Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down on April 23, and its replacement, is already up and running.

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

st: gradient in -ml- producing type mismatch?

From   Misha Spisok <>
Subject   st: gradient in -ml- producing type mismatch?
Date   Sat, 8 May 2010 03:16:14 -0700

Hello, Statalist!

I'm getting the following error (-ml- program and example that
reproduces error is included below; extra details about the function
being maximized and its gradient are at the end):

type mismatch

end of do-file


after running my maximum likelihood function.

I suspect it's due to the gradient.  Can I use -mlvecsum- to form
pieces that will be used subsequently?  If so, how?

For example, (borrowed heavily from -multin- by P. Guimaraes), given
that -mlvecsum- multiplies the expression by x_{ij} (as stated in the
manual), consider the following:

program define mycondlog
    version 10.1
    args todo b lnf g negH
    tempvar theta sumvij prob ysum a1 a2 partone newpiece sumexb myfrac obsg
    mleval `theta'=`b'
    local by $GROUP /* GROUP includes all choices for agents of a given type */
    sort `by'
    quietly {
    by `by': egen double `sumvij' = sum(exp(`theta'))
    gen double `prob' =  1/(exp(-`theta')*`sumvij')
    mlsum `lnf' = $ML_y1*(`theta'-log(`sumvij'))
    if (`todo' == 0 | `lnf' >= .) exit

    /* if I could use mlvecsum to form pieces... */

    mlvecsum `lnf' `partone' = $ML_y1 /* This forms the first part */
    mlvecsum `lnf' `newpiece' = exp(`theta') /* which forms
x_{ij}*exp(x_{ij}b) */
    by `by': egen double `sumexb'= sum(`newpiece')
    gen double `myfrac' = `sumexb'/`sumvij'
    gen double `obsg' = `partone' - $ML_y1*`myfrac'
    matrix `g' = `obsg'

Now I can do the following:

use, clear
expand 2
bys id: replace x2 = x2/count if _n==2
multin count x1 x2, gr(ycount) /* This assigns $GROUP; could also
write: global GROUP ycount */

poisson ycount x1 x2, nocons

mat b0 = e(b)
local initopt init(`b0', skip)

ml model d1 mycondlog (ycount = x1 x2, nocons), init(`b0')
ml maximize

type mismatch

end of do-file


-----Extra Details-----

For the conditional logit, p_{ij} = exp(x_{ij}b)/[sum_{l} exp(x_{il}b)],

where i indexes choice-making agents and j indexes alternatives, so
that this is the probability that agent i chooses alternative j.

The log-likelihood for the conditional logit is

ll = y_{ij}*p_{ij}

ll = y_{ij} * ln( exp(x_{ij}b)/[sum_{l} exp(x_{il}b)] )

The derivative of p_{ij} with respect to the parameter vector, b, can
be written as

p_{ij}*[ x_{ij} - xbar_{i} ]

where xbar_{i} = sum_{l} p_{il}*x_{il}.

The gradient can be written as

g = sum_{i}sum_{j} y_{ij}*[ x_{ij} - xbar_{i} ]
so for a given observation, it is
g_{ij} = y_{ij}*[ x_{ij} - xbar_{i} ]

(This is all from Cameron and Trivedi, Microeconometrics, p.524.)

Another way to write the gradient is

g = sum_{i}sum_{j} y_{ij}*[ x_{ij} -
(sum_{l}x_{il}exp(x_{il}b))/(sum_{l}exp(x_{il}b)) ]
so for a given observation, it is
g_{ij} = y_{ij}*[ x_{ij} - (sum_{l}x_{il}exp(x_{il}b))/(sum_{l}exp(x_{il}b)) ]

-mlvecsum- would work for the first part of g,

y_{ij}*x_{ij}, but what about the second part (in either formulation)?

Thanks for your time.

*   For searches and help try:

© Copyright 1996–2015 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index