The Effects of the Choice of Meta Analysis Model on the Overall Estimates for Continuous Data with Missing Standard Deviations

Abstract

The choice between the fixed and random effects model for providing an overall meta analysis estimate in continuous data may affect the accuracy of these estimates. For studies with complete information, the Cochrane’s Q-test could provide some guide on the choice, although the power of this test is quite low. If the study- level standard deviations (SDs) are not completely reported or “missing”, selection of meta analysis model should be done with more caution. Many studies suggest that imputation is a good way of recovering the lost information in the effect size estimate and the corresponding standard error. In this article, we compare empirically, the effects of imputation of the missing SDs on the overall meta analysis estimates based on both the fixed and random effect model. The results suggest imputation is recommended to estimate the overall effect size. However, to estimate its corresponding standard error (SE), imputation is recommended for the estimates based on the random effect model. If the fixed effect model is used, imputation may lead to bias estimates of the SE.