Clarification about how alpha is estimated for CoxNetSurvivalAnalysis

[nnursimulu]

First of all, thank you for the great documentation -- your Python package is great.

Can you please clarify how the various values of alpha are estimated when a CoxnetSurvivalAnalysis object is fitted? I want to know if it makes sense for me to provide my own set of alphas or not.

I look forward to hearing from you.

[sebp]

It's the same as in the original R implementation:

When automatically generated, the alphas sequence is determined by alpha_max and alpha_min_ratio. The latter is the ratio of smallest value of the generated alphas sequence (say alpha_min) to alpha_max. The program generatesn_alphas values linear on the log scale from alpha_max down to alpha_min. alpha_max is not user-specified but is computed from the input X and y: it is the smallest value for alpha such that all the coefficients are zero.

[nnursimulu]

Thank you for your swift reply! Can you please tell me what the expected behaviour of the model would be if I were to set alphas=[0] only? Would it default to a non-penalized model?

[nnursimulu]

Thank you. I do get an error about inverting a singular matrix when I try this. I have seen suggestions of setting a small alpha to offset this. Would you have any recommendations regarding this?

[sebp]

That's what I was referring to. Setting a small alpha might work, just give it a try.

[nnursimulu]

Thank you for your help. I still get some errors but it could be that my values need to be normalized as they can be vary from very large to very small. Do you have any other recommendations?

[sebp]

The impact of alpha depends on the scale of features, hence standardization might help.

Alternatively, you can try CoxnetSurvivalAnalysis with l1_ratio close to zero. This will provide you with a list of alpha values from which you could try the smallest one.

[nnursimulu]

Okay, thank you very much!