Assessing the efficacy of the trim and fill method in adjusting for publication bias in meta analysis
One of the most common critics leveled against meta-analysis pertains to the issue of publication bias. A publication bias may be encountered if a meta analysis is restricted to the combination of results obtained from studies which have been published. The bias occur when unpublished studies are as...
| Main Authors: | , |
|---|---|
| Format: | Conference or Workshop Item |
| Language: | English English |
| Published: |
2012
|
| Subjects: | |
| Online Access: | http://irep.iium.edu.my/25593/ http://irep.iium.edu.my/25593/1/irie_2012_1211.pdf http://irep.iium.edu.my/25593/4/cert_of_participation.pdf |
| Summary: | One of the most common critics leveled against meta-analysis pertains to the issue of publication bias. A publication bias may be encountered if a meta analysis is restricted to the combination of results obtained from studies which have been published. The bias occur when unpublished studies are associated with their outcomes, namely, studies which produced small effects or negative results are more likely not to get published. The most common method used to adjust for publication bias is the Duvall and Tweedie’s Trim and Fill method. The method relies on the scrutiny of one side of a funnel plot for asymmetry which is assumed to be due to publication bias. In most of the studies investigating the performance of this method, binary meta analysis data were used, utilizing the effect size measures such as the odds ratio, or risk relative. The main aim of this study is to evaluate the efficacy of the Trim and Fill method in correcting the publication bias in meta analysis of continuous data with standardized mean difference as the measure of effects.
The data used in this study were simulated using R. The characteristics and assigned values of the simulated meta-analysis, namely the size of treatment effects, study variances and the number of primary studies used, were based on a meta-analyses from medical research. Three levels of primary studies ,N, associated with small, medium and large meta analysis are N =10, 30 and 50/100, respectively, and six levels of percentage of missing publications ,x%, were induced, corresponds to high, medium and low degree of bias at 70%/50%, 30%/20% and 10%/5%, respectively. Each combination of meta-analysis generated was replicated 10,000 times.
The inverse-variance weighted random effects estimate were computed for both unadjusted and adjusted meta analysis. The assessment of the effect estimates were evaluated in terms of statistical bias and the standard errors for point estimates, and in terms of confidence band coverage for the interval estimates.
|
|---|