The robust error meta-regression method for dose-response meta-analysis

Dose-response meta-analysis has been widely employed in evidence-based decision-making. Currently, the most popular approach is the one or two-stage generalized least squares for trend model. This approach however has some drawbacks, and therefore, we compare the latter with a one-stage robust error meta-regression (REMR) model, based on inverse variance weighted least squares regression and cluster robust error variances for dealing with the synthesis of correlated dose-response data from different studies.

We apply both methods to three examples (alcohol and lung cancer, alcohol and colorectal cancer, and BMI and renal cancer). The analysis of the three datasets reveals that the one-stage REMR approach may result in better error estimation and a better visual fit to the data than the generalized least squares approach with the added benefit of not needing to impute covariances from the data.

The one-stage REMR approach is easily executed in Stata with codes given in this article. We therefore recommend that REMR models be considered for dose-response meta-analysis and suggest further comparison of these two methods in future studies to conclusively determine the benefits and pitfalls of each.