Augmented inverse probability weighing analysis

[amentias]

What is the difference between:

1- Augemented IPW with proc causaltreat statment 

2- Doing multivariable cox regression with proc phreg with Adding a weight statment with IPW weight?

proc phreg data=***;

weight ipw;

model Time_to_event*mortality(0)= variable1 variable2 ...... / rl;

run;

I prefer the second one because it takes into account time to event analysis, while the first way seems to not be taking into account time to event. 

Any advice?

thanks

[MichaelL_SAS · Accepted solution]

Good question, and you’ve already identified the most important difference, namely that the estimation methods PROC CAUSALTRT supports are not designed specifically for time to event analysis. In particular, PROC CAUSALTRT does not account for censoring or support the distributions most often used to model time to event data.

The difference between the augmented inverse probability weighting (AIPW) method in PROC CAUSALTRT versus just inverse probability weighting involves how the model for the outcome is used. The AIPW approach solves a set of estimating equations that include an “augmentation” term that gives the method it’s doubly robust property. Basically, that term involves predicted values from the outcome model, assuming the different treatment assignments, multiplied by some weighted error from the treatment model. The details on the estimating equations solved by the AIPW are available in the Doubly Robust Estimation section of the PROC CAUSALTRT documentation.