choice-metrics.com

Hi there,

I am using Nlogit to estimate the employment preferences of community health workers. The survey instrument included 5 attributes (supervision format; training frequency; monthly financial benefit; recognition format; and employment structure) and we included an unforced choice (i.e., with an opt-out option) followed by a forced choice. We reached a large sample (471).

The MNL model and mixed MNL model run fine for the unforced choices, yet I run into issues for the forced choices – the results return ‘fixed parameters’ for several attributes.

Is there any way to overcome this? I haven’t found any errors in my data and haven’t been able to find any solutions in the manual or online. Any help or feedback would be very appreciated.

My code for the MNL model and results are copied here:

Code: Select all

|-> sample;all$
|-> REJECT ;cset=3 $
|-> Nlogit
    ;lhs = choice
    ;choices = jobA, jobB
    ;checkdata
    ;model:
    U(jobA) = a + supervis*supervis + training*training + twntyfv*twntyfv + hndrd*hndrd + thrhnd*thrhnd + endrsmnt*endrsmnt + award*award + report*report + employme*employme /
    U(jobB) = supervis*supervis + training*training +  twntyfv*twntyfv + hndrd*hndrd + thrhnd*thrhnd + endrsmnt*endrsmnt + award*award + report*report + employme*employme
    $


Code: Select all

No bad observations were found in the sample

Hessian is not definite at current values.
Switching to BFGS (gradient based) method.
(Not a failure. Just looking for a better algorithm.)
Line search at iteration  8 does not improve the function
Exiting optimization

-----------------------------------------------------------------------------
Discrete choice (multinomial logit) model
Dependent variable               Choice
Log likelihood function     -3868.65298
Estimation based on N =   5652, K =  10
Inf.Cr.AIC  =   7757.3 AIC/N =    1.372
---------------------------------------
            Log likelihood R-sqrd R2Adj
Constants only  -3917.6250  .0125 .0108
Note: R-sqrd = 1 - logL/Logl(constants)
---------------------------------------
Chi-squared[ 9]          =     97.94413
Prob [ chi squared > value ] =   .00000
Response data are given as ind. choices
Number of obs.=  5652, skipped    0 obs
--------+--------------------------------------------------------------------
        |                  Standard            Prob.      95% Confidence
  CHOICE|  Coefficient       Error       z    |z|>Z*         Interval
--------+--------------------------------------------------------------------
       A|    -.00718         .99971     -.01  .9943    -1.96658   1.95223
SUPERVIS|    -.00919         .23278     -.04  .9685     -.46542    .44705
TRAINING|     .02913    .....(Fixed Parameter).....
 TWNTYFV|     .20481    .....(Fixed Parameter).....
   HNDRD|    -.24500         .63703     -.38  .7005    -1.49356   1.00355
  THRHND|    -.01591    .....(Fixed Parameter).....
ENDRSMNT|    -.08617    .....(Fixed Parameter).....
   AWARD|     .01065    .....(Fixed Parameter).....
  REPORT|     .10456    .....(Fixed Parameter).....
EMPLOYME|     .09209        7.55602      .01  .9903   -14.71744  14.90162
--------+--------------------------------------------------------------------



Thanks

I forgot to mention that I'm estimating the 'monthly financial benefit' attribute categorically rather than continuously because the latter returns a parameter of -0.00 for both forced and unforced choices. When categorical, the parameters reveal a more interesting pattern (higher preference for lower amount)...

This is an Nlogit question that you posted on a forum for Ngene questions. I am not famliar with Nlogit and cannot answer your question. You could try asking on the "Choice experiments - general" forum, but it is best if you post your question on the Nlogit discussion board since Nlogit a product from a different company: https://www.limdep.com/listserver/

Note that you can estimate a single joint model for forced and unforced choices, where you have a mix of choice set sizes 2 and 3. I know that Nlogit users often do this. The only thing you will need to account for is heteroscedasticity by estimating a scale parameter for the forced choice (while normalising for the unforced choice to 1) since different choice set sizes influence scale.

Michiel

Thanks Michiel