Propensity score adjustment using multiple imputation (and missing indicator)

[jemima spencer]

Hi,

Could I possibly have some help using multiple imputation within a propensity score model?

Sorry for lack of dataex - my STATA work is all on a secure server so copied here.

I have read around the topic, and I think(!) what I should be doing (in view of the pattern of missingness and presence of unmeasured confounding), is to use a combination of multiple imputation and the missing indicator method to develop my propensity score (which I then plan to adjust for in a logistic regression model to compare with multivariable adjustment).

So far, I have a multiple imputation model for ethnicity, and a separate propensity score model.

My propensity score model is as follows:

stepwise, pr(0.2): ologit i.ckdstage_grp i.ageyears_grp ethnic_white obese diabetes i.smoker gender cvd

predict propensity_NST

And my multiple imputation model (if relevant) is:

mi set wide

mi register imputed ethnic_white

mi register regular obese ckdstage_grp ageyears_grp COPD diabetes smoker gender cvd

mi impute chained (logit) ethnic_white = obese ckdstage_grp gender, add(20) rseed(2232) (I added obese, ckdstage_grp and gender as highly correlated with missingness)


Could you possibly tell me how I would rephrase my propensity model to include the multiply imputed ethnicity data (if this sounds vaguely correct!)? I think I can manage the missing indicator aspect.

Thankyou so much

Jemima

[daniel klein]

My propensity score model is as follows:

stepwise, pr(0.2): ologit i.ckdstage_grp i.ageyears_grp ethnic_white obese diabetes i.smoker gender cvd

predict propensity_NST

That code is not going to work. You cannot use factor variable notation for the dependent (right-hand-side) variable.

Assuming the code did work, you would only get the predicted probabilites to fall into the first category of ckdstage. What about the other categories? How exactly are you planning to include the predictions (note plural!) for multiple categories (note plural!) as propensity scores (note plural!) in the regression models?

Also, I would never use stepwise here. Why not include all variables in the matching model?

And my multiple imputation model (if relevant) is:

mi set wide

mi register imputed ethnic_white

mi register regular obese ckdstage_grp ageyears_grp COPD diabetes smoker gender cvd

mi impute chained (logit) ethnic_white = obese ckdstage_grp gender, add(20) rseed(2232) (I added obese, ckdstage_grp and gender as highly correlated with missingness)

The imputation model looks too small. It should include at least all variables that are used in the substantive analyses. Otherwise, the associations with excluded variables are biased towards zero.

[jemima spencer]

Thanks Daniel

(And sorry - you seem to be doing all my analyses!)

a) Correct - I didn't include the factor variable in my propensity score model, sorry I incorrectly added it here when copying.

b) Honest answer is that this is the method that others in my working group have used, so I didn't look too much further into it. The variables (not all listed here) are those that are associated with exposure and outcome, or outcome only, which I believe is right but am ready to be corrected on.

c) MI model - I've only included variables that were highly correlated with missingness. Unsure where I read that, and now reading extensively to the contrary!