Poor man's metaregression with lmer ?
Le samedi 23 mai 2009 ? 07:50 +1000, David Duffy a ?crit :
On Fri, 22 May 2009, Emmanuel Charpentier wrote:
One obvious solution is mima : transform each failure rate f in logit (log(f/(1-f)) with estimated variance 1/(n*f+0.5)+1/(n*(1-f)+0.5). With a hiccup : three of the series have no failure or no success (yes, both can happen ...), which gives an infinite log(OR). Ouch ! Excluding them is not acceptable, and adding 0.5 to all counts (aka "continuity correction") will anyway be judged passe by the panel ...
You just have to explain that you are using maximum penalized likelihood, with the Jeffreys invariant prior as penalty function ;).
:-)) *I* won't face a panel : the student who (foolishly ? ) seeked my help will...
See the brlr and logistf packages and refs. If you did have time to failure, there is also a coxphf. We usually present results from both the fixed and random effects metaanalysis -- your lmer code looks OK to me.
Thank you for your review ! So lmer() it probably will be (but I'll check the mima() approximation...). What a pity that lme4 doesn't (yet ...) have covariance and correlation modelling functions... Emmanuel Charpentier