[R] maximum likelihood convergence reproducing Anderson Blundell 1982 Econometrica R vs Stata

[peter dalgaard]

On May 9, 2011, at 13:40 , Alex Olssen wrote:

> Hi Mike,
> 
> Mike said
> "is this it, page 1559?"
> 
> That is the front page yes, page 15*6*9 has the table, of which the
> model labelled 18s is the one I replicated.
> 

However, the R code you posted will at best replicate model 18. For 18s, you need to impose symmetry conditions.

The fit of model 18 does seem to give similar results to those in Table II, but not exactly. 

Even more puzzling, if you start from the published coefficients, nlm seems to be getting stuck at a non-concave point of the negative likelihood.


> nlm(p=ab, f=lnl, hessian=TRUE, y1=y1, y2=y2,
+ x1=x1, x2=x2, x3=x3, gradtol=1e-15)
$minimum
[1] -76.60681

$estimate
[1] -0.64701504  0.27284992 -0.13713084 -0.07205753  1.28200342 -0.24596766  0.13502787  0.06001181

$gradient
[1]   2.782425  12.648629  -1.830244 -29.296795  -2.577734  -4.954042   1.495517  11.016190

$hessian
          [,1]     [,2]     [,3]      [,4]      [,5]     [,6]     [,7]     [,8]
[1,]  197435.9  1977197  1741724  805132.7  135821.9  1358013  1196714   553746
[2,] 1977197.4 19712076 17383474 8064540.2 1359856.0 13527720 11935305  5544903
[3,] 1741724.0 17383474 15336008 7121047.7 1197868.5 11929667 10529439  4896008
[4,]  805132.7  8064540  7121048 3315377.8  553733.1  5534923  4889523  2279404
[5,]  135821.9  1359856  1197868  553733.1  271511.7  2713733  2391332  1106572
[6,] 1358012.7 13527720 11929667 5534923.0 2713732.9 26955636 23784039 11054174
[7,] 1196714.3 11935305 10529439 4889522.6 2391331.9 23784039 20993413  9765022
[8,]  553746.0  5544903  4896008 2279404.1 1106571.9 11054174  9765022  4552693

$code
[1] 2

$iterations
[1] 4


-- 
Peter Dalgaard
Center for Statistics, Copenhagen Business School
Solbjerg Plads 3, 2000 Frederiksberg, Denmark
Phone: (+45)38153501
Email: pd.mes at cbs.dk  Priv: PDalgd at gmail.com