Currently I'm working with ecological studies, where my response is a count variable. I need to estimate several models, each one represents a city. Afterwards I aggregate them to obtain meta-analysis and forestplots. Naturally I'm using Poisson models within GAMs but I noticed that some of the data is overdispersed. Neither I can't find satisfactory answer nor figure out if I can use Poisson for some cities and Quasi-Poisson for the others. Is it erroneous to use both type of models for one meta-analysis? Is it just the matter of the estimated error?

  • 2
    $\begingroup$ Why not fit all the models using quasi-Poisson? $\endgroup$
    – mdewey
    Jun 4, 2020 at 10:02
  • $\begingroup$ I commited an understatement. Some Quasi-Poisson models cannot be fit returning "step size truncated due to divergence" error in mgcv::gam. All Poisson models can be fit without any problem. $\endgroup$
    – Tom
    Jun 4, 2020 at 11:15
  • $\begingroup$ In that case I am at a loss I am afraid. $\endgroup$
    – mdewey
    Jun 4, 2020 at 11:59
  • 2
    $\begingroup$ If your question is if you can have groups-specific overdispersion parameters, the answer is yes. See stats.stackexchange.com/questions/564285/…, stats.stackexchange.com/questions/418777/… $\endgroup$ Jun 16, 2022 at 14:30


Your Answer

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