GMM "missing values encountered in analytic gradient r(416);"

[Simon Schmickler]

Hi Statalisters,


you could save my life!

I want to estimate a model of the form

y = exp(x1+x2+constant) * error.

with GMM. The error has mean 1.
I set up moment conditions of the type (E is the expectation operator)

E[error|x1,x2] = 1 => E y exp(-x1-x2-constant) - 1 = 0, etc.

When I estimate this, I get the error: missing values encountered in analytic gradient r(416);


To illustrate my problem, here is an example of what happens.
When my y is actually generated by my model, the estimation goes through (Example 1).
When my y is just some unrelated random variable, I get the error above (Example 2).
I would have expected that GMM would just give me insignificant parameter estimates instead of this error.
Even more surprising to me, Example 2 does go through when I multiply y with 100 before estimating.

So, why would GMM fail because the dependent variable is close to 0?
Why would GMM fail because the model is bad?
Why do I get missing values in the analytic gradient, even though there are no missing values?

Thanks a lot for any help!


Simon



*****************************Example 1 - works
clear
set seed 1
set obs 1000

gen x1 = runiform()
gen x2 = runiform()
gen error = runiform() * 2

***
gen y = exp(x1+x2+1) * error
***

gmm(y* exp(-{x: x1 x2 _cons}) - 1), ///
conv_maxiter(100) instruments(x1 x2) deriv(/x = -y*exp(-{x:}))

***********************Example 2 - does not

clear
set obs 1000

gen x1 = runiform()
gen x2 = runiform()
gen error = runiform() * 2

***Only these 3 lines are different!
gen y = runiform()
egen Ty = sum(y)
replace y = y/Ty
***

gmm(y* exp(-{x: x1 x2 _cons}) - 1), ///
conv_maxiter(100) instruments(x1 x2) deriv(/x = -y*exp(-{x:}))

************************

[Simon Schmickler]

Update: At the initial parameter values, the gradient is small and so the optimization routine makes a large step. This lets the exponential function evaluate to missing, because it only has support -8e+307 to 709 in Stata. So now I need to find a way to make the optimizer update more conservatively.