Aim: 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.
Methods and results: 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.
Conclusion: 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.