However, we caution that when the number of studies is relatively small, association tests based on random-effects meta-analysis may be deflated, as the between-study variance τ2 will be poorly estimated. This is illustrated in Figure 2, which shows quantile-quantile plots for fixed-effect and random-effects meta-analyses of data from PanScan collaboration, which involves 13 studies in the initial GWAS scan. For the random effects analysis, the genomic-control “inflation factor” is in this case more aptly named a “deflation factor:” λGC=0.84, indicating that the random effects p-values are larger than expected under the assumption that the vast majority of SNPs are not associated with pancreatic cancer. Fixed-effect meta analysis is arguably more appropriate as an initial screening test for associated markers, although because fixed-effect analysis can be highly significant when only one (relatively large) study shows evidence for association, analyses that incorporate effect heterogeneity such as random effects meta-analysis should be reported for highly significant markers from fixed-effect analyses.
