Our results demonstrate how a multilevel model-based framework can be used to explore the social geography of premature mortality in an urban setting based on routinely collected public health surveillance data. The multilevel framework allows us to partition observed variation in premature mortality at the neighborhood, CT, and within-CT levels and to model this variation as a function of fixed and random effects. As a result, this framework allows for: (a) statistical smoothing in the estimation of small area rates; (b) estimation of variance at each of the specified levels; and (c) adjustment for multiple covariates. We have also presented a set of measures and corresponding 95% credible intervals based on Monte Carlo simulation, including model-based direct standardized rates, standardized premature mortality ratios, and population attributable fraction that can be easily calculated based on the model and mapped at the CT and neighborhood levels.
Together, our results provide evidence of the role of CT poverty in patterning premature mortality within and across neighborhoods. In the most afflicted areas, whether neighborhoods or CTs, for every 100 premature deaths, 25 to 30 of the deaths would not have occurred if residents of these areas had the same age-specific premature mortality rates as residents of the city’s least impoverished neighborhoods or CTs. To our knowledge, the population attributable fraction of premature mortality in relation to CT poverty has not previously been reported in the U.S. at either the neighborhood or CT level.
Before providing interpretation of additional aspects of our findings, however, we note a set of caveats relevant to interpretation of the causal effect of poverty on premature mortality. These include ecologic bias, the modifiable areal unit problem, etiologic period, and omitted variable bias.
Ecologic bias can occur when both the dependent and independent variables are group-level data, and confounding is introduced through the grouping process.44,45
This can lead to fallacious inferences, particularly when the confounded group-level effect is used as a proxy for the corresponding individual level effect. We are careful to note, therefore, that our estimates of the CT poverty gradient are to be interpreted as a group-level association that likely reflects a complex combination of individual-level and area-level processes. However, we further note that the CT poverty effect is in the expected direction and is unlikely to reflect the sort of substantial cross-level confounding that can lead to a change in the direction of the association between poverty and premature mortality. Furthermore, in previous analyses for which we had individual as well as CT-level socioeconomic data, we have found that the direction of the socioeconomic gradient was the same, and also of similar magnitude, for both the individual-level and CT-level socioeconomic measures.11,21,23
The modifiable areal unit problem pertains to the sensitivity of spatial patterns to the choice of area level units in the analysis.43
The multilevel approach we have used addresses one aspect of this problem, as it permits modeling of spatial variation in premature mortality at each of the three levels (age cell, CT, and neighborhood) at which we have data. Another aspect concerns whether or not CTs and neighborhoods are meaningful entities (rather than “arbitrary” spatial units) relevant to characterizing socioeconomic conditions and to shaping population health. We note that CT boundaries are drawn to be socially meaningful by local census committees46
and that CTs serve as administrative areas relevant to resource allocation and for such programs as “Urban Empowerment Zones,” “Medically Underserved Areas,” and “Qualified Census Tracts” for the purpose of the low-income housing tax credit.15,47,48
Moreover, in Boston, the Boston Public Health Commission uses CTs to define Boston neighborhoods for health programs, thereby having real-life implications for the health and quality of life of their residents.
Regarding etiologic period, in the case of premature mortality, CT characteristics at the time of death are likely to be temporally relevant to outcomes with a short etiologic period. These would include deaths among infants and children, as well as deaths among adults due to “preventable and immediate causes” or for which survival is drastically curtailed by inadequate access to health care.6
Examples of the latter include deaths due to motor vehicle and other accidents; asthma, pneumonia and influenza; suicide, homicide and legal intervention; complications of pregnancy, diabetes, tuberculosis, and HIV/AIDS; as well as deaths due to temporally proximate triggers of heart attacks, e.g., pollution.24
In 2000, these causes of death accounted for nearly 40% of potential years of life lost before age 75 in the U.S.49
Finally, regarding omitted variable bias, we note that while there are certainly covariates (e.g., smoking) that may also account for the observed disparities in premature mortality by CT poverty, these are likely to be in the causal pathways linking socioeconomic position to premature mortality, such that it may not be appropriate to control for them when the population impact of CT poverty is of interest.44
We also emphasize that the primary aim of our analyses was to model the poverty–mortality relationship from a monitoring perspective, rather than engage in an exhaustive etiologic investigation of all potential determinants of premature mortality.
Assuming, then, that our findings present a plausible depiction of the patterning of premature mortality in Boston in relation to CT poverty, we offer some preliminary interpretation of additional aspects of our findings. First, our analysis confirms previously published observations of a socioeconomic gradient in premature mortality in the U.S., as reported in both the handful of other analyses using CT poverty,7,14,23
as well as those employing economic indicators based on larger geographic areas (e.g., neighborhoods, metropolitan areas, and counties).4–6
Only two of these prior studies used a multilevel analysis7,23
; the rest used the more conventional approach of comparing rates based on aggregating the death and population data into strata defined by area-based socioeconomic position. Although this latter approach, by ignoring spatial variation, could potentially yield biased estimates, we note in our study that this did not occur, as shown by the similarity of the socioeconomic gradients estimated in Tables and .
Interestingly, our estimates for Boston of a rate ratio of ~1.4, comparing premature mortality in census tracts with 20–100% poverty to those with <5% poverty, was equal to the analogous rate ratio we computed for Boston for 1988–1992, i.e., 1.4 (95% CI 1.3–1.5),50
and less than the analogous rate ratio of ~2.2 we computed for this same time period for Massachusetts as a whole.14
The smaller magnitude of the CT socioeconomic gradient for premature mortality in Boston compared to MA overall is likely due to their differing poverty levels. In 2000, the proportion of the population living in CTs with ≥20% below poverty (42%) was 2.5 times greater than for the state as a whole (16%). Given well-established patterns of wealthier suburban areas surrounding cities with more concentrated poverty,51,52
it is possible that rates of premature mortality in Boston’s least impoverished neighborhoods may nevertheless be higher than in their MA counterparts outside of Boston. If true, the net impact would be to reduce the Boston range of premature mortality associated with CT poverty level.
Second, we found that while adjusting for CT poverty overall tended to attenuate the estimated premature mortality ratio in most neighborhoods and CTs towards the null, as expected, in some cases it shifted the estimated premature mortality ratios away from the null. Thus, South Boston had an unusually high premature mortality rate despite its low poverty rate, whereas Allston/Brighton had an unusually low premature mortality rate despite its relatively high poverty rate.
Recognizing that interpretation of these anomalous results requires local knowledge of the social geography of Boston neighborhoods, we drew on the accumulated knowledge and experience of our local and state public health department partners to propose possible explanations. For the unusually high residual premature mortality rates in South Boston, hypotheses include: (1) a spike in heroin-related overdose deaths during the study period; (2) a high overall rate of mortality due to substance use, including both drugs and alcohol; (3) excess mortality due to lung cancer and cardiovascular disease, likely reflecting high rates of smoking; and (4) the residential concentration of relatively well-paid police officers and firefighters, who may experience high premature mortality but do not live in impoverished CTs.53
Conversely, Allston/Brighton is a neighborhood chiefly populated by students, who may not be gainfully employed but who do not experience high mortality. These interpretations, which point to the difference of student vs. family poverty and also the existence of relatively well-paid but risky jobs, are conjectural and require further study as possible explanations for the observed patterns. They also underscore the need for thoughtful interpretation of the relationship between CT poverty and risk of premature mortality, even as the overall findings clearly demonstrate that overall increasing CT poverty levels are associated with increased risk of premature mortality.
Lastly, an additional contribution of our study is to improve visual representation and interpretation of the social geography of premature mortality in relation to CT poverty. Previous literature on statistical mapping has noted that maps of standardized incidence ratios are interpretable only relative to the mean rate across the study population and, unlike directly standardized rates, are not readily comparable to externally published rates or policy relevant targets.54
On the other hand, traditional direct standardization tends to perform poorly in small areas because it requires stable age-specific rates. At the small area level (e.g., CTs), this is made further problematic by sparseness of the data and “empty” age cells in the denominators. Moreover, traditional direct standardization does not incorporate any smoothing at the area-level.
As an alternative to the standardized premature mortality ratio map, we have proposed a model-based direct standardized rate that overcomes the problems with traditional direct standardization while permitting comparison with published standardized rates (e.g., the Healthy People 2010 objectives).55
Our small area rate estimates incorporate model-based smoothing and are robust to sparse age cells. Moreover, as we show in Fig. a and b, the method can be extended to yield predicted rates given a “set” level for the covariate CT poverty. This is a useful tool, for example, if one wishes to predict the effect of an intervention by “setting” the covariate level equal to some target in the model-based calculation.56
Nevertheless, we note that care should be taken in interpreting the estimated population attributable fraction as an “etiological” fraction, since this requires strong assumptions about the causal relationships between area socioeconomic conditions, premature mortality, and other unmeasured covariates. For this reason, we stress that the population attributable fraction should simply be regarded as an “excess” fraction.57
We also note that, in the current analysis, Model 1a has assumed a set of consistent age effects across all areas. Therefore, the spatial patterning of the standardized premature mortality ratios and model-based direct standardized rates at the CT level is the same, although the former are centered around the “null” ratio of 1.0, rather than expressed on the rate scale. This corresponds to the condition of “no age-area interactions” under which indirectly and directly standardized rates are identical.54
However, in our model-based framework, this assumption could be relaxed to allow age x
area interactions (a “random slopes” model for age), or even age x
covariate interactions that would be reflected in the model-based standardized rate, thereby yielding estimates that would be more similar to the purely non-parametric directly standardized rates than the indirectly standardized rates. In future analyses, we plan to explore these alternative modeling strategies and also investigate whether the poverty–premature mortality relationship varies by neighborhoods, race/ethnicity, gender, and specific causes of death and also whether it exhibits any non-linearities.
In conclusion, we believe that wider use of our proposed methodology can enable local and state health departments, as well as academic researchers, to use routinely available data to generate meaningful, policy-relevant analyses and depictions of socioeconomic inequities in premature mortality. As recognized in the U.S.,8
and the United Kingdom,59–61
the premature mortality rate is a highly informative, easily calculated, and easily understood single measure that captures social disparities in community health. When considered in the geographical context, it also has enormous potential for identifying acute, small area clusters especially burdened by premature death and encouraging exploration of the specific correlates of area deprivation (including environmental conditions, housing, education, material resources, behavioral factors, access to care, etc.) that contribute to social disparities in health.