Problem with Maximum Likelihood model

[Leyla Guluyeva]

Hello, everyone.

I have run the following code and got the r(133) error 'unknown function ()'. It is my first time running a Maximum Likelihood model; therefore, I don't really know where to look. I have removed the spaces, tried to modify the last part, and write
ml model lf myll (d: _cons) (l1: _cons) (l2: _cons)
ml maximize

but nothing worked.

The code is:


capture program drop myll

use first_spell_for_ml.dta, clear


summarize e0i, meanonly
local u = 1 - r(mean)

program define myll
version 18.1
args d l1 l2

quietly {
* 1) k and λ0
local k = `l1'/(`d'+`l2')
local l0 = `d'*( (1-`u')/`u' )

* 2) temporaries
tempvar F0 f0 F1 f1 mu logli

* CDF & density at w0i
gen double `F0' = ((1+`k')*Ghat0) / (1 + `k'*Ghat0)
gen double `f0' = (1+`k') / ( (1 + `k'*Ghat0)^2 ) * ghat0

* CDF & density at w1i
gen double `F1' = ((1+`k')*Ghat1) / (1 + `k'*Ghat1)
gen double `f1' = (1+`k') / ( (1 + `k'*Ghat1)^2 ) * ghat1

* 3) hazard μ_i
gen double `mu' = cond(e0i==1, `d'+`l2'+`l1'*`F0', `l0')

* 4) survival term
gen double `logli' = - `mu' * t0i
replace `logli' = `logli' + (cs0i==0)*ln(`mu')

* 5) branch on transition
* job→unemp
replace `logli' = `logli' + ///
(e0i==1 & cs0i==0 & e1i==0)*ln( `d'/`mu' )

* job→job
replace `logli' = `logli' + ///
(e0i==1 & cs0i==0 & e1i==1)*(
ln((`l2'+`l1')/`mu') + ln(`f1')
)

* nonemp→job
replace `logli' = `logli' + ///
(e0i==0 & cs0i==0 & e1i==1)*(
ln(`l0') + ln(`f1')
)

* 6) total log‐likelihood
quietly summarize `logli', meanonly
return scalar lf = r(sum)
}
end

*where the problem seems to occur
ml model lf myll (d(l1(l2
ml maximize



I have tried a different version of the code, and the problem seemed to be that Stata didn't recognize my parameters and saw them as variables. I used to get error r(198) '05 invalid name'. The code I used was:




use first_spell_with_gdensity.dta, clear
capture program drop loglf
program define loglf

args lnf d l0 l1 l2


capture drop double kappa F0 lambdaE lambdaU rate S h pEE pEU f1 logL

// steady‐state mapping
gen double kappa = `l1' / (`d' + `l2')
gen double F0 = (1 + kappa)*Ghat / (1 + kappa*Ghat)

// hazard rates

gen double lambdaE = `d' + `l2' + `l1'*(1-F0) // employed spells


gen double lambdaU = `l0' // unemployed spells



// 1) survival up to t0i
gen double rate = cond(e0i==1, lambdaE, lambdaU)
gen double S = exp( - rate * t0i )

// 2) density of exit at t0i
gen double h = rate * S

// 3) transition‐probabilities for job‐spells
gen double pEE = (`l2' + `l1'*F0)/lambdaE
gen double pEU = `d'/lambdaE

// 4) offer‐density at w1 (for UE‐ or job→job spells)
gen double f1 = ((1 + kappa)/((1 + kappa*Ghat)^2)) * ghat_density ///
if cs0i==0 & e1i==1



gen double logL = .
quietly {
// case 1: censored spells (any initial state)
replace logL = ln(S) if cs0i==1

// case 2: uncensored job‐spells
replace logL = ln(h) + ln( cond(e1i, pEE, pEU) ) if e0i==1 & cs0i==0

// case 3: add f1 for job→job transitions
replace logL = logL + ln(f1) if e0i==1 & cs0i==0 & e1i==1

// case 4: uncensored nonemp spells (must take a job‐offer)
// h=λ0 S ; then one job‐offer density f1
replace logL = ln(h) + ln(f1) if e0i==0 & cs0i==0
}

// pass the result back to ml
replace `lnf' = logL
end

*where the problem seems to occur
ml model lf loglf ///
(d:0.05 l0:0.10 l1:0.10 l2:0.05) ///
, maximize difficult

[Joseph Coveney]

There is an ml trace option that you set on, which might help you locate the problem.

As an aside, you seem to be referring by name to variables in the dataset from within the evaluator program. Maybe look into trying to pass them through ml perhaps via nonlinear equations.

[Nick Cox]

The code in #1 is very hard to follow. At one point you seem to be defining a program within code that defines another program.

That aside, the error message you got may be because a program is evaluating a local macro that it can't see.