lincom after mi estimate: stcox

[Julia Chou]

Hi,

I ran a Cox regression on multiply imputed data that includes multiplicative interaction terms between a continuous variable (BMI) and levels of a categorical variable (race). I am trying to obtain the hazard ratios and 95% confidence intervals for a 1-unit difference in the BMI for each level of race. For example, I am trying to combine the coefficients for BMI and say BMI*race2 and report them as hazard ratios with 95% CIs. To do that, I tried using -mi estimate, post: lincom- but received an error saying that lincom is not a supported postestimation command for -mi estimate-. I would greatly appreciate any insight or suggestions on how I may calculate point estimates and confidence intervals of linear combinations of coeffcients after -mi estimate: stcox- in Stata.

Thank you very much!
Julia Chou

[Dave Airey]

Does the mi estimate postestimation section in the MI manual address this? It says the post estimation commands are mi test, mi testtransform, mi predict, and mi predictnl. I think the last commands has options for confidence intervals, skimming the help.

[Timothy Mak]

lincom is not officially supported for mi analyses. I think this is because there isn't a clear way to calculate appropriate degrees of freedom for results arising from lincom. If you are not worried about small sample adjustments, though, you can include the post option in your mi estimate, and use lincom afterwards. Using the example in the help file:

webuse mheart1s20
mi estimate, dots post: logit attack smokes age bmi hsgrad female
lincom age + bmi

Note that while the original results give t-statistics, the lincom results give z-statistics. This is because the small sample adjustments are ignored. You can have some idea of whether this is likely to be a problem by examining e(df_mi) after mi estimate. If all the values are fairly large (e.g. > 20 or so), it's probably not going to be a problem. (Don't quote me on this though!)

[Timothy Mak]

Actually, perhaps a better strategy is to mi estimate your lincom results, e.g.:

webuse mheart1s20
mi estimate (_b[age] + _b[bmi]), dots: logit attack smokes age bmi hsgrad female

Note that the syntax used is that of nlcom, not lincom.

[Julia Chou]

Thank you! Using mi estimate itself to combine the coefficients worked perfectly. Thank you again!