PROC ARIMA with binary response

[jp134711]

Is it possible to combine a binary response variable (e.g. whether or not a patient is readmitted) with a time series model using PROC ARIMA?

Thanks!

[Reeza]

No idea, but it sounds more like survival analysis.

What doesn't survival analysis cover that you would want in ARIMA, some sort of seasonality?

[jp134711]

I'm interested in investigating the effect of an intervention on readmission rate, after controlling for patient-level covariates. I have about 3 years of historical/pre-intervention data and 1.5 years of post-intervention data. I want to incorporate time trend to account for changes in medical practice over time and its relation with readmission.

[SteveDenham]

That sounds, at least to me, more like a survival analysis with a time-dependent covariate, as proposed by @Reeza.  Think about what the ARIMA model would be fitting--a long string of zeroes, a single 1, perhaps some more 1's (if you model as still admitted), then another long string of zeroes.  That is not a good dataset for fitting an ARIMA model.  Instead, time to re-admission, with a covariate that describes the intervention status, strikes me as something that would work.  Check out PROC PHREG.

Steve Denham

[Reeza]

I would add some indicator variables, possible time dependent to account for the changes in practice.

You'll have to be careful with the pre-intervention/post-intervention data to make sure they're handled appropriately, but survival analysis is what you're looking for.

[SteveDenham]

, that's what I was trying to say.  I think a record would look like:

subjid      date      admission_status      intervention_status      covariate1      covariate2      (other covariates of interest).

That should set it up for a survival analysis, as per Example 67.7 Time-Dependent Repeated Measurements of a Covariate.

Steve Denham