Background: Individual participant data meta-analysis (IPDMA) is considered to be the gold standard in epidemiologic research. When IPDMA are affected by missing data, several strategies exist to obtain summary statistics.
Objectives: To compare the possible strategies for conducting IPDMA in the presence of missing data.
Methods: We first conducted a simulation study to compare various strategies for meta-analyzing study results (through one-stage or two-stage meta-analysis) and dealing with missing values (through complete case analysis, within-study imputation, stratified imputation or hierarchical imputation). By doing this we evaluated the bias and coverage of pooled study results, as well as the bias of estimated between-study heterogeneity. We then illustrated the implementation of each strategy in an empirical example where we meta-analyzed the predictive value of C-reactive protein in diagnosing community acquired pneumonia. Finally, we provide recommendations on the implementation of imputation and meta-analysis models in an IPDMA.
Results: We found that stratified imputation was most problematic in terms of bias and coverage. Although complete case analysis and within-study imputation performed adequately, the best results were obtained by hierarchical imputation. When summarizing the study results, one-stage and two-stage meta-analysis methods performed roughly the same. Finally, we found that recent recommendations on the order of combining imputed datasets in a two-stage IPDMA were detrimental, and that the reverse ordering was more appropriate.
Conclusions: We recommend hierarchical imputation followed by one-stage meta-analysis in an IPDMA with missing data, rather than analyzing each dataset separately or including dummy variables to adjust for potential between-study heterogeneity. Two-stage meta-analysis with within-study imputation is a viable alternative when sharing of IPD is difficult, e.g. due to confidentiality agreements. Each of the imputed datasets should first be meta-analyzed, and the resulting estimates should then be combined using Rubin's rule.