Missing variability in meta-analysis : is imputing always good?
This paper examines the implications of the present approaches in handling missing variability in meta analysis on the overall standard error (SE) of the estimate. The approaches are (1) exclusion of the studies with missing standard deviations (SDs) and (2) imputation of the missing SDs. The d...
| Main Authors: | , |
|---|---|
| Format: | Conference or Workshop Item |
| Language: | English |
| Published: |
2006
|
| Subjects: | |
| Online Access: | http://irep.iium.edu.my/5555/ http://irep.iium.edu.my/5555/ http://irep.iium.edu.my/5555/1/ICSTIE.uitm.pdf |
| Summary: | This paper examines the implications of the present approaches in handling missing variability in meta analysis on the overall standard error (SE) of the estimate. The approaches are (1) exclusion of the studies with missing standard deviations (SDs) and (2) imputation of the missing SDs. The data was simulated with the SDs assumed to be missing according three scenarios, namely, missing completely at random (MCAR), missing at random (MAR) and not missing at random (NMAR). The study demonstrates that imputation is preferable over excluding the studies with missing standard deviations if the the missing variability occurs completely at random, or if the mechanism for missing variability depends on the size of the studies. However if studies with larger variability measures tend not to report the standard deviations, then imputation will lead to a bias in the standard error of the estimates. As the later case is impossible to ascertain, it is thus recommended that an analysis based upon studies with full available data and imputed data be carried out, and comparison between the two results are made. |
|---|