Disagreements in statistical inferences from the randomeffects metaanalyses model based on different betweenstudy variance estimators
*Kristian Thorlund, Copenhagen trial
Keywords: metaanalysis, betweenstudy variance estimator, statistical inferences
Previous studies found inconsistencies between available betweenstudy variance estimators in metaanalyses, but implications to metaanalytical inference in practice are limited since they only measure bias and mean square errors of the estimators. Metaanalytical inferences are chiefly based on treatment effect estimates, confidence intervals, and pvalues. We compared metaanalytical inferences from three selected estimators by reanalyzing 297 heterogeneous metaanalyses. Disagreements in estimates of treatment effects and confidence intervals were analyzed pair wise using mixed model linear regression. Disagreements in pvalues were determined with respect to the 5% and 1% threshold. Large systematic disagreements in treatment estimates and standard errors were found. Pvalues were similar. The findings provide useful guidelines for interpretation of sensitivity analysis.
