Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

I've been trying to read up on Poisson regression models, and it looks like it is possible to estimate such a model with a binary outcome. This has come up before on this site here (and somewhat here and there). I am still a bit confused about how to interpret the coefficients when the outcome is binary, and how to specify an offset to facilitate this interpretation. Let's assume that $E[y|x]=\exp(a+\beta x + \gamma d)$, where $x$ is continuous, and $d$ and $y$ are binary. Let's say my random sample is $N=10,000$ and $y=1$ happens 10% of the time and $d=1$ 25% of the time.

  1. I believe (but would like to verify) that for small values of $\beta$, I can interpret it as an elasticity, so if $\hat \beta=.05$, that's a ~5% increase in $\Pr (y)=1$ for an additional unit of $x$. When $\beta$ is larger, it is more accurate to exponentiate it, so if $\hat \beta=.5$, that corresponds to a 65% increase in $\Pr (y=1)$. With binary $d$ going from 0 to 1, the marginal effect is $\exp(\hat \gamma)-1,$ so for $\hat \gamma=0.3$, we would get a 35% increase.
  2. Can I interpret $\exp(\alpha)$ as the baseline probability for those with $d=0$ and $x=0$?
  3. If the outcome $y$ was actual counts or even continuous, I can just change the end to read "increase in $y$" rather than "increase in $\Pr (y=1)$".
share|improve this question
Why would you want to use Poisson regression for a dichotomous dependent variable? – Peter Flom Oct 3 '12 at 17:15
I am also estimating some related count data models, and it would be convenient to have the same underlying model for comparison purposes. Binary also seems like a special case of the count. – Dimitriy V. Masterov Oct 3 '12 at 17:59
Poisson regression is not such a great idea for binary data, and is a very bad idea if the binary data is not numerical (as is almost always the case, e.g. "Did you vote for candidate $x$?"). Not to mention, under the model, non-zero mass would be given to points other than 0 and 1. Echoing @PeterFlom - why would you want to do this? Is something wrong with using standard techniques for binary data (e.g. logistic regression)? – Macro Jan 17 '13 at 14:08
I was trying this in the context of an instrumental variable estimator as an alternative to using a maximum-likelihood two-equation probit model, which was performing poorly. Using the Poisson GMM-IV gave me very reasonable estimates. My attempt was inspired a bit by the two papers I mentioned in, which use the Poisson with a continuous outcome. I hoped the P/QMLE intuition would carry over to the binary case, but I have not worked that one out. – Dimitriy V. Masterov Jan 17 '13 at 18:31
You can use poisson regression with a continuous $Y$ variable because poisson regression provides consistent estimators as long as $E(Y_i|X_i)=exp(X_i\beta)$ --- an amazing, useful, and much-overlooked fact. However, you can't do this for binary $Y$ in general. The reason is the same reason you can't use OLS for binary $Y$ in general. There is nothing constraining $exp(X_i\beta)$ to be between 0 and 1. And, since $E(Y)=P(Y=1)$ for binary $Y$, we must constrain $E(Y)$ to be between 0 and 1. For binary $Y$, the crucial assumption legitimizing poisson regression is false, in general. – Bill May 9 '13 at 16:21

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Browse other questions tagged or ask your own question.