telasso does not work with mi estimate

[Felix Bittmann]

Dear all,
I am fully aware of the fact that mi estimate does not officially support telasso, yet I think this seems to me like a programming bug and not a question of statistical validity. The setup is as follows:

clear all
version 18.0
sysuse nlsw88

* Imputation
mi set flong
mi register imputed union wage hours tenure south
mi impute chained (pmm, knn(4)) union wage hours tenure south ///
    , add(3) rseed(123) dots
    
* Estimation
mi estimate, cmdok dots: ///
    telasso (wage tenure south hours, linear) (union tenure south, logit)

The output is this:

. clear all

. version 18.0

. sysuse nlsw88
(NLSW, 1988 extract)

. 
. * Imputation
. mi set flong

. mi register imputed union wage hours tenure south
(379 m=0 obs now marked as incomplete)

. mi impute chained (pmm, knn(4)) union wage hours tenure south ///
>         , add(3) rseed(123) dots
note: variables wage south contain no soft missing (.) values; imputing nothing

Conditional models:
             hours: pmm hours wage south tenure union , knn(4)
            tenure: pmm tenure wage south hours union , knn(4)
             union: pmm union wage south hours tenure , knn(4)

Performing chained iterations:
  imputing m=1 through m=3 ... done

Multivariate imputation                     Imputations =        3
Chained equations                                 added =        3
Imputed: m=1 through m=3                        updated =        0

Initialization: monotone                     Iterations =       30
                                                burn-in =       10

             union: predictive mean matching
              wage: predictive mean matching
             hours: predictive mean matching
            tenure: predictive mean matching
             south: predictive mean matching

------------------------------------------------------------------
                   |               Observations per m             
                   |----------------------------------------------
          Variable |   Complete   Incomplete   Imputed |     Total
-------------------+-----------------------------------+----------
             union |       1878          368       368 |      2246
              wage |       2246            0         0 |      2246
             hours |       2242            4         4 |      2246
            tenure |       2231           15        15 |      2246
             south |       2246            0         0 |      2246
------------------------------------------------------------------
(Complete + Incomplete = Total; Imputed is the minimum across m
 of the number of filled-in observations.)

.         
. * Estimation
. mi estimate, cmdok dots: ///
>         telasso (wage tenure south hours, linear) (union tenure south, logit)

Imputations (3):
  .invalid syntax
r(198);

Apparently, the process fails after the first imputation. I then used the nois option to check what is going on. The concrete error:

file /tmp/St02964.000001.stxer already exists

Of course, deleting this file manually does not solve the issue as a new one is created in the estimation loop over all imputed samples every time. I wonder what happens here, as mi estimate only collects the results over the various samples, I don't understand why telasso needs to keep some file in memory once the statistics of interest have been generated. What one can do is to estimate telasso separately for each imputed sample, collect the statistics of interest manually and apply Rubin's rules. But this should work in the automated way as well.

[Jeff Wooldridge]

Because the asymptotic arguments underlying LASSO are pretty intricate even when using complete cases, perhaps there are statistical reasons for not allowing multiple imputation with telasso. If it is more of a coding bug then I'd be curious as to how the uncertainly from imputation is being handled.

[Enrique Pinzon (StataCorp)]

Dear Felix and Jeff,

The command is not supported with -mi estimate-. The issue in the code above, is that Felix used the option -cmdok- which bypasses the checks we put in place. -cmdok- is used to bless an unsupported command, a blessing that should be bestowed carefully and that we did not provide originally. If you run the code without the option -cmdok-, it will error out telling you that -telasso- is not supported.