3

I have plotted the effects of a model glm( A ~ B, family=poisson, data=data) both with the sjplot R package and with the effects package. However, my A variable is a percentage (min=0, max=100) and the effect plot shows A values higher than 100. What should I do?

mkt
  • 11,770
  • 9
  • 51
  • 125
Sara
  • 41
  • 3
  • 2
    I've changed your title because it seems like the problem is not your plot, but the underlying model. Please check the edit to see if it captures your meaning. – mkt Jul 22 '19 at 13:53
  • 5
    A Poisson regression is suitable when the outcome is a count. In your case, the outcome `A` is a percentage. In that case, I'd recommend using logistic regression or beta regression. – COOLSerdash Jul 22 '19 at 13:54
  • Related: https://stats.stackexchange.com/q/418584/121522 – mkt Jul 24 '19 at 08:08

1 Answers1

4

If the prediction limits or confidence intervals exceed some limit beyond which predictions/inference become unphysical, then your model is misspecified.

The most straightforward example of this is using linear regression to perform prediction of a continuous covariate's affect on a binary outcome. Eventually, the model will predict something above/below 1/0.

The poisson's mean is supported on the positive reals, and so if your outcome is bounded to the unit interval, you may consider a logistic regression or a beta regression. To say more, we would need to know what you are modelling and perhaps see some data.

Demetri Pananos
  • 24,380
  • 1
  • 36
  • 94