I am using a zero-inflated negative binomial regression model for my data analysis, where the dependent variable is the number of retweets. I have three categorical variables called emotionally_charged (0 if not 1 if yes), hashtag_used (0 if not 1 if yes), and media_used (0 if not 1 if yes). The base reference for all three categorical variables is 1. I need to understand how to interpret the effect of all three on the number of retweets (a count variable) according to the signs of coefficients. Following are my partial results:

VARIABLE | Coefficient | p-value

1.emotionally_charged | -0.5825642 | 0.000

1.hashtag_used | -0.6320729 | 0.000

1.media_used | 0.1608499 | 0.206


1 Answer 1


In both Poisson and negative binomial regression, if you exponentiate the raw coefficients, you get an incidence rate ratio. That is not being emotionally charged (if I read your post correctly, even though it seems to contradict the output) is associated with an IRR of $e^{-0.5825642} = 0.558$. You need to note that this is the IRR conditional on not being a structural zero.

That is, a zero-inflated model estimates the probability of always responding a zero, and the coefficients of a count model conditional on not being a zero. You didn't show the zero-inflated part of the model, but those coefficients are for the log-odds of being a structural zero. You interpret those just like a logistic regression.

A zero-inflated model output is not that complicated to interpret. However, you should ask yourself a) is a zero-inflated model plausible, and b) even if a is true, is a zero-inflated model a better model for the data than a normal count model? That is, are there some people in the sample who will never retweet? This seems like a justifiable assumption, but it isn't always. Even if so, say the marginal distribution of the outcome has a lot of zeros. That could be adequately explained by a very low rate.


Your Answer

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

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