Calculating inverse probability weights after multiple imputation

[Jasper Tromp]

Dear all,

I am trying to calculate inverse probability weights after multiple imputations due to missingness in a number of linear confounders. I am hoping for some feedback on my approach, given that this is my first try and nobody around me has any experience. I have done an extensive search on stata list as well as stack exchange and went through the whole stata manual on mi, but could unfortunately not find a solution. The code that I am currently using is the following:

##MI settings
mi set mlong

##registering missing continuous variables
mi register imputed age bmi sytolic_bp

##imputing missing data
mi impute chained (regress) age bmi systolic_bp = country copd sex kidney_disease treatment, add(5) rseed(1235) 

mi estimate, eform saving(saved) : logistic treatment age bmi systolic_bp country copd sex kidney_disease 

#predicting linear predictors
mi predict pi using saved

#calculating probabilities
quietly mi xeq: generate phat2 = invlogit(p3)

#calculating inverse probability weights
quietly mi xeq: gen ipw2=. 
quietly mi xeq: replace ipw2= 1/phat2 if ace_bi==1
quietly mi xeq: replace ipw2= 1/(1-phat2) if ace_bi==0

#setting survival settings
mi stset time [pweight = ipw2], failure(death) 

#running survival analysis
mi estimate, eform: stcox i.treatment

[Andrea Discacciati]

Jasper Tromp

Dear Jasper, the authors of this recent publication http://journals.sagepub.com/doi/full...62280217713032 suggest the approach where "The treatment effect estimates from each imputed dataset are then combined to obtain an overall estimate" as opposed to the approach where "the propensity scores are combined across the imputed datasets" (which seems the one you adopted in the example above). What do you think?

(As a side note, this approach complicates things if applied to the situation where, instead of using IPW, one wishes to use PS matching to estimate the ATT. Of course, it's always possible to use SMR weights to estimate the ATT).