2
$\begingroup$

I am using emmeans to produce estimated marginal means from GLMMs with Poisson or negative binomial distributions, but for a few of my models the confidence interval is in the negative. I was not expecting this given the distribution. Can I simply set the lower bound to zero, or am I missing something?

m1<-glmmTMB(count ~ treatment + (1|nestBoxSlot), data=data, offset=log(noIndividual), ziformula=~1, family="nbinom2")

emmean <- emmeans(m1, "treatment", type="response", offset=0)

Result of emmean

  treatment  response        SE  df   lower.CL   upper.CL
1         C 8.1624947 8.0114700 528 -7.5757742 23.9007636
2         M 0.2043217 0.2637268 528 -0.3137608  0.7224043
3        MD 0.4341052 0.5491020 528 -0.6445876  1.5127979
$\endgroup$
2
$\begingroup$

This can happen if your software back-transforms the estimates before computing the intervals, and then uses the SEs of the back-transformed estimates (usually obtained by the delta method). You can avoid this by computing the intervals on the link scale (e.g., log), and then back-transforming the endpoints.

$\endgroup$
6
  • 1
    $\begingroup$ BTW, the latter is what the emmeans R package does by default. So I am not sure how this happened. Can you show the code you used, and the results? $\endgroup$
    – Russ Lenth
    May 10 '20 at 12:55
  • $\begingroup$ Indeed, that is why I was confused. Edited above with an example. $\endgroup$ May 10 '20 at 16:02
  • $\begingroup$ Odd. Can you tell if the answers are reasonable? What do you get on the link scale? And keep in mind that the estimates don't include the ZI part, I believe. (However, if Ben has revamped it to include the ZI part and estimate the mean response, that would involve back-transforming, and would explain everything. $\endgroup$
    – Russ Lenth
    May 10 '20 at 16:46
  • $\begingroup$ The estimated response is reasonable, but these are counts of seconds. When I set zero inflation to 0 they are considerably less into the negative. $\endgroup$ May 10 '20 at 16:58
  • $\begingroup$ I have not been able to find out how emmeans handles calculating CI from zero inflation models in glmmTMB $\endgroup$ May 12 '20 at 16:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.