Hi Michael, Thank you for your input. Actually the idea of using the model to estimate a number of sick people by age group came later in the project, and I missed the issue with the ecological nature of the analysis.
Initially, I intended to use the model to predict a single prevalence for each country at the median age of the country population. Should I go back to this approach? If I'm not mistaken, this could be interpreted as age-adjusted pooled prevalence at the country level. In this case, should I use the model with the study-level random effect? Tkanks again for your time, Julien On 11 April 2014 13:15, Michael Dewey <i...@aghmed.fsnet.co.uk> wrote: > At 13:17 10/04/2014, Julien Riou wrote: > >> > >> > Message: 11 >> > Date: Wed, 09 Apr 2014 18:39:30 +0100 >> > From: Michael Dewey <i...@aghmed.fsnet.co.uk> >> > To: Julien Riou <julien.rio...@gmail.com>, r-help@r-project.org >> > Subject: Re: [R] Meta-analysis of prevalence at the country level with >> > mgcv/gamm4 >> > Message-ID: <zen-1wxwtl-0005q4...@smarthost01a.mail.zen.net.uk> >> > Content-Type: text/plain; charset="us-ascii"; format=flowed >> > >> >> >> Hi Michael, >> >> Thank you for taking the time to help me. >> >> So the first UK study with a median age of 25 is going to be used to >> > estimate prevalence over a range of ages? You are going to have to >> > make some very strong assumptions here which I personally would not >> > want to make. >> > >> >> I'm a little confused by this. In my understanding, the mixed-effects >> model >> does not do that. The slope of the relation between age and prevalence >> will >> be estimated from the full pool of studies, and the country-level random >> intercept will be estimated from all studies in the country. So the >> assumption here is that the relation between age and incidence is the same >> in every country, which is quite reasonable. Of course, there will be more >> uncertainty with the estimation of the random intercept if there is few >> studies in a country, or if there is a strong inter-study variance in a >> country. This will influence the confidence interval of the random >> intercept, and so the CI of the predicted prevalence for this country. >> > > Your studies are ecological. You are estimating the relationship between > prevalence and being in a study of median age X which is not necessarily > the same as the relationship between prevalence and being a person of age X. > > > > Is there any possibility that in the real dataset you can fit your >> > model to those studies which do provide age-specific prevalences and >> > then use that to impute? >> > >> > You do not say when these studies were published but I would ask the >> > authors of the primary studies if they can make the information >> > available to you. You may have already done that of course. I referee >> > quite a few papers on systematic reviews and my impression is that >> > some authors are amenable to doing the work for you. You mileage may >> > vary of course. >> > >> >> Yes, it would be easier to have prevalence for age subgroups of studies, >> but we did not have access to that information for most studies even after >> contacting the authors. >> >> >> > >*Standard random-effect meta-analysis* with package meta. >> > > >> > >I used metaprop() to get a first estimate of the prevalence in each >> > country >> > >without taking age into account, and to obtain weights. As expected, >> > >heterogeneity was very high, so I used weights from the random-effects >> > >model. >> > >> > Which will be nearly equal and so hardly worth using in my opinion >> > but again your mileage may vary. >> > >> >> The weights from the random-effects method were actually far from equals, >> as sample size was highly variable between studies. With the RE method, >> small studies have much more impact. >> >> >> > I am afraid that is the way with systematic reviews, you can only >> > synthesise what you find, not what you would like to have found. >> > Anyone who has done a review will sympathise with you, not that that >> > is any consolation. >> > >> >> I'm not sure I'm following your point. My objective is to synthesise the >> included studies, while taking the age factor into account, since it is >> strongly linked to prevalence and very heterogeneous. The alternative is >> to >> only include studies with low median age, but I would lose a lot of >> information. >> >> Thank you again, >> Julien >> >> > >> > >> >> [[alternative HTML version deleted]] >> > > Michael Dewey > i...@aghmed.fsnet.co.uk > http://www.aghmed.fsnet.co.uk/home.html > > [[alternative HTML version deleted]] ______________________________________________ R-help@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
