How to debug "could not calculate numerical derivatives -- flat or discontinuous region encountered"

[Henry Strawforrd]

I have the following custom ml problem.

cap program drop censored_likelihood
program define censored_likelihood
    version 17.0
    args lnf sigma lambda gamma0 gamma1
    tempvar m epsilon Iy0 Iy010 Iy10
    local y "$ML_y1"
    gen double `m' = (1 + (`lambda' - 1) * x * (`gamma0' - exp(-`gamma1' * z)))
    gen double `epsilon' = `y' - `m'
    qui replace `epsilon' = 0 if `y' == 0
    
    gen double `Iy0' = (`y' == 0)
    gen double `Iy010' = (`y' > 0 & `y' < 10)
    gen double `Iy10' = (`y' == 10)
    
    replace `lnf' = -0.5 * ln(2*c(pi)*`sigma'^2) - (`y'-`m')^2/(2*`sigma'^2)  ///
                + `Iy0'   * ln(normal(-`m'/`sigma')) ///
                + `Iy010' * ln(normal((10-`m')/`sigma') - normal(-`m'/`sigma')) ///
                + `Iy10' * ln(1 - normal((10-`m')/`sigma'))
    
end

ml model lf censored_likelihood (b1: y = x z, nocons) (lambda:) (gamma0:) (gamma1:) (sigma:)
ml check
ml maximize

I received the error: "could not calculate numerical derivatives -- flat or discontinuous region encountered". How do you debug this?

[Joseph Coveney]

I received the error: "could not calculate numerical derivatives -- flat or discontinuous region encountered". How do you debug this?

What did ml check give you beforehand? It was OK?

Typically, ml check is the major diagnostic method for debugging, and it is at that step (making ml check happy) where you do most of the debugging of the evaluator.

[Henry Strawforrd]

Yes, ml check did not give any error. I also used ml search actually, did not put it in the code.

[Joseph Coveney]

As I mentioned in your parallel thread, you'll want to make sure that the arguments passed from ml model and what your likelihood evaluator expects align.

One tactic you might want to consider is to start with a simpler likelihood function, get it working and build up progressively.