Application of intersectionality framework and area-level indicators in machine learning analysis of depression disparities in all of us research program data

Abstract

Background/objective

Depression is a complex mental health disorder influenced by various social determinants of health (SDOH) at individual and community levels. Area-level indicators and the intersectionality framework, which considers overlapping personal identities, are used in this paper to get a nuanced picture of depression disparities.

Methods

This cross-sectional study uses electronic health records data from the All of Us research network. Our study cohort includes 33,994 individuals who completed the SDOH surveys in All of Us and had at least one inpatient visit, with 31.7% diagnosed with depression between 2020 and 2025. We used depression diagnosis as an outcome, while independent variables include area-level variables from US Religious Census (religious adherents, %), American Community Survey response (birthrates by living arrangement), and All of Us (median area income and deprivation index); individual-level variables from All of Us surveys: age group, income, gender, sexual orientation, race/ethnicity, and others. We also added the interactions of the individual-level variables with each other and with area-level variables. The association between depression and the variables is reported by fitting the mixed effects logistic regression model on the subset of variables identified by the LASSO method with 3-digit ZIP code as a random effect.

Results

Protective factors included male sex at birth among straight (OR 0.71, 95% CI: 0.59–0.86) and non-binary (OR 0.42, 95% CI: 0.24–0.75) individuals, older adults with active military service (OR 0.64, 95% CI: 0.53–0.77) or health insurance (OR 0.80, 95% CI: 0.72–0.88), and younger adults with 1–2 children (OR 0.72, 95% CI: 0.60–0.86). Risk factors were strongest for inability to work (OR 2.01, 95% CI: 1.76–2.31), further impacted by medium per capita income (OR 1.39, 95% CI: 1.13–1.70). Housing concerns (OR 1.24, 95% CI: 1.09–1.41), frequent disrespect (OR 1.41, 95% CI: 1.17–1.69), and high area deprivation in older adults (OR 1.31, 95% CI: 1.12–1.53) increased risk. Other intersectional risks included non-binary identity in older adults (OR 5.14, 95% CI: 1.94–13.62), non-binary individuals born outside the USA (OR 10.23, 95% CI: 2.07–50.48), women with children (OR 1.39, 95% CI: 1.25–1.55), and Hispanic active-duty members (OR 1.90, 95% CI: 1.32–2.74).

Discussion/conclusion

Our findings, while being limited by 3-digit ZIP code aggregation that might have obscured some area-level effects, suggest that clinical and public health strategies for depression prevention and intervention should account for overlapping identities while incorporating place-aware approaches.

Introduction

Depression is a multifaceted mental health disorder influenced by various social determinants of health (SDOH) that encompass individual, community, and societal factors. SDOH domains include Economic Stability, Education Access & Quality, Health Care Access & Quality, Neighborhood & Built Environment, and Social & Community Context [1]. Research has consistently shown that sociodemographic factors, social support, and adverse life experiences significantly contribute to the prevalence and severity of depressive symptoms across diverse populations [2–4].

Intersectionality theory, which was originally developed to acknowledge the combined effects of sex, gender, race, and ethnicity on life outcomes [5], can be projected into the domain of human health to show that different intersections of identity, social position, processes of oppression or privilege, and policies or institutional practices contribute to health outcomes. The intersectionality framework provides researchers with a perspective to view individuals’ overlapping identities that interact to shape their experiences, in addition to traditional demographic characteristics [6]. In the mental health gender disparities area, the framework may reveal the well-known difference between females and males in depression prevalence [7], A study by Persmark et al. emphasizes the importance of contextualizing health issues within an intersectional framework in the context of the U.S. opioid crisis, where individual risk factors often overshadow underlying systemic causes of health disparities [8]. This perspective is echoed by Ogungbe et al., who highlight the need for intersectionality to address implicit biases in healthcare that contribute to disparities [9].

The intersectionality framework has been applied to various health contexts, revealing how specific groups experience compounded disadvantages. For example, a study on disability trajectories illustrates how race, ethnicity, and gender intersect to influence health outcomes, suggesting that health disparities are not uniform but rather contingent upon the specific identities individuals hold [6]. This is further supported by research that examines how sexual orientation and race intersect to affect health risk factors, indicating that individuals with multiple marginalized identities face greater health challenges [10], with specific instruments developed to quantify the intersectional effects, for example, related to discrimination [11]. Intersectionality aligns with a broader socioecological framework, recognizing that health outcomes emerge from interactions between individual, social, and environmental factors [12].

The practical implementation of intersectionality research in healthcare is hindered by specific modeling challenges – namely, the growth of the number of independent variables as a researcher accounts for the interaction effects of a significant number of demographic and social variables. For example, a model with 10 sociodemographic independent variables, each having only 3 levels, results in 200 binary encoded variables when all possible interactions are considered. The growth of the number of variables leads to issues, such as the model not being generalizable when the number of observations per each variable in the minority class is small [13]. Thus, statistical modeling of intersectionality requires special techniques.

This paper explores the LASSO (Least Absolute Shrinkage and Selection Operator) variable selection method [14] to reduce the dimensions of the design matrix and highlight the most significant intersectional variables that affect depression. LASSO is an efficient method to subset variables and shows superior performance to stepwise methods [15]. It works by adding a penalty term to the classical regression objective, shrinking less important coefficients to zero, essentially performing variable selection. The procedure typically involves fitting the model with cross-validation to select the optimal penalty parameter (lambda). LASSO can be effective for modeling intersectionality, since many of the smaller subsamples (e.g., personal identities with very small counts), which do not have a statistically significant impact on the model output, and multicollinear variables are removed from the design matrix while retaining model performance and gaining additional statistical power for the remaining and most impactful variables.

According to the socioecological perspective [12], in addition to individual-level, SDOH operate at the community level, and these area-level indicators should be considered in the model as they may have differentiated effects that uniquely interact with personal identity. This publication aims to utilize the LASSO method to explore the simultaneous influence of identity factors and several area-level indicators (median household income and related area deprivation index [16], % of women who gave birth in the past 12 months decomposed by family structure type, and religious adherents as a share of the total population) on depression through variable interactions while, in addition, modeling geographic factor as a random effect in the model. These four area-level indicators were chosen based on the knowledge of their importance in depression at the individual level, but a specific hypothesis we have is that they would also be influential at the area level.

It is known that lower individual SES is associated with an increased prevalence of depression due to factors such as financial stress, limited access to healthcare, and reduced educational opportunities [17, 18]. The correlation of household income with depression prevalence is apparent when two maps are compared side by side (Fig. 1A and B), with regions with high depression (e.g., Appalachian) correlating with lower-income regions. At the same time, living in areas where the median income is lower than the national or state average can affect depression through indirect pathways, even if individual or family income is on par with national averages. For example, a middle-class family living in a deprived area might be exposed to higher levels of neighborhood crime and illicit drug exposure, factors associated with lower-income areas [19]. In contrast, one may also consider a salubrious effect through mechanisms of social comparison (the opposite of relative deprivation). However, it is important to avoid ecological fallacy when interpreting Fig. 1A and D: macro-level correlations are not shown here to make inferences at the individual level. Instead, in this paper, we analyze how personal identity factors, such as higher personal income, interact with area-level indicators.

Fig. 1.

US County-level data for (A) Depression prevalence, %; B Median household income in USD; C Percentage of religious adherents; D Percentage of women who gave birth in an intact family. Source: Authors’ own work using ACS Survey and US Religion Census

In addition to income and area deprivation as geographic factors, the percentage of women who give birth stratified by living arrangement and religious activity is explored in this paper. These family-structure and religious variables were included as contextual measures of social cohesion, an important SDOH aspect in depression [20, 21]. Studies highlight that individuals with strong social networks and supportive relationships tend to experience lower levels of depressive symptoms [22, 23]. Positive religious coping is known to provide emotional support and a sense of meaning, which may buffer against depressive symptoms [24, 25]. At the same time, structural stigma embedded in community norms can influence mental health outcomes beyond individual experiences [26]. For example, Hatzenbuehler et al. demonstrated that lesbian, gay, and bisexual youths living in counties with more supportive religious climates exhibited significantly fewer alcohol abuse symptoms and sexual risk behaviors [27]. Interestingly, differences in family compositions and religious adherence form distinct geographic spots, shaping areas that can be called “family islands” and “religious islands” (Fig. 1C and D).

We hypothesize in this study that area religious activity, higher childbirth rates in families with a spouse present, higher area income, and lower deprivation would be associated with decreased depression prevalence, even after controlling for the same or closely related variables at individual level.

To summarize, previous studies have not fully explored the interaction of individual-level characteristics, while at the same time accounting for area-level indicators, limiting our understanding of intersectional risk factors for depression. The present study combines individual-level sociodemographic indicators from The All of Us Research Program and several area-level variables derived from All of Us 3-digit ZIP code level data, the American Community Survey (ACS), and the 2020 US Religion Census. We aim to identify individual, area-level, and interaction predictors of depression using LASSO for variable selection and a mixed-effects logistic regression model with 3-digit ZIP code as random effect to account for clustering. Individual-level factors such as age, race/ethnicity, depression diagnosis, and other relevant individual sociodemographic data are obtained from the All of Us platform. The All of Us program’s commitment to diversity and inclusion is essential to capture a broad range of intersectional health outcomes in this study [28]. All of Us employs multiple strategies, including participant incentives, to encourage broad and diverse survey participation. This has resulted in high-quality, reliable responses across domains, including the Social Determinants of Health (SDOH) survey [29].

Methods

Our general approach follows these steps: (1) Identify individuals in the All of Us dataset with SDOH survey responses and record if they were diagnosed with depression between January 1st, 2020, and April 13th, 2025; (2) Link area-level indicators to individual data using a 3-digit ZIP code; (3) Perform LASSO variable selection; (4) Fit a random effects regression model using variables selected by LASSO with 3-digit ZIP code as random effect; and plot statistically significant variables from this model after p-value adjustment for false discovery rate.

Dataset

Data were obtained from the All of Us Research Program, which includes EHR and survey data from a diverse U.S. cohort [29]. Identification of people with depression was performed using the All of Us platform phenotyping algorithm, which suggests diagnoses based on the term entered (“depression”). We limited our cohort to patients with at least one in-office visit, who have EHR data present, and who have completed the SDOH survey at any point in time. The SDOH survey contains a range of questions related to social and community context, economic stability, education, neighborhood and built environment, health, and health care.

We marked patients as having depression if they had a diagnostic code suggested for the term “depression” and filtered to exclude unrelated conditions (e.g., bipolar disorder, myocardial depression – see Table 1). The depression codes were taken from participants’ EHR data, and the diagnosis date was January 1st, 2020, up to April 13th, 2025. These cut-off dates were selected because SDOH surveys in All of Us were completed by participants during 2021 or later; thus, earlier mental health outcomes could have been impacted by earlier sociodemographic factors and changes.

Table 1.

Depression phenotype in the all of us platform

Diagnosis	OMOP ID	SNOMED Code
Acute depression	37,016,718	712,823,008
Chronic depression	4,103,574	192,080,009
Depressive disorder*	440,383	35,489,007
Depressive episode	3,656,234	871,840,004
Major depression in full remission	4,269,493	63,412,003
Major depression in partial remission	4,148,630	30,605,009
Major depression in remission	4,176,002	42,810,003
Major depression single episode, in partial remission	4,323,418	70,747,007
Major depression, single episode	4,282,096	36,923,009
Major depressive disorder	4,152,280	370,143,000
Mild depression	4,149,320	310,495,003
Mild major depression	4,336,957	87,512,008
Mild major depression, single episode	4,195,572	79,298,009
Mild recurrent major depression	4,228,802	40,379,007
Minimal depression	36,713,698	718,636,001
Mixed anxiety and depressive disorder	4,338,031	231,504,006
Moderate depression	4,151,170	310,496,002
Moderate major depression	4,307,111	832,007
Moderate major depression, single episode	4,049,623	15,639,000
Moderate recurrent major depression	4,077,577	18,818,009
Moderately severe depression	36,717,092	719,593,009
Reactive depression (situational)	4,314,692	87,414,006
Recurrent depression	4,098,302	191,616,006
Recurrent major depression	4,282,316	66,344,007
Recurrent major depression in full remission	4,263,748	46,244,001
Recurrent major depression in partial remission	4,141,454	33,135,002
Recurrent major depression in remission	433,991	68,019,004
Severe depression	4,149,321	310,497,006
Severe major depression	42,872,722	450,714,000
Severe major depression without psychotic features	4,327,337	75,084,000
Severe major depression, single episode	42,872,411	251,000,119,105
Severe major depression, single episode, without psychotic features	441,534	76,441,001
Severe recurrent major depression	43,531,624	281,000,119,103
Severe recurrent major depression without psychotic features	435,220	36,474,008
Single episode of major depression in full remission	4,025,677	19,527,009

*This diagnostic grouping includes various forms of depression, such as dysthymia, atypical depression, and others not separately mentioned above

Independent variables

Area-level indicators are described in Table 2. Because All of Us provides individual participant location using a 3-digit ZIP code, we had to aggregate ACS and US Religion Census data into this higher geographical level using the mean values approach. We also binned all area-level indicators into “Low” (less than one SD from the nationwide mean for a given indicator), “High” (higher than one SD from the nationwide mean), and “Average” (in between Low and High).

Table 2.

Area-level indicators

Indicator	Source	Variable name (in the source)
Median income in the area, US$	All of Us 3-digit ZIP code level data	median_income
Area deprivation index	All of Us 3-digit ZIP code level data	deprivation_index
Women who gave birth in the past year (total), %	5-year (2018–2022) ACS	B13004_002
Women who gave birth in the past year, with a spouse present, %	〃	B13004_003
Women who gave birth in the past year, in an unmarried partner household, %	〃	B13004_004
Women who gave birth in the past year, in other living arrangement, %	〃	B13004_005
Share of religious adherents (all denominations/religious groups), %	US Religion Census 2020	TOTRATE_2020

Area religious activity was operationalized as the percentage of religious adherents in each 3-digit ZIP code, derived from the U.S. Religion Census. This similar measure has been used in prior research as a contextual indicator of community norms, social climate, and potential sources of social cohesion [30, 31]. Birthrate-related indicators were similarly linked from the US Religion Census 2020. They reflect broader patterns of partnered versus non-partnered parenting, and we were interested in whether they potentially create a specific area-level climate that could impact the depression of individuals living in such areas. We describe these area-level indicators in the Introduction section (Fig. 1).

We used these area-level indicators, together with individual-level predictors from the All of Us platform, and intersectional variables created by specifying two-way variable interactions of a select number of individual predictors known to be correlated with depression prevalence (Age group; Per capita income, calculated as Household income divided by number of people in household; Gender identity; Marital status; Sexual orientation; Race/ethnicity; and Birthplace, In- or outside the US) with each other and with other predictors. The individual, area-level, and variables selected for interaction, corresponding number of participants, and univariable odds ratios are described in Table 3 (Results section).

Table 3.

Sociodemographic profile of the study cohort

			No Depression, N(%)	Depression, N(%)	OR (univariable)
	Total, N (%)		23,224 (68.3)	10,770 (31.7)
Demographics	Age group†	Adults	10,097 (43.5)	5752 (53.4)	Ref.
		Older adults (> 70 years old)	10,990 (47.3)	3583 (33.3)	0.57 (0.54–0.60, p < 0.001)
		Younger adults (< 40 years old)	2137 (9.2)	1435 (13.3)	1.18 (1.09–1.27, p < 0.001)
	Per capita income in household†	High (> 75 K$)	4843 (20.9)	1052 (9.8)	Ref.
		Low (< 17 K$)	5262 (22.7)	4331 (40.2)	3.79 (3.51–4.10, p < 0.001)
		Medium	13,119 (56.5)	5387 (50.0)	1.89 (1.76–2.04, p < 0.001)
	Household size	Single person household	4241 (18.3)	2502 (23.2)	Ref.
		Three and more person household	6982 (30.1)	3911 (36.3)	0.95 (0.89–1.01, p = 0.107)
		Two person household	12,001 (51.7)	4357 (40.5)	0.62 (0.58–0.65, p < 0.001)
	Biological Sex At Birth	Female	13,449 (57.9)	7663 (71.2)	Ref.
		Intersex	**	**	2.63 (0.75–10.30, p = 0.134)
		Male	9767 (42.1)	3099 (28.8)	0.56 (0.53–0.58, p < 0.001)
		Sex At Birth None Of These	**	**	0.88 (0.12–4.50, p = 0.880)
	Home ownership	Own	17,424 (75.0)	6492 (60.3)	Ref.
		Other Arrangement	1083 (4.7)	860 (8.0)	2.13 (1.94–2.34, p < 0.001)
		Rent	4717 (20.3)	3418 (31.7)	1.94 (1.85–2.05, p < 0.001)
	Health insurance	No	270 (1.2)	197 (1.8)	Ref.
		Yes	22,954 (98.8)	10,573 (98.2)	0.63 (0.53–0.76, p < 0.001)
	Living Situation: People Under 18 in Household	0	19,022 (81.9)	8165 (75.8)	Ref.
		1 or 2	3648 (15.7)	2161 (20.1)	1.38 (1.30–1.46, p < 0.001)
		3 or more	554 (2.4)	444 (4.1)	1.87 (1.64–2.12, p < 0.001)
	Stable House Concern	No	21,905 (94.3)	9139 (84.9)	Ref.
		Yes	1319 (5.7)	1631 (15.1)	2.96 (2.75–3.20, p < 0.001)
	Current Marital Status†	Married	14,429 (62.1)	5151 (47.8)	Ref.
		Divorced	2976 (12.8)	2072 (19.2)	1.95 (1.83–2.08, p < 0.001)
		Living With Partner	1084 (4.7)	618 (5.7)	1.60 (1.44–1.77, p < 0.001)
		Never Married	2999 (12.9)	1945 (18.1)	1.82 (1.70–1.94, p < 0.001)
		Separated	296 (1.3)	276 (2.6)	2.61 (2.21–3.09, p < 0.001)
		Widowed	1440 (6.2)	708 (6.6)	1.38 (1.25–1.51, p < 0.001)
	Race/ethnicity†	White	18,799 (80.9)	8690 (80.7)	Ref.
		Other	934 (4.0)	367 (3.4)	0.85 (0.75–0.96, p = 0.010)
		Black	1930 (8.3)	975 (9.1)	1.09 (1.01–1.18, p = 0.032)
		Hispanic	1561 (6.7)	738 (6.9)	1.02 (0.93–1.12, p = 0.629)
	Birthplace†	USA	21,001 (90.4)	10,065 (93.5)	Ref.
		Outside USA	2223 (9.6)	705 (6.5)	0.66 (0.61–0.72, p < 0.001)
	Sexual Orientation†	Non-heterosexual	1719 (7.4)	1401 (13.0)	Ref.
		Straight	21,505 (92.6)	9369 (87.0)	0.53 (0.50–0.58, p < 0.001)
Other identity factors	Active Duty Serve Status	No	20,372 (87.7)	9217 (85.6)	Ref.
		Yes	2852 (12.3)	1553 (14.4)	1.20 (1.13–1.29, p < 0.001)
	Do you speak a language other than English at home	No	20,265 (87.3)	9500 (88.2)	Ref.
		Yes	2959 (12.7)	1270 (11.8)	0.92 (0.85–0.98, p = 0.014)
	Education Level	College Graduate	6971 (30.0)	2886 (26.8)	Ref.
		Advanced Degree	8759 (37.7)	2798 (26.0)	0.77 (0.73–0.82, p < 0.001)
		College One to Three	5134 (22.1)	3472 (32.2)	1.63 (1.54–1.74, p < 0.001)
		Unfinished school	**	**	1.63 (1.37–1.94, p < 0.001)
		Never Attended	**	**	1.21 (0.17–6.19, p = 0.828)
		Twelve Or GED	2019 (8.7)	1384 (12.9)	1.66 (1.53–1.80, p < 0.001)
	Employment	Employed For Wages	8918 (38.4)	3958 (36.8)	Ref.
		Homemaker	631 (2.7)	405 (3.8)	1.45 (1.27–1.65, p < 0.001)
		Out Of Work Less Than One	387 (1.7)	300 (2.8)	1.75 (1.49–2.04, p < 0.001)
		Out Of Work One Or More	601 (2.6)	494 (4.6)	1.85 (1.63–2.10, p < 0.001)
		Retired	9340 (40.2)	3123 (29.0)	0.75 (0.71–0.80, p < 0.001)
		Self Employed	1785 (7.7)	526 (4.9)	0.66 (0.60–0.74, p < 0.001)
		Student	358 (1.5)	229 (2.1)	1.44 (1.21–1.71, p < 0.001)
		Unable To Work	1204 (5.2)	1735 (16.1)	3.25 (2.99–3.53, p < 0.001)
	Gender Identity†	Man	9723 (41.9)	3060 (28.4)	Ref.
		Non-binary	109 (0.5)	167 (1.6)	4.87 (3.82–6.23, p < 0.001)
		Woman	13,392 (57.7)	7543 (70.0)	1.79 (1.70–1.88, p < 0.001)
	Within the past 12 months were you worried whether your food would run out before you got money to buy more	Never true	21,323 (91.8)	8337 (77.4)	Ref.
		Often true	403 (1.7)	653 (6.1)	4.14 (3.65–4.71, p < 0.001)
		Sometimes true	1498 (6.5)	1780 (16.5)	3.04 (2.82–3.27, p < 0.001)
	How often do you go to religious meetings or services	None/Little	11,473 (49.4)	5621 (52.2)	Ref.
		Often	6094 (26.2)	2558 (23.8)	0.86 (0.81–0.91, p < 0.001)
		I am not religious	5657 (24.4)	2591 (24.1)	0.93 (0.88–0.99, p = 0.019)
Discrimination	In your day to day life how often are you treated with less respect than other people	None/Little	17,084 (73.6)	6388 (59.3)	Ref.
		Often	1738 (7.5)	1869 (17.4)	2.88 (2.68–3.09, p < 0.001)
		A few times a year	4402 (19.0)	2513 (23.3)	1.53 (1.44–1.62, p < 0.001)
	How often are you treated with less respect than other people when you go to a doctor s office or other health care provider	None/Little	21,130 (91.0)	8751 (81.3)	Ref.
		Often	2094 (9.0)	2019 (18.7)	2.33 (2.18–2.49, p < 0.001)
Area-level indicators*	Median income in the area	Average	17,549 (75.6)	8505 (79.0)	Ref.
		High	3939 (17.0)	1221 (11.3)	0.64 (0.60–0.69, p < 0.001)
		Low	1736 (7.5)	1044 (9.7)	1.24 (1.14–1.35, p < 0.001)
	Area deprivation index	Average	15,697 (67.6)	8199 (76.1)	Ref.
		High	4441 (19.1)	1526 (14.2)	0.66 (0.62–0.70, p < 0.001)
		Low	3086 (13.3)	1045 (9.7)	0.65 (0.60–0.70, p < 0.001)
	Share of women in the area who had a birth in the past 12 months (overall)	Average	17,865 (76.9)	7605 (70.6)	Ref.
		High	2692 (11.6)	1848 (17.2)	1.61 (1.51–1.72, p < 0.001)
		Low	2667 (11.5)	1317 (12.2)	1.16 (1.08–1.25, p < 0.001)
	Share of women in the area who had a birth in the past 12 months, Has a spouse present	Average	12,937 (55.7)	6372 (59.2)	Ref.
		High	4215 (18.1)	2128 (19.8)	1.03 (0.97–1.09, p = 0.421)
		Low	6072 (26.1)	2270 (21.1)	0.76 (0.72–0.80, p < 0.001)
	Share of women in the area who had a birth in the past 12 months, Partner in an unmarried partner household	Average	16,210 (69.8)	6988 (64.9)	Ref.
		High	3884 (16.7)	2333 (21.7)	1.39 (1.31–1.48, p < 0.001)
		Low	3130 (13.5)	1449 (13.5)	1.07 (1.00-1.15, p = 0.041)
	Share of women in the area who had a birth in the past 12 months Other living arrangement	Average	16,179 (69.7)	7865 (73.0)	Ref.
		High	4337 (18.7)	1814 (16.8)	0.86 (0.81–0.91, p < 0.001)
		Low	2708 (11.7)	1091 (10.1)	0.83 (0.77–0.89, p < 0.001)
	Rate of religious adherence, county level	Average	15,668 (67.5)	6888 (64.0)	Ref.
		Low	4597 (19.8)	1988 (18.5)	0.98 (0.93–1.04, p = 0.590)
		High	2959 (12.7)	1894 (17.6)	1.46 (1.37–1.55, p < 0.001)

*Counts for area-level indicators represent the number of patients in a 3-digit ZIP that had a residence in an area with average/low/high values for a given characteristic. **AllofUs requires masking counts for elements where the participant count of less than 20 can be deducted

† denotes variables selected for modeling of interactions (for intersectional analysis). High/Low/Medium bins for area-level were defined relative to the study population means (+- 1 SD). P-values not adjusted. Odds ratios (OR) with P-values < 0.05 are statistically significant

Analytical approach

The categorical variables were encoded using one-hot encoding into binary predictors. The LASSO regression method is then used for variable selection using R glmnet package.

We used the logistic regression mixed effect model (R lme4 package) with 3-digit ZIP code as random effect on the set of predictors identified by the LASSO to calculate and then present our main result – the impact of area and individual level predictors and their interactions on the presence of depression diagnosis (binary variable) – through an odds ratio diagram where we display only significant predictors (p-value < 0.05 after adjustment using Benjamini-Hochberg procedure using R p.adjust).

The supplement provides information about the AIC criterion of the regression model, and demonstrates that the model created with LASSO variables provides the best fit and identifies more significant variables compared to the model with all predictors (Supplemental Table 2).

Results

The matching All of Us cohort had 33,994 individuals; their demographics, along with individual and area-level variables used in the analysis, are presented in Table 3. In our cohort, 31.7% had a depression diagnosis since 2020.

The initial list for modeling contained 1465 binary independent variables (that represents categorical variables from Table 3 transformed using one-hot encoding and the specified interactions between them, and includes 562 3-digit ZIP codes encoded as binary variables). LASSO method identified 218 important variables (ZIP codes were not included in variable selection and retained as random effect variable). The results of logistic regression mixed-effects modeling on the variables selected by LASSO method are displayed in Fig. 2, which shows only the 18 predictors with p-adjusted value < 0.05 (and hence confidence intervals not overlapping with 1 on the odds ratio scale). The model intercept corresponded to baseline odds of 0.40 (95% CI: 0.30–0.52, adjusted p < 0.001), reflecting the baseline probability of depression. Reference levels are described in the caption and Table 3.

Fig. 2.

Odds ratio analysis for two-way interactions with p-value < 0.05 (Benjamini-Hochberg adjusted). Values greater than one indicate a higher chance of the occurrence of a depression diagnosis. 95% CI displayed. “x”-shaped points indicate unadjusted (univariable) odds ratios. Colors represent specific social groups (except yellow, which represents area factor, and grey represents unspecific/other factors). The border around the column indicates the interaction between area and individual variables, or between overlapping identity factors. Reference levels: for Yes/No questions – No, Race/ethnicity – White, area-level indicators – Medium, per capita income (individual) – High, Age – Middle aged adults, Employment – Employed For Wages, Gender – Man, Sex at birth – Female, Sexual orientation – Non-heterosexual, Household – Single person household, How often/frequency questions – None/Never/Little, People under 18 in household – 0

Protective factors (OR < 1) included male sex at birth among straight individuals (OR 0.71, 95% CI: 0.59–0.86) and non-binary individuals (OR 0.42, 95% CI: 0.24–0.75), as well as older adults with active military service (OR 0.64, 95% CI: 0.53–0.77) or health insurance (OR 0.80, 95% CI: 0.72–0.88). Younger adults with 1–2 children under 18 also showed reduced odds (OR 0.72, 95% CI: 0.60–0.86).

Risk factors (OR > 1) were prominent in employment, food/housing insecurity, and discrimination domains. Being unable to work was strongly associated with higher odds (OR 2.01, 95% CI: 1.76–2.31), with additional risk for medium per capita income individuals (OR 1.39, 95% CI: 1.13–1.70). Housing concerns (OR 1.24, 95% CI: 1.09–1.41) and frequent experiences of disrespect (OR 1.41, 95% CI: 1.17–1.69) increased risk, as did high area deprivation in older adults (OR 1.31, 95% CI: 1.12–1.53). Intersectional risks were notable for non-binary gender in older adults (OR 5.14, 95% CI: 1.94–13.62) and those born outside the USA (OR 10.23, 95% CI: 2.07–50.48), women with children under 18 (OR 1.39, 95% CI: 1.25–1.55), and Hispanic individuals with active military service (OR 1.90, 95% CI: 1.32–2.74). Straight individuals who are not religious showed modestly elevated odds (OR 1.13, 95% CI: 1.05–1.22).

Supplementary Table 1 provides a complete regression table of predictors selected by LASSO, not just statistically significant ones.

Discussion

Our statistical learning analysis using LASSO for variable selection and a mixed-effects logistic regression model with 3-digit ZIP code random effect contributes to the growing literature on intersectional mental-health disparities and socioecological framework by jointly modeling individual-level identities while adding area-level contextual indicators within a large, diverse national cohort. Our findings align with prior work showing strong associations between individual socioeconomic disadvantage, discrimination, and depression [32, 33], while also highlighting several intersectional patterns—such as elevated risk among non-binary individuals and those unable to work—that have been less frequently examined. At the same time, our results identified only limited interactions between individual and area-level factors; we provide several possible explanations later in this section.

Several individual and interaction predictors showed strong associations with depression. Employment-related disadvantage was the most consistent predictor: individuals unable to work had more than twice the odds of depression (OR ≈ 2.01), and those out of work for one or more years also faced elevated risk (OR ≈ 1.34). In addition, individuals unable to work with a medium per capita household income showed an additional increased risk (OR ≈ 1.39), reflecting a potential intersection of disability and income [34].

Gender identity (non-binary) demonstrated an exceptionally strong effect (OR ≈ 10.22) when interacting with immigration status (born outside the USA), potentially mirroring research on mental health struggles of minority immigrant groups [35]. In addition, non-binary self-identification demonstrated a strong interaction with the older adults group (OR ≈ 5.1), contrasting existing research, which shows that older age may offer some protection against psychological difficulties [36]. While it is known that non-binary and non-heterosexual individuals face higher rates of depression [37–39], the present study found an additional risk (OR ≈ 1.13) for heterosexual individuals who identify as not religious, which may reflect research demonstrating religion may mitigate depression risks via social support [40]. It is worth mentioning one more time that our religious adherence measure captures the proportion of religious adherents at the community level, it serves as a contextual indicator of local social norms rather than an individual-level measure of religiosity.

The protective factor of being male is well described in the literature [7], but in our study it was found to be significant for both heterosexual males (OR ≈ 0.71) and for non-binary identifying males (OR ≈ 0.42). Women with two or more children under 18 in the household also showed elevated odds (OR 1.38), possibly due to the compounded stress of caregiving responsibilities [41]. However, younger adults living with children showed a modest protective effect (OR ≈ 0.71).

Active-duty status (OR ≈ 1.90) was another significant factor, highlighting potential occupational and structural stressors. The interaction between Active Duty Serve Status – Yes and Race – Hispanic may suggest the compounded stress faced by Hispanic individuals on active duty. This could be linked to unique challenges such as discrimination, deployment-related stressors, or limited access to culturally competent mental health support within military settings [42].

Age-related interactions were also important. Older adults experiencing food insecurity (OR ≈ 1.47) or housing instability (OR ≈ 1.33) had higher odds of depression, highlighting the potential role of basic needs and social discrimination in mental health outcomes [43, 44]. However, some factors emerged as protective for older adults: active duty status (OR ≈ 0.64) and the presence of health insurance (OR ≈ 0.79).

Two discrimination measures emerged as significant — being treated with less respect often (OR ≈ 1.41) or a few times a year (OR ≈ 1.17), as did housing insecurity (OR ≈ 1.24). These findings further emphasize that social exclusion and unmet basic needs are critical determinants of mental health.

We considered a broad range of area-level variables and their interactions with individual factors. Among those, most effect sizes were small and not statistically significant after Benjamini-Hochberg procedure. Non-the-less one interaction reached significance: older adults living in areas with a high deprivation index had elevated odds of depression (OR ≈ 1.31). Thus, our hypothesis that the area-level indicators would be important predictors of depression was only partially confirmed. We have several explanations for this observation. First, individual socioeconomic and psychosocial factors exert more immediate influence on mental health outcomes, whereas contextual effects tend to be more diffuse. Additionally, the use of 3-digit ZIP codes may reduce geographic variability, limiting the ability to detect subtle contextual interactions. The LASSO procedure further penalizes small or unstable interaction effects, making it less likely that weaker area-level interactions would be retained. Thus, the limited number of significant interactions likely reflects both methodological and substantive factors. In addition, adding more meaningful area-level variables to the model could have revealed more potential interactions.

These findings have several implications for public health. First, interventions should adopt an intersectional approach, addressing unique risks faced by non-binary individuals, those unable to work, and older adults experiencing material hardship. Second, strategies must target structural determinants such as employment barriers, housing instability, and discrimination. Finally, fostering social inclusion—whether through community engagement or anti-discrimination initiatives—can serve as a protective factor against depression.

Limitations

Not all possible variable interactions were explored in this analysis. The initial design matrix had only a finite curated list of demographic and social variables from the All of Us platform. In addition, we investigated only a few area-level predictors in this study linked from census data, while there are other important area-level indicators, such as ecological, safety, and housing, to name a few.

Other limitations of this study include potential misclassification due to aggregating data into 3-digit ZIP codes, which may obscure neighborhood-level variation, and categorizing area-level measures into ‘low,’ ‘high,’ and ‘average,’ which may oversimplify continuous distributions and may ignore state-by-state variations.

Although 3-digit ZIP-code random effects were included, this geographic unit may obscure neighborhood-level variation; results should not be interpreted as individual-level effects.

Additionally, the study period (January 2020–December 2023) overlaps with the COVID-19 pandemic, which likely influenced mental health, particularly among populations of color, due to disproportionate social and economic impacts.

Remission depression diagnoses were included because these codes represent individuals with a documented clinical course of major depressive disorder rather than absence of disease. Depression is often recurrent, and remission status reflects prior episodes and ongoing vulnerability.

Finally, a major limitation is the cross-sectional design, which precludes causal inferences due to temporal ambiguity between exposures (e.g., area-level indicators, individual characteristics) and depression outcomes. Longitudinal studies are needed to establish causality.

Conclusions

These findings highlight the value of integrating intersectional perspectives into both clinical and public health practice. At the individual level, clinicians should interpret screening guidelines in the context of unique area-level characteristics and patients’ overlapping identities. Similarly, from the population health perspective, incorporating area-level indicators such as deprivation or local social norms can help identify populations facing structural disadvantage. Together, these approaches support more targeted, equity-oriented strategies for depression prevention and intervention.

Authors’ contributions

All authors contributed to the study conception and design. Material preparation, data collection, and analysis were performed by Dmitry Scherbakov through consultations on data analysis with Michael Marrone and Alexander Alekseyenko. Leslie Lenert provided consultations on the usage and limitations of EHR data. The first draft of the manuscript was written by Dmitry Scherbakov, and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

Dr. Scherbakov is supported by grant T15 LM013977, Biomedical Informatics and Data Science for Health Equity Research (SC BIDS4Health) Training Program. Dr. Alekseyenko is support by grant T15 LM013977.

The All of Us Research Program is supported by the National Institutes of Health, Office of the Director: Regional Medical Centers: 1 OT2 OD026549; 1 OT2 OD026554; 1 OT2 OD026557; 1 OT2 OD026556; 1 OT2 OD026550; 1 OT2 OD 026552; 1 OT2 OD026553; 1 OT2 OD026548; 1 OT2 OD026551; 1 OT2 OD026555; IAA #: AOD 16037; Federally Qualified Health Centers: HHSN 263201600085U; Data and Research Center: 5 U2C OD023196; Biobank: 1 U24 OD023121; The Participant Center: U24 OD023176; Participant Technology Systems Center: 1 U24 OD023163; Communications and Engagement: 3 OT2 OD023205; 3 OT2 OD023206; and Community Partners: 1 OT2 OD025277; 3 OT2 OD025315; 1 OT2 OD025337; 1 OT2 OD025276.

Data availability

The data underlying this article is available for registered users on the All of Us platform. The code required to fetch the data and produce the results in this paper will be made available via GitHub [https://github.com/scherbakovdmitri/AllofUsDepressionDisparitiesLASSO/](https://github.com/scherbakovdmitri/AllofUsDepressionDisparitiesLASSO/).

Declarations

Ethics approval and consent to participate

This research aligns with the ethical principles set forth by the All of Us Research Program, including respect for persons, beneficence, and justice. This involves treating participants as autonomous agents, safeguarding their well-being, and ensuring fair distribution of research benefits and burdens.

The Institutional Review Board of the All of Us Research Program approved the collection of survey data and the procedures for sharing the data with verified and qualified researchers. Consent to participate was obtained from participants. The authors who conducted the analysis have gained controlled tier access rights per the All of Us policy.

Consent for publication

Not Applicable.

Competing interests

The authors declare no competing interests.