I am conducting a systematic review and meta analysis of the prevalence of lost to follow-up among MDR TB patients. I am using STATA 14 for the analysis. I have performed the pooled estimate using the metan command and I have got high heterogeneity (I2=94.4%, p
The I square value if more than 50% so it shows very high level of heterogenity.Thus you have to go for sub-group analysis or meta-regression to find out the cause of your heterogenity.For further guidance ple email me.
Cochrane's Handbook suggests that it's more important to find out whether a meta-analysis is valid rather than worrying about how to do it. I'd suggest doing heterogeneity analysis and looking at variables like (rstudent, weight of studies, and others).
I suggest that you learn R to get use of packages like "meta" and "metafor".
metafor package has tests that show you which studies influence your final estimate the most. There could be one study that's causing all the heterogeneity.
In a study that my team are working on right now, we removed one study out of 5, and the heterogeneity went down from 89% to 32%. But Cochrane suggest that there should be a logical reason for "clinical heterogeneity" and removing the study just because the numbers were nice isn't justified. I suggest you go read more about this in their handbook.
You should see plots like this, using the "metafor" package, spend some hours on learning it or get someone to do it for you.
There is a high variation in study outcomes, so you need to decide what caused this. Once you adjust the different studies for bias, such as different duration or methodology, you might find that Iˆ2 decreases, hopefully
I^2 is a relative measure of heterogeneity so it indicates that between study variability is much more variable than within-study heterogeneity (and that is accounting for most of the variability in effect sizes). it is still possible (though probably unlikely in a prevalence study) that the overall variability is modest. Explaining this variation by systematic factors (moderators included in a meta-regression) would be useful - especially if you have moderators that theoretically should be related to the measurement or rate of prevalence. This can reduce heterogeneity and increase the value of the analysis.
Sensitivity checks are a good idea, but don't rely too much on tools to detect or deal with publication bias (like funnel plots) as they are known to generally be poor. Ideally you reduce publication bias by obtaining a relatively unbiased set of studies to include (though that's hard!).
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis