Logit option in teffects outcome model returns r322 error unexpectedly

[Hannah Beilby]

Hello,

Prior to today, the below code ran fine. This morning I ran some updates, and am now running into the same error, which disappears when logit is removed (i.e. a linear outcome model)

. teffects ipwra (nllonely $xlist, logit)           ///
>                            ($treatment $xlist pnq3), vce(robust)        
outcome model: perfect predictions detected; the model, as specified, is not identified
    1.source#1.singleparent != 0 (n=7) predicts failure perfectly
    1.source#1.otherhhtype != 0 (n=2) predicts failure perfectly
    1.source#c.dseifa4 != 0 (n=13) predicts failure perfectly
    1.source#c.dseifa8 != 0 (n=10) predicts failure perfectly
    1.source#c.dseifa9 != 0 (n=25) predicts failure perfectly

I don't really understand why this is the case.. does anyone know of a work around (noting "asis" option in logit will not run within teffects)?

Many thanks,
Hannah

[Hannah Beilby]

Hi all,

Just want to clarify that my covariates are not interacted:

global xlist "male dagecat2 dagecat3 dagecat4 dagecat5 dagecat6 fulltime parttime unemployed postgrad cert someuni livesalone couplewdeps singleparent houseshare otherhhtype $seifa"

Therefore I don't think I should be receiving the error with interacted variables?

Also, the logit outcome model is running fine separately:

logit nllonely $xlist

Iteration 0:  Log likelihood = -2449.0117  
Iteration 1:  Log likelihood =  -2397.881  
Iteration 2:  Log likelihood =  -2394.827  
Iteration 3:  Log likelihood = -2394.8159  
Iteration 4:  Log likelihood = -2394.8159  

Logistic regression                                     Number of obs =  8,011
                                                        LR chi2(26)   = 108.39
                                                        Prob > chi2   = 0.0000
Log likelihood = -2394.8159                             Pseudo R2     = 0.0221

------------------------------------------------------------------------------
    nllonely | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
        male |    -.04833   .0830371    -0.58   0.561    -.2110797    .1144196
    dagecat2 |  -.1936724    .135027    -1.43   0.151    -.4583205    .0709757
    dagecat3 |  -.1880247   .1385637    -1.36   0.175    -.4596047    .0835552
    dagecat4 |  -.2746864   .1330054    -2.07   0.039    -.5353722   -.0140006
    dagecat5 |  -.2550917   .1461777    -1.75   0.081    -.5415947    .0314112
    dagecat6 |   -.345141   .1604027    -2.15   0.031    -.6595245   -.0307575
    fulltime |  -.3730588   .1165145    -3.20   0.001     -.601423   -.1446945
    parttime |  -.0995582   .1192863    -0.83   0.404    -.3333551    .1342387
  unemployed |   .2794165   .2121323     1.32   0.188    -.1363553    .6951882
    postgrad |  -.6269923   .2150008    -2.92   0.004    -1.048386   -.2055986
        cert |    .087169   .0946687     0.92   0.357    -.0983782    .2727163
     someuni |   .0057425   .1117026     0.05   0.959    -.2131905    .2246755
  livesalone |    .480646   .1101299     4.36   0.000     .2647954    .6964966
 couplewdeps |   .0763031   .1141998     0.67   0.504    -.1475245    .3001306
singleparent |   .5657626   .1434381     3.94   0.000     .2846291    .8468961
  houseshare |  -.1859757   .4008341    -0.46   0.643    -.9715962    .5996447
 otherhhtype |   .2334798   .2010009     1.16   0.245    -.1604747    .6274344
     dseifa1 |   .3739698   .1795043     2.08   0.037     .0221479    .7257918
     dseifa2 |   .3077496   .1743917     1.76   0.078    -.0340519    .6495511
     dseifa3 |   .1942783   .1747092     1.11   0.266    -.1481453     .536702
     dseifa4 |    .123598   .1834911     0.67   0.501    -.2360378    .4832339
     dseifa5 |  -.0943419   .1978045    -0.48   0.633    -.4820315    .2933478
     dseifa6 |   .0362833   .1829781     0.20   0.843    -.3223473    .3949138
     dseifa7 |   .1445419   .1816156     0.80   0.426    -.2114181     .500502
     dseifa8 |   .0703278   .1845394     0.38   0.703    -.2913627    .4320183
     dseifa9 |  -.1868921   .1972172    -0.95   0.343    -.5734306    .1996465
       _cons |  -2.181734   .1964907   -11.10   0.000    -2.566849   -1.796619

Therefore, suspect it may have something to do with a teffects ipwra update?

Perhaps related to the update 10may2022 entry:

13. predict after logit and probit, when a double-precision variable was omitted because of perfect prediction,
could return missing values for observations where it should not have been missing. This has been fixed.

Can anyone reassure either way whether my initial results were valid? They are below (apologies for screenshot could no longer C+P the code):

Any wisdom at all is greatly appreciated.

Kind regards,
Hannah

[Hannah Beilby]

Hi again,

FYI - I’ve had a response from Stata in relation to this…“This relates to one of our previous bugs that happened when collinear predictors exist in the model. The old behavior reported incorrect results. This has been fixed and an error message will be given in this scenario.”

Relevant update from 04oct2022: teffects with an outcome or treatment model logit, probit, or hetprobit that contained perfect predictors either reported incorrect estimates or did not converge. teffects now issues an error if perfect predictors occur in either the treatment or the outcome model. A warning is issued for the outcome model if completely determined outcomes are detected.

Kind regards,
Hannah