Associations of perceived discrimination with health outcomes and health disparities in the All of Us cohort

Abstract

Objectives

The goal of this study was to investigate the association of perceived discrimination with health outcomes and disparities.

Materials and Methods

The study cohort consists of 60 180 participants from the 4 largest self-identified race and ethnicity (SIRE) groups in the All of Us Research Program participant body: Asian (1291), Black (4726), Hispanic (5336), and White (48 827). A perceived discrimination index (PDI) was derived from participant responses to the “Social Determinants of Health” survey, and the All of Us Researcher Workbench was used to analyze associations and mediation effects of PDI and SIRE with 1755 diseases.

Results

The Black SIRE group has the greatest median PDI, followed by the Asian, Hispanic, and White groups. The Black SIRE group shows the greatest number of diseases with elevated risk relative to the White reference group, followed by the Hispanic and Asian groups. Perceived discrimination index was found to be positively and significantly associated with 489 out of 1755 (27.86%) diseases. “Mental Disorders” is the disease category with the greatest proportion of diseases positively and significantly associated with PDI: 59 out of 72 (81.94%) diseases. Mediation analysis showed that PDI mediates 69 out of 351 (19.66%) Black-White disease disparities.

Discussion

Perceived discrimination is significantly associated with risk for numerous diseases and mediates Black-White disease disparities in the All of Us participant cohort.

Conclusion

This work highlights the role of discrimination as an important social determinant of health and provides a means by which it can be quantified and modeled on the All of Us platform.

Introduction

The expanding focus on the social determinants of health (SDOH), also referred to as social drivers of health, in biomedical research is reflected in the growing literature on SDOH in the United States and beyond.^1,2 The World Health Organization defines SDOH as “the nonmedical factors that influence health outcomes.” Social determinants of health can include income, education, healthcare access, housing stability, and social support, among other factors related to the conditions and environments under which individuals live and work.³ Social determinants of health are estimated to influence up to 60% of health outcomes,⁴ with adverse SDOH exposures being associated with poorer health outcomes. For instance, infant mortality is significantly higher in rural and poorer areas.⁵ Furthermore, mortality and prevalence rates for specific diseases such as COVID-19, cancer, diabetes, and cardiovascular disease are greater for those of lower socioeconomic status (SES) and for those residing in deprived neighborhoods.^6–9

Discrimination—the unequal treatment of individuals on the grounds of physical characteristics or membership to particular demographic groups—is an SDOH with negative effects on health outcomes. Such negative effects often manifest as mental health disorders, as individuals who experience discrimination demonstrate greater risk of ailments such as psychosis, suicidal ideation, and depression.^10–12 Although to a generally lesser extent, discrimination is also a predictor for poorer physical health.¹³ One potential mechanism through which discrimination may worsen physical health outcomes is increasing the likelihood of engaging in behaviors that negatively impact health, such as substance abuse.^14–16 Discrimination may also impact physical health through more direct routes, such as by increasing red blood cell oxidative stress and affecting cortisol levels in the body.^17,18

Discrimination may contribute to health disparities. The National Institute on Minority Health and Health Disparities (NIMHD) defines health disparities as “health difference[s] that adversely affect disadvantaged populations.”^19,20 Health disparities are widespread in the United States, with minority racial groups faring worse than the majority White population across several health measures, including life expectancy, disease prevalence, and perceived health.^21,22 Racial discrimination may contribute to these disparities by driving racial differences in SES, a known driver of health disparities.²³ Racial discrimination may also result in minority racial groups receiving poorer quality of health care compared to White patients.²⁴ Furthermore, as racial minority groups in the United States generally report greater levels of perceived discrimination than their White counterparts, such groups may be more vulnerable to the negative health effects that accompany discrimination.^25,26 As such, discrimination may help explain the disproportionately negative health outcomes experienced by disadvantaged groups.

Associations between perceived discrimination and health disparities/outcomes have been previously investigated in prior studies. These studies often use surveys such as the Everyday Discrimination Scale to measure perceived discrimination.²⁷ Principal component analysis is a common instrument used to quantify survey responses, though other studies may analyze the sums or means of ordinally coded survey responses.^28–31 Such investigations have demonstrated that perceived discrimination is associated with poorer health outcomes and with racial disparities in disease burden. For instance, perceived discrimination may contribute directly to increased risk of cardiovascular disease and may discourage individuals belonging to minority and ethnic groups from pursuing health care.^32–35

We hypothesized that a connection between discrimination and health disparities may be uncovered using the All of Us Research Program’s (abbreviated as All of Us hereafter) data on perceived discrimination, demography, and health outcomes. All of Us is a federal initiative launched in 2015 with the goal of advancing precision and health equity through the collection of genetic, health, and demographic data from over 1 million American volunteers.³⁶ Among the program’s priorities is the recruitment of individuals from demographic groups that have been historically underrepresented in biomedical research. The resulting diversity of its participant body, coupled with the program’s wealth of data on health status and discrimination, make it a compelling platform from which to study the confluence of discrimination and health disparities.

The data on perceived discrimination present on the All of Us platform takes the form of participant responses to a series of survey questions. The first aim of this study was to use these survey responses to derive a quantitative measure of perceived discrimination that can be associated with health outcomes. The second aim of this study was to use this derived metric to assess how perceived discrimination is associated with health outcomes and health disparities across the All of Us participant cohort. The metric produced in this study can aid future research on discrimination in the All of Us cohort by providing a means by which perceived discrimination can be quantified and modeled. The accompanying analyses elucidate the role of perceived discrimination in driving health disparities and provide a broad picture of how discrimination affects health and disease.

Materials and methods

Study cohort

The cohort for this study was assembled and analyzed using participant data obtained from the cloud-based All of Us Researcher Workbench. Volunteers can enroll in the program online through JoinAllofUs.org or through a participating healthcare provider. Enrollment is restricted to individuals who are 18 years of age or older and to those residing in the United States or in a US territory. Individuals who are incarcerated or unable to provide consent are not eligible to enroll in the program.

All of Us participant data were taken from the Registered Tier Dataset v7 (curated version R2022Q4R9), from which participant demographic, electronic health record (EHR), and survey data were obtained. Enrollment began on May 31, 2017. The cutoff date for v7 of the Registered Tier Dataset is July 1, 2022. Extracted demographic data consisted of participant self-identified race and ethnicity (SIRE), date of birth, and sex at birth. International Classification of Diseases codes (ICD-9-CM and ICD-10-CM) were extracted from participant EHR data and mapped to 1755 disease phecodes to designate participants as either a case or control for each disease.³⁷

The study cohort was restricted to participants who had SIRE data, were assigned either male or female at birth, as to facilitate the use of sex as a covariate in regression analyses, who have EHR data available, and who have provided responses to survey questions on perceived discrimination.

Race and ethnicity

In the All of Us survey titled “The Basics,” participants are asked to select 1 or more of 7 racial and ethnic categories that they most closely identify as: (1) American Indian or Alaska Native, (2) Asian, (3) Black, (4) Hispanic or Latino, (5) Middle Eastern or North African, (6) Native Hawaiian or Pacific Islander, and (7) White. Participants are also given the option to respond with “None of these fully describe me” or “Prefer not to answer.” The All of Us Researcher Workbench codes these data as SIRE categories, following the US Office of Management and Budget Standards. Self-identified race and ethnicity data are currently not available for those who selected the “American Indian or Alaska Native” category. To maximize statistical power, the study cohort was limited to the 4 largest SIRE groups in the All of Us participant body. Asian, Black, and White participants were defined as those who selected the corresponding category as their sole category, and Hispanic participants were defined as individuals who selected “Hispanic or Latino” as either one of or their only category, consistent with the Office of Management and Budget (OMB) standards.

Quantifying perceived discrimination

A metric for quantifying the discrimination that All of Us participants perceive was derived from participant responses to the program’s “Social Determinants of Health” survey (Table S1). In the section of the survey centered on discrimination, participants are asked to state the frequency with which they experience 9 different forms of discrimination, corresponding to 9 different questions. The questions read, “In your day-to-day life, how often do any of these happen to you?”: (1) you are treated with less courtesy than other people are, (2) you are treated with less respect than other people are, (3) you receive poorer service than other people at restaurants or stores, (4) people act as if they think you are not smart, (5) people act as if they are afraid of you, (6) people act as if they think you are dishonest, (7) people act as if they’re better than you are, (8) you are called names or insulted, and (9) you are threatened or harassed. Participants may respond with 1 of 6 options corresponding to ascending levels of frequency: (1) never, (2) less than once a year, (3) a few times a year, (4) a few times a month, (5) at least once a week, and (6) almost every day. Furthermore, participants are asked to cite an attribution for these experiences: “What do you think is the main reason for these experiences.” This survey section is adapted from the Everyday Discrimination Scale, which was developed by Williams and colleagues to measure perceived discrimination.^27,38

A normalized first eigen vector of a cross-correlation matrix was constructed from ordinally coded survey responses and used as a metric for each participant’s perceived discrimination. Matrix construction was performed using version 2.4.1 of the psych package in R version 4.3.1.³⁹ The resulting metric is termed the perceived discrimination index (PDI). Additional details on the creation and validation of the PDI can be found in the Supplementary Material S1.

Statistical analysis

All statistical analyses in this study were performed on R version 4.3.1. Logistic regression models were constructed using the stats package’s glm function. Logistic regression models were used to model disease status (case = 1, control = 0) as a function of PDI, PDI as a function of SIRE, and disease status as a function of different combinations of SIRE and PDI. When modeling disease status, participant age and sex at birth were used as covariates to mitigate bias produced by the overrepresentation of older and female volunteers in the All of Us participant body. Further details on statistical analyses performed in this study can be found in the Supplementary Material S1.

Results

Study cohort

Data were available for a total of 413 457 participants in the All of Us Registered Tier Dataset v7. The study cohort was limited to participants who had both demographic and EHR data available and who provided responses to questions regarding discrimination in the “Social Determinants of Health” survey (Figure S1). The study cohort was further limited to participants belonging to the Asian, Black, Hispanic, and White SIRE groups and who were assigned either male or female at birth. These restrictions were employed to attain adequate statistical power and to allow adjustment by sex at birth. The final study cohort featured a total of 60 180 participants, whose mean age is 60.7 years and 65.0% were assigned female at birth (Table 1).

Table 1.

Study cohort.

Characteristic	Full cohort	Asian	Black	Hispanic	White
No. of participants	60 180 100%)	1291 (2.15%)	4726 (7.85%)	5336 (8.87%)	48 827 (81.13%)
Mean age (SD)	60.66 (15.81)	62.23 (15.48)	58.17 (13.77)	50.87 (15.54)	62.23 (15.48)
Female (%)	39 118 (65.00)	828 (64.14)	3498 (74.02%)	3844 (72.04%)	30 948 (63.38)
Male (%)	21 062 (35.00)	463 (35.86)	1228 (25.98%)	1492 (27.96%)	17 879 (36.62)
Mean PDI (SD)	0.15 (0.16)	0.19 (0.15)	0.24 (0.21)	0.18 (0.18)	0.14 (0.15)

Perceived discrimination index

A PDI for All of Us participants was derived from participant responses to 9 questions regarding discrimination in the “Social Determinants of Health” survey (Table S1). Individual values for PDI were computed for 107 723 participants, of whom 60 180 met the study inclusion criteria. The PDI is a composite metric of discrimination that All of Us participants perceive, incorporating frequencies that participants report for each of 9 questions describing experiences of discrimination. The questions capture different forms of discrimination that an individual may experience. Reported frequencies were coded as ordinal values, with higher values representing higher frequencies. Polychoric correlation (ρ) tests performed on each pair of questions reveal that reported frequencies for these questions are highly and positively correlated. Correlation scores range from ρ = 0.49 for questions 2 and 5 to ρ = 0.91 for questions 3 and 4 (Figure 1A). The mean correlation score for all pairwise correlations was ρ = 0.63.

Figure 1.

Development of the perceived discrimination index (PDI). (A) Pairwise polychoric correlation scores for ordinally coded responses to the 9 discrimination-related questions from the “Social Determinants of Health” survey. (B) The proportion of variance in participant responses to the 9 discrimination-related questions explained (y-axis) by each of the first 5 principal components produced through principal component analysis of the polychoric correlation matrix (x-axis). (C) Principal component loadings for each of the 9 discrimination-related questions (y-axis). (D) Distribution of participant PDI values (x-axis) for the study cohort.

Principal component analysis was applied to the ordinal-coded frequencies to generate single, participant-specific values that capture the variation in reported frequencies across the 9 questions. The first principal component (PC1) explains 67.54% of the variance in participants’ reported frequencies, followed by 7.82% for PC2 and 6.90% for PC3 (Figure 1B). The variable loadings for PC1 are all positive, with higher loading values corresponding to greater degrees of perceived discrimination. The greatest loading value observed was 0.89 for question 4 and the lowest loading value observed was 0.70 for question 5 (Figure 1C). The mean loading value across all 9 questions was 0.82. As PC1 explained the greatest amount of variance in reported frequencies and demonstrated consistently positive variable loadings, values for PC1 were used to calculate PDI. Min-max normalization was applied to PC1 to produce a continuum of PDI values ranging from 0 to 1, with higher values corresponding to greater degrees of perceived discrimination (Figure 1D).

Differences in the degree of discrimination perceived by individuals belonging to different SIRE groups were assessed through modeling PDI as a function of SIRE, with participant age and sex at birth as covariates. Using White as the reference SIRE group, positive and significant associations were observed between PDI and the Asian (β = 1.84e-2, P = 1.61e-5) and Black (β = 8.74e-2, P = 1.91e-312) SIRE groups (Table S2). These associations remain when covariates are removed (Table S3). No significant associations were observed between PDI and the Hispanic SIRE group. The patterning in these model coefficient values is reflected in the median PDI values for the 4 SIRE groups. The Black SIRE group has the highest median PDI of 0.20, followed by the Asian, Hispanic, and White SIRE groups, with scores of 0.18, 0.14, and 0.11, respectively (Figure 2). The percentages of participants in each SIRE group were stratified across low, medium, and high PDI terciles (Table S4). The percentages of Asian and Black participants increase monotonically across PDI terciles, whereas the percentage of White participants decreases across PDI terciles. The percentages of Asian, Black, and Hispanic participants are all greatest in the high PDI tercile, whereas the greatest percentage of White participants is found in the lowest PDI tercile.

Figure 2.

Distribution of perceived discrimination index (PDI) scores for participant self-identified race ethnicity (SIRE) groups. Boxplots show the median, interquartile ranges, and outliers, with median values included.

The reasons that participants gave for experiencing discrimination that were most significantly and positively associated with PDI values are “Some Other Aspect of Your Physical Appearance” (β = 7.33e-2, P = 5.76e-195), followed by “Your Weight” (β = 5.75e-2, P = 1.87e-150), and “Your Race” (β = 5.08e-2, P = 1.49e-128) (Table S5). The reasons for experiencing discrimination were further stratified by participant SIRE (Table S6). The most frequently listed reason for discrimination among minority SIRE groups was race. Gender and age were the second and third most common reasons given for discrimination among minority SIRE groups. Age was the most common reason given for discrimination for the White group followed by gender; race was not listed as a common reason for discrimination for the White group.

Perceived discrimination and health outcomes

To assess the association between perceived discrimination and health outcomes across the All of Us participant body, logistic regression models were used to model disease status (case = 1, control = 0) as a function of PDI for a total of 1755 diseases (Figure 3A). Participant age and sex at birth were included as covariates in all models, as prevalence for a number of diseases are known to differ across sex and age groups. Of the 1755 diseases modeled, 489 (27.86%) were positively and significantly (Bonferroni adjusted P < 2.85×10⁻⁵) associated with PDI and 20 (1.14%) were negatively and significantly associated with PDI. “Mental Disorders” is the disease category with the largest percent of diseases significantly and positively associated with PDI, with a total of 59 out of 72 (81.94%) PDI-associated diseases, followed by “Neurological Disorders,”, which has a total of 43 out of 82 associated diseases (52.43%) (Figure 3B). The disease most strongly and positively associated with PDI is posttraumatic stress disorder (β = 3.51, P = 1.19e-231), followed by major depressive disorder (β = 1.94, P = 1.41e-195) and mood disorders (β = 1.94, P = 7.66e-169) (Table 2). Similar results are achieved when perceived discrimination is represented as a categorical variable, such as in previous studies conducted by Forde et al. (Figure S2).^30,31 When not adjusting for age or sex, a total of 255 (14.53%) diseases are positively and significantly associated with PDI and 257 (14.64%) diseases are negatively and significantly associated with PDI. “Mental Disorders” remains the category with the largest percent of diseases significantly and positively associated with PDI, with a total of 55 out of 72 (76.39%) PDI-associated diseases (Figure S3 and Table S7).

Figure 3.

Phenome-wide associations between disease status and the perceived discrimination index (PDI). (A) Beta coefficients for PDI effect size estimates for associations with individual diseases are shown on the y-axis. Values greater than 5 or less than −5 were coerced to 5 and −5, respectively. Colors represent different disease categories, as indicated by the color key. (B) Proportion of diseases that show significant and positive (red), not significant (gray), and significant and negative associations with PDI.

Table 2.

Diseases most strongly and positively associated with perceived discrimination index (PDI).^a

	Logistic regression model coefficients
Phecode/Disease	Estimate	SE	z-value	P-value
300.9/Posttraumatic stress disorder	3.51	1.08e-1	32.50	1.19e-231
296.22/Major depressive disorder	1.94	6.50e-2	29.83	1.41e-195
296/Mood disorders	1.94	7.02e-2	27.70	7.66e-169
318/Tobacco use disorder	2.09	7.86e-2	26.63	3.19e-156
296.1/Bipolar	3.06	1.17e-1	26.08	5.83e-150
300.1/Anxiety disorder	1.48	6.29e-2	23.49	5.03e-122
316/Substance addiction and disorders	2.29	9.79e-2	23.42	2.65e-121
300/Anxiety, phobic and dissociative disorders	1.66	7.12e-2	23.29	4.90e-120
278.11/Morbid obesity	1.80	7.83e-2	23.05	1.46e-177
278.1/Obesity	1.32	5.93e-2	22.21	2.59e-109

^aModel specification: Disease ∼ PDI + age + sex.

The cohort was stratified into low, medium, and high PDI terciles to evaluate how disease prevalence varies with PDI. The most prevalent conditions and their prevalence estimates can be found in Table S8.

Health disparities

To identify health disparities as diseases for which disease burden differs across SIRE groups in the All of Us participant body, case and control status for 1755 diseases were modeled as a function of SIRE with age and sex at birth as covariates (Figure 4). Using White as the reference SIRE group, these analyses revealed that membership to the Asian SIRE group was positively and significantly associated with 22 (1.25%) diseases, and negatively and significantly associated with 150 (8.55%) diseases. Membership to the Black SIRE group was positively and significantly associated with 351 (20%) diseases, and negatively and significantly associated with 89 (5.07%) diseases. Membership to the Hispanic SIRE group was positively and significantly associated with 132 (7.52%) diseases, and negatively and significantly associated with 146 (8.32%) diseases. Glaucoma (β = 9.62e-2, P = 4.97e-23), intestinal disaccharide deficiencies (β = 7.14e-1, P = 9.55e-16), and viral hepatitis B (β = 1.53, P = 5.52e-12) are the diseases for which case status was the most highly and positively associated with membership to the Asian SIRE group. Hypertension (β = 1.32, P = 7.55e-255), essential hypertension (β = 1.17, P = 3.00e-247), and type 2 diabetes (β = 1.17, P = 2.86e-227) are the diseases for which case status was the most highly and positively associated with membership to the Black SIRE group. Type 2 diabetes (β = 8.43e-1, P = 5.53e-106), Helicobacter pylori infections (β = 1.96, P = 4.27e-92), and type 2 diabetes with ophthalmic manifestations (β = 1.52, P = 1.16e-81) are the diseases for which case status was the most highly and positively associated with membership to the Hispanic SIRE group (Table S9). Different subsets of diseases are identified when not controlling for either age or sex (Figure S4 and Tables S10 and S11).

Figure 4.

Phenome-wide associations between self-identified race and ethnicity (SIRE) and disease status. Model specification: disease ∼ SIRE+age+sex. Points represent individual diseases. Point positions on y-axis represent −log₁₀  P-value for the association between each disease-SIRE combination. Red lines represent the Bonferroni-adjusted −log₁₀  P-value threshold of 4.545. Colors represent different disease categories, as indicated in the color key.

Perceived discrimination and health disparities

To compare the relative effects of participant race/ethnicity vs perceived discrimination on health outcomes, case and control status for 1755 diseases were modeled as a function of either SIRE or PDI, with age and sex at birth as covariates, and pseudo-R² values for the SIRE and PDI models were compared for each disease (Figure S5A). However, SIRE explains more of the variance in disease status for 74.76% of diseases, PDI explains more of the variance in 24.44% of diseases, whereas, the relative variance in disease status explained by SIRE vs PDI varied by disease category. Perceived discrimination index explains more of the variance in 59 (81.94%) mental disorders, compared to 13 (18.06%) mental disorders where SIRE explains more of the variance (Figure S5B). For neoplasms, SIRE explains more of the variance in disease status for 113 (83.70%) cancers, compared to 22 (16.30%) cancers where PDI explains more of the variance (Figure S5C).

As described in the “Methods” section, the contribution of perceived discrimination to SIRE disparities in disease burden was quantified with percent attenuation, $\mathrm{PD}{\mathrm{I}}_{\mathrm{attenuation}}$. Following this metric, perceived discrimination contributes the greatest to disparities in disease burden observed between the Black and White SIRE groups. The disease for which the greatest percent reduction was observed in the Black SIRE coefficient following adjustment for PDI is suicidal ideation ($\mathrm{PD}{\mathrm{I}}_{\mathrm{attenuation}}$ = 98.91%) followed by chronic pain syndrome ($\mathrm{PD}{\mathrm{I}}_{\mathrm{attenuation}}$ = 69.36%) and psychosis ($\mathrm{PD}{\mathrm{I}}_{\mathrm{attenuation}}$ = 64.38%). For each of these diseases, the Black SIRE coefficient was positive and significant in its association with disease status prior to PDI adjustment but was no longer significant following adjustment (Bonferroni adjusted P < 1.42×10⁻⁴). This loss in significance was observed for a total of 43 (12.25%) out of the 351 diseases for which membership to the Black SIRE group was significantly and positively associated with disease status (Figure 5). The most common disease category represented by these diseases was “Digestive,”, comprising 9 of the 43 (20.93%) diseases. This loss of significance was not observed in either the Asian or Hispanic SIRE groups, for whom values of $\mathrm{PD}{\mathrm{I}}_{\mathrm{attenuation}}$ were 6.08% or below for all diseases whose case status is positively and significantly associated with membership to either group (Table 3).

Figure 5.

Association of Black SIRE with disease before (unadjusted) and after (adjusted) controlling for perceived discrimination index (PDI). Unadjusted model specification: disease ∼ SIRE+age+sex. Adjusted model specification: disease ∼ SIRE+PDI+age+sex. Dashed line represents y = x. Points represent individual diseases and are colored by significance after adjustment for PDI: red (not significant after adjustment for PDI) and gray (remain significant after adjustment for PDI).

Table 3.

Attenuation of associations between self-identified race and ethnicity (SIRE) and disease after adjustment by perceived discrimination index (PDI).

Phecode/Disease	% Reduction in estimate	Raw change in estimate	Significant after adjustment
Asian
250.42/Other abnormal glucose	4.40	−0.0160	Yes
366.2/Senile cataract	3.68	−0.0172	Yes
365.11/Primary open angle glaucoma	3.68	−0.0469	Yes
250.4/Abnormal glucose	3.66	−0.0192	Yes
70.2/Viral hepatitis B	3.07	−0.0458	Yes
Black
297.1/Suicidal ideation	98.91	−0.47	No
355.1/Chronic pain syndrome	69.36	−0.28	No
295.3/Psychosis	64.38	−0.38	No
327.3/Sleep apnea	61.38	−0.13	No
327.1/Hypersomnia	53.14	−0.16	No
Hispanic
525/Other diseases of the teeth and supporting structures	6.08	−0.0306	Yes
250.6/Polyneuropathy in diabetes	5.54	−0.0330	Yes
70.3/Viral hepatitis C	5.32	−0.0267	Yes
278.11/Morbid obesity	4.53	−0.0123	Yes
401.21/Hypertensive heart disease	4.19	−0.0236	Yes

These attenuation effects were further evaluated using mediation analyses, which were performed for each of the 43 diseases for which a loss of significance in the Black SIRE coefficient was observed after adjusting for PDI. The disease for which the indirect effect of PDI is the greatest is suicidal ideation, whose indirect effects of PDI constituted 97.54% of the total effects of PDI and SIRE on case status (Table 4). This is followed by chronic pain syndrome and sleep apnea, whose indirect effects accounted for 66.97% and 62.99% of their total effects, respectively. A strong correlation (Pearson’s r = 0.99) is observed between the indirect effect of PDI from mediation analysis and $\mathrm{PD}{\mathrm{I}}_{\mathrm{attenuation}}$ when the former is expressed as a percentage of the total effect of membership to the Black SIRE group on case status (Figure S6).

Table 4.

Mediation analyses for Black-White disease disparities.

Phecode/Disease	Direct effect a	Indirect effect b	Total effect c	% Indirect d
297.1/Suicidal ideation	1.12e-2 (1.12e-1)	3.64e-1 (2.56e-2)	3.73e-1 (1.02e-1)	97.54
355.1/Chronic pain syndrome	1.20e-1 (9.01e-2)	2.50e-1 (1.63e-2)	3.74e-1 (8.55e-2)	66.97
327.3/Sleep apnea	8.18e-2 (1.31e-1)	1.37e-1 (2.60e-2)	2.17e-1 (1.23e-1)	62.99
295.3/Psychosis	2.09e-1 (4.90e-2)	3.03e-1 (8.11e-3)	5.10e-1 (4.78e-2)	59.50
327.1/Hypersomnia	1.40e-1 (5.61e-2)	1.57e-1 (1.33e-2)	2.96e-1 (5.56e-2)	53.22

The direct effect of self-identified race and ethnicity (SIRE) on disease is mediated by perceived discrimination index (PDI). Results are shown for diseases with the top 5 indirect effects.

^aDirect effect of SIRE on disease.

^bIndirect effect of SIRE on disease, mediated by PDI.

^cTotal effect of SIRE on disease: direct+indirect.

^dPercent of total effect of SIRE on disease mediated by PDI.

Discussion

The PDI developed in this study provides a means of quantifying the degree of discrimination that All of Us participants experience. The composite metric captures most of the variance in participant responses to questions related to experiences of discrimination in the All of Us “Social Determinants of Health” survey. Furthermore, the metric is highly correlated with participant responses to each of the 9 questions. These characteristics suggest that the PDI may be an effective means of allowing All of Us researchers to incorporate perceived discrimination as a variable in disease modeling.

The significant associations found among PDI values, SIRE, and disease status further underscore the potential of the PDI to be a proxy for perceived discrimination. Values for PDI tend to be higher for individuals identifying as Asian, Black, or Hispanic than those identifying as White. These results are consistent with those from prior studies on race and discrimination, which show that racial minority groups report more perceived discrimination than their White counterparts.^25,26 The relative levels of PDI for each minority group are consistent with survey data from the Pew Research Center. According to Pew, 75% of Black adults have experienced racial discrimination compared to 58% of Asians and 31% of Hispanics.^40–42

Perceived discrimination index was found to be positively and significantly associated with diseases from a wide range of disease categories. The disease category with the greatest proportion of positive associations with PDI is “Mental Disorders.” These results validate past research on discrimination and health, which have identified perceived discrimination as a potential predictor for a number of physical and especially mental disorders.^6–13 While not traditionally classified as such, the influence of perceived discrimination on health outcomes demonstrated by these results and prior studies validate its classification as an SDOH.

The results of the attenuation analyses performed provide evidence that discrimination contributes to racial disparities in disease burden in the All of Us participant body. These results are consistent with previous findings, which have demonstrated that perceived discrimination contributes to health disparities in other contexts. For instance, perceived discrimination has been demonstrated to have stronger associations with negative indicators such as poorer self-reported health status and other poor outcomes among African Americans than other groups.^35,43 There is also evidence to suggest that perceived discrimination contributes to the disproportionately high risk of cardiovascular disease among African Americans.³⁴ These findings are consistent with our attenuation analyses suggesting that perceived discrimination contributes to worse health outcomes among individuals identifying as Black. Other studies suggest that perceived discrimination may contribute to health disparities among other SIRE groups not captured by this study, such as those between Māori and European populations in New Zealand and between Aboriginal and Non-Aboriginal populations in Australia.^44,45 Our findings are reinforced by the accompanying mediation analyses performed, whose results also show that a number of SIRE-disease associations are partially attributable to perceived discrimination. However, while PDI values for both the Asian and Black SIRE groups were found to be significantly higher than the White SIRE group, perceived discrimination was found to only meaningfully contribute to health disparities between the Black and White SIRE groups. This may be a result of there being fewer health disparities between the Asian and White SIRE groups overall. Additionally, there may be other factors beyond perceived discrimination contributing to disparities between these groups.

While the results of the attenuation and mediation analyses performed may suggest that discrimination drives racial health disparities, the methods used are insufficient to discount the possibility of reverse causality. Trauma and depressive symptoms, such as those that occur in mental disorders such as depression and posttraumatic stress disorder, may lead to heightened perceptions of discrimination, as these diseases may impair an individual’s ability to properly assess threats.^46,47 More rigorous analyses are needed to elucidate the true nature of the PDI-disease associations observed in the All of Us participant body.

Limitations

There are several important limitations to this study. The Everyday Discrimination Scale that was used to derive the PDI assesses perceived discrimination in societal life.²⁷ As such, it reflects participants’ subjective experiences, which may not fully represent their actual lived experiences, and it may be expected to vary over time depending on experiences that occurred closer to the time of enrollment. Furthermore, as this is an observational study, there is the potential for unobserved confounding variables to influence study results. Indeed, PDI covaries with several measures of socioeconomic status—education, health insurance, and income—although the correlations are modest suggesting that multicollinearity would not be a major issue for these particular variables. Nevertheless, there may be other unmeasured variables that covary more highly with PDI and better explain the health outcomes studied here.

Other limitations are tied to the nature of the data collected by All of Us. As the All of Us participant body is comprised of volunteers, it is not a representative sample of the US population. As a result, the participant body differs from the general population in several fundamental ways. The cohort has a greater prevalence of both common and rare diseases compared to the general population. Additionally, All of Us participants are more likely to be female, older, have more educational attainment, and identify as either Black or Hispanic than the general population.⁴⁸ As this study encompasses only the 4 most common SIRE groups and everyday discrimination, its generalizability to other racial/ethnic groups and discrimination in other contexts may be limited. Furthermore, there may be generational and cultural differences in how individuals may perceive and interpret discrimination that may not be adequately captured through controlling for age and SIRE.^49,50 Beyond limitations inherent to the study cohort, there are important caveats in the use of the Everyday Discrimination Scale on which the discrimination section in the All of Us “Social Determinants of Health” survey is based. It has been previously reported that the aspects of perceived discrimination that this scale captures may be limited, potentially necessitating the use of other discrimination measures in conjunction with this scale.⁵¹ The findings reported here are therefore the most generalizable to the All of Us participant body and to the dimensions of discrimination captured by the Everyday Discrimination Scale.

Conclusions

These limitations notwithstanding, the methodology in calculating the PDI outlined here may support future discrimination-related research on the All of Us platform by providing a means through which perceived discrimination can be quantified. The associations between this metric and the case status for diseases spanning a variety of different categories validate the role of discrimination as an important SDOH. We also demonstrate the potential for feelings of perceived discrimination to contribute to racial disparities in disease burden on the All of Us platform. The PDI metric may further facilitate research on health disparities by enabling the assessment of how perceived discrimination mediates and interacts with other SDOH to influence disparities in health outcomes. The PDI metric may also be useful in policy interventions aimed at reducing discrimination-related health disparities. The American College of Physicians recommends that policymakers address such disparities by implementing measures that curb the adverse health effects caused by discrimination and enhancing health care delivery for those affected by discrimination.⁵² Perceived discrimination index provides a means of quantifying discrimination, potentially enabling policymakers to target and prioritize affected populations with relevant policy measures. Efforts to mitigate racial health disparities may therefore benefit from addressing the disproportionate amount of discrimination suffered by socially disadvantaged groups.

Author contributions

Vincent Lam (Data curation, Formal analysis, Methodology, Visualization), Sonali Gupta (Data curation, Formal analysis, Methodology, Visualization), I. King Jordan (Conceptualization, Funding acquisition, Methodology, Project administration, Visualization), and Leonardo Mariño-Ramírez (Conceptualization, Funding acquisition, Methodology, Project administration, Supervision, Visualization)

Supplementary material

Supplementary material is available at Journal of the American Medical Informatics Association online.

Funding

V.L., S.G., and L.M.-R. were supported by the Division of Intramural Research of the NIMHD at National Institutes of Health (NIH) (award number: 1ZIAMD000018). L.M.-R. was supported by the NIH Distinguished Scholars Program. I.K.J. was supported by the by the IHRC, Inc.-Georgia Tech Applied Bioinformatics Laboratory (award number: RF383).

Conflicts of interest

None declared.

Data availability

All code and data used in this study are made available on the All of Us Researcher Workbench.