Social Relationships and Mortality Risk: A Meta-analytic Review

Identification of Studies

To identify published and unpublished studies of the association between social relationships and mortality, we used three techniques. First, we conducted searches of studies from January 1900 to January 2007 using several electronic databases: Dissertation Abstracts, HealthSTAR, Medline, Mental Health Abstracts, PsycINFO, Social Sciences Abstracts, Sociological Abstracts via SocioFile, Academic Search Premier, ERIC, and Family & Society Studies Worldwide. To capture the broadest possible sample of relevant articles, we used multiple search terms, including mortality, death, decease(d), died, dead, and remain(ed) alive, which were crossed with search words related to social relationships, including the terms social and interpersonal linked to the following words: support, network, integration, participation, cohesion, relationship, capital, and isolation To reduce inadvertent omissions, we searched databases yielding the most citations (Medline, PsycINFO) two additional times. Next, we manually examined the reference sections of past reviews and of studies meeting the inclusion criteria to locate articles not identified in the database searches. Finally, we sent solicitation letters to authors who had published three or more articles on the topic.

Inclusion Criteria

We included in the meta-analysis studies that provided quantitative data regarding individuals' mortality as a function of social relationships, including both structural and functional aspects. Because we were interested in the impact of social relationships on disease, we excluded studies in which mortality was a result of suicide or injury. We also excluded studies in which the only measurement of social support was an intervention provided within the context of the study (e.g., support group), the source of social support was nonhuman (e.g., a pet or higher power), or the social support was provided to others (i.e., giving support to others or measures of others' benefit from the support provided) rather than to the individual tracked for mortality status. We coded studies that included participant marital status as one of several indicators of social support, but we excluded studies in which marital status was the only indicator of social support. We also excluded studies in which the outcome was not explicitly and solely mortality (e.g., combined outcomes of morbidity/mortality). Reports with exclusively aggregated data (e.g., census-level statistics) were also excluded. Manuscripts coded were all written in English, which accounted for 98% of the total retrieved. See Figure 1 for additional details.

Data Abstraction

To increase the accuracy of coding and data entry, each article was initially coded by two raters. Subsequently, the same article was independently coded by two additional raters. Coders extracted several objectively verifiable characteristics of the studies: (a) the number of participants and their composition by age, gender, marital status, distress level, health status, and pre-existing health conditions (if any), as well as the percentage of smokers and percentage of physically active individuals, and, of course, the cause of mortality; (b) the length of follow up; (c) the research design; and (d) the aspect of social relationships evaluated.

Data within studies were often reported in terms of odds ratios (ORs), the likelihood of mortality across distinct levels of social relationships. Because OR values cannot be meaningfully aggregated, all effect sizes reported within studies were transformed to the natural log OR (lnOR) for analyses and then transformed back to OR for interpretation. When effect size data were reported in any metric other than OR or lnOR, we transformed those values using statistical software programs and macros (e.g., Comprehensive Meta-Analysis). In some cases when direct statistical transformation proved impossible, we calculated the corresponding effect sizes from frequency data in matrices of mortality status by social relationship status. When frequency data were not reported, we recovered the cell probabilities from the reported ratio and marginal probabilities. When survival analyses (i.e., hazard ratios) were reported, we calculated the effect size from the associated level of statistical significance, often derived from 95% confidence intervals (CIs). Across all studies we assigned OR values less than 1.00 to data indicative of increased mortality and OR values greater than 1.00 to data indicative of decreased mortality for individuals with relatively higher levels of social relationships.

When multiple effect sizes were reported within a study at the same point in time (e.g., across different measures of social relationships), we averaged the several values (weighted by standard error) to avoid violating the assumption of independent samples. In such cases, the aggregate standard error value for the lnOR were estimated on the basis of the total frequency data without adjustment for possible correlation among the averaged values. Although this method was imprecise, the manuscripts included in the meta-analysis did not report the information necessary to make the statistical adjustments, and we decided not to impute values given the wide range possible. In analyzing the data we used the shifting units of analysis approach which minimizes the threat of nonindependence in the data while at the same time allowing more detailed follow-up analyses to be conducted (i.e., examination of effect size heterogeneity).

When multiple reports contained data from the same participants (publications of the same database), we selected the report containing the whole sample and eliminated reports of subsamples. When multiple reports contained the same whole sample, we selected the one with the longest follow-up duration. When multiple reports with the same whole sample were of the same duration, we selected the one reporting the greatest number of measures of social relationships.

In cases where multiple effect sizes were reported across different levels of social relationships (i.e., high versus medium, medium versus low), we extracted the value with the greatest contrast (i.e., high versus low). When a study contained multiple effect sizes across time, we extracted the data from the longest follow-up period. If a study used statistical controls in calculating an effect size, we extracted the data from the model utilizing the fewest statistical controls so as to remain as consistent as possible across studies (and we recorded the type and number of covariates used within each study to run post hoc comparative analyses). We coded the research design used rather than estimate risk of individual study bias. The coding protocol is available from the authors.

The majority of information obtained from the studies was extracted verbatim from the reports. As a result, the inter-rater agreement was quite high for categorical variables (mean Cohen's kappa = 0.73, SD = 0.13) and for continuous variables (mean intraclass correlation = 0.80, SD =.14). Discrepancies across coding pairs were resolved through further scrutiny of the manuscript until consensus was obtained.

Aggregate effect sizes were calculated using random effects models following confirmation of heterogeneity. A random effects approach produces results that generalize beyond the sample of studies actually reviewed. The assumptions made in this meta-analysis clearly warrant this method: The belief that certain variables serve as moderators of the observed association between social relationships and mortality implies that the studies reviewed will estimate different population effect sizes. Random effects models take such between-studies variation into account, whereas fixed effects models do not. In each analysis conducted, we examined the remaining variance to confirm that random effects models were appropriate.