Issue with teffects IPWRA-

[Michael Andersonnnn]

I am hoping someone can help with this, I am using the following code with a hospital administrative dataset with rare events:
teffects ipwra (readmission alter sex i.charlindex i.imddecile) (ind alter sex i.charlindex i.imddecile i.n_hrg3, probit) , atet

and I keep getting the following result:

missing values encountered in analytic gradient
gmm estimation failed

The covariates are coded in a manner whereby there aren't missing values, so I am not sure what STATA means by "missing values encountered in analytic gradient"?

[William Lisowski]

The posts at

https://www.statalist.org/forums/for...gradient-r-416

describe a similar problem and suggests a possible source:

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.

Perhaps the discussion in

help gmm##optimization_options

will give you a few ideas of things to try.

[Michael Andersonnnn]

HI William, thanks for replying, and very helpful to see the optimization options.I must admit I am a bit lost reviewing the help file. Do you have any possibly suggestions for any specifications of optimize (), which may help me overcome this issue?

[William Lisowski]

Actually, my answer in post #2 was confused. While the error message you received referred to gmm, which supports the optimize options, apparently gmm was invoked by the teffects ipwra command which supports the maximize options. So to be honest, I don't know what to suggest.

On reflection, I think this might be a question best directed to Stata Technical Services through

https://www.stata.com/support/tech-support/

My own feeling on this is - if the diagnosis quoted in #2 is in fact relevant in this case - that Stata should be detecting the problem and re-trying with a smaller step. The alternative is to start from somewhere closer to the optimum.

And on further reflection, if sex is a categorical variable (as opposed to, say, frequency ) you might do well to indicate that by specifying i.sex rather than allow it to be treated as continuous. That may have an effect on the workings of the optimization code.