I was just wondering if Poisson regression has an error term? Can a Poisson regression have random effects and an error term? I am confused about this point. In logistic regression, there is no error term because your outcome variable is binary. Is that the only glm model that doesn't have a residual term?

  • $\begingroup$ Of course. This question is what the folks at Stack Overflow call unconstructive because it can be answered in a lot of different ways. Be more specific with your question. $\endgroup$
    – ndoogan
    Commented Apr 9, 2013 at 0:35
  • $\begingroup$ I was just wondering because logistic regression does not have an error term but I get since counts is a numerical value, then it does have an error term. $\endgroup$
    – phil12
    Commented Apr 9, 2013 at 0:39
  • 7
    $\begingroup$ Logistic regression DOES have an error term. Logistic regression predictions are in the form of a probability, not 1 or 0. $\endgroup$
    – ndoogan
    Commented Apr 9, 2013 at 0:40
  • 1
    $\begingroup$ so would it be correct to say: log(count) = b1x1 + b2x2 + e where e is an error term? $\endgroup$
    – phil12
    Commented Apr 9, 2013 at 0:47
  • 4
    $\begingroup$ a logistic regression has error yes, but the variance of the latent variable is fixed (otherwise the model is unidentified). the poisson only takes one parameter which describes the mean and variance. $\endgroup$
    – Zach
    Commented Apr 9, 2013 at 1:10

1 Answer 1


I think the problem that's confusing you is that you're used to having an additive error. Most models won't.

Think of linear regression not as a linear mean with an additive error, but as the response being conditionally normal:

$$(Y|X) \sim \operatorname{N}(X\beta,\sigma^2I)$$

Then the similarities to GLMs, in particular, to Poisson regression and logisitic regression are more clear.

Because of the nice properties of the normal, the normal case can be written in terms of the mean and an additive error. This is not always so nice with other models and it makes sense to stick to the distributional form for the model, or at least to write about the mean and variance of $(Y|X)$ rather than writing a model for $\operatorname{E}(Y|X)$ and trying to describe the characteristics of $Y-\operatorname{E}(Y|X)$.

[You can take any particular combination of predictors and write the response variable in terms of its expectation and a deviation from that - an 'error' if you will - but it's not particularly enlightening when its a different object from every other combination of predictors. It's usually more informative and more intuitive to just write the response as a distribution that is a function of the predictors than in deviation-from-expectation form.]

So while you can write it 'with an error term' it's just less convenient and conceptually harder to do so than to do other things.


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.