Neighborhood characteristics and incident myocardial infarction in US older adults: evaluation in 2 nationwide cohorts

Abstract

Inconsistent findings on the associations of adverse neighborhood context with myocardial infarction (MI) or racial disparities in MI may reflect publication bias or chance. We compared results from harmonized analyses of the REasons for Geographic and Racial Differences in Stroke (REGARDS, n = 25 196, aged ≥45 years, 42% Black; 2003-2018) study and the Health and Retirement Study (HRS, n = 14 191, aged >50 years, 13% Black; 2004-2018). For both cohorts, we used 51 American Community Survey (ACS) census tract variables to predict hazard of incident MI using Cox models. We evaluated consistency of coefficients and Black-White differences in coefficients between cohorts. Cumulative MI incidence in REGARDS (6.2% over 11.5 years median follow-up) was similar to HRS (7.1% over 13.1 years median follow-up). Of 51 ACS predictors evaluated, the log(HR) for incident MI differed by ≤0.05 between REGARDS and HRS for 34 variables. Of the 12 census tract predictors with significantly different associations with MI for Black versus White respondents in REGARDS, none showed interactions with race in HRS at the P <.05 threshold. Neighborhood socioeconomic associations with MI across 2 national studies were largely replicable. Racial differences in associations were inconsistent.

Introduction

Residence in a disadvantaged neighborhood has been associated with elevated incidence of cardiovascular disease (CVD).^1-6 This disadvantage has been variously characterized by low socioeconomic status (SES), labor force earning potential, or housing conditions.^7,8 Many investigations linking the local environment with CVD disparities report that racially marginalized people are overrepresented in disadvantaged neighborhoods; the excess risk for CVD events such as myocardial infarction (MI) among non-Hispanic (NH) Black adults has therefore been partially attributed to racial segregation.^9-11

However, between-sample differences may impact robustness of main and racially stratified estimates, particularly in predominantly White samples. Complex sampling and recruitment patterns influenced by correlates of neighborhood exposures may induce misleading associations, which may be especially problematic when evaluating racial differences. Violations of the positivity assumption will likely arise from analyses conducted in samples where the small proportion of Black participants are disproportionately concentrated in under-resourced, majority Black neighborhoods and are unlikely to be “exposed” to socioeconomically advantaged, racially integrated neighborhoods.^12,13

These methodologic challenges are exacerbated by potential cherry-picking of results in the neighborhood effects literature. For example, a recent meta-analysis on the impact of area-level characteristics on individual-level educational outcomes indicated that there is a statistically implausible proportion of studies with findings just above a 0.05 significance threshold.¹⁴ Across a large number of theoretically relevant metrics of neighborhood disadvantage and reasonable individual-level covariate sets, the specific combinations that meet conventional statistical significance criteria may disproportionately be published.¹⁵ The published characterizations of uncertainty, such as CIs, are almost uninterpretable in research on neighborhood disadvantage, without the context of the multiple comparisons that were examined but not comprehensively presented. Potential publication bias against null or unexpected results makes the existing literature unreliable for identifying the most relevant measures from the large set of plausible candidates for characterizing neighborhood disadvantage. Few studies have applied consistent analytic approaches to a large number of neighborhood factors across independent data sets to assess replicability. Additionally, assessing the impact of adverse neighborhood context on MI without sufficiently accounting for known individual-level social determinants of health may misestimate associations for older, low-SES, and rural adults most vulnerable to excess morbidity and mortality.^16,17 Systematically evaluating associations across an array of conceptually related exposures at the same spatial unit (here, the census tract) and from the same data source (here, the Census Bureau) can strengthen the research base investigating neighborhood determinants of CVD.

A well-developed theoretical base motivates our evaluation of neighborhood disadvantage and CVD disparities.^18-21 Racially marginalized people may engage in suboptimal health behaviors to cope with stressors in their local environment, losing the self-efficacy that facilitates management of CVD risk factors^22-24 and benefit less from higher SES (Figure 1).^25,26 The neighborhood factors conducive to improved CVD outcomes for socially dominant groups may similarly not confer the same benefit to the CVD health of their marginalized counterparts. Elucidating which neighborhood features are uniquely associated with lower MI incidence among Black adults remains vital to reduce racial disparities.^27,28

Figure 1.

Conceptual diagram of hypothesized pathways between adverse neighborhood context and disparities in myocardial infarction (MI).

Motivated by these evidence gaps, we assessed whether associations of neighborhood factors and incident MI were consistently observed among older adults participating in 2 national US cohorts: the REasons for Geographic and Racial Differences in Stroke (REGARDS) study and the Health and Retirement Study (HRS). We hypothesized that associations would vary in direction as well as magnitude across 2 national cohorts due to differences in sample composition, outcome ascertainment, and parent study sampling schema.

Methods

Study population

REGARDS is a closed cohort of 30 239 community-dwelling self-identified NH White and NH Black adults in the conterminous US states at least 45 years in age at 2003-2007 enrollment.²⁹ Black adults (42%) and residents of 8 southeastern states (North Carolina, South Carolina, Georgia, Arkansas, Tennessee, Alabama, Mississippi, and Louisiana, 56%) were oversampled using a random-stratified design. Approximately 1 month after collecting demographic information and medical history with computer-assisted telephone interviews, study staff conducted physical measurements, collected biospecimens, and conducted a medication inventory during a home exam. We calculated follow-up in REGARDS as the time between the date of baseline questionnaire completion and whichever of the following dates came first: MI event diagnosis, death, or December 31, 2018. REGARDS participants were asked if they had ever had a heart attack or coronary revascularization procedure at baseline; those who responded “no” to these questions, who had no ECG evidence of MI, and with available linkage to Census data were included in the current analytic sample (n = 25 196). We excluded 287 participants with no census tract linkage in all analyses, and 3821 with prevalent MI (Figure 2).

Figure 2.

Study inclusion criteria.

The HRS is a nationally representative open cohort of noninstitutionalized US adults over 50 years old along with their partners of any age.³⁰ The study uses a 4-stage sampling strategy of counties and metropolitan statistical areas within the conterminous United States. Participants completed biennial telephone and in-person questionnaires about their marital status, finances, and health conditions. For consistency with REGARDS, we defined the 2004 HRS wave as baseline for these analyses. Follow-up was from a baseline first interview in 2004 through time to reported event, the midpoint between the last interview without MI, and the first interview where MI was reported for those with missing date information (25%), date of death, date of the last interview, or December 31, 2018. The analytic sample included HRS participants who were interviewed in 2004 (excluding n = 1401 noninterviews and 1802 proxy interviews), identified as NH White or NH Black (excluding n = 1899 who identified as some other racial/ethnic group), reported being free of MI at the 2004 interview (excluding n = 508 reporting prevalent or unknown MI status), and had an address linked to 2005-2009 census data (final analytic n = 14 191). 325 HRS participants were lost to follow-up without an incident reported MI post-2004 (Figure 2).

2018 was chosen as the end point for follow-up because we wanted to isolate pre-COVID-19 pandemic effects of neighborhood-level factors on MI. REGARDS and HRS received approval from their respective Institutional Review Boards (IRBs). Both studies obtained written or verbal informed consent from all participants. We followed the STROBE (Strengthening The Reporting of OBservational studies in Epidemiology) guidelines.³¹

Ascertainment of incident MI

Regards

Telephone follow-ups of REGARDS participants or their proxies at 6-month intervals allowed for regular surveillance of CVD events. Hospitalization records for both nonfatal and fatal MI events were collected and adjudicated by medical reviewers following contemporaneous guidelines with disagreements resolved by consensus.^32,33 Following next of kin interviews, study staff adjudicated fatal MI events using all available information, including baseline and follow-up events data, medical records, death certificates, autopsy reports, and the National Death Index.

HRS

We ascertained nonfatal MI status in HRS across 7 biennial survey waves with the question “Has a doctor told you that you have had a heart attack?” Among participants who died, an analogous question was asked of first-degree family member proxies to ascertain fatal MIs.

Neighborhood-level exposures

The Census Bureau randomly samples housing units for participation in the American Community Survey (ACS).³⁴ We used residential census tracts as the closest local area surrogate for neighborhood context available in both samples. We evaluated ACS variables collected between 2005 and 2009, which have been aggregated as recommended by the Census Bureau for more robust estimates and interpolated to 2010 boundaries to account for annual changes, as the neighborhood characteristics most proximal to participant study baseline. To maintain participant anonymity, HRS supplies researchers a premanaged array of certain ACS variables from Social Explorer proprietary products that are commonly used in indicators of disadvantage³⁵; we derived the equivalent 51 ACS variables in REGARDS. Because missingness in exposure variables ranged from 0% to 3%, we replaced missing values in census tract variables with the median value for each state to reduce selection bias.

Neighborhood demographic composition

We included as demographic exposures variables that commonly are included in indices to represent populations vulnerable to chronic disease outcomes, such as older adults and children, and variables that are indirectly included in disadvantage indices, such as the proportion of women.³⁶ We assessed age distribution with the following proportional variables: under 18, between 18 and 64, and at least 65. We assessed proportions of those reporting the following racial and ethnic identities, which are often included in measures of segregation and deprivation: Hispanic or Latino, NH Asian, NH Black, NH White, and NH Other or multiple. To capture patterns of immigration and acculturation that correspond with marginalization and limited opportunities for upward social mobility, we assessed proportions of US nativity, those in the full population as well as those at least 65 years old speaking only English at home, those speaking English well or better at home, and household linguistic isolation (defined as all at least 14 years in the household speak a language other than English and none speak English very well).

Neighborhood socioeconomic status

We assessed highest educational attainment with the following proportional variables: no high school diploma, the equivalent of a high school diploma, and college degree. We assessed household income, adjusted for inflation in 2018 dollars. We also assessed factors commonly used in indices of area deprivation or disinvestment in the local socioeconomic infrastructure: female-headed households, household income falling below the poverty line, and households receiving public assistance income.

Neighborhood labor market conditions

We assessed the following proportional variables related to the civilian labor force at least 16 years in age, which can correspond with suppressed local economies: labor force participation, unemployment rate, employment in management and professional occupations, and employment in the manufacturing industry. Main commuting mode to work among the civilian labor force was determined with proportional variables for those driving alone, taking public transportation, and walking or biking.

Neighborhood housing

We assessed variables related to housing conditions, which can reflect poor housing quality or an influx of high-socioeconomic residents who can inflate cost of living for existing residents. This includes the proportion of householders who lived alone (among the full population, those aged 65, and by sex), as well as the proportion of households that were occupied by owners. We also considered proportion of housing units that were vacant (among all housing units, as well as among units for sale or for rent), which housed more than 1 person per room to indicate crowding, year built for housing units (before 1980, between 1980 and 1999, and after 1999), year that owners or renters moved into the census tract (after 1999 or after 2004), and single-unit homes. We assessed housing costs via the median house value and median gross rent, adjusted for inflation in 2018 dollars.

Individual-level and state-level factors related to adverse neighborhood context and incident coronary heart disease

We adjusted for individual-level age in 5-year intervals, sex as the closest available approximation of gender, as well as self-reported social determinants of coronary heart disease (CHD) assessed at baseline: highest educational attainment (less than high school, high school diploma or equivalent, some college, college degree, or not reported), marital status (currently married/partnered, never married/partnered, divorced, widowed, or other), insurance status (currently insured, not currently insured, or not reported), and annual household income in thousands of dollars (under 20, 20-34.9, 35-74.9, at least 75, or not reported). In HRS, we used imputed household income data provided by the RAND Cooperation, based on spousal demographic, economic, and health data, to minimize systematic error and sample size loss due to nonresponse.³⁷ We used an indicator for self-reported NH Black versus NH White racial and ethnic identity as a proxy for societal treatment. To account for social and policy context for neighborhood characteristics, we included in all models a fixed term for state of residence.

Statistical analyses

We first described baseline characteristics of the REGARDS sample (2003-2007), and the HRS sample (2004), stratified by Black versus White race. To increase comparability of estimates, we standardized all census tract variables to have a mean of zero and SD of 1, and centered the individual-level participant age around the mean in each cohort. We further reduced the skew in continuous variables for housing costs and household income by log-transforming these variables prior to standardization.³⁸ We conducted parallel Cox proportional hazard models of each census tract variable and incident MI separately in REGARDS and HRS. Each hazard model was stratified for baseline age in 5-year intervals and state of residence. For each multiply adjusted model, we assessed violations of the proportional hazard assumption with the Grambsch-Therneau test.³⁹ In each cohort, we compared estimates that were minimally adjusted for sex with estimates further adjusted for educational attainment, marital status, household income, and current insurance coverage. We also applied available weights for the probability of being eligible for parent study inclusion so that HRS estimates were representative of the noninstitutionalized US population over 50 in 2004.

To evaluate the impact of publication bias against paradoxical findings, we compared whether associations in REGARDS and HRS were in the same direction and met conventional thresholds for significance (P <.05). Additionally, we considered estimates to be inconsistent across studies if the estimate in HRS differed by at least 0.05 on the logarithmic scale from the corresponding estimate in REGARDS. The log(0.05) threshold for between-cohort differences in point estimates was chosen to balance the goals of detecting substantively important divergence in results against avoiding overinterpreting trivial fluctuations due to finite samples.

We evaluated heterogeneity in estimated effects among White and Black participants according to the interaction between participant race and each standardized census tract variable using a 0.05 P-value threshold for the interaction term. We assessed whether the racial composition of neighborhoods was differentially correlated with neighborhood housing, economic, labor force, and demographic variables in either cohort. To do so, we compared with Pearson correlation coefficients, density plots, and binned scatter plots the distributions between the proportion of NH White or NH Black residents and census tract factors for which we observed racial differences in the association with incident MI. We conducted all analyses using Stata version 18.5 and R version 4.3.2.

Results

Participants were followed for a median of 11.5 years (IQR: 6.5-13.6) in REGARDS and 13.1 (8.3-14.1) years in HRS. Cumulative incidence of MI was similar in REGARDS (6.2%) and HRS (7.1%). By applying the sampling weight, all HRS estimates are meant to represent 3 949 063 self-reported MIs among 62 397 106 noninstitutionalized US adults over 50 years. Compared to those who identified as White, Black REGARDS and HRS participants were on average more likely to be female, unmarried, of lower income and educational attainment, and uninsured. The REGARDS participants were more likely than their HRS counterparts to have attended college or attained a college degree (Table 1). Compared to REGARDS participants, HRS participants on average lived in census tracts with higher proportions of White residents and lower educational attainment (Table 2). Participants overwhelmingly resided in census tracts that corresponded with their own racial identity. Census tract-level household income, median rent, and home values were also higher in HRS. Black adults in both cohorts lived in census tracts with more residents of lower SES, fewer owner-occupied housing units, more housing units built before 1980, and more female-headed households with children (Table 2). The proportional hazard assumption was not violated for individual- or census tract-level covariates (Figure S1).

Table 1.

Baseline characteristics and myocardial infarction (MI) incidence of REasons for geographic and racial differences in stroke (REGARDS) and health and retirement study (HRS) participants initially free of myocardial infarction, stratified by racial and ethnic identity.

	REGARDS, 2003-2007 (n = 25 196)		HRS, 2004 (n = 14 191)
	Non-Hispanic White (n = 14 678)	Non-Hispanic Black (n = 10 518)	Non-Hispanic White (n = 11 981)	Non-Hispanic Black (n = 2210)
Characteristic
Age, years (IQR)	64 (58-72)	63 (57-70)	67 (60-75)	65 (58-72)
Male, n (%)	6933 (47)	3865 (37)	5194 (43)	827 (37)
Educational attainment
Less than high school, n (%)	938 (6)	1977 (19)	1840 (15)	807 (36)
High school diploma, n (%)	3514 (24)	2923 (28)	4564 (38)	686 (31)
Some college, n (%)	3985 (27)	2840 (27)	2726 (23)	439 (20)
At least college graduate, n (%)	6233 (43)	2769 (26)	2854 (24)	278 (13)
Missing, n (%)	8 (<1)	9 (<1)	N/A	N/A
Relationship status
Married/partnered, n (%)	10 038 (68)	4768 (45)	8498 (71)	1083 (49)
Divorced, n (%)	1680 (11)	2108 (20)	1059 (10)	470 (21)
Widowed, n (%)	2284 (16)	2358 (23)	2136 (17)	524 (24)
Never married/partnered, n (%)	554 (4)	819 (8)	286 (2)	128 (6)
Other/missing, n (%)	111 (1)	455 (4)	5 (<1)	5 (<1)
Currently insured
Yes, n (%)	13 998 (95)	8116 (89)	11 514 (96)	2021 (91)
No, n (%)	670 (5)	1050 (10)	469 (4)	185 (8)
Missing, n (%)	10 (<10)	13 (<1)	1 (<1)	4 (<1)
Household income
Less than $20 K, n (%)	1621 (11)	2694 (26)	2477 (21)	979 (44)
$20 K-34.9 K, n (%)	3232 (22)	2776 (26)	2544 (21)	416 (19)
$35 K-$74.9 K, n (%)	4853 (33)	2748 (26)	3774 (32)	533 (24)
At least $75 K, n (%)	3183 (22)	986 (9)	3189 (27)	281 (13)
Not reported, n (%)	1789 (12)	1314 (13)	N/A	N/A
Incident MI, n (%)	972 (7)	599 (6)	836 (7)	166 (8)

Educational attainment categories were based on the comparatively less granular options in REGARDS. In HRS, “some college” indicates either some years of reported college attendance or the attainment of an Associate Degree or technical degree and “at least college graduate indicates the attainment of at least a 4-year degree, including bachelor’s degree.” Both studies ascertain household income data using unfolding sequential brackets. HRS only provides income for respondents and their spouses. In collaboration with the RAND Cooperation, HRS provides imputed household income data to minimize systematic error and nonresponse based on spousal demographic, economic, and health data.

Table 2.

Baseline census tract characteristics of (REGARDS) and HRS participants initially free of MI, stratified by participant-level racial and ethnic identity.

	REGARDS, 2003-2007 (n = 25 196)		HRS, 2004 (n = 14 191)
Characteristic	Non-Hispanic White (n = 14 678)	Non-Hispanic Black (n = 10 518)	Non-Hispanic White (n = 11 981)	Non-Hispanic Black (n = 2210)
Unique census tracts, n	9653	4495	3855	895
Gender distribution
Men, %	48	46	49	47
Women, %	52	54	51	53
Age demographics
Aged under 18, %	25	26	24	26
Aged 18-64, %	62	61	62	62
Aged 65 and over, %	13	13	14	12
Racial and ethnic categories
NH Asian and Pacific Islander, %	<1	<1	1	<1
NH Black, %	14	78	2	59
NH White, %	76	14	85	16
NH Other or Multiple, %	3	3	2	2
NH Hispanic or Latino, %	3	2	4	4
Non-US nativity, %	4	3	5	4
Speak only English at home, %	95	96	92	93
Aged 65+ speak only English at home, %	97	98	95	96
Speak English well or better at home, %	99	99	99	99
Aged 65+ speak English well or better at home, %	100	100	100	100
Linguistically isolated households, %	<1	<1	<1	<1
Educational attainment
No high school degree, %	15	22	12	21
High school degree, %	32	34	62	61
College graduates, %	27	19	22	15
Labor force participation rate, %	64	60	66	62
Unemployed, %	4	7	6	11
In management or professional occupations, %	30	24	32	24
In manufacturing industry, %	11	9	11	9
Commuting method among labor force
Drove alone, %	81	75	81	75
Public transportation, %	<1	4	1	4
Walked or bicycled, %	1	1	2	2
Households receiving public assistance, %	1	3	1	3
Female-headed households with children, %	11	22	5	13
Families with children in poverty, %	9	19	5	15
In poverty, %	13	24	9	22
Aged 65 and over in poverty, %	9	15	7	14
Living alone, %	10	11	10	10
Aged 65 and over living alone, %	7	9	6	7
Men aged 65 and over living alone, %	20	21	19	19
Women aged 65 and over living alone, %	<1	<1	4	4
Owner-occupied housing units, %	63	50	77	59
Vacant housing units, %	10	14	9	12
Vacant housing units for rent, %	17	22	5	8
Vacant units for sale, %	9	7	2	2
Single-unit housing units, %	70	67	72	63
More than 1 person per bedroom, %	1	2	1	3
Housing unit construction
Units built before 1980, %	60	83	66	80
Units built between 1980 and 1999, %	30	12	27	15
Units built after 1999, %	7	4	7	5
Homeowners moved after 1999, %	37	29	39	33
Homeowners moved after 2004, %	14	10	14	13
Renters moved after 1999, %	84	80	84	81
Renters moved after 2004, %	52	47	52	48
Median household income, $1000s (IQR)	53 (40-72)	38 (29-53)	60 (47-79)	57 (44-76)
Median home value, $1000s (IQR)	159 (106-258)	112 (82-195)	187 (122-314)	136 (90-239)
Median rent (IQR)	837 (683-1071)	834 (678-1035)	896 (705-1188)	892 (706-1156)

In estimates minimally adjusted for age, sex, and state of residence, 10 of the 51 census tract variables evaluated were associated with incident MI in both REGARDS and HRS at conventional levels of significance (Figure 3). These variables were generally indicators of socioeconomic status, housing costs, or the civilian labor force. Additionally, 12 census tract factors were only associated with MI in 1 cohort. As an example, the percentage of NH Black adults was positively associated with higher incident MI in HRS (HR per 1 SD = 1.14 [95% CI, 1.05-1.23]) but not in REGARDS (HR per 1 SD = 1.03 [95% CI, 0.98-1.09]).

Figure 3.

Minimally adjusted* associations between census tract (CT) factors and incident myocardial infarction in REasons for Geographic and Racial Differences in Stroke (REGARDS) and Health and Retirement Study (HRS). *models adjusted for sex, and stratified by baseline age in 5-year intervals as well as and state of residence. All CT variables are standardized around the mean. REGARDS n = 25 196; HRS n = 14 191.

After further adjustment for individual-level covariates, the estimated effects of neighborhood factors were attenuated. Median home value, median household income, the proportion of individuals aged 65 in poverty, and proportion of families with children in poverty were the only census tract factors that remained consistently associated with the incidence of MI across cohorts. For 8 ACS variables, estimated effects with incident MI met conventional thresholds for significance in only 1 cohort. As an example, the percentage of adults with less than a high school diploma was associated with higher incident MI more robustly in REGARDS (HR per 1 SD = 1.10 [95% CI, 1.04-1.17]) than HRS (HR per 1 SD = 1.06 [95% CI, 0.96-1.17]). Census tract exposures related to age, sex, racial and ethnic composition, language use, housing, and the labor force were not consistently associated with MI incidence in either cohort.

Furthermore, we observed that 34 of 51 multiply adjusted associations between census tract variables and MI in HRS differed from corresponding REGARDS associations by 0.05 or less on a logarithmic scale. (Figures 4 and 5). In both cohorts, census tract-level factors that indicated high area-level socioeconomic status—such as median home value (REGARDS HR per 1 SD = 0.91 [95% CI, 0.84-0.99; HRS HR = 0.87 [0.77-0.99]) and the median household income (REGARDS HR per 1 SD = 0.93 [95% CI, 0.86-0.99; HRS HR = 0.89 [0.81-0.99])—were inversely associated with MI. The estimated effect of the percentage of residents in poverty on MI was weaker in REGARDS (HRs per 1 SD = 1.05 [95% CI, 1.05-1.00-1.11]) than the corresponding estimate in HRS (HR per 1 SD = 1.17 [95% CI, 1.17-1.08-1.27]).

Figure 4.

Multiply adjusted* associations between CT factors and incident myocardial infarction in REGARDS and HRS. *models adjusted for gender, identify as NH Black versus NH White, educational attainment, categories of household income, and insurance coverage, and stratified by baseline age in 5-year intervals as well as state of residence. All CT variables are standardized around the mean. REGARDS (2003-2018) n = 25 196; HRS (2004-2018) n = 14 191.

Figure 5.

The CT variables with at least a 0.05 difference in the relative hazard (logarithmic scale) of MI in REGARDS and HRS*. *models adjusted for gender, educational attainment, categories of household income, and insurance coverage, and stratified by baseline age in 5-year intervals as well as state of residence. All CT variables are standardized around the mean. REGARDS (2003-2018) n = 25 196; HRS (2004-2018) n = 14 191.

We observed racial differences in associations with incident MI for 12 census tract factors in REGARDS (Figure 6). None of these 12 factors replicated to show an interaction with race in HRS at the P <.05 threshold. As an example, in REGARDS, a higher proportion of households receiving public assistance was positively associated with MI among White (HR, 95% CI, 1.12-1.03-1.21) but not Black participants (HR, 95% CI, 1.01-0.94-1.10, P-value for interaction: 0.023). Conversely, the corresponding estimated effect of households receiving public assistance being associated with incident MI in the full HRS sample was only observed among the Black participants (HR, 95% CI, 1.17-1.03-1.33) but not White participants (HR, 95% CI, 0.99-0.88-1.12, P-value for interaction: 0.133). In HRS, only 1 census tract factor, the proportion of housing units built between 1980 and 1999, showed significant interaction with race, which was not replicated in REGARDS.

Figure 6.

The CT variables with the greatest (P for interaction = <.05) Black-White differences in associations with incident myocardial infarction in REGARDS or HRS. *Models adjusted for gender, educational attainment, categories of household income, insurance coverage, and stratified by baseline age in 5-year intervals as well as state of residence. All CT variables are standardized around the mean. REGARDS (2003-2018) n = 25 196; HRS (2004-2018) n = 14 191.

Differences in statistical power and correlation structures between cohorts may together contribute to the lack of consistency in main or interactive estimates effects observed between cohorts. The census tract factors that were most collinear in both samples were the proportions of NH White and NH Black adults (Figure S1 and S2: r among White REGARDS participants = 0.95, among Black REGARDS participants = –0.94, among White HRS participants = –0.63, and among Black HRS participants- =0.82). White HRS participants lived in census tracts with lower proportions of Black residents than their REGARDS counterparts (median 2% vs 14%), so a census tract factor that disproportionately elevated risk in majority Black neighborhoods would be more strongly interrelated with MI incidence among Black than White adults in HRS. However, there did not appear to be a meaningful pattern that explained the differences in racial heterogeneity between cohorts.

Conclusion

In one of the first studies to systematically compare neighborhood characteristics as predictors of health outcomes in 2 large, diverse, and national US samples, we found some evidence for robust effects of neighborhood characteristics associated with disadvantage on incident MI. Census tract variables related to socioeconomic status had stronger, more consistent associations with incident MI than area-level demographics, housing, or labor force characteristics. Racial differences in census tract-level risk factors did not replicate across cohorts.

The proportion of NH Black residents in a census tract is often included in composite scales that measure structural inequity and disadvantage.^7,20,40 Given strong racial residential segregation in the United States, we observed older adults disproportionately live in census tracts whose composition corresponds with their racial identity. However, we found little evidence of independent associations between racial composition of a census tract and MI after adjustment for individual-level covariates. Our findings correspond with those from the Multi-Ethnic Study of Atherosclerosis (MESA), where less income, lower educational attainment, nonmanagement or professional occupations, and non-US nativity contributed to 1.6 more CVD per 1000 person years (–8.4 to 11.6) among participants of lower socioeconomic status.¹ In MESA, participants of lower socioeconomic status were disproportionately exposed to adverse neighborhood conditions, which also held true in our analysis. Similarly, among 4096 Black adults in the Jackson Heart Study, living in neighborhoods with higher disadvantage (using a composite of 8 indicators including poverty and fewer residents without a high school degree) was associated with higher incidence of coronary heart disease or stroke among women (HR per 1 SD, 95% CI, 1.23-1.04-1.45) but not men (HR per 1 SD, 95% CI, 1.03-0.73-1.36). Many of the indices commonly used to summarize adverse neighborhood context include collinear factors available from publicly available national datasets such as the Census Bureau.^7,8 The explanatory power of these indices appears to be driven by the few variables with the highest variance.^38,41 Taken in context with our results, where many associations had similar estimates and overlapping confidence intervals, future work on neighborhood determinants of incident CVD should quantify the additional explanatory power gained by using multiple correlated components in commonly used indices of disadvantage. Our Pearson correlation plots indicate that equivocal evidence on the impact of racial segregation on CVD may arise from positivity assumption violations of individual- and neighborhood-level factors, where Black adults are disproportionately concentrated in majority Black census tracts.⁴²

Our results show that in both REGARDS and HRS, several commonly used indicators of adverse neighborhood SES context are associated with MI. However, the same measures of neighborhood often did not replicate in magnitude between the 2 studies and racial differences were not corroborated. As a result, any composite index summarizing across multiple domains of neighborhood deprivation would likely show associations in either data source. However, our systematic comparison calls into question the interpretation of any specific indicator of deprivation. This is critical as the field moves from describing the association of neighborhood deprivation and health to developing specific intervention targets to reduce CVD among older adults, such as providing more affordable housing or increasing household earning power for vulnerable populations.

Strengths and limitations

We coordinated analyses of 2 large, national samples to allow direct comparisons of associations of commonly used census tract-level demographic, labor force, economic, and housing conditions with incident MI. By evaluating interactions by race, we considered how differences in MI incidence by race look relative to the neighborhood characteristics of these participants. Though we took steps to make our study samples comparable, differences between REGARDS and HRS remain. Follow-up time for REGARDS participants began at least 6 years earlier in life (45 vs 51 years). Differences in results between studies may be partially attributable to HRS’ use of self- or proxy-reported MI events versus medically adjudicated events in REGARDS. Estimates may be systematically biased to the null in HRS, as events and causes of death were not adjudicated in HRS. Notably, though prior studies demonstrated high concordance between reported and medically adjudicated CVD events,⁴³ clinical advancements such as high sensitivity of troponin assays may detect increasingly smaller MI events that older patients of low SES may not report.⁴⁴ A recent study among HRS participants at least 67 years old indicated that the sensitivity of Medicare claims-diagnosed MI with self-report was moderate (67%) though claims are not available for younger segments of the sample.⁴⁴

Though a greater proportion of REGARDS participants identified as Black than in HRS, there did not seem to be systematic differences in individual-level covariates or coverage across census tracts that would explain the results. Because Black people comprised a larger fraction of the REGARDS participants than HRS participants, racial interaction analyses in HRS may be underpowered. Additionally, the parent studies only make available census tractlink ages for participants in mid-to-late life. Neighborhood factors ascertained at one point in time may be consequences of poor CVD health or individual-level socioeconomic status instead of factors contributing to MI incidence. Future work should incorporate data from across the lifecourse, and where possible, study designs that better approximate random exposure assignment to better infer causality between neighborhood disadvantage and MI.^45,46

Conclusion

The equivocal evidence base linking adverse neighborhood conditions to MI may be partially attributable to between-study variation in how demographic, housing, and labor force neighborhood characteristics co-occur in racially segregated areas. We observed that the multistage sampling strategy of HRS may have induced more severe violations of the positivity assumption than observed in REGARDS by more frequently including Black participants with lower educational attainment residing in majority Black, lower SES areas. Choice of covariate adjustment as well as between sample differences in racial composition and outcome ascertainment (adjudicated versus self-reported) also appear to influence associations between neighborhood characteristics and incident MI events.

Our work illustrates that the association of many neighborhood characteristics with neighborhood racial composition complicates the interpretation of findings on the relationship of neighborhood context with health outcomes. To reduce the publication bias in this research area, we recommend contextualizing null or unexpected results with explicit evaluations of the sampling strategy and sample composition. After retirement, older adults spend on average more time in their neighborhood of residence. Rigorous research triangulating evidence from multiple, large samples is therefore needed to identify features in the local environmental context that promote MI prevention and reduce racial disparities among middle-aged and older adults.

Supplementary material

Supplementary material is available at the American Journal of Epidemiology online.

Funding

This work was funded by National Institutes of Health grants (K99AG083121 to K.D.S., 2P30AG015272-21 to K.D.S., T.B.N., and J.K.J., T32AG049663 to K.D.S., R56AG049970 to G.S.L., R01HL165452 to M.M.S., and R01AG072681 to M.M.G.).

Conflict of interest

M.M.G. is on the Editorial Board at AJE.

Disclaimer

This research project is supported by cooperative agreement U01NS041588 cofunded by the National Institute of Neurological Disorders and Stroke (NINDS) and the NIA. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NINDS or the NIA. Representatives of the NINDS were involved in the review of the manuscript but were not directly involved in the collection, management, analysis or interpretation of the data.

Data availability

REGARDS data are available to researchers after approval from the parent study. HRS data are available after registration and approval to use restricted spatial data. Analytic code can be provided by the first author upon reasonable request.