
Office for Human Research Protections. (2010, January). 45 CFR 46 [Text].
http://www.hhs.gov/ohrp/regulations-and-policy/regulations/
45-cfr-46/index.html.
O’Neil, C., & Schutt, R. (2013). Doing data science: Straight talk from the frontline
(1st ed.). O’Reilly Media.
Ooms, J., Temple Lang, D., & Hilaiel, L. (2016). jsonlite: A robust, high
performance json parser and generator for r [Computer software manual].
Retrieved from https://CRAN.R-project.org/package=jsonlite (R
package version 1.0)
Ospina, R., & Ferrari, S. L. (2010). Inflated beta distributions. Statistical Papers ,
51 (1), 111–126.
Ospina, R., & Ferrari, S. L. (2012). A general class of zero-or-one inflated beta
regression models. Computational Statistics & Data Analysis , 56 (6),
1609–1623.
Pappas, N., & Zelcer, M. (2014). Politics and philosophy in Plato’s Menexenus (1st
ed.). Milton Park, Abingdon, Oxon; New York: Routledge.
Plummer, M. (2016). rjags: Bayesian graphical models using mcmc [Computer
software manual]. Retrieved from
https://CRAN.R-project.org/package=rjags (R package version 4-6)
148
Plummer, M., Best, N., Cowles, K., Vines, K., Sarkar, D., Bates, D., . . .
Magnusson, A. (2015). coda: Output analysis and diagnostics for mcmc
[Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=coda (R package version 0.18-1)
Porter, M. F. (1980). An algorithm for suffix stripping. Program, 14 (3), 130–137.
Princeton University. (2010). About WordNet. https://wordnet.princeton.edu.
R Core Team. (2015). foreign: Read data stored by minitab, s, sas, spss, stata,
systat, weka, dbase, ... [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=foreign (R package version
0.8-66)
R Core Team. (2016). R: A language and environment for statistical computing
[Computer software manual]. Vienna, Austria. Retrieved from
https://www.R-project.org/
Revolution Analytics, & Weston, S. (n.d.). foreach: Provides foreach looping
construct for r [Computer software manual].
Rodler, C., Kirchler, E., & Ho¨lzl, E. (2001, December). Gender stereotypes of
leaders: An analysis of the contents of obituaries from 1974 to 1998. Sex
Roles , 45 (11-12), 827–843. doi: 10.1023/A:1015644520770
Rosel, N. (1978). Toward a social theory of dying. Omega: Journal of Death and
Dying , 9 (1), 49–55. doi: 10.2190/FGXD-GFGU-H3QM-GCR0
Russell, J. G. (2012, September). Copyright and the obit.
http://www.legalgenealogist.com/blog/2012/09/12/
copyright-and-the-obit/.
Saucier, G. (2009). What are the most important dimensions of personality?
Evidence from studies of descriptors in diverse languages. Social and
Personality Psychology Compass , 3 (4), 620–637. doi:
10.1111/j.1751-9004.2009.00188.x
Schmid, H. (2013). Probabilistic part-of-speech tagging using decision trees. In
Proceedings of International Conference on New Methods in Language
Processing (pp. 1–9).
Schutten, G.-J., & hong Chan, C. (2016). readods: Read ods files [Computer
software manual]. Retrieved from
https://CRAN.R-project.org/package=readODS (R package version
1.6.2)
149
Schwartz, S. H. (1992). Universals in the content and structure of values:
Theoretical advances and empirical tests in 20 countries. Advances in
experimental social psychology , 25 , 1–65.
Schwartz, S. H. (2012). An overview of the Schwartz theory of basic values. Online
Readings in Psychology and Culture, 2 (1), 1–20.
Schwartz, S. H., Melech, G., Lehmann, A., Burgess, S., Harris, M., & Owens, V.
(2001, January). Extending the Cross-Cultural Validity of the Theory of
Basic Human Values with a Different Method of Measurement. Journal of
Cross-Cultural Psychology , 32 (5), 519–542. doi:
10.1177/0022022101032005001
Schwartz, S. H., & Rubel, T. (2005). Sex differences in value priorities:
Cross-cultural and multimethod studies. Journal of Personality and Social
Psychology , 89 (6), 1010–1028. doi: 10.1037/0022-3514.89.6.1010
Scime, A., & Murray, G. R. (2013, May). Social science data analysis: The ethical
imperative. In H. Rahman & I. Ramos (Eds.), Ethical data mining
applications for socio-economic development (pp. 131–147). Idea Group Inc
(IGI).
Skaug, H., Fournier, D., Nielsen, A., Magnusson, A., & Bolker, B. (2016).
glmmadmb: Generalized linear mixed models using ’ad model builder’
[Computer software manual]. (http://glmmadmb.r-forge.r-project.org,
http://admb-project.org)
Smithson, M., & Verkuilen, J. (2006). A better lemon squeezer?
Maximum-likelihood regression with beta-distributed dependent variables.
Psychological methods , 11 (1), 54.
Snopes.com. (2016). http://www.snopes.com/.
Solove, D. J. (2007). ”I’ve Got Nothing to Hide” and Other Misunderstandings of
Privacy. San Diego Law Review , 44 (GWU Law School Public Law Research
Paper No. 289), 745.
The Associated Press. (2013, September). Son who wrote vicious obit for Reno
mom insists ’everything in there was completely true’.
http://www.nydailynews.com/news/national/son-wrote-vicious-obit
-reno-mom-insists-completely-true-article-1.1454890.
Tsvetovat, M., & Kouznetsov, A. (2011). Social Network Analysis for Startups:
Finding connections on the social web (1st ed.). Sebastopol, CA: O’Reilly
Media.
150
United States Bureau of the Census. (2015a). 2010-2014 American Community
Survey 5-Year Estimates (Generated by Jacob Levernier using American
FactFinder) (Tech. Rep.).
United States Bureau of the Census. (2015b). Small Area Income and Poverty
Estimates (Tech. Rep.).
United States Bureau of the Census Population Division. (2015, June).
CC-EST2014-ALLDATA-[ST-FIPS]: Annual County Resident Population
Estimates by Age, Sex, Race, and Hispanic Origin: April 1, 2010 to July 1,
2014 (Tech. Rep.).
University of California, Los Angeles Institute for Digital Research and Education.
(2016). Stata FAQ: How does one do regression when the dependent variable
is a proportion?
http://www.ats.ucla.edu/stat/stata/faq/proportion.htm.
Urbanek, S. (2016). rjava: Low-level r to java interface [Computer software
manual]. Retrieved from https://CRAN.R-project.org/package=rJava
(R package version 0.9-8)
Valeski, J. (2012, November). Social media: Erasable ink? In Q. E. McCallum
(Ed.), Bad data handbook: Cleaning up the data so you can get back to work
(1st ed., pp. 213–224). Sebastopol, CA: O’Reilly Media.
van der Loo, M. (2016). stringdist: Approximate string matching and string
distance functions [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=stringdist (R package version
0.9.4.1)
Wei, T., & Simko, V. (2016). corrplot: Visualization of a correlation matrix
[Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=corrplot (R package version
0.77)
White, N. (2016). Rneo4j: Neo4j driver for r [Computer software manual].
Retrieved from https://CRAN.R-project.org/package=RNeo4j (R
package version 1.6.4)
Wickham, H. (2014a). reshape2: Flexibly reshape data: A reboot of the reshape
package. [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=reshape2 (R package version
1.4.1)
Wickham, H. (2014b). reshape: Flexibly reshape data. [Computer software
manual]. Retrieved from https://CRAN.R-project.org/package=reshape
(R package version 0.8.5)
151
Wickham, H. (2015). stringr: Simple, consistent wrappers for common string
operations [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=stringr (R package version
1.0.0)
Wickham, H. (2016a). gtable: Arrange ’grobs’ in tables [Computer software
manual]. Retrieved from https://CRAN.R-project.org/package=gtable
(R package version 0.2.0)
Wickham, H. (2016b). plyr: Tools for splitting, applying and combining data
[Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=plyr (R package version 1.8.4)
Wickham, H. (2016c). rvest: Easily harvest (scrape) web pages [Computer software
manual]. Retrieved from https://CRAN.R-project.org/package=rvest
(R package version 0.3.2)
Wickham, H., & Chang, W. (2016). ggplot2: An implementation of the grammar of
graphics [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=ggplot2 (R package version
2.1.0)
Wickham, H., & Francois, R. (2016). dplyr: A grammar of data manipulation
[Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=dplyr (R package version 0.5.0)
Wilke, C. O. (2016). cowplot: Streamlined plot theme and plot annotations for
’ggplot2’ [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=cowplot (R package version
0.6.3)
with contributions from Yunchuan Kong, F. L. (2016). zoib: Bayesian inference for
beta regression and zero-or-one inflated beta regression [Computer software
manual]. Retrieved from https://CRAN.R-project.org/package=zoib (R
package version 1.4.1)
Xie, Y. (2016). knitr: A general-purpose package for dynamic report generation in
r [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=knitr (R package version 1.13)
Zeileis, A., Cribari-Neto, F., Gruen, B., & Kosmidis, I. (2016). betareg: Beta
regression [Computer software manual]. Retrieved from
https://CRAN.R-project.org/package=betareg (R package version
3.1-0)
152