I am looking to build a Cox proportional survival hazard model in SAS on time dependent covariates. Ultimately, I will put this model in production and score on recurring basis to target the right clients for cross-selling a home loan product.
My data is unique at Client Quarter level. I have an unbalanced panel data and each client can have anywhere between 1 to 8 records (2 years of time series) depending upon whether client was censored or observed the event (home loan). I have start and stop values (1 and 90 for Q1, 91 and 180 for Q2, 181 and 270 for Q3 etc.) at each Client Quarter level (basically each row) combination. My SAS model code looks some thing like
proc Phreg data = abc; model (start, stop) * home_loan (0) = X1 X2 X3...; run;
While I have some hypothesis, I am also leveraging a lot of other client attributes to test. In all I have close to 200 variables. My questions are:
- What bivariate variable selection can one apply in this situation? Correlation, Info value etc.
- How can multi collinearity be controlled?
- What could be the model performance and validation metrics? How can one determine if the model fits well and generalizes well?