I am to run logistic regression on approximately 900,000 observations (each a unique customer) in order to determine propensity to purchase a given product or propensity to hold a given product. I would prefer building propensity to buy models but am not sure if I have enough data/how to best construct the outcomes. I must build models for 7 different products, and those models are to be treated independently (i.e. the outcome period for each model can be different).

The following table shows the products that I'm going to model and the respective count of unique customers that either purchased the products from December 13-Jan 14, Nov 13-Jan 14, Oct 13-Jan 14 or simply hold the product irrespective of date.

Product holding table

I have almost no data for most products in the months Dec-Jan and Nov-Jan, but I start seeing more if I look at the months Oct-Jan.

Here - my questions, given the total of 900,000 observations:

1) Is there any sense at all in building propensity to buy models for those products where I have data on product purchase between Oct-Jan (i.e. setting the dependent variable of those customers who bought a given product between Oct-Jan to 1 ("event")), or do I simply have too little data and must therefore rely on propensity to hold models for all products?

2) How do I determine the "non-event" (e.g. 0) segment of my dependent variable? Should I randomly sample from the population of customers who do not hold the product (or who did not buy the product if I can use propensity to buy models)? How many observations should I draw (i.e. do I want a 50-50 breakdown of events and non-events, or should I use more non-events, as I have more information on them? How many more non-events should I use?)?


Your Answer

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

Browse other questions tagged or ask your own question.