1
$\begingroup$

Lets say I have a random variable $y$ that is expected to be:

$$y_i \sim poisson(\lambda_i)$$

$$log(\lambda_i) = \beta_0 + \beta_1x_i$$

But, for every reasons, I am running the following linear regression using OLS:

$$log(y_i) = \beta_0 + \beta_1log(x_i)$$

Given the log-log transformation, I could than say "a 1% change in $x$ causes a $\beta_1$% change in $y$". Switching back to the poisson context, how would I set up a regression model where a similar interpretation (% change in $x$ causes $\beta_1$% change in $y$) is possible?

Since the log is the common link-function for a poisson glm, I do not need to transform the lhs of the equation. So transforming $x_1$ should be sufficient?

$\endgroup$
1
  • 1
    $\begingroup$ The Poisson GLM and the log-log (elasticity) model are generally two different approaches. They can both work but the choice depends on the problem. An answer to a similar question has been provided here. $\endgroup$
    – Digio
    Jul 30, 2017 at 20:16

3 Answers 3

4
$\begingroup$

The transform on $X$ is not a key difference between the two methods, because, like you have noticed, you can also do it in Poisson regression without a problem. The essential difference is about $Y$: transform is not link (GLM). You see the difference clearly when you write the formulas as a conditional mean.

Linear model transformed with log is:

$$E(\log(Y)|X)=\beta_0+\beta_1X$$

GLM with a log link is (as in Poisson regresion):

$$\log(E(Y|X))=\beta_0+\beta_1X$$

Even if they look the same, they are not the same at all (because $\log$ is not linear).

If you are interested in having no bias on $E(Y|X)$ then GLM is the model to choose. With a transformed linear model, there is a (usually strong) bias. The subtlety happens in the way the noise ($\epsilon$) is transformed in a non linear way. The noise has mean 0, but when transformed by $\log$ it modifies the mean of the estimation.

$\endgroup$
1
  • 1
    $\begingroup$ +1. I would only like to point out that the log-log model is not unimportant at all. It is a fundamental model in econometrics which gives a different coefficient interpretation than Poisson regression. $\endgroup$
    – Digio
    Jul 31, 2017 at 8:40
2
$\begingroup$

Actually, this is a pretty interesting topic in price elasticity modeling. Real-world sales data is often contaminated with zero values in the dependent variable due to no-sales days. Since the log-log regression cannot handle zero values properly without applying the log(c+x) transformation, Poisson regression or any other count data models seem to be a relevant alternative but cannot be interpreted as easily as log-log regressions.

$\endgroup$
2
$\begingroup$

I agree that the interpretation for the Poisson regression is not clearly straightforward, especially when there are categorical or binary covariates involved. For example, if there is a binary independent variable involved in the Poisson regression the coefficient should be exponentiated before accurately interpreted. If the regression output resulted in a 0.56 for the coefficient on the binary variable, it should be treated with (exp(0.56)-1)*100 in order to obtain the proper value.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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