topic Re: Two part model for healthcare costs in Statistical Procedures

Two part model for healthcare costs

CC13 — Tue, 07 Aug 2018 19:41:07 GMT

Hi,

I am analyzing healthcare costs. There are so many zero values in the data. I would prefer to use two part model. Is there anyone familiar with this model code in sas?

My outcomes is the copay for the insurance and my covariates would be the plan ID. I want to investigate the relationship between copay values and the list of covariates.

Thanks!

Re: Two part model for healthcare costs

cau83 — Wed, 08 Aug 2018 12:42:32 GMT

You may want to use some kind of regression model suitable for zero-inflated data (ZIP or ZINB, where P and NB are Poisson or Negative Binomial distribution).

Example using proc genmod here may be helpful

https://stats.idre.ucla.edu/sas/dae/zero-inflatedpoisson-regression/

Re: Two part model for healthcare costs

CC13 — Wed, 08 Aug 2018 18:37:36 GMT

Hi,

Thank you for the response. Just wondering are there any difference between ZIP and ZINP as the distribution type of the model? Can we use the FMM (probit + gamma/log) instead?

Thanks!

Re: Two part model for healthcare costs

cau83 — Wed, 08 Aug 2018 18:46:25 GMT

Visually I don't see a lot of difference b/w ZIP and ZINB-- I have not done extensive work where it ultimately mattered (only some exploratory analysis). Google will return a lot of information if you search for it.

I am not familiar with the FMM proc. If t explicitly handles these distributions then it may be worth a shot. Regardless of PROC/method, the modeling ZI data should be a two part process as you describe, where it's estimating separately whether or not it's zero; and if not, then estimating the non-zero value.

A similar concept exists with forecasting methods for intermittent count data (where zeroes often occur). I have more experience with this, but not in SAS as I do not have SAS Forecasting Server which is where those procedures live.

Re: Two part model for healthcare costs

CC13 — Sat, 11 Aug 2018 03:05:22 GMT

Thanks! That works.