Evaluating Confounding Bias when Designing Meta-analyses of Dietary Risk Factors with Weak Associations: a Systematic Review of Risk Factors for Type 2 Diabetes

INTRODUCTION: Planning meta-analyses of weakly associated dietary factors is complicated by the fact that epidemiological studies often measure diet only at baseline and have limited capacities to control for known confounders in studies of long duration.  Because weak associations are sensitive to uncontrolled confounding, we present a systematic method for identifying and rank ordering potential confounders of dietary risk factors and type 2 diabetes (T2D) prior to meta-analysis.

METHODS: A systematic review of meta-analyses of prospective studies of T2D identified the strength of association (including dose-response relationships) and usual practices for modeling risk factors in the primary studies.

RESULTS: 45 risk factors for T2D were identified from 47 published meta-analyses. Body size was the strongest risk factor (RR pooled of obesity = 7.3 and  RR pooled of overweight = 2.9). In addition to body size, meta-analyses adjusted for lifestyle covariates with established dose-response relationships: e.g. smoking (RR pooled=1.5-1.6), physical inactivity (RR pooled =1.4-1.5), and moderate alcohol consumption (RR~0.60-0.87).

Dietary risks of T2D were the focus of 28 meta-analyses revealing weak associations with RRs < 1.5, including diets high in processed meats (RR pooled =1.2-1.5) and eggs (RR pooled =1.42).  All meta-analyses of diet and T2D combined risk estimates from studies that adjusted for body size (including macronutrients fructose, sucrose and dietary sugars) except for a meta-analysis of sugar-sweetened beverages (RR pooled  = 1.26).

CONCLUSIONS: Because T2D risk is associated with many dietary factors that are considerably weaker than lifestyle factors, the relative strength of known confounders should be documented prior to summarization.  Current practice does not include this critically important step.  As a result, risk estimates from current meta-analyses may provide a precise estimate of average confounded effects rather than accurate estimates of effect.