>Is there in Stata a non-parametric regression procedure that allow
>non-parametric curve fitting of dose-response relationships with adjustment
>(parametrically or not) for covariates. The intent is to perform a
>non-parametric logistic regression analysis.
I believe if you interact all the covariates with each other, you will get
a nonparametric logistic regression. For variables you want as proportional
hazards (or parametric), _don't_ interact them with other variables.
The outcome (response) variable is "exit", the treatment (dose) variable is
"treat", T2-T12 are duration indicators (time-from-diagnosis=duration)
which are then interacted with "treat" to get a fully-interacted baseline
hazard, and all the other covariates in the model are assumed to act
proportionately.
Another way is to use the "if" option and only run the model conditional on
a particular treatment (or dose), say. This creates a fully-interacted
baseline hazard model without creating all the interacted covariates. But,
this does not allow you to formally test for differences across treatments
(doses) except perhaps by bootstrapping.
The outcome (response) variable is "exit", and the treatment (dose)
variable is "treat". In this way "treat" is fully interacted with all the
covariates.
Remember if you have more than one observation per individual, then you'll
want to cluster on ID.
Others might have easier solutions. I hope this is helpful.
