Clarification requested on “fatal failure” in JamesSteinEncoder (binary model)

[xypherhead]

The documentation for JamesSteinEncoder and the source file header both include the note:

“The estimation of parameters of this model is tricky and sometimes it fails fatally. In these situations, it is better to use the beta model.”

Could the maintainers please clarify what this fatal failure specifically refers to?

• Which part of the binary model's source code is said to possibly fail fatally- is it the optimizer or something else?
• Under what data or numerical conditions is this issue known to occur?
• If you could give a sample that produces this error?

[xypherhead]

@PaulWestenthanner

[PaulWestenthanner]

Hi @xypherhead
thanks for raising this question. I don't have an example for when the failure happens and also can only look at the code myself. I think you are right that it might happen in the optimizer, possibly related to zero division errors.