Racial Disparities in Context: A Multilevel Analysis of Neighborhood Variations in Poverty and Excess Mortality Among Black Populations in Massachusetts

Abstract

Objectives. We analyzed neighborhood heterogeneity in associations among mortality, race/ethnicity, and area poverty.

Methods. We performed a multilevel statistical analysis of Massachusetts all-cause mortality data for the period 1989 through 1991 (n=142836 deaths), modeled as 79813 cells (deaths and denominators cross-tabulated by age, gender, and race/ethnicity) at level 1 nested within 5532 block groups at level 2 within 1307 census tracts (CTs) at level 3. We also characterized CTs by percentage of the population living below poverty level.

Results. Neighborhood variation in mortality across CTs and block groups was not accounted for by these areas’ age, gender, and racial/ethnic composition. Neighborhood variation in mortality was much greater for the Black population than for the White population, largely because of CT-level variation in poverty rates.

Conclusions. Neighborhood heterogeneity in the relationship between mortality and race/ethnicity in Massachusetts is statistically significant and is closely related to CT-level variation in poverty.

Despite documentation of substantial variations in US mortality, whether by state and county or by race/ethnicity, gender, and socioeconomic position,1–10 few analyses have examined mortality variation at the local geographic level, for example, by neighborhood. Even fewer have attempted to elucidate whether the well-known average racial/ethnic disparities in mortality vary across neighborhoods. We offer a systematic examination of such contextual heterogeneity, defined as geographic variation in the individual relationship between race/ ethnicity and mortality, conditional upon adjustment of other individual covariates.

Why would an empirical evaluation of contextual heterogeneity by population subgroups (e.g., White and Black) be important? Patterns of all-cause mortality are shaped by a complex constellation of individual as well as contextual factors that may conceivably vary for Whites and Blacks, as suggested, for example, by the fact that leading causes of death differ for different racial/ethnic groups.11 An investigation of race/ethnicity–specific contextual heterogeneity in mortality at the neighborhood level can give insight into the relative importance to mortality of individual and contextual factors for different racial/ethnic populations. From a methodological standpoint, addressing this question requires that context be included as an intrinsic part of analytical strategies, as achieved by multilevel statistical models.12,13

To date, however, multilevel methods in public health research have principally been applied to estimate the “average” effect of a predictor measured at an area level on individual-level health outcomes, including mortality.14,15 While this approach is important, it can potentially obscure the contextual heterogeneities underlying such average effects. We examined contextual heterogeneity in mortality in relation to race/ethnicity and area poverty by analyzing 1989–1991 Massachusetts mortality data, prepared for the US Public Health Disparities Geocoding Project,16 in conjunction with 1990 census data. Specifically, we addressed 3 questions: (1) What is the neighborhood variation in mortality rates for different racial/ethnic groups? (2) What is the magnitude of neighborhood variation in the average Black–White disparity? (3) Does neighborhood variation in poverty account for the racial/ethnic contextual variation and racial/ethnic disparities in mortality?

To characterize individuals’ neighborhood context, we employed 2 levels of census geography: the census tract (CT), which on average contains 4000 persons, and the census block group (BG), a subdivision of the CT, which on average contains 1500 persons.17 The 2-fold appeal of CTs is that “when first delineated, [they] are designed to be homogeneous with respect to population characteristics, economic status, and living conditions,”17^{(ppG10–G11)} and that, once created, they constitute administrative units used by federal, state, and local governments, including public health departments, to characterize jurisdictions, determine eligibility for diverse programs, and allocate resources.18–20 Therefore, CTs have real-life implications for their residents.

METHODS

Data

The Massachusetts Department of Public Health provided data for all deaths occurring between January 1, 1989, and December 31, 1991, to Massachusetts residents (n = 156 366). The denominator data were obtained from the 1990 US census for Massachusetts (n = 6016425). Mortality records were geocoded to the CT and BG level with an accuracy of about 96%.21 Death records missing information on age, gender, or race/ ethnicity were excluded from the analyses, along with records geocoded to BGs with populations of 0. The final data consisted of 142 836 death records corresponding to 18049275 person-years.

Analytical Structure

The data had a hierarchical multilevel structure of 79813 cells at level 1, consisting of individuals in numerators and denominators cross-tabulated by age, gender, and race/ ethnicity, which were nested within 5532 BGs at level 2, nested within 1307 CTs at level 3. Each BG had between 1 and 30 cells, with the numerator specifying the number of persons who died and the denominator providing the total population used to calculate the proportion of deaths in each cell. In the analytical model we considered both BG and CT levels; however, given the greater policy appeal for health monitoring of populations, our substantive interest was at the CT level. Structurally, our models were identical to models with individuals at level 1.22

Response and Predictors

The response, mortality, is defined as a proportion: the number of deaths as a proportion of the total population for each cell. The cell predictor variables relate to 5 age categories (0–14 years, 15–24 years, 25–44 years, 45–64 years, and 65 years and older), 2 gender categories (male and female); and 3 racial/ ethnic categories (White, Black, and all others, each including non-Hispanics and Hispanics; these 3 groups made up, respectively, 89.9%, 4.9%, and 5.1% of the population enumerated in Massachusetts in the 1990 census). Because of the heterogeneity of the population in “all others,” we focused our interpretation mainly on the White–Black comparisons. On the basis of the cross-tabulation of age, gender, and racial/ethnic categories, we obtained 30 unique population groups, or cells, as shown in Table 1 ▶.

TABLE 1—

Description of Data Used for Multilevel Models Analyzing Neighborhood Heterogeneity in Associations Between Mortality, Race/Ethnicity, and Area Poverty: Massachusetts, 1989–1991

Predictor	No. of Cells (% of Total)	No. of Deaths	Person-Years	Mortality Rate per 100 000 (SD)
Overall	79 813	142 836	18 049 275	967 (2644)
Level 1: cell (n = 79 813)
Age 0–14 y
White male	5243 (6.6)	675	1 496 835	61 (293)
Black male	1284 (1.6)	147	116 898	103 (556)
Other male	1849 (2.3)	120	135 519	88 (625)
White female	5234 (6.6)	498	1 416 618	44 (241)
Black female	1274 (1.6)	116	115 398	104 (673)
Other female	1886 (2.4)	96	131 454	61 (457)
Age 15–24 y
White male	5210 (6.5)	860	1 196 304	99 (401)
Black male	1076 (1.3)	158	80 268	159 (675)
Other male	1486 (1.9)	86	93 798	92 (487)
White female	5230 (6.6)	289	1 198 560	34 (212)
Black female	1110 (1.4)	37	82 476	39 (364)
Other female	1511 (1.9)	23	98 478	24 (265)
Age 25–44 y
White male	5431 (6.8)	4265	2 681 712	196 (415)
Black male	1718 (2.2)	549	151 275	297 (918)
Other male	2391 (3.0)	373	161 760	214 (822)
White female	5460 (6.8)	1807	2 748 753	76 (253)
Black female	1653 (2.1)	263	155 595	119 (514)
Other female	2525 (3.2)	130	164 478	83 (593)
Age 45–64 y
White male	5379 (6.7)	12 570	1 500 516	1080 (1450)
Black male	999 (1.3)	597	58 539	825 (1870)
Other male	1312 (1.6)	228	48 906	470 (1570)
White female	5375 (6.7)	8029	1 627 536	598 (878)
Black female	1019 (1.3)	429	73 383	449 (1170)
Other female	1424 (1.8)	143	54 435	282 (1440)
Age ≥ 65 y
White male	5233 (6.6)	46 725	900 960	6079 (5160)
Black male	492 (0.6)	726	21 948	3133 (4650)
Other male	460 (0.6)	191	14 193	1177 (3140)
White female	5341 (6.7)	61 535	1 467 528	4636 (4320)
Black female	620 (0.8)	965	35 238	2560 (3880)
Other female	588 (0.7)	206	19 914	939 (2470)
Level 2: block group (n = 5532)	. . .	. . .	. . .	. . .
Level 3: census tract (n = 1307)	. . .	. . .	. . .	. . .
% of CT population below poverty	base: CT poverty 0%–4.9% (n = 492; 37.6%); contrast: CT poverty 5%–9.9% (n = 386; 29.5%); CT poverty 10%–19.9% (n=225; 17.2%); CT poverty 20%–100% (n=204; 15.6%)

Note. CT = census tract. Data on deaths of Massachusetts residents occurring between January 1, 1989, and December 31, 1991, are from the Massachusetts Department of Public Health (n = 156 366). Denominator data were obtained from the 1990 US census for Massachusetts (n = 6 016 425). Mortality rate was calculated on the basis of the means of the proportion of deaths for each cell type across all block groups and CTs. The data on CT poverty came from the 1990 US Census.

CTs were characterized by percentage of the population living below the poverty line, with the following cutoff points: 0%–4.9%, 5%–9.9%, 10%–19.9%, and 20%–100%. The rationale for considering area poverty was based on previous research showing the consistency with which area poverty detected socioeconomic gradients across a range of health outcomes, including mortality.16,23,24 The federal definition of “poverty areas”25 as areas where more than 20% of persons are living below the poverty line (in 1989, equal to $12575 for a family of 2 adults and 2 children)26 was a key consideration in determining the cutoff points. Although data on educational level were available from the death certificates, 1990 census population data stratified simultaneously by age, gender, race/ethnicity, and educational level were available only at the county—not CT or BG—level, precluding the use of education in specifying the cell structure.

Statistical Analysis

We employed multilevel statistical procedures27–29 because of their ability to model complex variance structures at multiple levels. The principles underlying multilevel modeling procedures are now well known,13 and in the context of the present analysis they allow estimation of the relationship between mortality and race/ethnicity, conditional on individual age and gender variations (“fixed parameters”) and CT- and BG-level variations (“random parameters”). They also enable an estimation of the extent to which the relationship between mortality and race/ethnicity varies across CTs (random parameters) and the degree to which CT poverty explains this variation (fixed parameters).

The response variable, proportion of deaths in each cell, was modeled with allowances made for the varying denominator in each cell.27 The fixed and random parameter estimates (along with their standard errors) for the 3-level binomial logit link model were calibrated using predictive/penalized quasi-likelihood procedures with second order Taylor series expansion,30 as implemented within the MLwiN program.31 We allowed for extrabinomial variations at level 1, because proportions may exhibit more or less variation than a binomial distribution.32 We calibrated 3 models.

Model 1.

Model 1 was a 3-level model of cells (level 1) within BGs (level 2) within CTs (level 3), with cell characteristics related to age, gender, and race/ethnicity specified in the fixed part of the model and a residual variation estimated at the CT and BG levels in the random part. The fixed-part specification included the main effects for the age, gender, and racial/ethnic categories, along with a second-order interaction between age–gender, age–race/ethnicity, and gender–race/ ethnicity. We did not find any empirical support for third-order interaction terms. The estimates from model 1 allowed an evaluation of the racial/ethnic differences in mortality and the magnitude of variation in mortality at the BG and CT levels, conditional on the relationship between mortality and age, gender, and race/ethnicity within each BG and CT.

Model 2.

Model 2 was similar to model 1, but it allowed the fixed racial/ethnic differential on mortality to vary across CTs in the random part to obtain differential CT-level variation in mortality for Whites, Blacks, and others. CT-level random parameters from model 2 were used to test the hypothesis pertaining to race/ethnicity–based contextual heterogeneity, that is, whether neighborhood-level variation in mortality was different for different racial/ethnic groups.

Model 3.

Model 3 was similar to model 2, but it included a fixed cross-level interaction effect between CT-level poverty and individual race/ethnicity. In this way, we ascertained the relationship between neighborhood poverty, individual race/ethnicity, and mortality, as well as the extent to which CT-level poverty accounted for the CT-level racial/ethnic variation in mortality.

RESULTS

When we controlled for age, the mortality odds ratio (OR) was 25% higher for men than for women (OR = 1.25, 95% confidence interval [CI] = 1.11, 1.41) and twice as high for Blacks as for Whites (OR = 1.96, 95% CI = 1.65, 2.33; Table 2 ▶, model 1), in the reference age group of 0-14 years. The odds ratio for the 3 interaction terms (age × gender, age × race/ethnicity, gender × race/ethnicity) represent the unique additional impact of these interactions, not including the main effects. Crucially, and of relevance to this study,the between-CT variance (σ²_υ0 = 0.09; Table 3 ▶, model 1) was statistically significant (P < .001) even after we took into account the fixed main effect and the 3 two-way interactions, of age, gender, and race/ethnicity on mortality and after we took into account the between-BG (within-CT) differences in mortality.

TABLE 2—

Odds Ratios and 95% Confidence Intervals for Fixed Parameters From Models 1, 2, and 3 Analyzing Neighborhood Heterogeneity in Associations Between Mortality, Race/Ethnicity, and Area Poverty: Massachusetts, 1989–1991

	OR (95% CI)
	Model 1	Model 2	Model 3
Individual-level predictors
Main effect
Age, y
0–14	1.00	1.00	1.00
15–24	0.64 (0.54, 0.75)	0.64 (0.54, 0.75)	0.64 (0.54, 0.75)
25–44	1.82 (1.63, 2.03)	1.81 (1.62, 2.02)	1.81 (1.62, 2.02)
45–64	13.92 (12.59, 15.38)	13.89 (12.56, 15.36)	13.87 (12.53, 15.36)
≥65	119.46 (108.36, 131.70)	119.22 (108.09, 131.51)	118.99 (107.77, 131.36)
Gender
Female	1.00	1.00	1.00
Male	1.25 (1.11, 1.41)	1.25 (1.11, 1.41)	1.25 (1.11, 1.41)
Race/ethnicity
White	1.00	1.00	1.00
Black	1.96 (1.65, 2.33)	1.30 (1.08, 1.56)	0.72 (0.53, 0.98)
Other	1.47 (1.21, 1.78)	0.98 (0.80, 1.21)	0.54 (0.39, 0.74)
Interaction effect
Male, 15–24 y	2.56 (2.12, 3.09)	2.57 (2.12, 3.10)	2.57 (2.12, 3.11)
Black	0.95 (0.75, 1.22)	0.97 (0.75, 1.24)	0.97 (0.75, 1.24)
Other	0.63 (0.47, 0.84)	0.66 (0.48, 0.90)	0.64 (0.47, 0.87)
Male, 25–44 y	1.93 (1.69, 2.21)	1.94 (1.70, 2.22)	1.95 (1.70, 2.23)
Black	0.90 (0.74, 1.08)	0.92 (0.76, 1.11)	0.93 (0.77, 1.12)
Other	0.73 (0.59, 0.90)	0.78 (0.63, 0.97)	0.81 (0.65, 1.00)
Male,45–64 y	1.38 (1.22, 1.56)	1.38 (1.22, 1.57)	1.38 (1.22, 1.57)
Black	0.42 (0.35, 0.51)	0.42 (0.35, 0.50)	0.43 (0.36, 0.51)
Other	0.28 (0.22, 0.34)	0.30 (0.24, 0.37)	0.32 (0.25, 0.40)
Male, ≥65 y	1.04 (0.93, 1.18)	1.05 (0.93, 1.18)	1.05 (0.93, 1.18)
Black	0.22 (0.18, 0.26)	0.21 (0.18, 0.26)	0.22 (0.18, 0.26)
Other	0.12 (0.10, 0.15)	0.13 (0.11, 0.17)	0.14 (0.11, 0.17)
Male, Black	1.01 (0.93, 1.09)	1.02 (0.94, 1.11)	1.02 (0.94, 1.11)
Male, Other	1.08 (0.95, 1.23)	1.10 (0.96, 1.25)	1.10 (0.96, 1.25)
CT-level predictor
Main effect
5%–9.9% poverty	. . .	. . .	1.05 (1.00, 1.10)
10%–19.9% poverty	. . .	. . .	1.23 (1.16, 1.30)
20%–100% poverty	. . .	. . .	1.42 (1.33, 1.51)
Cross-level interaction effect (individual race and CT poverty)
White, 0%–4.9% poverty	. . .	. . .	1.00
Black, 5%–9.9% poverty	. . .	. . .	1.67 (1.24, 2.26)
Black, 10%–19.9% poverty	. . .	. . .	2.34 (1.76, 3.12)
Black, 20%–100% poverty	. . .	. . .	3.00 (2.28, 3.94)
Other, 5%–9.9% poverty	. . .	. . .	1.32 (0.95, 1.84)
Other,10%–19.9% poverty	. . .	. . .	2.26 (1.67, 3.07)
Other, 20%–100% poverty	. . .	. . .	3.37 (2.56, 4.45)

Note. OR = odds ratio, CI = confidence interval, CT = census tract. ORs for interaction terms are the unique (additional) effect of a particular variable and do not include the main effect.

TABLE 3—

Random Parameters at the Census Tract (CT) and Block Group (BG) Levels From Models 1, 2, and 3 Analyzing Neighborhood Heterogeneity in Associations Between Mortality, Race/Ethnicity, and Area Poverty: Massachusetts, 1989–1991

	Estimate (SE)
	Model 1	Model 2	Model 3
Between-CT variation
Constant/constant (σ2υ0)	0.095 (0.005)	0.085 (0.005)	0.066 (0.004)
Constant/Black (σ2υ0υ1)	. . . (. . .)	0.050 (0.015)	−0.017a (0.012)
Black/Black (σ2υ1)	. . . (. . .)	0.337 (0.053)	0.158 (0.034)
Constant/other (σ2υ0υ2)	. . . (. . .)	0.054 (0.016)	−0.034 (0.013)
Black/other (σ2υ1υ2)	. . . (. . .)	0.136 (0.047)	0.028a (0.030)
Other/other (σ2υ2)	. . . (. . .)	0.325 (0.064)	0.120 (0.040)
Between-BG variation
Cell-level dispersion	1.399 (0.007)	1.405 (0.007)	1.442 (0.007)
Constant/constant (σ2u0)	0.111 (0.004)	0.111 (0.004)	0.1083 (0.004)

Note. The parameter σ²_υ0 represents the variance for the base category (Whites), whereas σ²_υ1 and σ²_υ2 represent the differential variance for Blacks and others, respectively. The parameters σ_υ0υ1 σ_υ0υ2, and σ_υ1υ2 present the covariance associated with the random variables, υ_0k,υ_1k, and υ_2k, associated with the race/ethnicity categories. The random parameters were tested with “Wald-like” tests,27,31 and P values were based on a χ² distribution. Tests were conducted for both individual random parameters and the entire random part at the CT level.

^a These terms were not significant at P ≤ .001; all other terms were.

Model 2 (Table 3 ▶) allows an assessment of the neighborhood heterogeneity (across CTs) in racial/ethnic disparities in mortality. We found a statistically significant CT-level variation (P< .001) in the individual relationship between mortality and race/ethnicity. The between-CT variation in mortality was substantially greater for Blacks (0.524) than for Whites (0.085; Table 4 ▶). Accounting for this substantial CT-level heterogeneity in mortality by race/ ethnicity reduced the fixed mortality differential for Blacks (aged 0–14 years) from a nearly 2-fold odds ratio (Table 2 ▶, model 1) to an odds ratio of 1.30 (95% CI=1.08, 1.56). Understandably, it did not attenuate any of the associated interaction effects.

TABLE 4—

Variation (on Logit Scale) in Mortality Between Census Tracts (CTs) by Race/Ethnicity, Before and After Taking Into Account CT-Level Poverty: Massachusetts, 1989–1991

	White	Black	Other
Before taking account of poverty	0.08	0.52	0.52
After taking account of poverty	0.07	0.19	0.12

Note. Estimates for Whites relate to the CT-level random parameter estimate associated with σ²_υ0 in Models 2 and 3 (see Table 3 ▶). Estimates for Blacks were not estimated directly but were based on the sum of the random parameter variance related to Whites (σ²_υ0), the differential variance associated with Blacks (σ²_υ1), and 2 times the covariance between these 2 variances (σ_υ0υ1 ) in Models 2 and 3. Estimates for “others” were not estimated directly but were based on the sum of the random parameter variance related to Whites (σ²_υ0) the differential variance associated with others (σ²_υ2) and 2 times the covariance between these 2 variances (σ_υ0υ2) in Models 2 and 3.

Results from model 3 (Table 3 ▶) supported the hypothesis that CT poverty accounts for the racial/ethnic-specific heterogeneity in mortality at the CT level. While between-CT variances in mortality declined only slightly for Whites (from 0.08 in model 2 to 0.07 in model 3), CT poverty accounted for 63% of the CT-level mortality variation for Blacks (from 0.52 in model 2 to 0.19 in model 3; Table 4 ▶).

The odds ratio for mortality increased with neighborhood poverty, and the relationship was substantially stronger for Blacks than for Whites (Table 2 ▶, model 3). When Whites living in the 3 higher CT-poverty strata were compared with the reference group (Whites living in CTs with lower than 5% poverty), their odds ratios were 1.05 (CTs with 5%–9.9% poverty), 1.23 (CTs with 10%–19.9% poverty), and 1.42 (CTs with 20% or higher poverty). For Blacks, however, compared with the same reference group (Whites living in CTs with lower than 5% poverty), the odds ratios for the unique interactions were, respectively, 1.67 (CTs with 5%–9.9% poverty), 2.34 (CTs with 10%–19.9% poverty), and 3.00 (CTs with 20% or higher poverty).

DISCUSSION

We found, first, that between-CT variation in mortality was some 6 times greater for Blacks than for Whites. Second, neighborhood poverty contributed substantially to the observed area variations in Black excess mortality. Indeed, if we consider the estimated variation between CTs as a “true” estimate of race/ ethnicity–specific contextual heterogeneity, then the mortality odds ratio for Blacks compared with Whites can range from 0.31 to 5.36 (with the “average” Black–White disparity 1.30). While the existing literature is conceptually (and to a large extent, empirically) rich in descriptions of average racial/ethnic differences in mortality and possible explanations for such differences,11,33,34 there has been little documentation—let alone investigation—of why the racial/ethnic disparities are much greater in some neighborhoods than others.

Our findings suggest a need to conceptualize contextual effects (e.g., neighborhood poverty) in more complex ways. Typically, the examination of contextual effects has proceeded with an assumption of “main contextual effects” (e.g., that neighborhood poverty affects individual mortality in a similar manner for all racial/ethnic groups). However, it is entirely reasonable (and perhaps more realistic) to anticipate that contextual differences as well as contextual effects inherently interact with individual characteristics. Our finding that neighborhood-level poverty directly contributes to the greater geographic heterogeneity in mortality rates for Blacks suggests that the consequences of neighborhood deprivation may be particularly exacerbated for Blacks, compared with Whites. That is, there is a joint and synergistic shaping of population patterns of mortality by individual (e.g., race/ethnicity) and contextual (e.g., CT poverty) factors.

The following caveats should be considered in interpreting the empirical findings of our study. First, there is a distinct possibility of a potential misspecification of the individual-level factors that are associated with mortality. Our specification of individual variables was constrained by the degree of cross-tabulation (i.e., by age, race/ethnicity, and gender) that was possible on the numerator and denominator populations simultaneously at the BG level. This constraint precluded estimation of the impact of other important (unmeasured) socioeconomic markers on clustering in mortality.

Had we constructed our cells by county, it would have been possible to include educational data at the individual level (thus increasing the “social resolution” at the individual level), but this would have come at the cost of ignoring the critical level of BGs and CTs (thus decreasing the “spatial resolution”). While it is difficult to evaluate the sensitivity of our findings, given the lack of data, it is likely that including additional socioeconomic markers might have attenuated both the magnitude of racial/ethnic disparities in mortality and the effect of CT poverty. It is unlikely, however, that such data would have accounted for the contextual heterogeneity in racial/ethnic disparities in mortality. For this to have happened, the omitted individual socioeconomic markers not only would have to be to perfectly collinear with individual race/ ethnicity but would also need to have exactly the same spatial clustering as race/ethnicity.

It is, however, worth noting that neighborhood effects continue to be thought of as an extension of individual effects on health, as is reflected in the overbearing concern for “controlling” for individual confounders. Such an approach, arguably, obscures the need to reconceptualize the very notion of “individual effects” that results from incorporating neighborhoods in our conceptual model to elucidate health disparities. The differential response to neighborhood environments by individual Blacks and Whites, demonstrated in our analysis, is a simple illustration of this point.

A second limitation is our focus on fixed, discrete, and hierarchical census-defined contexts to define neighborhoods. While these are important, there may be other nonspatial contexts (e.g., households, nongeographic communities) and nonhierarchical contexts (e.g., workplaces and subjectively defined neighborhoods) that are also important to understanding the patterning of mortality. Thus, in addition to the issue of missing covariates, the issue of missing “levels” remains, especially in multilevel models.

Third, our characterization of neighborhoods was based on a single dimension—poverty. Future researchers might consider systematically developing a “typology” of neighborhoods to take into account the complex interaction of multiple characteristics.

A fourth concern is that it is problematic to demonstrate a “neighborhood effect” on mortality on the basis of cross-sectional observational data, since the magnitude of both neighborhood heterogeneity and neighborhood poverty is for individuals at the time of their death; therefore, the issue of potential mobility (social as well as spatial) over the life course is ignored.

Fifth, the partitioning of variation by different levels (e.g., individual, BG, and CT) in logistic models is not straightforward.35 Specifically, the magnitude of variance at the BG and CT levels cannot be evaluated for its size in relation to the individual-level variance, since the latter is a known function in logistic models. Thus, the magnitude of CT-level contextual heterogeneity in mortality by race/ ethnicity reported here is simply a conditional (not absolute or relative) estimate, that is, conditional on the fixed part and the random part of the model at the BG level.

Finally, our findings can be generalized only to populations (of neighborhoods and individuals) that are similar in their characteristics to those in Massachusetts. The replicability of our results in other states or geographic settings is an empirical question warranting further research.

Despite these challenges, quantifying and studying the distribution of mortality across different local geographic units disaggregated by key sociodemographic and economic markers may provide important input into social and health policy as well as sharpen etiological analysis of variations in mortality. For example, our results suggest that while the routine monitoring of how population groups are doing needs to be area-specific, the monitoring of how areas are doing (a prerequisite for developing area-based public health interventions) needs to be population group specific. Such multilevel thinking has yet to permeate the research geared toward public health practice. Such considerations, we believe, may lead to a fairer, and clearly more sensitive and realistic, means of target-setting and evaluation for population groups and areas. If we rely only on across-the-board, average targets for areas, it is difficult to distinguish improvements consisting of changes only in groups that are already “better-off” in those areas. The analytical approach developed here offers a framework for understanding such public health issues.

In conclusion, our analysis points to the importance of generating a quantitative description of the varying impact of neighborhood on health, in relation to diverse population subgroups, rather than focusing solely on average differences in health disparities across these groups. Extending this approach to investigating health patterns over time could provide opportunities to quantify the impact of various social policies on health inequalities between population groups and areas. Indeed, the complex geographic and social variation in important health outcomes such as mortality is an intrinsically important attribute of society that needs to be routinely described and understood.