2
$\begingroup$

My main goal is to model some count data. The dataset is a time series dataset that if I were to perform linear regression, I would difference to make sure the data became stationary. However, if I want to use Poisson GLM, would I still difference it or try to make stationary? My main concern is if I were to apply a transformation, I wouldn't be modeling counts anymore, rendering Poisson irrelevant. In addition, differencing can lead to negative values, and Poisson is only for positive values.

$\endgroup$
-1
$\begingroup$

If you need to model longitudinal data where the outcome is a count, I'd recommend using a mixed model approach or a Generalized Estimating Equations (GEE) approach. Both of these methods will account for the serial correlation of your outcomes and you will not have to worry about transformations or taking differences of measurements. A search on this site for GEE/Poisson or Mixed Models/Poisson will lead you in the right direction.

| cite | improve this answer | |
$\endgroup$
  • 1
    $\begingroup$ I'm just trying to keep it in GLM though. Not differencing time series data might make predictions pretty bad. At the same time, out of all of the GLM family options, Poisson makes the most sense $\endgroup$ – Lzydude Aug 15 '16 at 18:10

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.