I am doing a mini-meta-analysis of four correlational studies.

In the estimated model I have three predictors and one DV.

Using the meta-package in R, I use the following code to test the link between one IV and DV.

m1 <- metacor(c(-.18, -.09, -.38, -.03), c(886, 1050, 478, 1043))

I have three questions.

  1. Should entered correlations be zero-order correlations between IV and DV or partial correlations where the other two predictors from my model are being controlled for?

  2. Should I enter correlations or should I use Fisher's z transformed correlations? My understanding is that they should be regular correlations, as in the details of the meta-analytical method of metacor function says it is using Fisher's z transformation of correlations, which I guess means that entered correlations are being transformed to Fisher’s z by default?

  3. The default estimator is DerSimonian-Laird estimator. Is it ok to use it?

Thanks! Aneta

  • 1
    $\begingroup$ I am voting to leave this open as questions 1 and 3 seem to me to demand statistical knowledge although I agree question 2 is about the options of an R package. $\endgroup$
    – mdewey
    Mar 26, 2021 at 16:31

1 Answer 1


Q1 - that depends on your scientific question. They measure different things, which is relevant for you?

Q2 - You can determine that by finding the default for the sm parameter but I think it transforms them for you which is what you want.

Q3 - it it probably better to use REML although it is hard to give an answer which applies across all situations. In a paper entitled Bias and efficiency of meta--analytic variance estimators in the random--effects model Viechtbauer discusses five different estimates of $\tau^2$ including DL and REML.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.