Error message when introducing factor variables in stcox on imputed data

[Sigrid Vikjord]

Dear Statalist,

I'm running stcox on multiply imputed data. I have used the same model and syntax on non-imputed data, and it works fine. But when I try to fit the model using mi estimate, I get the following error message:

mi estimate, hr: stcox i.HADS_A_gli c.PartAg_gli if Sex==0

HADS_A_gli:  factor variables may not contain noninteger values
an error occurred when mi estimate executed stcox on m=1
r(452);

It works fine when I run HADS_A_gli as a continuous variable. I also tried to specify the reference group using ib1.HADS_A_gli, without luck. Do anyone know why it's not possible to run stcox with a factor variable?



The imputation output looks like this:

. mi impute mvn HADS_A_gli HADS_D_gli Sex PartAg_gli SmoStatPackYrs_gli_missing SES_gli physact_gli alcoh
> ol_gli COPDcat_gli CVD_gli cancer_gli chrondisADL_gli diabetes_gli musc_skel_gli if (glicopd_HUNT==1 | 
> glicopd_HUNT==2), add(10) rseed (53421)
note: variables Sex PartAg_gli COPDcat_gli CVD_gli cancer_gli chrondisADL_gli diabetes_gli
      musc_skel_gli contain no soft missing (.) values; imputing nothing

Performing EM optimization:
  observed log likelihood = -6277.4486 at iteration 13

Performing MCMC data augmentation ... 

Multivariate imputation                     Imputations =       10
Multivariate normal regression                    added =       10
Imputed: m=1 through m=10                       updated =        0

Prior: uniform                               Iterations =     1000
                                                burn-in =      100
                                                between =      100

------------------------------------------------------------------
                   |               Observations per m             
                   |----------------------------------------------
          Variable |   Complete   Incomplete   Imputed |     Total
-------------------+-----------------------------------+----------
        HADS_A_gli |       1962          586       586 |      2548
        HADS_D_gli |       2169          379       379 |      2548
               Sex |       2548            0         0 |      2548
        PartAg_gli |       2548            0         0 |      2548
    SmoS~i_missing |       2434          114       114 |      2548
           SES_gli |       2365          183       183 |      2548
       physact_gli |       2532           16        16 |      2548
       alcohol_gli |       2383          165       165 |      2548
       COPDcat_gli |       2548            0         0 |      2548
           CVD_gli |       2548            0         0 |      2548
        cancer_gli |       2548            0         0 |      2548
    chrondisADL_~i |       2548            0         0 |      2548
      diabetes_gli |       2548            0         0 |      2548
     musc_skel_gli |       2548            0         0 |      2548
------------------------------------------------------------------
(complete + incomplete = total; imputed is the minimum across m
 of the number of filled-in observations.)

. 
end of do-file

Best regards,
Sigrid

[Carlo Lazzaro]

Sigfrid:
I would check whether the imputation model created =>1 non-integer values in -i.HADS_A_gli-.

[daniel klein]

Carlo is right. Using mvn (multivariate normal) as the imputation method, you will most likely end up with all non-integer imputed values. If you want to impute categorical variables, you should use a chained equations approach (chained) with logit, ologit, or mlogit models.

Also, since you want to use a Cox model, you should probably include an estimate for the cumulative baseline hazard in the imputation model (White and Royston 2009). I have no personal experience with this, so I cannot advise further.

Best
Daniel


White, I., R., and Royston, P. 2009. Imputing missing covariate values for the Cox model. Statistics in Medicine, 28(15), 1982--1998.

[Sigrid Vikjord]

Thank you for your helpful input, Carlo and Daniel! I will try to solve it using your advice, and post an update if I manage to get through with it.