Request for Std. Error,t value, Pr(>|t|) in Regression

User 1828 | 4/29/2015, 2:14:20 AM


While doing linear regression, I want to know whether the coefficient estimate of the features are significant or not. Therefore, values like Std. Error,t value, Pr(>|t|) would be helpful. Is it possible to have that as well as a part of the regression diagnostics.



User 91 | 4/29/2015, 5:30:51 PM

This has definitely been something we have talked about.

Its easy enough to do when you have few coefficients in your problem i.e < 1000 but as soon as you have many coefficients it becomes much more computationally intensive to do. In applications with text data (e.g. sentiment analysis) and a large number of categorical variables ( e.g. click-through-rate prediction), it is very easy to exceed tens of thousands of coefficients which then makes the computation of p-values and std-errors a bit too expensive to do by default.

Would you want a feature like this to be done on demand or by default?

User 1828 | 4/29/2015, 7:06:24 PM

Great. Thanks for answering my question. I understand your point about being computationally expensive to calculate. If this is something you can produce by demand also it would be good enough for me.

User 3167 | 2/1/2016, 6:00:54 PM

I just saw this (months later) and I agree with the original poster. Does this feature exist yet? This would be extremely useful.