Starting values for QuantReg

[Jo Lu]

Hello,
Recently I wondered why the provision of starting values to the Quantreg().fit() function is impossible. After looking into the code I found the reason and a very simple solution. Now it works perfectly. However, I am new on the developer group and wonder whether I sould make the canges or simply hand my replacement code to an experienced statsmodels developer who pastes it into the github copies. This would save a lot of time since my replacement code contains less than ten lines of code.

Any advice?

Thank you and best regards.

[Jo Lu]

Dear all,

I am a little surprised that nobody responds to my proposal for a simple fix of a missing feature of the QuantReg class. There seems to be small interest in getting help from the community???

By the way: accepting starting values to the QuantReg.fit() method is important since one usually estimates a large number of quantile regressions for a narrow grid of quantiles 0.01, 0.02, 0.03, ..., 0.99 in order to get an estimate for the entire conditional distribution of the dependent. And the parameter estimates for quantile q are usually good starting values for quantile q+delta (if delta is small). In my tests the provision of starting values reduced the computing time by about 50-70 percent.

Best regards,

  Johannes

[josef...@gmail.com]

Hi,

Sorry for the delay, Sometimes messages go into the gmail spam folder.

open issue is https://github.com/statsmodels/statsmodels/issues/3774

It's a bit tricky to add start_params. GLM and RLM have similar irls loops and both of those allow for start_params.

You could open a pull request, or add your code changes to the issue.

Thanks,

Josef

--
You received this message because you are subscribed to the Google Groups "pystatsmodels" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pystatsmodel...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/pystatsmodels/55503889-3511-431f-aa58-0750e1c76259n%40googlegroups.com.

[Jo Lu]

Thanks. Okay. As already mentioned, I  know why adding start params is a little bit tricky. I will make a further minor changes. Currently the auxiliary OLS regressions are performed using the Moore-Penrose Inverse instead of directly solving the normal equations which would be faster and numerically more stable.

There will, however, be some delay since I will be on holiday in the next week.

Best regards,

  Johannes

[Jo Lu]

Hi,
right now I pushed my small amendments to QuantReg.fit(). I tested the code on my local copy but I don't know how to add and push a testing script. Please let me know whether this is OK.

Best regards,

   Johannes