zero-inflated and right-censored count data

[Makoto Okawara]

Dear all statalists,
Thank you for clicking on my post.

What is the proper way to deal with zero-inflated and right-censored count data?

I have written a histogram of count data that takes integer values from 0 to 30, and found excess zero and bunching at 30.
I have used the zinb command to deal with zero inflation, but there does not seem to be a ul(30) option as implemented in cpoisson.
Please let me know how to deal with this in a practical way.

Thank you for taking the time to read my post.

[Joao Santos Silva]

It would be helpful if you could show us the histogram.

Best wishes,

Joao

[Makoto Okawara]

It would be helpful if you could show us the histogram.

Best wishes,

Joao

Dear Mr. Silva,

Thank you for replying my post.
I have attached a histogram of the three groups.



sincerely,

[Joao Santos Silva]

Thanks for providing the image.

Is 30 the largest value the variable can take, or is it just the maximum in the sample? If it is just the sample maximum, I would start with a simple Poisson regression (the histogram does not show any evidence of zero inflation and cases where that approach is suitable are rare). If it is the maximum the variable can take, I suggest you use a fractional logit unless you need to compute the probabilities of some events.

Best wishes,

Joao

[Makoto Okawara]

Thanks for providing the image.

Is 30 the largest value the variable can take, or is it just the maximum in the sample? If it is just the sample maximum, I would start with a simple Poisson regression (the histogram does not show any evidence of zero inflation and cases where that approach is suitable are rare). If it is the maximum the variable can take, I suggest you use a fractional logit unless you need to compute the probabilities of some events.

Best wishes,

Joao

Dear Mr. Silva,

Thank you for your kindness and supportive advice.
When I divide the count data by 30, which is the largest value the variable can take, and use fractional logistic regression, it seems that the goodness of fit of the model is indeed good.

If it's not too much trouble, please let me know if you have any additional information.

If I still want to use the zero-inflated model, is there any option to deal with right-censored like cpoisson?

Sincerely,

[Joao Santos Silva]

Please check the help file of the flogit command for references.

I do not think there is a way to impose an upper bound when using zip, but note that your data appears to be bounded, not censored.

Best wishes,

Joao

[Maarten Buis]

You could try a zero and one inflated beta model (zoib available on SSC). However, even though I wrote it, I am actually gravitating away from that. I find the model often too fragile, whereas a fractional logit is often more robust. So I would only do that if I was really interested in those 0s and 1s (well 30s in your case), and even than I would be careful.

[Makoto Okawara]

Please check the help file of the flogit command for references.

I do not think there is a way to impose an upper bound when using zip, but note that your data appears to be bounded, not censored.

Best wishes,

Joao

Dear Mr. Silva,

Thank you for adding to my understanding.

Sincerely,
Makoto

[Makoto Okawara]

You could try a zero and one inflated beta model (zoib available on SSC). However, even though I wrote it, I am actually gravitating away from that. I find the model often too fragile, whereas a fractional logit is often more robust. So I would only do that if I was really interested in those 0s and 1s (well 30s in your case), and even than I would be careful.

Dear Mr. Buis,

Thank you for your kindness and new point of view.
I'll consider it again.

[Weiwen Ng]

You could try a zero and one inflated beta model (zoib available on SSC). However, even though I wrote it, I am actually gravitating away from that. I find the model often too fragile, whereas a fractional logit is often more robust. So I would only do that if I was really interested in those 0s and 1s (well 30s in your case), and even than I would be careful.

Maarten,

Just curious, Stata has had a beta regression (betareg) since version 14 (I believe). That native command can only handle values of the DV between 0 and 1 exclusive, i.e. an observation can't have proportions of 0 or 1. Is your version different?

[Maarten Buis]

Yes, the zeros and ones have their own part in the model.