Causal Inference Approaches for Understanding the Social Determinants of Substance Abuse and Depression in the 21st Century: The Opioid Epidemic ...

Please do not remove this page

Causal Inference Approaches for Understanding
the Social Determinants of Substance Abuse and
Depression in the 21st Century: The Opioid
Epidemic and the Great Recession
Swift, Samuel L.
https://scholarship.miami.edu/discovery/delivery/01UOML_INST:ResearchRepository/12355360940002976?l#13355512550002976

Swift, S. L. (2019). Causal Inference Approaches for Understanding the Social Determinants of Substance
Abuse and Depression in the 21st Century: The Opioid Epidemic and the Great Recession [University of
Miami].
https://scholarship.miami.edu/discovery/fulldisplay/alma991031447520902976/01UOML_INST:ResearchR
epository

Embargo
Downloaded On 2021/10/22 14:56:27 -0400

Please do not remove this page

UNIVERSITY OF MIAMI

CAUSAL INFERENCE APPROACHES FOR UNDERSTANDING THE SOCIAL
DETERMINANTS OF SUBSTANCE ABUSE AND DEPRESSION IN THE 21ST
 CENTURY: THE OPIOID EPIDEMIC AND THE GREAT RECESSION

 By

 Samuel L. Swift

 A DISSERTATION

 Submitted to the Faculty
 of the University of Miami
 in partial fulfillment of the requirements for
 the degree of Doctor of Philosophy

 Coral Gables, Florida

 May 2019

©2019
 Samuel L. Swift
All Rights Reserved

UNIVERSITY OF MIAMI

 A dissertation submitted in partial fulfillment of
 the requirements for the degree of
 Doctor of Philosophy

 CAUSAL INFERENCE APPROACHES FOR UNDERSTANDING THE SOCIAL
 DETERMINANTS OF SUBSTANCE ABUSE AND DEPRESSION IN THE 21ST
 CENTURY: THE OPIOID EPIDEMIC AND THE GREAT RECESSION
 Samuel L. Swift

Approved:

________________ _________________
Adina Zeki Al Hazzouri, Ph.D., M.Sc. Tali Elfassy, Ph.D., M.P.H.
Assistant Professor Assistant Professor
Department of Epidemiology Division of Epidemiology
Mailman School of Public Health Department of Public Health
Columbia University Sciences

________________ _________________
Hermes Florez, M.D., Ph.D., M.P.H. Zinzi Bailey, Sc.D., M.S.P.H.
Professor Assistant Scientist
Division of Epidemiology Jay Weiss Institute for Health
Department of Public Health Sciences Equity,
 Sylvester Comprehensive Cancer
 Center

________________ ________________
Daniel J. Feaster, Ph.D. Guillermo Prado, Ph.D.
Professor Dean of the Graduate School
Division of Biostatistics
Department of Public Health Sciences

SWIFT, SAMUEL L. (Ph.D., Epidemiology)
 (May 2019)
Causal Inference Approaches for Understanding the Social Determinants of
Substance Abuse and Depression in the 21ST Century: The Opioid Epidemic and
the Great Recession

Abstract of a dissertation at the University of Miami.

Dissertation supervised by Assistant Professor Adina Zeki Al Hazzouri.
No. of pages in text (83)

 The great recession and the opioid epidemic are two defining public health

events of the 21st century. Within the context of these two events, the 21st

century has been marked by increases in mortality due to suicide, alcohol

misuse, and drug overdose among Americans, and especially among white

Americans, warranting further epidemiological investigation. For example, in

recent years a black-white disparity in opioid pain reliever (OPR) misuse has

emerged such that white Americans are more likely to suffer opioid overdoses or

opioid related hospitalizations than black Americans. There is evidence to

support that much of the opioid epidemic is related to overprescribing of OPR

medications, and that whites are more likely to be prescribed these medications

than blacks. In aim 1 of this dissertation I build on prior work which suggests that

discrimination in medical settings may actually be protective against iatrogenic

outcomes in medicine. I postulate that discrimination could result in less

prescribing of OPR medications for blacks and thus a lower risk of OPR misuse.

While most epidemiologic studies rely on a self-reported measure of OPR

misuse, these measures are often an underreport of true OPR misuse when

compared with gold standard measures. Aim 2 of this dissertation examines the

possible bias introduced to the observed associations of aim 1 which uses self-

reported OPR misuse. Finally, in aim 3 of this dissertation, I examined the great

recession of 2008, another defining public health event of the 21st century which

resulted in shocks to financial wellbeing that were felt by persons worldwide, and

its effect on drug misuse, alcohol misuse, and depression. During the great

recession in the United States shocks to financial wellbeing such as

unemployment drastically increased, while median income and household assets

decreased. While shocks to financial wellbeing have been shown to be

associated with increased depressive symptoms, the associations between these

shocks and alcohol and drug use behaviors are less understood.

 As a whole, this dissertation aims to better understand how the opioid

epidemic and the great recession have played out with regards to the social axes

of class and race, using causal inference methods. All three of my aims were

evaluated using the Coronary Artery Risk Development In young Adults study

(CARDIA) which is an ongoing prospective longitudinal cohort study of black and

white adults who were ages 18 to 30 when they were enrolled in the study in

1985. This sample is ideal for these dissertation aims because the participants

were adults during the opioid epidemic and at the peak of their working years

during great recession of 2008 (mean ages 45-50). Additionally, this study

collects rich data on discrimination, as well as socioeconomic data related to

employment, income, debts and assets.

In aim 1 of the dissertation, I first assess the black vs white disparity in

OPR misuse, and then I evaluate whether racial discrimination in medical care

settings mediates this disparity. In my framework, race is the exposure, racial

discrimination in a medical care setting is the mediator, and OPR misuse is the

outcome. I used causal mediation analysis to compute the total effects and

controlled direct effects, which are as follows: the total effect (TE) is the effect of

race on OPR misuse, adjusted for measured confounders; and the controlled

direct effect (CDE) is the effect of race on OPR misuse when discrimination is set

at the same level for all participants, adjusted for measured confounders. I

calculated inverse probability treatment weights (IPTW) for discrimination to

address confounders that also act as mediators. I found that black participants

were more likely to report discrimination in a medical setting (20.3% vs 0.9%)

and less likely to report OPR misuse (OR=0.71, 95% CI= 0.55, 0.93), in the total

effect model. In my final model adjusting for the IPTW of the

confounding/mediating variables, my controlled direct effect suggests that when

everyone is set to being not discriminated against, the disparity is wider such that

black persons are further less likely to report OPR misuse (OR=0.63, 95%

CI=0.45, 0.89) compared to their white counterparts. These findings suggest

racial discrimination in medical settings is a risk factor for OPR misuse rather

than protective.

 In aim 2 of the dissertation, I evaluate how potential measurement error in

discrimination could have influenced the findings of aim 1. Validation studies that

compare self-reported misuse with urine or hair analysis show that self-reporting

of OPR misuse is usually underreported, often at very low levels of sensitivity. In

this aim I use probabilistic bias analysis applying different values of the sensitivity

of self-reported OPR misuse to the observed bivariate relationships between

race and OPR (from aim 1). Assuming a non-differential misclassification and a

sensitivity of 75% among blacks and 90% among whites, which is higher than

that found in validation studies among high risk OPR users, our aim 1 black-white

disparity (i.e. race – OPR association) reflects an overestimate of the true

disparity (bias corrected OR =0.94, 95% CI =0.91, 0.97) vs. observed OR=0.73,

95% CI= 0.57, 0.93). When we change the sensitivity among blacks to a very low

level (50%) with the white sensitivity at 75%, the observed black-white disparity

would actually flip (bias corrected OR= 2.13, 95% CI=1.51, 3.30 vs. observed

OR= 0.73, 95% CI= 0.57, 0.93).

 In aim 3 of the dissertation, I use a fixed effects design to examine how

within person change in financial well-being at time of the great recession

influenced depressive symptoms, alcohol and drug misuse. I consider three

types of shocks to financial wellbeing: unemployment, drops in income, and a

shift in the ratio of household debts to assets. In my final models, I found that

unemployment was associated with an increase in depressive symptoms (β=2.0,

95% CI=0.6, 3.3). A shift in debts to assets ratio was also associated with an

increase in depressive symptoms (β=1.3 95% CI= 0.1, 2.5) but a decrease in

daily mL of alcohol consumed (β=-4.1, 95% CI= -6.7, -1.5). While the

associations between shocks to financial wellbeing and drug misuse were not

statistically significant, the odds ratios suggest that these shocks were a possible

risk for drug misuse.

 Taken together, these three aims illustrate methodological approaches to

evaluating research questions pertaining to some of the most currently pressing

public health issues. The results of aim 1 suggest that racial discrimination in a

medical setting is a risk factor for OPR misuse rather than being protective, and

thus could not explain the seen black-white disparity in OPR misuse. This finding

is consistent with previous literature on discrimination and substance misuse

among other health outcomes. Further research is needed to investigate the

mechanisms behind the black–white disparity in OPR misuse, including the

source of OPR medications. In aim 2 I found that differential misclassification of

OPR misuse at levels similar to those found in validation studies has the potential

to significantly bias observed associations. In aim 3 I found that unemployment

and shifts in debts to assets ratio during the great recession increased

depressive symptoms. A shift in debts to assets ratio also decreased the total mL

of alcohol consumed. These results highlight the complex relationship between

recessions and health and the importance of examining different types recession

related shocks.

Dedication

For Neal Stewart and for all the people in Albuquerque.

 iii

Acknowledgments

 In chronological order of influence on my life: I would like to acknowledge

my parents and their spouses for inspiring me to seek knowledge and believe in

justice. I am grateful to my sister for being an inspiration to my life constantly and

often editing my writing. Academically and professionally I have also had several

mentors who changed my life forever, and it is likely that I would not be here

without them. My high school photography teacher Annie Bromberg taught me

that the world is full of art and beauty if you look at it right. My paramedic field

trainers Randy Anderson (rest in peace) and Sean Robertson took a big chance

on me and gave me my start at Albuquerque Ambulance, which is where I came

of age. Dr. Celia Iriart, my MPH mentor taught me the importance of using public

health as a lens to make larger critiques about society. At the New Mexico

Department of Health Dr. Michael Landen and Dr. Lois Haggard taught me how

to put that worldview into real life public health practice.

 Especially for this dissertation I would like to acknowledge the people on

my committee who are helping me cross this major milestone in my life. Dr.

Feaster and Dr. Florez have supported me through their courses and also as

knowledgeable experts who I look up to in this field. Dr. Bailey is the type of

epidemiologist I would like to be someday, and I am lucky to have had her as a

collaborator and as committee member. Dr. Elfassy went above and beyond

guiding me through multiple analyses and challenges in the program and I am

very grateful to her for going the extra mile to help me learn so much. Finally, my

dissertation chair Dr. Adina Zeki Al Hazzouri has definitely changed the course of

 iv

my life through her mentorship in this program. Under her mentorship I have not

only learned an incredible amount of epidemiology, but more importantly I believe

in myself much more than I did before, which is the best lesson a mentor can

teach. I sincerely hope to take what everyone has taught me and continue to

grow moving forward in my career.

 v

Table of Contents

LIST OF FIGURES……………………………………………...……………………..vii

LIST OF TABLES…………………………………………………………………...…viii

Chapter 1: Introduction…………………………………………………………………1

Chapter 2: Methodological Overview………………………………………………….7

Chapter 3: The Opioid Epidemic and Racial Discrimination in a Medical Care

Setting…………………………………………………………………………………..11

Chapter 4: Probabilistic Bias Analysis for Misclassification……………………….31

Chapter 5: The Great Recession, Depressive Symptoms, Alcohol Misuse and

Drug Misuse…..………………………………………………………………………..46

Chapter 6: Conclusions……………………………………………………………….70

References……………………………………………………………………………..74

Appendix A: Aim 1 Supplemental: SAS Code………………………………………81

Appendix B: Aim 2 Supplemental: STATA Code…………………………………..82

 vi

LIST OF FIGURES

Figure 3.1…..…………………………………………………………………………..12

Figure 3.2 ………………………………………………………………………………13

Figure 3.3..……………………………………………………..………………………18

Figure 4.1..……………………………………………………..………………………35

Figure 4.2..……………………………………………………..………………………35

Figure 5.1..……………………………………………………..………………………53

Figure 5.2..……………………………………………………..………………………54

Figure 5.3..……………………………………………………..………………………62

 vii

LIST OF TABLES

Table 3.1. ……………………………………………………………………………....20

Table 3.2……………………………………………………………………………..…22

Supplementary Table 3.1………………………………………..……………………24

Supplementary Table 3.2………………………………………..……………………25

Supplementary Table 3.3………………………………………..……………………26

Supplementary Table 3.4………………………………………..……………………27

Table 4.1………………………………………..………………………………………37

Table 4.2………………………………………..………………………………………39

Table 4.3………………………………………..………………………………………41

Table 4.4………………………………………..………………………………………41

Table 5.1………………………………………..………………………………………58

Table 5.2………………………………………..………………………………………59

Table 5.3………………………………………..………………………………………60

Table 5.4………………………………………..………………………………………61

 viii

Chapter 1: Introduction

I. Overview
 In the United States the 21st century has been marked by two notable

public health challenges, the opioid epidemic and the great recession. As is

shown by Case and Deaton, after decades of improving all-cause mortality rates

in the United States, in the 21st century all-cause mortality actually increased

among whites, which is a trend not seen in other developed nations. Case and

Deaton attribute this phenomena to increases in drug overdoses, suicides, and

alcohol related mortality, which offset decreases in cancer and heart disease

mortality1. Case and Deaton paint a bleak portrait of this current trend in mortality

in the United States, naming these deaths “deaths of despair”. The broad

objective of this dissertation is to better understand the social determinants of

this modern despair, disentangling how the current social conditions of the United

States may be influencing these unprecedented trends in substance misuse and

depression related outcomes.

 Substance misuse and depressive symptoms are particularly sensitive to

social stressors such as discrimination and shocks to financial wellbeing. Class

and race are two (of many) important axes to consider when examining the ways

that social subordination, inequity, and other unjust conditions existing in modern

life influence health outcomes. In this dissertation I examine two specific aspects

of social determinants related to racial discrimination and financial wellbeing,

understanding that the broader influence of these large constructs falls outside of

the scope of a single dissertation. First, I examine the role of racial discrimination

 1

2

in a medical care setting in the black-white disparity in Opioid Pain Reliever

(OPR) misuse. Second, I examine how measurement error may influence

observed disparities in OPR misuse, which is an important consideration for

epidemiologists wanting to make stronger inferences. Finally, I examine the

relationship between shocks to financial wellbeing during the great recession and

depressive symptoms, alcohol misuse, and drug misuse. As these are difficult

and important questions to ask using observational epidemiological data, I

explore methodological approaches for improving the inferences I can make

about my results. Taken together, the three aims of this dissertation use

emerging causal inference methodologies in epidemiology to provide a critique of

some of the defining public health issues of the current generation.

II. The Need for Causal Inference Methods

 Our three research aims are not only poised to ask difficult questions

about the public health challenges of the United States in the 21st century, they

are designed to explore methodological approaches to exploring these questions.

When evaluating the influence of social factors like discrimination and shocks to

financial wellbeing, traditional methods in observational epidemiology may not be

sufficient. As exposures these constructs are often hard to define, difficult to

measure, unethical to randomize, and subject to confounding, both observed and

unobserved. Meeting causal assumptions is often a challenging task when

investigating the relationships between measures of social or economic factors

and health. When evaluating the role of racial discrimination in a medical setting

in the relationship between race and OPR misuse, careful consideration of the

3

mediation analysis approach and causal structure of this relationship is required.

When assessing relationships using self-reported variables, careful consideration

of the influence of measurement error is also required in assessing this

relationship. Finally, a true controlled experiment in which participants are

randomly assigned different levels of financial status would be difficult and likely

unethical to conduct. In lieu of a truly randomized experiment, the great

recession provides researchers with a natural experiment in which I can exploit a

fixed effects design to examine the associations between shocks to financial

wellbeing and depressive symptoms, substance misuse, and drug misuse,

making stronger inferences about the results. Causal inference methods,

including causal mediation analysis (chapter 3), probabilistic simulation studies to

assess measurement error (chapter 4), and fixed effect models (chapter 5) are all

tools that can fill the gap of better understanding contemporary issues in social

epidemiology.

III. Dissertation Aims

 As explained above, one goal of this dissertation is to provide an example

of how emerging epidemiological methods can be better applied to observational

data to understand the context of social issues in public health today. More

specifically, this dissertation will use causal mediation to examine the mediating

role of racial discrimination in the opioid epidemic, simulation methods to assess

measurement bias in measurement of OPR misuse, and finally fixed effects

methods to examine the ways that the great recession may have influenced

4

mental health, alcohol misuse, and drug misuse. My specific research aims are

as follows:

Aim 1 (chapter 3): Evaluate whether a black–white disparity in OPR misuse

exists, and if so whether racial discrimination in a medical setting mediates such

disparity in OPR misuse, using causal mediation methods.

 H1.1. Black or African Americans will be at a lower risk of OPR misuse

 than whites in my sample.

 H1.2 Discrimination will act as a mediator in this relationship and will

 explain the black–white disparity in OPR misuse. This will be evaluated by

 looking at the controlled direct effect of race on OPR misuse.

Aim 2 (chapter 4): Assess measurement error in self-reported OPR Misuse

across categories of race and racial discrimination in a medical setting, using

probabilistic bias analysis.

 H2.1: When assessing the role of differential misclassification of OPR

 misuse across categories of race, the introduction of significant bias will

 change the effect estimate observed in the black-white disparity in OPR

 misuse and its interpretation. .

 H2.1: When assessing the role of differential misclassification of OPR

 misuse across categories of racial discrimination in a medical setting, the

 introduction of significant bias will change the effect estimate in the

 bivariate relationship between racial discrimination in a medical setting

 and OPR misuse and its interpretation.

5

Aim 3 (chapter 5): Evaluate the influence of three shocks to financial wellbeing

during great recession (employment status, income and debt to assets ratio) on

outcomes related to depressive symptoms, alcohol misuse and drug misuse

using a fixed effects design.

 H3.1. Persons who experienced unemployment, an income drop, or a shift

 in debt to assets ratio during the great recession will have increased

 depressive symptoms following the great recession.

 H3.2. Persons who experienced unemployment, an income drop, or a shift

 in debt to assets ratio during the great recession will have decreased total

 alcohol consumption and lower odds of binge drinking behavior following

 the great recession.

 H3.3. Persons who experienced unemployment, an income drop, or a shift

 in debt to assets ratio during the great recession will have higher odds of

 drug misuse following the great recession.

IV. Summary

 The opioid epidemic and the great recession are two defining American

public health issues in early 21st century. Taken together, research on the health

consequences of these two issues helps us make a deeper critique of issues of

disparity, racial discrimination, and economic inequity in American life, diving

deeper into the conditions influencing Case and Deaton’s “deaths of despair”.

Through the dual lenses of the opioid epidemic and great recession, this

dissertation will explore racial disparities in opioid misuse and the role of racial

6

discrimination in such disparities, and the effect of the great recession on

depressive symptoms, alcohol misuse, and drug misuse outcomes. These

research questions pose serious methodological challenges. Epidemiology has

shifted from the infectious disease focused perspectives of the 19th and 20th

centuries to a focus on chronic disease, substance abuse, and social

epidemiology. Just as threats to public health have evolved, so must my

methodological approaches to understanding the causes and consequences of

social maladies. Causal inference methods, including causal mediation analysis,

probabilistic simulation studies to assess measurement error, and fixed effect

models are all tools that can fill the gap of better understanding contemporary

issues in social epidemiology.

Chapter 2: Methodological Overview
I. Overview

 In this chapter I review the cohort study I employ in this dissertation, I

broadly review the methodological approaches that I use to evaluate each aim.

The measures, methods and statistical approaches for each aim are discussed in

greater detail in each study’s respective chapter 3, 4 and 5.

II. Study Sample

 We evaluated all three aims using the Coronary Artery Risk Development

in young Adults (CARDIA) sample. The CARDIA study is an ongoing multisite

prospective cohort study of the determinants of clinical and subclinical

cardiovascular disease. In 1985 and 1986 a total of 5,114 black and white adults

were recruited into this study from four field centers which include: The University

of Alabama at Birmingham, The University of Minnesota, Northwestern

University, and Kaiser Permanente in Oakland California. Participants were

followed from 1985-86 to the present with follow up visits in 1987-88, 1990-91,

1992-93, 1995-96, 2000-01, 2005-06, 2010-11, and 2015-16. Details of the study

design are described elsewhere2. The study was IRB approved at each field

center.

 The CARDIA Study is especially well suited to answer the questions

posed in this dissertation because it includes a sample of black and white adults

who were going into their 30s, 40s and early 50s during the opioid epidemic and

the great recession. As is noted by Case and Deaton, the largest increases in

“deaths of despair” were among white males in their 40s3, making this sample

 7

8

well suited for assessing these research questions. Furthermore, the CARDIA

study includes a wealth of information on the social determinants of health, such

as discrimination and financial wellbeing. These determinants are described in

greater detail in the methods sections of chapters 3, 4 and 5.

III. Aim 1 (Chapter 3) Methodological Overview

 In Aim 1 I employ mediation models using causal mediation structures for

binary outcomes as outlined by VanderWeele and Vansteelandt4. Causal

mediation approaches build on traditional mediation approaches outlined by

Baron and Kinney5 by allowing for exposure mediator interactions and using

counterfactual logic to assess more causal effects. For this Aim, I will first

examine the black-white disparity in OPR misuse. Second, I will employ a causal

mediation framework to examine whether racial discrimination in a medical

setting may be mediating such disparity. I use marginal structural models and

construct inverse probability treatment weights to account for time variant

confounding variables that may also act as mediating variables.

IV. Aim 2 (Chapter 4) Methodological Overview.

 When compared to gold standard instruments like hair or urinalyses6-11,

previous research shows that underreporting is common when OPR misuse

(outcome of interest in Aim1) is self-reported. The objective of this aim is then to

examine different situations related to misclassification of OPR misuse, and how

9

different levels of sensitivity of self-reported OPR misuse could be biasing the

observed effect sizes of the relationships in Aim 1.

 A full discussion of the methodological approach for Aim 2 is provided in

more detail in chapter 4. I examined OPR misclassification scenarios across

categories of race and racial discrimination in a medical setting. I employ the

episensi package in STATA12, which is a tool made for these types of

probabilistic bias analyses. The episensi package allows me to quantify the size

of the bias, inputting various levels of sensitivity and specificity (bias parameters)

on the classification of my outcome (OPR misuse) and examining the influence

that these levels of bias parameters has on my observed effect. The episensi

package does this by allowing users to specify a probability density function for

each bias parameter, and then uses Monte Carlo simulations to generate

simulated datasets based on summary data from the measured data. The

episensi package then averages effect sizes across a user specified number of

simulated datasets, producing “bias adjusted estimates”.

V. Aim 3 (Chapter 5) Methodological Overview

 Recessions provide researchers with an opportunity to take advantage of

a “natural experiment” using observational data. The structure of the fixed effects

design is described in detail in some texts13, but the overall idea of this design is

rather simple. By examining within person variation rather than between person

variation during an exogenous event such as a recession, the fixed effects

design compares individuals to themselves at different time points. Said

differently, in a fixed effects design individuals act as their own controls, which

10

offers several advantages from a causal inference perspective. By having

individuals act as their own controls, this design removes the threat of time

invariant confounding, both measured and unmeasured. Additionally, if an

exposure is truly exogenous, the fixed effects design allows for epidemiologists

to say a study is quasi-experimental rather than observational, also allowing for

stronger causal inferences.

Chapter 3: The Opioid Epidemic and Racial Discrimination in a Medical Care
 Setting

I. Chapter 3 Introduction

 Drug overdose death rates within the United States nearly tripled between

1999 and 201514. Vital records data, opioid sales data, and other sources

suggests that a large portion of this increase in deaths is attributable to Opioid

Pain Relievers (OPRs) such as oxycodone, hydromorphone and

oxymorphone15,16. The National Institute on Drug Abuse (NIDA) defines misuse

of prescription drugs as “taking a medication in a manner or dose other than

prescribed; taking someone else’s prescription, even if for a legitimate medical

complaint such as pain; or taking a medication to feel euphoria (i.e., to get

high)”17.

 During the prescription opioid epidemic, death rates attributed to OPR

overdose among whites have been more than twice that of the rates seen among

blacks16, and whites have had significantly higher rates of hospitalization due to

OPRs than blacks18. This interesting pattern black vs white substance misuse is

reflected in all cause drug overdose death can be seen in figure 1.1, which was

produced using vital records data. From the 1990s to early 2010s OPR misuse

was attributable mostly or in part to patterns of OPR overprescribing in the United

States15. The sales of OPRs within the United States increased by almost 500%

between 1999 and 2011 and the same time the rates of OPR related deaths

were tripling15, as can be seen in figure 1.2. Given this knowledge, it would follow

that any racial/ethnic disparities in OPR misuse could be related to disparities in

 11

12

OPR prescribing, based on the premise that less prescribing for minorities might

result in fewer opportunities for misuse.

Figure 3.1 Drug overdose death by year and race within the United States, 1999-
2016

Source: Centers for Disease Control14

 A recent review paper found that across 70 studies of prescribing patterns,

racial and ethnic minorities received less treatment for acute and chronic pain

than their white counterparts19. A study of disparities in opioid prescribing in pain-

related emergency room visits also showed blacks to have a lower odds of

receiving OPRs than whites20. These disparities pose several interesting

questions related to discrimination in medical care and drug misuse in the United

States. Perceived interpersonal discrimination (henceforth called discrimination)

13

Figure 3.2 Opioid sales, deaths and treatment admissions within the United
States, 1999 to 2010.

Source: Kolodny et al, 201515

due to race within the medical care setting has been associated with negative

healthcare service outcomes and underutilization of preventive services21,22.

Using an experimental design and actors playing cardiac patients, one study

found that given identical clinical presentations for white and black patients,

physicians were more likely to recommend white patients for cardiac

catheterization than black patients23. In another study, Krieger et al. hypothesized

that the lack of prescription of hormone therapy to black women, due to

discrimination, was likely the reason why black women had lower rates of breast

14

cancer in the 1990s before the results of the Women’s Health Initiative Study

showed an association between these hormones and breast cancer24.

 In this chapter I will examine the relationship between race and OPR

misuse (i.e. the black-white disparity), and whether discrimination in a medical

setting acts as a mediator in this relationship. Following the evidence from the

studies noted above23,24, I hypothesized that racial discrimination in a medical

setting would result in under prescribing of OPR to blacks, which would then

result in lower OPR misuse among blacks and thus explain the black-white

disparity.

II. Chapter 3 Methods

a. Mediator: Perceived discrimination in a medical setting prior to year 2000

 Discrimination is measured using the Experiences of Discrimination (EOD)

instrument, which has been previously validated25. The discrimination questions

were asked in years 1992, 2000, 2005, 2010 and 2015. Study participants were

asked “have you ever experienced discrimination, been prevented from doing

something, or been hassled, or made to feel inferior in any if the following

situations because of your race?”. They were then presented with

situations/domains, which included, “at school,” “getting a job,” “getting housing,”

“at work,” “at home,” “getting medical care,” or “on the street in a public setting”.

In this study, I used discrimination because of race in “getting care in a medical

setting” (and not in any other setting). Since the question reflects lifetime

experience, I used discrimination from exam years 1992 and 2000 for my

15

exposure of interest, thus reflecting lifetime discrimination prior to year 2000.

These two exam years were combined to increase the sample size of persons

who experienced discrimination in a medical setting, while I used other

approaches/definitions in sensitivity analyses (described later).

b. Outcome: Nonmedical Opioid Pain Reliever (OPR) Misuse from 2005 to 2015

 CARDIA participants were asked about lifetime illicit drug misuse in all

examination years. For this analysis, I combined lifetime OPR misuse in exam

years 2005, 2010 and 2015 as my outcome, excluding people who reported OPR

or opiate misuse in years 2000 and prior. This allowed us to capture new users

and that OPR misuse took place after my exposure period (defined above). In

exam year 2005, the question combined heroin and prescription opioids, with a

follow-up question that specifically asked if a participant used “Heroin,”

“Dilaudid,” Morphine,” “Demerol,” “Oxycodone,” “Hydromorphone,” or “other

prescription opioids” for non-medical reasons. For exam year 2005, participants

that used only heroin and no other opioids were not included as being OPR

misusers. In all subsequent exam years OPR misuse (non-medical use or

prescription opioids) was asked about in a question separate from non-

pharmaceutical opiate misuse (heroin and other non-prescription opiates), so

these exclusions were not necessary.

16

c. Covariates

Participants reported their race (white vs. black), sex, age, years of educational

attainment, total household income (collected in 8 brackets and treated

continuously using the midpoint dollar amount of bracket), and health insurance

status (insured vs. not). Parental SES was operationalized measured as the

highest level of education in years achieved by any parent or guardian.

Depressive symptoms were measured using the Center for Epidemiological

Studies depression scale (CES-D) instrument, which includes 20 items and

ranges from 0-6026. All covariates were measured in year 1992 (study baseline,

and when discrimination was first measured), except for depressive symptoms

which was measured in year 1990.

d. Statistical Analysis

 Of the 4617 participants who were present at any exam between 1992

and 2015, I excluded 264 participants without a measure of lifetime

discrimination in 1992/2000. Since the OPR question is also a lifetime measure, I

further excluded 318 participants who by exam year 2000 had misused OPR at

least once in their lifetime, and further excluded 507 participants with missing

OPR measures in 2005, 2010 and 2015. This left a final analytical sample of

3528 persons. As such, for this analysis, I established temporality between

exposure and outcome by defining the exposure period as lifetime medical

discrimination prior to year 2000, and the outcome period as any OPR misuse in

years 2005 through 2015.

17

 I first compared black and white participants across my covariate

characteristics using chi-squared tests and t-tests. Next, I used a causal

mediation framework to assess whether there is a racial disparity in OPR misuse,

and if so, whether this disparity could be explained by discrimination. The

directed acyclic graph (DAG) for this research question is presented in Figure 1.

Race as the exposure/ treatment (white =0 (reference), black=1), discrimination

in a medical care setting as the mediator (M=1 discrimination, M=0 no

discrimination), and OPR misuse as the outcome (Y=1 OPR misuse, Y=0 no

OPR misuse).

 I performed causal mediation analysis as described by Valeri and

VanderWheele27, a method that uses counterfactual contrasts to compute a total

effect (TE) and a controlled direct effect (CDE) for the relationship of interest.

The total effect (TE) is the effect of race on OPR misuse, adjusted for measured

confounders; and the controlled direct effect (CDE) is the effect of race on OPR

misuse when discrimination is set at the same level for all participants (either

M=0 or M=1), adjusted for measured confounders. I focus on the CDE, setting M

to 0 (i.e. no discrimination), because it makes little sense to imagine the racial

inequality if both whites and blacks all encountered racial discrimination in

medical care. I consider the following variables as confounders of the

discrimination–OPR path: parental SES, sex, age, and study site. Variables such

as education, income, depressive symptoms, and insurance status are

confounders of the discrimination–OPR path but are also descendants of race

and thus mediators of the race–OPR path (see Figure 3.3).

18

Figure 3.3. Directed Acyclic Graph for Proposed Relationship Between Race,
Discrimination in a Medical Care Setting, and OPR Misuse.

 Since my outcome was rare (

19

race as describing the racial inequality in OPR misuse, and the controlled direct

effect (CDE, setting the mediator to zero) as the magnitude of the racial

inequality I would expect if discrimination were absent for both blacks and whites.

 Due to the nature of mediation models, I conducted the following

sensitivity analyses. First, I performed an analysis excluding participants who

reported lifetime discrimination in 1992 but not in 2000, as this pattern is

indicative of inconsistent responses to the question. Second, I conducted an

analysis examining “recent” discrimination rather than lifetime discrimination, by

including only persons who reported discrimination in a medical setting in 2000

and not in 1992 or prior. Third, since the opioid epidemic has been characterized

by increases in prescribing of OPRs for various conditions related to both chronic

and acute pain15, I conducted an analysis adjusting for self-reported pain in year

2000 as a proxy variable for pain related conditions that may influence OPR

misuse. Participants were asked the following question: “During the past 4

weeks, how much did pain interfere with your normal work (including work

outside the home and housework)?”; and responses were coded as “no pain, a

little bit of pain to moderate pain, or quite a bit of pain to extreme pain”. In a final

sensitivity analysis, I analyzed the effect of discrimination on the racial disparity

in OPR misuse outside of a causal mediation framework. The latter was done by

assessing the association between discrimination and OPR misuse, stratified by

race. All analyses were conducted using SAS version 9.4 and SAS University

Edition Software29.

20

III. Chapter 3 Results

 As shown in table 3.1, black participants were much more likely to report

experiencing discrimination in a medical setting (20.3% vs 0.9%) and less likely

to report OPR misuse (5.8% vs 8.0%). Black participants were more likely to be

female, have lower incomes, lower educational attainment, and lower parental

SES, and were more likely to have elevated depressive symptoms (CES-D≥16).

Table 3.1. Baseline Characteristics of Study Participants by Race, CARDIA
Study, 1992-2015
 Black White P
 n= 1717 n= 1811 value
 Age (years), mean, (SD) 31 (3.8) 32 (3.4)

21

CI=0.55, 0.93) compared to their white counterparts. Adjusting for all measured

confounders, the controlled direct effect in model 2 suggests that when everyone

is set to being not discriminated against (M=0), black persons would have 0.63

times the odds of reporting OPR misuse (OR=0.63, 95% CI=0.45, 0.89)

compared to their white counterparts. Adjusting for all measured confounders as

well as confounders/mediators, the controlled direct effect in model 3 suggests

that when everyone is set to being not discriminated against (M=0), black

persons would have 0.55 times the odds of reporting OPR misuse (OR=0.55,

95% CI=0.38, 0.80) compared to their white counterparts. Finally, the controlled

direct effect in model 4 which uses stabilized weights to account for measured

confounding, suggests that when everyone is set to being not discriminated

against (M=0), black persons would have 0.65 times the odds of reporting OPR

misuse (OR=0.65, 95% CI=0.46, 0.91) compared to their white counterparts.

 The results of my four sensitivity analyses can be seen in Supplementary

Tables 3.1, 3.2, 3.3, and 3.4. In my first sensitivity analysis (Supplemental Table

3.1), I assessed measurement error in the lifetime discrimination measure by

excluding 127 participants who inconsistently responded to the discrimination

measure across years 1992 and 2000, and my results were largely similar to

those seen in my main analysis (Table 3.2). In my second sensitivity analysis

(Supplemental Table 3.2) including only “recent” experience of discrimination in a

medical setting, my sample size of persons who experienced discrimination was

drastically reduced (n=116) and thus I were unable to draw conclusions from

these models, however the point estimates were in the same direction as my

Table 3.2. Relationship of Race, Discrimination in a Medical Setting, and OPR Misuse Using Causal Mediation Methods,
CARDIA Study (N=3528)
 Total effect CDE: Adjusted for CDE: Adjusted for Marginal
 (model 1) discrimination and discrimination, structural model
 confounders confounders, and (model 4)
 (model 2) confounders/mediators
 (model 3)
 OR 95% CI OR 95% CI OR 95% CI OR 95% CI

Black vs. white 0.71 (0.55, 0.93) 0.63 (0.45, 0.89) 0.55 (0.38, 0.80) 0.65 (0.46, 0.91)

No discrimination
vs. discrimination 0.54 (0.34, 0.86) 0.50 (0.31, 0.83) 0.51 (0.31, 0.83)

Model 1 is adjusted for race; Model 2 is adjusted for race, medical discrimination and confounders (parental SES, age, sex, and study site); Model
3 is adjusted for race, medical discrimination, confounders (parental SES, age, sex, and study site), and confounders/mediators (education,
income, depressive symptoms, and insurance status); Model 4 is adjusted for race and medical discrimination, and uses stabilized inverse
probability weights to account for the confounders and confounders/mediators (see supplemental SAS code).

 22

23

main analysis. In my sensitivity analysis adjusting for self-reported pain as a

confounder (Supplemental Table 3.3), the results were also similar to those seen

in table 3.2. In my final sensitivity analysis (Supplemental Table 3.4), when the

sample is stratified by race, I see that discrimination in a medical setting is

associated with significantly greater odds of OPR misuse among blacks (OR=

1.74, 95% CI= 1.06, 2.86), and similarly a greater odds among whites, though not

statistically significant (OR= 3.15, 95% CI=0.87, 11.45).

IV. Chapter 3 Discussion

Our findings are counter to my original hypothesis that discrimination in a medical

setting would be protective against iatrogenic effects of healthcare-related opioid

misuse and thus a mediator of the black-white disparity. While my total effect of

0.71 suggests that blacks have lower odds of OPR misuse compared to whites –

which is in line with the established rates of racial disparity in OPR misuse16,18 –

in my final model my controlled direct effect of 0.65 became more protective (i.e.

further away from the null), thus widening the disparity and suggesting that

discrimination in a medical setting is a risk factor, rather than a protective factor,

for OPR misuse.

We hypothesized that racial discrimination in a medical setting would result in

less treatment and thus under prescribing of OPR to blacks, which would then

result in lower OPR misuse among blacks, and thus explain the observed black-

white disparity. As my findings were counter to my hypothesis, I must consider

the possibility that misused OPR medications originated from potential sources

Supplementary Table 3.1: Relationship of Race, Discrimination in Medical Settings, and OPR Misuse Using Causal
Mediation Methods, Restricted to Participants Whose Answer to Lifetime Discrimination was Consistent Across Years
1992 and 2000, CARDIA Study (N=3,401)
 Total effect CDE: Adjusted for CDE: Adjusted for Marginal structural model
 (model 1) discrimination and discrimination, (model 4)
 confounders confounders, and
 (model 2) confounders/mediators
 (model 3)
 OR 95% CI OR 95% CI OR 95% CI OR 95% CI
 Black vs. white 0.71 (0.55, 0.93) 0.64 (0.45, 0.9) 0.54 (0.37, 0.79) 0.65 (0.46, 0.91)

 No discrimination
 vs. discrimination 0.61 (0.34, 1.01) 0.52 (0.28, 0.95) 0.52 (0.29, 0.95)
Model 1 is adjusted for race; Model 2 is adjusted for race, medical discrimination and confounders (parental SES, age, sex, and study site); Model
3 is adjusted for race, medical discrimination, confounders (parental SES, age, sex, and study site), and confounders/mediators (education,
income, depressive symptoms, and insurance status); Model 4 is adjusted for race and medical discrimination, and uses stabilized inverse
probability weights to account for the confounders and confounders/mediators (see supplemental SAS code).

 24

Supplementary Table 3.2: Relationship of Race, Recent Discrimination in Medical Settings, and OPR Misuse Using
Causal Mediation Methods, CARDIA Study (N=3,528)
 Total effect (model CDE: Adjusted for CDE: Adjusted for Marginal structural model
 1) discrimination and discrimination, (model 4)
 confounders confounders, and
 (model 2) confounders/mediators
 (model 3)
 OR 95% CI OR 95% CI OR 95% CI OR 95% CI
 Black vs. white 0.71 (0.55, 0.93) 0.70 (0.51, 0.98) 0.61 (0.43, 0.88) 0.72 (0.52, 0.98)

 No discrimination
 vs. discrimination 0.77 (0.35, 1.72) 0.66 (0.29, 1.51) 0.66 (0.29, 1.49)
Recent discrimination was constructed by restricting to participants answering “Yes” to Lifetime Discrimination in 2000 but “No” in 1992.
Model 1 is adjusted for race; Model 2 is adjusted for race, medical discrimination and confounders (parental SES, age, sex, and study site); Model
3 is adjusted for race, medical discrimination, confounders (parental SES, age, sex, and study site), and confounders/mediators (education,
income, depressive symptoms, and insurance status); Model 4 is adjusted for race and medical discrimination, and uses stabilized inverse
probability weights to account for the confounders and confounders/mediators (see supplemental SAS code).

 25

Supplementary Table 3.3: Relationship of Race, Discrimination in Medical Settings, and OPR Misuse Using Causal
Mediation Methods, additionally adjusted for Self-Reported Pain in Models 3 and 4 (N= 3,528)
 Total effect (model CDE: Adjusted for CDE: Adjusted for Marginal structural model
 1) discrimination and discrimination, (model 4)
 confounders confounders, and
 (model 2) confounders/mediators
 (model 3)
 OR 95% CI OR 95% CI OR 95% CI OR 95% CI
 Black vs. white 0.71 (0.55, 0.93) 0.63 (0.45, 0.9) 0.56 (0.38, 0.84) 0.63 (0.44, 0.90)

 No discrimination
 vs. discrimination 0.54 (0.34, 0.86) 0.53 (0.31, 0.90) 0.52 (0.31, 0.88)
Model 1 is adjusted for race; Model 2 is adjusted for race, medical discrimination and confounders (parental SES, age, sex, and study site); Model
3 is adjusted for race, medical discrimination, confounders (parental SES, age, sex, and study site), and confounders/mediators (education,
income, depressive symptoms, insurance status and self-reported pain); Model 4 is adjusted for race and medical discrimination, and uses
stabilized inverse probability weights to account for the confounders and confounders/mediators including self-reported pain.

 26

Supplementary table 3.4. Association of Discrimination with OPR Misuse, Overall and by Race, CARDIA Study
 Overall Black White
 (n=3528) (n=1717) (n=1811)
 OR 95% CI OR 95% CI OR 95% CI
 Unadjusted
 No discrimination - - -
 (ref)
 Discrimination 1.27 (0.86, 1.88) 1.56 (0.99, 2.46) 2.49 (0.71, 8.78)
 Adjusted
 No discrimination - - -
 (ref)
 Discrimination 1.48 (0.97, 2.28) 1.74 (1.06, 2.86) 3.15 (0.86-11.45)
*Models adjusted for sex, age, study site, and parental SES

 27

28

other than the medical setting, such as the diversion of OPR medications and

black market sale of these medications. A recent paper estimated that 42% of

OPRs prescribed in the emergency department may be ultimately misused,

either by the patient or persons close to them30. Given these diversion rates,

prescribing practices may have little impact on the actual availability of OPR

medications for persons who would misuse them. This could be a larger problem

in the future, as recent research points to a new shift in the opioid epidemic since

2010, wherein the large disparity in black versus white OPR misuse from the

1990s to 2010 (largely driven by prescription OPRs) has narrowed in recent

years due to the return to synthetic and illicit OPRs31.

 Our findings suggest that discrimination in a medical setting is a risk factor

for OPR misuse. This is further illustrated in my stratified sensitivity analysis,

where discrimination in a medical setting increased the odds of OPR misuse

among blacks as well as whites. The relationship between discrimination in a

medical care setting and OPR misuse I found is likely similar to the relationship

observed between discrimination in any other setting and negative outcomes. In

this cohort, experiences of discrimination in any setting have been associated

with several negative health outcomes, including lower birthweight among

mothers who reported such experiences32, greater waist circumference33, and

increased sedentary behaviors34. There is also a large body of evidence

supporting the relationship between perceived discrimination across all settings

and increased substance misuse and risky behaviors among both whites, blacks

and other groups35-42, suggesting that substance misuse could be a coping