[R] wrong signs using MNP.

[Kosuke Imai]

Dear Bill,
  You might want to check whether the parameterization of MNP differs from
the book you mention, which I don't have. In particular, I would check how
the choice specific covariate is calculated. More importantly, I would
also make sure that your chain has converged. Based on my experiences of
the multinomial probit model, 1500 draws is not sufficient. Of course,
this depends on one's model, data, and starting values. On this issue, you
might consult the accompanying paper as well as my Journal of Econometrics
article (both available at
http://www.princeton.edu/~kimai/research/MNP.html). The former will walk
you through how one might use MNP and coda packages to conduct convergence
diagnostics using multiple chains via some detailed examples.
Hope this helps,
Kosuke

---------------------------------------------------------
Kosuke Imai               Office: Corwin Hall 041
Assistant Professor       Phone: 609-258-6601 
Department of Politics    eFax:  973-556-1929
Princeton University      Email: kimai at Princeton.Edu
Princeton, NJ 08544-1012  http://www.princeton.edu/~kimai
---------------------------------------------------------

> Hi,
>   I recently found the MNP package.  Out of curiosity, I tried to reproduce
> results from Greene (Econometric Analysis, fourth edition) on page 874.
> 
> The signs of the estimates are all opposite those of Greene's table.  Might
> anyone be able to tell me what I am doing wrong?
> 
> I have attached the function call, the coefficients, and a few rows of the
> data.  The dataset has 210 observations.
> 
> Thanks,
> 
> --Bill
> 
> 
> res1<-mnp(mc ~ 1,
>   choiceX = list(a=cbind(ttme.1,gc.1,hinc),
>                  b=cbind(ttme.2,gc.2,0),
>                  c=cbind(ttme.3,gc.3,0),
>                  d=cbind(ttme.4,gc.4,0)
>             ),
>   cXnames = c("ttme","gc","hinc"),
>   data = wide,
>   n.draws = 1500,
>   burnin = 50,
>   base = "d",
>   verbose = TRUE
> )
> 
> Coefficients:
>                    mean  std.dev.      2.5%  97.5%
> (Intercept):a -1.003078  0.508752 -2.292813 -0.207
> (Intercept):b -1.511963  0.582518 -2.875069 -0.705
> (Intercept):c -1.241368  0.411084 -2.231260 -0.646
> ttme           0.022973  0.007515  0.011386  0.042
> gc             0.002175  0.001981 -0.001518  0.006
> hinc          -0.030616  0.006030 -0.042062 -0.018
> 
> Covariances:
>        mean std.dev.    2.5%  97.5%
> a:a  1.0000   0.0000  1.0000  1.000
> a:b -0.8026   0.4655 -1.8569 -0.252
> a:c -0.7246   0.3721 -1.8966 -0.305
> b:b  1.4315   1.9552  0.1265  7.973
> b:c  1.1526   1.7011  0.1585  7.276
> c:c  1.1281   1.6327  0.2184  7.682
> 
> Number of alternatives: 4
> Number of observations: 210
> Number of Gibbs draws: 1450
> 
> 
> The data:
> 
> mc ttme.1 ttme.2 ttme.3 ttme.4 gc.1 gc.2 gc.3 gc.4 hinc
>  d     69     34     35      0   70   71   70   30   35
>  d     64     44     53      0   68   84   85   50   30
>  d     69     34     35      0  129  195  149  101   40
>  d     64     44     53      0   59   79   81   32   70
>  d     64     44     53      0   82   93   94   99   45
>  b     69     40     35      0   70   57   58   43   20
>  a     45     34     35      0  160  213  167  125   45
>  d     69     34     35      0  137  149  146  135   12
>  d     69     34     35      0   70   71   70   40   40
>  d     69     34     35      0   65   69   68   30   70
>  d     64     44     53      0   68   70   73   36   15
>  d     64     44     53      0   79   90   91   44   35
>  d     64     44     53      0   63   81   83   41   50
>  d     64     44     53      0   72   85   86   58   40
>  d     64     44     53      0  109  214  189  199   26
>  b     69     20     35      0   73   55   72   52   26
> 
> .
> .
> .
> 
> 
> 
> 
>