Adjusting Adjustments: Using External Data to Estimate the Impact of Different Confounder Sets on Published Associations

Background: A 2013 meta-analysis observed a protective association between overweight body mass index (BMI) (vs. normal BMI) and all-cause mortality that was particularly strong in people aged ≥65. Estimates informing this meta-analysis were highly heterogeneous, and critics raised insufficient or inappropriate confounder adjustment in many studies as an explanation for the protective summary association. Using this topic as an example, we demonstrate a novel approach for external adjustment of individual studies for a uniform and sufficient confounder set before meta-analysis. Methods: We abstracted summary data on the 33 associations comprising the age ≥65 stratum of the 2013 meta-analysis. Using an external dataset (NHANES III), we derived covariates used in each study’s multivariable model of the overweight–mortality association. We then calculated a bias factor to quantify the direction and magnitude of displacement of the ratio measure of association after changing from the original adjustment set to a sufficient adjustment set. After applying bias factors to adjust original associations, we compared summary results from random-effects meta-analyses with and without such adjustment. Results: We reproduced the original meta-analysis of overweight–mortality estimates among older participants and found a protective association similar to that reported in 2013 (summary relative risk = 0.88; 95% confidence interval: 0.84, 0.92; I 2 = 38.4%). After we simulated uniform adjustment of all 33 associations for a minimally sufficient confounder set (age, sex, and smoking status), the meta-analysis showed a similar summary association (summary relative risk = 0.90; 95% confidence interval: 0.86, 0.94), but with reduced heterogeneity (I 2 = 34.6%). Conclusion: Simulated uniform adjustment for a sufficient confounder set may improve rigor and promote consensus in meta-analysis.