Marginal effect after cmmixlogit

[Jean Dewals]

Dear collegues,

I'm currently working on a consumers survey. In this survey, we asked consumers to rank their preferred labelling scheme among a pool of nine alternatives (and a no-alternative option). For each consumers, we thus have their most prefered option among 10.

After that, I wanted to analyse the answers provided to better understand how consumers' preferences are structured. To do so, i used the cmmixlogit command.

My data are organised as follow :

- ID identifies each respondent (here, i present data only for my first two respondents)
- Rang: Equal to 1 for the chosen option, otherwise equal to 0
- Alternatives = Associated alternatives
- The remaining variables are related to individual characteristics.

For exemple, my first respondent ranked the health allegation as his/her favourite scheme. The second consumer did the same.

I used the followind code line to launched my mixed logit choice model

cmset ID alternatives
cmmixlogit rang, casevars(i.recode_age i.genre deplittoral prix_crit enfant univ tradi prefgeo) basealternative (10)

I get consistent results. But after this step, i want to measure the marginal effect of explanatory variables on the preference expressed.

I thus tried to use first the following command

margins, dydx(*)

Unfortunately, I get the following error message :

alternative Com.equi is not in the fitted model; error in option outcome()
r(458);

I then tried to used the same command, but using only one explanatory variable

margins, dydx(prix_crit)

I also get an error message :

no cases remain after removing invalid observations
r(2000);


So I'm unable to analyse my marginal effects, and I don't see how to get around these problems. Has anyone had this problem before? Do you have any potential solutions for me?

Thank you in advance for your time

Best regards

Jean-François

[Joerg Luedicke (StataCorp)]

Is alternatives a string variable? If so, use help encode to create a numeric variable from it, then run your analysis again using the numeric variable. In theory, it is allowed to use a string variable there but there might be something weird about these particular strings, something we will have to look into.

That said, notice that if you exclusively use case-specific covariates, the mixed logit model boils down to the special case of simple logit. Have a look at the following example, the results from cmmixlogit and mlogit are equivalent:

. webuse inschoice
(Fictional health insurance data)

. cmset id insurance

     Case ID variable: id
Alternatives variable: insurance

. cmmixlogit choice, casevars(income)

Fitting fixed parameter model:

Fitting full model:

Iteration 0:   log likelihood = -394.22042  
Iteration 1:   log likelihood = -394.22042  

Mixed logit choice model                       Number of obs      =      1,250
Case ID variable: id                           Number of cases    =        250

Alternatives variable: insurance               Alts per case: min =          5
                                                              avg =        5.0
                                                              max =          5

Integration points:              0                Wald chi2(4)    =       5.25
Log likelihood =        -394.22042                Prob > chi2     =     0.2629

------------------------------------------------------------------------------
      choice | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
Health       |
      income |   .0762443   .1333404     0.57   0.567    -.1850981    .3375868
       _cons |  -.6201366   .7138245    -0.87   0.385    -2.019207    .7789337
-------------+----------------------------------------------------------------
HCorp        |  (base alternative)
-------------+----------------------------------------------------------------
SickInc      |
      income |  -.1176294    .129578    -0.91   0.364    -.3715976    .1363388
       _cons |   .4319901   .6677559     0.65   0.518    -.8767873    1.740768
-------------+----------------------------------------------------------------
MGroup       |
      income |  -.1652967   .1353851    -1.22   0.222    -.4306466    .1000531
       _cons |   .5032474   .6894884     0.73   0.465     -.848125     1.85462
-------------+----------------------------------------------------------------
MoonHealth   |
      income |  -.2203374   .1508352    -1.46   0.144    -.5159689    .0752941
       _cons |   .4075541   .7539384     0.54   0.589    -1.070138    1.885246
------------------------------------------------------------------------------

. mlogit insurance income if choice == 1

Iteration 0:   log likelihood = -396.92859  
Iteration 1:   log likelihood = -394.22974  
Iteration 2:   log likelihood = -394.22042  
Iteration 3:   log likelihood = -394.22042  

Multinomial logistic regression                         Number of obs =    250
                                                        LR chi2(4)    =   5.42
                                                        Prob > chi2   = 0.2472
Log likelihood = -394.22042                             Pseudo R2     = 0.0068

------------------------------------------------------------------------------
   insurance | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
Health       |
      income |   .0762443   .1333404     0.57   0.567    -.1850981    .3375868
       _cons |  -.6201366   .7138245    -0.87   0.385    -2.019207    .7789338
-------------+----------------------------------------------------------------
HCorp        |  (base outcome)
-------------+----------------------------------------------------------------
SickInc      |
      income |  -.1176294    .129578    -0.91   0.364    -.3715976    .1363388
       _cons |   .4319901   .6677559     0.65   0.518    -.8767873    1.740768
-------------+----------------------------------------------------------------
MGroup       |
      income |  -.1652967   .1353851    -1.22   0.222    -.4306466    .1000531
       _cons |   .5032474   .6894884     0.73   0.465     -.848125     1.85462
-------------+----------------------------------------------------------------
MoonHealth   |
      income |  -.2203374   .1508352    -1.46   0.144    -.5159689    .0752941
       _cons |   .4075541   .7539384     0.54   0.589    -1.070138    1.885246
------------------------------------------------------------------------------

[Jean Dewals]

Dear Collegues,

Thanks for your help, I managed to get my marginal effects!

Best regards,

Jean-François