Negative binomial estimate output

Question

I am doing a negative binomial regression:

m1 <- glm.nb(Trip ~ Origin + Destination + Distance, data)

I then wanted to get the predicted values to compare with the actual values

PredictTrip <- predict(m1)

My question is, I read on several websites that the predicted values need to be exponentiated because it is in log form. However, if I exponentiate the predicted values, it becomes ridiculously large (i.e. E+20, or even Infinity). If I did not exponentiate it, the values are close and comparable to the actual values. Which one is right? Thanks.

Answer 1

You don't necessarily need to exponentiate.

When you execute predict(m1), what is called is the predict.glm() function. Here is its help page:

Usage:

     ## S3 method for class 'glm'
     predict(object, newdata = NULL,
                 type = c("link", "response", "terms"),
                 se.fit = FALSE, dispersion = NULL, terms = NULL,
                 na.action = na.pass, ...)

Note in particular the type parameter, which governs what type of predictions or fits you will get. By default (type="link"), you get a prediction on the level of the link function. What you want is a mean response prediction, so do

predict(m1, type="response")

You may be interested in Regression predictions show far less variance than expected.