Hi everyone, I am trying to run a meta-analysis on a series of studies where the key result is a regression slope. I am getting this coefficient and its SE from lm() function in R. I was wondering if I can enter this information directly into the metafor package to run the meta-analysis. Something like...
meta
A meta-analysis uses summary information from studies with associated precision information to come up with an overall summary. It is crude in that entire studies are reduced to 2 or 3 numbers.
It sounds like you have access to the original data for the studies you want to summarise. It would be more powerful to combine the original data into one big dataset and run your regression on that. That will give you an overall summary slope with associated confidence interval. You will also be able to compare the studies within your data with interaction terms. If that is possible then that is what i would do.
True. I liked the idea of a meta-analysis because it was a very simple and graphical means to show the general pattern of results and the degree of heterogeneity. It is one of these situations where I think that a forest plot is worth a thousand words. But it is true that it ignores a lot of information. In the end, I might do both things.
Although Marks idea seems simple and feasable it will loose the individuality of each study, a specially important fact in studies where there any kind of randomisation. This can generate different results and is called the Simpson's paradox, and it is fundamentally wrong
Hi Miguel. If you are analysing your own data, then I'd rather transform the effect size to D and use Cummings' spreadsheet to calculate the overall effect. It gives you the forest plot and everything.
https://thenewstatistics.com/itns/esci/
Filipe makes a good point about Simpson's paradox, but this is only a problem if the data from the raw studies are combined without any information about which study the data come from. This would be wrong for a few reasons. The interaction terms between treatment arm and study i suggested will prevent Simpsons paradox. Incidentally, a mixed effect model should be applied to account for between study variation.
Thank you guys for this advice. Following the recommendation of Mark, I was thinking of using a linear mixed model with a random intercept and slope for study. If there is no heterogeneity across studies, then this model (with random intercept + slope) will not perform significantly better than an equivalente model without the random i+s. Is this correct?
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis