7

I have a time series count data by customers that I would like to regress on past months items (count) sold and promotional effects (current and past). Below is an example, and the dataset has one record per customer.

Y= number of items (count) sold

Customer 1: Y/Count data (current month) = Y/count data(past month) + promo data (current month) + promo data (past month)

Customer 2: Y/Count data (current month) = Y/count data(past month) + promo data (current month) + promo data (past month)

I'm not sure what is the right model to use. I could use Dynamic regression/ARIMAX if my response was a continous variable. But the response variable in my hand is a count data, so I'm not sure if ARIMAX is the right model for this problem. In addition, I have lot of zeros in my response variable (similar to zero inflated poisson regression).

I did search in standard textbooks and the internet, I did not come across any regrssion model with time series count data. I would really appreciate if anyone could recommend an appropriate model for this problem.

Thanks

Gavin Simpson
  • 37,567
  • 5
  • 110
  • 153
forecaster
  • 7,349
  • 9
  • 43
  • 81
  • 1
    Take a look at Trivedi's notes on [Poisson model with exponential feedback (sec. 23)](http://www.stata.com/meeting/mexico10/mex10sug_trivedi.pdf) and the Windmeijer paper he cites. They may be helpful. – dimitriy Sep 16 '13 at 20:54

1 Answers1

3

It looks like a zero-inflated Poisson or negative binomial dynamic model would be adequate. If you are using R, you might want to look at package ZIM.

F. Tusell
  • 7,733
  • 19
  • 34