2
$\begingroup$

I am using the "metafor" package to do a multivariate meta-regression in "R".

I have 6 predictors and I am able to run the full model (all the predictors simultaneously in the model) just fine.

However, I would like to do a backward deletion meta-regression. I cannot figure out how to do it, and I am wondering if this is a possibility.

Thank you.

$\endgroup$
  • 2
    $\begingroup$ Why would you want to do such a thing? Are you trying to demonstrate why it is a bad idea? $\endgroup$ – Peter Flom - Reinstate Monica Mar 29 '14 at 10:36
  • $\begingroup$ Not really, but I would be happy if you could explain briefly why it is a bad idea. To me and to some others with already published papers using this technique. $\endgroup$ – Spyros Mar 29 '14 at 10:56
  • 2
    $\begingroup$ If you use stepwise regression: 1) Your parameter estimates will be biased away from 0 2) Your standard errors will be too small 3) Your p-values will be too small. See e.g. (this thread)[stats.stackexchange.com/questions/69452/… for more. $\endgroup$ – Peter Flom - Reinstate Monica Mar 29 '14 at 11:05
3
$\begingroup$

There is an extensive account of how to do model selection using the metafor and glmulti packages on the metafor project home pages here. Note that this does not use step-wise selection which has been widely criticised but fits all possible models from a specified set (all main effects, all two-way interactions and main effects, and so on). The main criterion used in the example is AICc but others are available. One advantage of this approach over step-wise is that it encourages examination of all (or a subset) of the models rather than having a focus just on the "best" one which usually reveals that there are a number of models virtually indistinguishable from the "best" one.

$\endgroup$

Your Answer

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

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