using constraints in ML

[marcel spijkerman]

Dear Statalisters,

I want to impose constraints in a self written ML model. I've looked how to do this on the internet but can't get it to work.

This is what I tried::

constraint define 1 [eq1:]x2=1

ml model d0 cloglog_d0_hs (dead1=x2 delta) ///
(dead2=x1 x2) /logitp1 /logitp2 /logitp3 /mu1 /mu2, constraint(1)

ml init b1 b2 0 0 0 -0.5 -0.5, copy

ml maximize, trace difficult search(norescale)


no error messages but also no contraints imposed. Any suggestions?

Regards,

Marcel

[Jeff Pitblado (StataCorp)]

Without your evaluator and a working dataset, it is hard to tell you what to fix, but I
imagine it is how you defined your constraint. It probably should be:

constraint define 1 [eq1]:x2=1

Here is a working example using a simple d0 evaluator for liner regression:

program myreg_d0
        args todo b lnl

        tempname xb lnsigma
        mleval `xb' = `b', eq(1)
        mleval `lnsigma' = `b', eq(2)

        mlsum `lnl' = lnnormalden($ML_y1 - `xb', exp(`lnsigma'))
end

sysuse auto

constraint 1 [xb]: turn = -.75

ml model d0 myreg_d0 (xb: mpg = turn trunk) /lnsigma, constraint(1)
ml init /lnsigma = 5.8
ml init xb:_cons = 21.5
ml max

Here is the log from running the above code:

. program myreg_d0
  1.         args todo b lnl
  2. 
.         tempname xb lnsigma
  3.         mleval `xb' = `b', eq(1)
  4.         mleval `lnsigma' = `b', eq(2)
  5. 
.         mlsum `lnl' = lnnormalden($ML_y1 - `xb', exp(`lnsigma'))
  6. end

. 
. sysuse auto
(1978 Automobile Data)

. 
. constraint 1 [xb]: turn = -.75

. 
. ml model d0 myreg_d0 (xb: mpg = turn trunk) /lnsigma, constraint(1)

. ml init /lnsigma = 5.8

. ml init xb:_cons = 21.5

. ml max

initial:       log likelihood = -497.21266
rescale:       log likelihood = -298.76197
rescale eq:    log likelihood = -242.60861
Iteration 0:   log likelihood = -1985.0117  (not concave)
Iteration 1:   log likelihood = -313.18718  (not concave)
Iteration 2:   log likelihood = -244.96237  (not concave)
Iteration 3:   log likelihood = -237.48694  (not concave)
Iteration 4:   log likelihood = -232.63018  (not concave)
Iteration 5:   log likelihood = -227.93703  
Iteration 6:   log likelihood = -205.83078  
Iteration 7:   log likelihood = -204.67864  
Iteration 8:   log likelihood = -204.67749  
Iteration 9:   log likelihood = -204.67749  

                                                  Number of obs   =         74
                                                  Wald chi2(1)    =       9.42
Log likelihood = -204.67749                       Prob > chi2     =     0.0021

 ( 1)  [xb]turn = -.75
------------------------------------------------------------------------------
         mpg |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
xb           |
        turn |       -.75  (constrained)
       trunk |  -.3229896   .1052305    -3.07   0.002    -.5292375   -.1167417
       _cons |   55.47707    1.51509    36.62   0.000     52.50755    58.44659
-------------+----------------------------------------------------------------
lnsigma      |
       _cons |   1.346974   .0821995    16.39   0.000     1.185866    1.508082
------------------------------------------------------------------------------

[Richard Williams]

As a sidelight, I often find that the coeflegend option can help you to get constraints right.

sysuse auto, clear
logit foreign price mpg, coefl
constraint 1 _b[mpg] = .25
logit foreign price mpg, constraint(1)
sysuse nhanes2f, clear
mlogit health i.female, b(1) coefl
constraint 1 _b[fair:1.female] = .40
mlogit health i.female, b(1) constraint(1)