Announcement

Collapse
No announcement yet.
X
  • Filter
  • Time
  • Show
Clear All
new posts

  • Iterations in Poission Regression

    I was running a code for negative binomial and poisson regression and the results were not converging despite and showed 'non concave' log likelihood iterations even after 200 iterations. I used the iteration option to limit the number of iterations to 20. I got the results but with the warning sign 'Warning: convergence not achieved'. I want to know how relaible is this result? Can I use these result for a publication?
    Emily Albom

  • #2
    Hi Emily,
    Short answer its no. Not convergence means that the solution you get is not the one that maximizes the likelihood. Thus, the coefficients and standard errors will not be correctly specified, because those depends on the assumption that there is a unique solution to the model (at least locally).
    My suggestion is to look into the output you get, and check if there are any coefficients with either very large or very small standard errors (or missing standard errors). Those variables are the most likely to be causing the problem because of high multicolinearity or almost perfect prediction.
    When using poison regression, you also have potential problems of "zeros" that can sometimes have consequences as the ones you describe.
    HTH
    F

    Comment


    • #3
      Emily (please consider https://www.statalist.org/forums/help#realnames and act on accordingly. Thanks):
      your results are not reliable and as such not good to be disseminated via a paper.
      The usual issue with MLE is to start everything from scratch adding one oredictor at a time and see when Stata starts gasping.
      What above implies double-checking beforehand that the model is correctly specified.

      PD: crossed in the cyberspace with Fernando's helpful (and partially similar) advice.
      Last edited by Carlo Lazzaro; 27 Sep 2021, 07:27.
      Kind regards,
      Carlo
      (Stata 16.0 SE)

      Comment


      • #4
        I agree with @Fernardo Rios. In my experience Poisson usually converges quickly if the model is a good or even fair fit. You don't really tell us anything about your data and you don't show your code but often when something like this happens the model is just too complicated for the data.

        I would emphasise that the results are not publishable.

        Comment


        • #5
          Dear stata sd (Emily?),

          To add to the excellent advice you received above, I suggest you try to estimate the Poisson regression with the user-contributed ppml command. My guess is that you either have a perfect predictor or your dependent variable has some very large values; ppml is able to deal with most of those cases. Please do let us know if this works and, if it does not, please provide further information about the data and model, as suggested above.

          Best wishes,

          Joao

          Comment


          • #6
            Thanks for the suggestions. I don't think that I have a specification problem in my model, since the poisson and negative binomial models with random effects is converging. But, I need to also get poisson and negative binomial models with fixed effects which is not converging.
            Emily Albom

            Comment


            • #7
              For Poisson with FE, please used ppmlhdfe; it should converge and it is very fast.

              Comment

              Working...
              X