Modelling Incomplete/Censored Data

[Rob Wood]

Hi There,

I have a statistical problem I would appreciate some advice on.

Let's say I have a sample of males and females. I've asked each subject how many calls they made in the last 12 months (as an example). I want to model how the number of calls differs between the male and female sample. However, the complication is that not all subjects have had access to a phone (landline or mobile) for a full 12 months. Some subjects have only had access to a phone for say 4 months (again as an example).

I want to include such subjects in my model. However, I'm cautious of linearly extrapolating the number of calls made in 4 months to a 12 month period.

What is the best method for incorporating all subjects. Would weighting the model by the inverse probability of the length of access to a phone (1 over 4 for the example) be a viable option?

Ideally I would've asked how many calls they've made in the period they've had access to a phone, but we didn't. The data we have are the length of period with access to a phone, and the number of calls made within the last 12 months.

All thoughts welcomed.

Thanks,

Rob.

[Carlo Lazzaro]

Rob:
I would take a look at -churdle- (just for a suggestion, at least).

[Rob Wood]

Thanks Carlo.

So I would be modelling whether or not subjects had access to a phone for the last 12 months and then separately how many calls they had made over the last 12 months?

Thanks,

Rob.

[Carlo Lazzaro]

Rob:
you're seemingly dealing with a hurdle model problem.
If the previous Stata command was not helpful, the following http://www.stata.com/bookstore/microeconometrics-stata/ (pages: 583-589) covers some examples on that topic.

[Joao Santos Silva]

Dear Rob,

Can't you just use the length of period with access as exposure?

Best wishes,

Joao

[Rob Wood]

Thank you both for your responses. I will investigate both options.

Thanks,
Rob.

[daniel klein]

My first intuition would also point to a count model with access as exposure (or covariate, if you do not want to constrain the coefficient) just as Joao Santos has suggested.

Best
Daniel