Article type
Year
Abstract
Objectives: To explore the influence of various methods for handling missing data on the effect estimates of an individual patient data meta-analysis, i.e 1) complete case analyses, 2) single imputation within trials, 3) single imputation over trials, 4) multiple imputation within trials, 5) multiple imputation over trials
Methods: In an existing individual patient data meta-analysis on the effectiveness of antibiotics in children with acute otitis media, the results of the interaction tests (beta-coefficients, standard errors, and p-values) and stratified subgroup analyses (rate ratios, rate differences, and their 95% confidence intervals) were compared for the five methods of handling missing data.
Results: The results of the interaction test differed between the five methods, in particular between imputation within and between trials. Significant interaction effects were found with complete case analyses, and imputation within trials, whereas imputation between trials often showed no significant interaction effect. Finally, no relevant differences were found in the stratified analyses between the five methods.
Conclusion: Imputation between trials might lead to bias, because associated variables within trials can determine the imputation of missing data between trials. In our empirical example, imputation within trials seems to be better in individual patient data meta-analyses.
Methods: In an existing individual patient data meta-analysis on the effectiveness of antibiotics in children with acute otitis media, the results of the interaction tests (beta-coefficients, standard errors, and p-values) and stratified subgroup analyses (rate ratios, rate differences, and their 95% confidence intervals) were compared for the five methods of handling missing data.
Results: The results of the interaction test differed between the five methods, in particular between imputation within and between trials. Significant interaction effects were found with complete case analyses, and imputation within trials, whereas imputation between trials often showed no significant interaction effect. Finally, no relevant differences were found in the stratified analyses between the five methods.
Conclusion: Imputation between trials might lead to bias, because associated variables within trials can determine the imputation of missing data between trials. In our empirical example, imputation within trials seems to be better in individual patient data meta-analyses.