[Limdep Nlogit List] \MNL and marginal effects with NLOGIT?

William Greene wgreene at stern.nyu.edu
Wed Aug 13 01:11:55 EST 2008


Kandice. In earlier versions of NLOGIT, you were only given the means
of the partial effects, in the form of a table that would give them for
the four levels in a template tree for a nested logit model. I.e.,
if you have an MNL, only the lowest level would be used and the effects
at the other levels would be zero.  In the version you are using, the
program notes that you are fitting an MNL which is a one level model, and
it does not use the artificial tree structure. Rather, it gives you both
means and sample standard deviations of individual specific partial effects.
When you switch to a true nested logit model, you will then get the 
expected table of effects for a tree structure.
I would note, you are computing derivatives (times 100) of the probabilities,
as your partial effects. These might not be all that meaningful. You can
get elasticities, which are more naturally scaled, by changing the brackets
to parentheses in your ;EFFECTS:... specification.
/B. Greene

----- Original Message -----
From: "Kandice L. Kleiber" <kleiberk at onid.orst.edu>
To: limdep at limdep.itls.usyd.edu.au
Sent: Tuesday, August 12, 2008 11:00:41 AM GMT -05:00 US/Canada Eastern
Subject: [Limdep Nlogit List] \MNL and marginal effects with NLOGIT?

Hi ... I'm new to LimDep/NLOGIT so please bear with me!  I'm trying to set
up a nested logit piece by piece (ie add in a few variables at a time).
Because I'm waiting on same data that is choice specific data, I'm basically
setting up a multinomial logit using NLOGIT.  The data that I'm using is
observation specific.

I'm trying to set up the model so it's ready to use once I get the rest of
the choice specific data, but until then I would like to keep using what I
have.  My problem is in calculating marginal effects.  I know that if I use
the LOGIT command for my MNL I just need ;marginal effects.  I think if I
had the choice specific data I could use NLOGIT and I would need ;effects
price[*] ... along those lines.  But when I use ;effects in NLOGIT my output
doesn't make much sense - I'm given mean and standard deviation, which does
not look anything like the LimDep manual has.  I think this is because I'm
basically taking the marginal effects of a MNL in NLOGIT.  But does this
matter?  My code and output is below ... price is contant across all choices
but is set up in a matrix  with the value down the diagonal.  Same with the
other variables.

Please help!  Is there a way to continue down this path, or should I go back
to LOGIT until I get my data to create a nested logit?
Thank you!!

--> READ
    ; FILE = D:\LIMDEPDATA\ATTEMPT201.XLS
    ; NVAR = 15 ; NOBS = 1000
    ; NAMES = 1$
--> NLOGIT
    ; LHS = PROP
    ; CHOICES = cORNCORN, SOYCORN, CORN, SOY
    ; RHS = ONE, PRATIO1, PRATIO2, PRATIO3, HEL1, HEL2, HEL3, SL011, SL012,
S...
    ; EFFECTS: PRATIO1[*] / PRATIO2[*] / PRATIO3[*]$



+---------------------------------------------+
| Discrete choice and multinomial logit models|
+---------------------------------------------+
Normal exit from iterations. Exit status=0.
+---------------------------------------------+
| Discrete choice (multinomial logit) model   |
| Maximum Likelihood Estimates                |
| Model estimated: Aug 12, 2008 at 10:54:59AM.|
| Dependent variable               Choice     |
| Weighting variable                 None     |
| Number of observations              106     |
| Iterations completed                  5     |
| Log likelihood function       -113.7578     |
| Number of parameters                 12     |
| Info. Criterion: AIC =          2.37279     |
|   Finite Sample: AIC =          2.40444     |
| Info. Criterion: BIC =          2.67431     |
| Info. Criterion:HQIC =          2.49500     |
| R2=1-LogL/LogL*  Log-L fncn  R-sqrd  RsqAdj |
| Constants only.  Must be computed directly. |
|                  Use NLOGIT ;...; RHS=ONE $ |
| Chi-squared[ 9]          =      1.82943     |
| Prob [ chi squared > value ] =   .99389     |
| Response data are given as proportions.     |
| Number of obs.=   106, skipped   0 bad obs. |
+---------------------------------------------+
+---------------------------------------------+
| Notes No coefficients=> P(i,j)=1/J(i).      |
|       Constants only => P(i,j) uses ASCs    |
|         only. N(j)/N if fixed choice set.   |
|         N(j) = total sample frequency for j |
|         N    = total sample frequency.      |
|       These 2 models are simple MNL models. |
|       R-sqrd = 1 - LogL(model)/logL(other)  |
|       RsqAdj=1-[nJ/(nJ-nparm)]*(1-R-sqrd)   |
|         nJ   = sum over i, choice set sizes |
+---------------------------------------------+
+--------+--------------+----------------+--------+--------+
|Variable| Coefficient  | Standard Error |b/St.Er.|P[|Z|>z]|
+--------+--------------+----------------+--------+--------+
 PRATIO1 |    65.1147789      113.523892      .574   .5663
 PRATIO2 |   -17.9414075      35.6312734     -.504   .6146
 PRATIO3 |    11.3707783      27.9534400      .407   .6842
 HEL1    |    -.85499804      2.44543336     -.350   .7266
 HEL2    |     .03044104       .67727779      .045   .9642
 HEL3    |     .00948316       .53309998      .018   .9858
 SL011   |   -3.44864601      7.49912020     -.460   .6456
 SL012   |    1.51466998      2.68596449      .564   .5728
 SL013   |    -.48910971      2.06371694     -.237   .8127
 A_CORNCO|   -64.9198907      111.509720     -.582   .5604
 A_SOYCOR|    15.4826457      34.7911622      .445   .6563
 A_CORN  |   -11.1148123      27.3848457     -.406   .6848
+---------------------------------------------------+
| Derivative (times 100) averaged over observations.|
| Attribute is PRATIO1  in choice CORNCORN          |
| Effects on probabilities of all choices in model: |
| * = Direct Derivative effect of the attribute.    |
|                                  Mean    St.Dev   |
| *     Choice=CORNCORN         84.6351   59.7107   |
|       Choice=SOYCORN         -12.9924    6.1500   |
|       Choice=CORN            -31.1550   25.5122   |
|       Choice=SOY             -40.4877   28.3915   |
+---------------------------------------------------+
+---------------------------------------------------+
| Derivative (times 100) averaged over observations.|
| Attribute is PRATIO1  in choice SOYCORN           |
| Effects on probabilities of all choices in model: |
| * = Direct Derivative effect of the attribute.    |
|                                  Mean    St.Dev   |
|       Choice=CORNCORN        -12.9924    6.1500   |
| *     Choice=SOYCORN         921.0375  157.5860   |
|       Choice=CORN           -376.5597   47.7207   |
|       Choice=SOY            -531.4850  115.8219   |
+---------------------------------------------------+
+---------------------------------------------------+
| Derivative (times 100) averaged over observations.|
| Attribute is PRATIO1  in choice CORN              |
| Effects on probabilities of all choices in model: |
| * = Direct Derivative effect of the attribute.    |
|                                  Mean    St.Dev   |
|       Choice=CORNCORN        -31.1550   25.5122   |
|       Choice=SOYCORN        -376.5597   47.7207   |
| *     Choice=CORN           1458.0870   66.6691   |
|       Choice=SOY           -1050.3730   91.3753   |
+---------------------------------------------------+
+---------------------------------------------------+
| Derivative (times 100) averaged over observations.|
| Attribute is PRATIO1  in choice SOY               |
| Effects on probabilities of all choices in model: |
| * = Direct Derivative effect of the attribute.    |
|                                  Mean    St.Dev   |
|       Choice=CORNCORN        -40.4877   28.3915   |
|       Choice=SOYCORN        -531.4850  115.8219   |
|       Choice=CORN          -1050.3730   91.3753   |
| *     Choice=SOY            1622.3460    3.2597   |
+---------------------------------------------------+


-
_______________________________________________
Limdep site list
Limdep at limdep.itls.usyd.edu.au
http://limdep.itls.usyd.edu.au



More information about the Limdep mailing list