Setting maxiter doesn't align with error message

[snowysnowcones]

Hi,
I have a model that fails to converge, and receive the error message:
"ERROR: failure to converge after 2 iterations. Coefficient update contains infinite values"

The firs think I try to do is set the maxiter higher. I see in the docs the default is supposed to be 100 iterations. However, when I set the maxiter higher:
fit_beta_model = fit(BetaRegressionModel,betamodel,beta_file;maxiter=5000)

I still receive the same error message:
"ERROR: failure to converge after 2 iterations. Coefficient update contains infinite values"

Did the max iterations fail to set? Is the error message incorrect? How do I know one way or another?

Thanks
Snowy

[ararslan]

Hi @snowysnowcones, thanks for using this package and apologies for the very late response. I don't know whether a response is even still useful for you at this point given that we're closing in on a year since you opened this issue, but in case it is...

maxiter is a bound, not a requirement; it controls the maximum number of iterations that may be attempted, but the actual number of iterations may be lower, as you observed. The actual number will never exceed maxiter.

The error message is telling you three things:

1. convergence failed,
2. two iterations were attempted, and
3. the coefficient update¹ contains ±Inf and/or NaN, which are "infinite" (i.e. !isfinite).

The number of iterations attempted is included in the error message for context, but the actual reason for the failure is the non-finite values. The first and second derivatives are checked for non-finite values and the error you received is thrown if any are found. When that occurs, the fitting process has already diverged; convergence is not possible even if you tried to continue through all maxiter iterations. If you had reached maxiter without achieving convergence, the error message would be different and would not mention infinite values.

Does that make sense? Do you think there's a way it could be made clearer, either in the documentation or the error message?

Btw, if you're experiencing convergence issues with your model, I would recommend the following:

• The package tries to respect the float precision of the inputs, so if you're using e.g. Float32, try converting to Float64 first.
• The default link for the precision parameter is the identity link, but a common alternative is the log link. (In fact, the latter is the default in the R package betareg, if I'm not mistaken.) You could try using that instead.

Footnotes

1. In case you're not familiar, "coefficient update" here refers to the quantity added to the parameter vector at a given iteration of the fitting procedure. The package uses Fisher scoring, which formulates the update as the inverse of the expected information matrix (second partial derivatives of the parameter vector) right multiplied by the score vector (first partial derivatives). ↩

[snowysnowcones]

Hi @ararslan thanks for the response and thank you for your package. I use it daily in my work (at a Fortune 50 company). Yes, I did figure out the error was due to NaN/Inf values. The problem was ultimately some dummy variables I had in my model only had single values (e.g. all 0s). Once I eliminated those the model worked perfectly. Thanks again!

[ararslan]

I use it daily in my work (at a Fortune 50 company).

😳 That's wild! I'm really glad to know it's been useful for you.

If you encounter any more issues, things that are unclear, or have desires for additional features, I encourage you to open issues here. If you ping me in the issue then I'm likely to see it sooner and can get back to you in less than a year. 😅