Estimated Deaths Attributable to Social Factors in the United States

Sandro Galea, MD, DrPH; Melissa Tracy, MPH; Katherine J. Hoggatt, PhD; Charles DiMaggio, PhD; Adam Karpati, MD, MPH


Am J Public Health. 2011;101(8):1456-1465. 

In This Article


To calculate the number of deaths attributable to social factors in the United States, we first estimated the relative risk (RR) of mortality associated with each social factor and obtained an estimate of the prevalence of each social factor in the United States. These estimates were used to calculate the population-attributable fraction of mortality for each factor, which was multiplied by the total number of deaths in the United States in 2000 to estimate the number of deaths attributable to each social factor.

Relative Risk Estimates

We conducted a MEDLINE search for all English-language articles published between 1980 and 2007 with estimates of the relation between individual- and area-level social factors and adult all-cause mortality. Individual-level social factors included education, poverty, health insurance status, employment status and job stress, social support, racism or discrimination, housing conditions, and early childhood stressors. Area-level social factors included area-level poverty, income inequality, deteriorating built environment, racial segregation, crime and violence, social capital, and availability of open or green spaces. We identified these articles to extract RR estimates from independent samples that could be combined through meta-analysis to obtain summary RR estimates for the relations between each social factor and mortality.

The included studies presented data sufficient for calculating an estimate of the association between at least 1 of the social factors of interest and adult all-cause mortality, using unweighted counts, RR estimates, regression coefficients, or mortality rates. For studies that concerned multiple levels of analysis, we included only those that appropriately used multilevel analytic methods. Figure 1 summarizes the studies that we considered, included, and excluded. We excluded articles presenting results from studies conducted outside of the United States, those limited to participants with a history of disease or using composite measures of socioeconomic status, and those that did not use adult all-cause mortality as an outcome measure. We also excluded review articles, articles providing insufficient data to calculate RR estimates, and articles presenting data only for proxy measures of the social factors of interest. This left a total of 120 eligible studies.

Figure 1.

Flow diagram of studies considered for meta-analyses to derive summary relative risk (RR) estimates for each social factor in relation to mortality.

Further criteria for inclusion in the meta-analyses included the presentation of SE or other variance estimates to allow calculation of an approximate 95% confidence interval (CI) for a dichotomous contrast in the social factor of interest (e.g., low vs high educational attainment). Additionally, RR estimates unadjusted for potential mediators of the relation between the social factor and mortality were desired; however, this requirement was relaxed for estimates of the effect of area-level social factors since nearly all estimates in the literature were adjusted. Finally, we decided a priori to limit meta-analyses to social factors for which at least 3 RR estimates from separate studies were available.

We extracted RR estimates from the remaining 60 studies for the following factors: education, poverty, social support, area-level poverty, income inequality, and racial segregation. Because meta-analyses must be conducted on nonoverlapping samples,[34] we excluded an additional 13 articles because they provided only estimates for samples already represented by other articles. When multiple articles provided data for the same sample, we selected estimates incorporating the largest sample size, the longest duration of follow-up, and the fewest restrictions on the sample in terms of age group or gender; additionally, we preferred estimates incorporating person–time data from longitudinal studies. The final 47 studies used in the meta-analyses are summarized in Table 1.

From each of the 47 articles, we extracted unadjusted RR estimates if provided; otherwise, we calculated RR estimates using unweighted or weighted counts, regression coefficients, or mortality rates according to standard methods.[34] The cutpoints used for dichotomous contrasts for each social factor, which are summarized in Table 2, were based on definitions most commonly used in the included studies and the literature on these social factors more generally.

When possible, we extracted age-specific estimates for 2 broad age groups, those aged 25 to 64 years at baseline and those aged 65 years or older at baseline. Although some evidence suggests that the relation between the social factors of interest and mortality decreases with age,[19] most deaths occur among older individuals. Altogether, we extracted 68 estimates from the 47 articles, as summarized in Figure 1. We calculated summary statistics for each social factor using Comprehensive Meta-Analysis version 2 (Biostat, Englewood, NJ). We used random-effects models for all summary estimates, taking into account unmeasured heterogeneity in effect estimates across studies and allowing greater weight to be given to studies conducted on smaller samples than when using fixed-effects models.[34]

Prevalence Estimates and Mortality Data

Estimates of the prevalence of each social factor in the US adult population (aged ≥25 years) were obtained from the 2000 US Census, [75,77] except for the prevalence of low social support, which was obtained from the Third National Health and Nutrition Examination Survey (NHANES III).[76] To derive prevalence estimates, we used cutpoints as similar as possible to those used when calculating dichotomous contrasts for each of the included studies. Table 2 summarizes the definitions and sources of data used to obtain prevalence estimates for each social factor, and prevalence estimates for each factor are presented in Table 3.

We obtained the total number of deaths in 2000 from all causes by age group from the National Vital Statistics Report.[81] Because the average duration of follow-up for studies providing RR estimates for samples aged 25 to 64 years at baseline was 10 years, we included deaths among persons aged 25 to 74 years for this age group, which was similar to the method used by Hahn et al.[30]

Calculation of Population Attributable Fraction and Sensitivity Analyses

We calculated the population-attributable fraction (PAF ) for each social factor using the following formula:

where RR is the summary RR estimate for mortality derived from the meta-analyses described and p is the prevalence of the social factor in the US population in 2000. The population-attributable fraction represents the proportion of all deaths that can be attributed to the social factor (i.e., the proportion of all deaths that would not have occurred in the absence of the social factor).[19,82–84] The population- attributable fraction was then multiplied by the total number of deaths in the relevant age group to arrive at the number of deaths attributable to the social factor in that age group.

We conducted sensitivity analyses to assess the robustness of the summary RR estimate for each social factor using alternate cut points or multiple categories (e.g., tertiles) rather than a dichotomous contrast in exposure levels.


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.