Tools for Those Who Summarize the Evidence Base
Hi guys, I have 2 quick questions regarding estimation of publication bias in meta-analysis using the Funnel plot.
1. Regarding continuous variables where SMD are used, Do we run logES (logSMD) against standard error (SE) of the effect size (ES/SMD) or it is logES against the log of the standard error (logSE)
2. How do I handle studies that appear twice in the funnel plot? where an intervention of interest is compared with control and another active control, do I take the study as a single entity or do I have to enter them twice just we do when meta-analyzing.
1. Funnel plots conventionally plot the weight (inverse SE or inverse variance) as the Y-variable and the effect size, SMD, as the X-variable. No logs are needed with SMD (it is the odds ratio that is routinely logged, not the SMD). Some platforms offer enhanced versions of the funnel plot, such as Stata's confunnel command, which adds significance contours to the plot.
2. Technically, plotting more than one SMD that are based on the same observations (e.g., the same participants) violates the classical meta-analysis assumption of non-independence, and in this case you are giving studies with repeated SMDs more weight, potentially much more weight. You need to create a rule to deal with this problem: Averaging the SMDs or selecting the one that is most like the others in your literature, methodologically. Or, do sensitivity analyses to ensure that duplicating the studies' SMDs does not affect any statistical inference. If it does change whether an average SMD is significant or whether a particular moderator is significant, then there is something about the results of the repeated SMD studies that is making a difference. If feel your meta-analysis needs to retain all the cases, then you need to find a way to model the interdependence of the cases (e.g., see the discussion here about meta-analyses in multi-level models).
Hope that helps! Your post might be better placed in meta-analysis.ning.com's analyses group or the one on publication bias, which are more specialized to analyses. If you join these groups, you will automatically receive notifications of new posts about analyses.
Thanks so much for walking me through the issue of dealing with repeated studies in a funnel plot. It is certainly helpful. I will see what additional inputs the publication bias group may have. Thanks a million.
It is better to use meta-regression with con founders....
Make it as two separate studies.......