So correct me if i'm wrong. The usual link function for poisson regression is log, so that you're performing regression on log(y)~x1+x2+x3+x4+...

The variable y is typically a count, meaning it is restricted to integers from 0 to positive infinity. The input variables x1...xn are not restricted to the positive integers.

So how does the regression proceed when y = 0? Is log(0) merely ignored?

Also, to be clear, this question is not about zero-inflated poisson regression (which distinguishes between different kinds of zeros).


2 Answers 2


The Poisson model is $$y = \exp \left(\alpha + \beta \cdot x + \varepsilon \right).$$

The way you get an outcome of zero is when the index $\alpha + \beta \cdot x + \varepsilon$ is large and negative. The coefficients do not come from a regression of logged outcome on the covariates, but from maximization of the log likelihood. You can also use this model on non-integer outcomes, though that is more controversial.

You can learn more about this model from this blog post, including the zeros issue and a comparison to logged outcome regression.

  • 1
    $\begingroup$ I don't think that is correct (what is $\epsilon$ in the above equation?) I believe $y \sim Poisson(\lambda=\alpha + \beta x)$ is a better way to communicate the idea. $\endgroup$ Jan 10, 2018 at 2:47
  • $\begingroup$ The most correct way is to add an expectation operator to y. @MatthewDrury $\endgroup$
    – SmallChess
    Jan 10, 2018 at 3:23
  • $\begingroup$ Right, like $E[y \mid x] = exp(\alpha + \beta x)$. But that doesn't work out correctly the way you have it written, with the random component $\epsilon$ inside the exponential. $\endgroup$ Jan 10, 2018 at 3:31
  • $\begingroup$ That's the DGP, not the conditional expectation. You can always rewrite that to make the error enter multiplicatively if that is easier to digest. $\endgroup$
    – dimitriy
    Jan 10, 2018 at 3:41
  • 1
    $\begingroup$ +1 I was sceptical, and have seen William Gould's post on using Poisson models before, but upon re-reading it have to agree that what you write here is an accurate account of his approach. $\endgroup$
    – whuber
    Jan 10, 2018 at 5:16

The poisson regression model is that the logarithm of the expected values can be modeled by a linear combination of predictors. The expected values of y are not 0, even though there may be 0 counts in the real data. The expectation is a positive real number.


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.