I'm doing meta-analysis for a treatment done in seeds. I have approximately 130 series with value of the control mean (Xc), its treated homologous mean (Xe) for crop yield and the number of replications (Nc and Ne) of each experiment which I'll use the number of replications as "sample size”. From those 130 studies, I got like 5% of them with diverse measures of variance: pooled standard deviation for control and treated, standard deviations, P values… (see the picture) that's why I think it’s not good idea to impute.

I believe that I can still do meta-analysis without standard deviation by calculating effect size (ES) with log response ratio (lnR) of means and later weight the studies using sample sizes as measure of variance.

I have also seen some studies that just defined the variance as I think the inverse of sample size, but I’m not sure how valid or accepted that is as a method.

Would lnR be the best solution? Anybody has done meta-analysis in this fashion or knows the best way to proceed?

I would be happy to hear any R codes that could help for the analysis. piece_data_meta-analysis Thanks in advance.

  • $\begingroup$ Questions that are about how to use R, are typically off topic here. However, 'how to do MA w/o the SD via ES' is a valid statistical question. You might edit your Q to emphasize its statistical nature & deemphasize R (you needn't necessarily delete it). You might also provide some example data for clarity & for people to work with. $\endgroup$ – gung - Reinstate Monica Oct 25 '16 at 17:01
  • $\begingroup$ @gung sorry for me inexperience in this web, my bad. I corrected it in consequence, and it's actually something more interesting to discuss now. $\endgroup$ – Carrillo-Reche Oct 25 '16 at 17:32
  • $\begingroup$ That's no problem, now you know. Can you provide some sample data? $\endgroup$ – gung - Reinstate Monica Oct 25 '16 at 17:38
  • $\begingroup$ There certainly have been suggestions of using sample size and the canonical source seems to be work by Hunter and Schmidt in their book which unfortunately I have not read. Looking at your data it does seem that these are very heterogenous which means that if you use a random effects model you will get approximately equal weights so you might choose that path instead $\endgroup$ – mdewey Oct 26 '16 at 12:23
  • 2
    $\begingroup$ In case anyone is still interested in this question. I have used the SD I had and nonparametric variance to calculate the missing SD: VlnR=(nt+nc)/(nt*nc). I have also compared the results of meta-analysis based on studies where I had a complete info (SD/SE) with results of meta-analysis based in nonparametric weighting meta analysis as part of sensitivity analysis and it seems to be consistant. I think it's the best solution. $\endgroup$ – Carrillo-Reche Nov 14 '16 at 20:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.