The authors of this article would like to provide the following update regarding Figure 6:

As the authors of this study, we draw attention to three considerations regarding Figure 6 in this article, which compares our results to the findings of other meta-analyses. The comparison meta-analyses in Figure 6 reported data in terms of odds ratios and risk ratios, and it has come to our attention that:
1. Although for Figure 6 we transformed data reported in the metric of risk ratio to the natural log of the odds ratio for one of the comparison meta-analyses with a death rate higher than 10%, we did not transform risk ratio data for five of the comparison meta-analyses that appeared to have low death rates (rational for this approach explained below in Consideration A).
2. When looking at the data in the comparison meta-analyses, we identified errors made by the authors of the comparison meta-analyses, such that the values reported in Figure 6 need to be revised (Consideration B and C explained below).
We emphasize that these errors involve the comparison data used in Figure 6, not our own data. Consultation with the Executive Editor of PLOS Medicine confirmed that posting this public Comment is the appropriate step, since these issues concern comparison data and not the data in our study.

Consideration A: Transformation of comparison meta-analytic data from risk ratio data to odds ratio data
Figure 6 depicts the results from 9 comparison meta-analyses reported in 8 articles. Our results were reported as odds ratios, and six of the nine comparison studies reported risk ratio data. At the time we created Figure 6, other published literature suggested that “They [risk ratios, rate ratios, and odds ratios] are all approximately equal when the disease is rare during the follow-up period (e.g. a cumulative incidence of less than 10%)” (Pierce, 1993, p. 1190) and “when the incidence of an outcome of interest in the study population is low (<10%), the odds ratio is close to the risk ratio” (Zhang & Kai, 1998, p. 1690). Based on this, we only transformed risk ratio data to odds ratio data for the comparison meta-analysis on smoking (Shavelle et al., 2008), with death rates obviously exceeding 10%. However, we had not calculated precise death rates for all comparison meta-analyses, and while overall death rates may be low, the values of individual studies vary. Not transforming risk ratio data to odds ratio data resulted in imprecise estimates. As a solution, we have performed those transformations. A revised version of Figure 6 is available upon request.

Consideration B: Five of the six comparison meta-analyses ostensibly reported risk ratio data that mistakenly included hazard ratio data
As we began transforming individual effect sizes from risk ratio data to odds ratio data, we noticed that hazards ratio data were frequently mislabeled or misinterpreted as risk ratio data. Five of the six comparison meta-analyses were not actually based on risk ratio data but an inaccurate combination of risk ratio and hazard ratio data.
As a solution, we calculated a risk ratio or odds ratio when relevant data were found in the original article, and we calculated a new omnibus effect size including solely the data transformed to the natural log of the odds ratio. Since some studies did not report data that could be transformed to odds ratios, fewer effect sizes were included in the omnibus analyses for our revised Figure 6, which has larger 95% confidence intervals than those originally reported in our Figure 6.

Consideration C: Two comparison meta-analyses combined different values of the independent variable
Another error we observed in two of the six comparison meta-analyses was that effect sizes were combined across different values of the independent variables.
1. The comparison meta-analysis of BMI used the largest BMI group in every study as the contrast group with lean participants, but some individual studies reported BMI values in the overweight category while others reported values denoting obesity. Our original Figure 6 had used the authors’ indication of “lean vs. obese,” but after looking at the individual studies, we needed to remove the effect sizes based on participants who were overweight but not obese.
2. The comparison meta-analyses of alcohol consumption evaluated estimates of the largest number of drinks/day reported in studies, which varied from 1.5 to more than 6 (compared with abstainers). To make the outcomes consistent, we calculated new estimates based on the value of 5 or more drinks/day. Our re-calculation fixed the error, but since fewer studies were included in the omnibus analyses, the corresponding 95% confidence intervals were larger than those originally reported in Figure 6.

Clarification and Interpretation
Regardless of the errors noted above, our original conclusion about Figure 6 in the text of the 2010 publication remains consistent: the magnitude of the effect of social relationships is on par with other widely recognized risk factors for mortality. However, our revised Figure 6 (available upon request) resulted in larger confidence intervals for several estimates.

Given that all comparison estimates in our original Figure 6 are now over 15 years old, we invite scholars to consider more up-to-date estimates when comparing the effects of social relationships on mortality with other risk factors for mortality.

Julianne Holt-Lunstad and Timothy B. Smith (authors)