Reputation
27,953
Next tag badge:
98/100 score
35/20 answers
Badges
1 47 108
Impact
~1.0m people reached

22h
revised Testing linear restriction of parameters of ordered logistic regression models
typo
1d
revised Testing linear restriction of parameters of ordered logistic regression models
edited tags
1d
answered Testing linear restriction of parameters of ordered logistic regression models
1d
comment Can Random Forest be used for Feature Selection in Multiple Linear Regression?
I don't know how to do a proper linear model "inside" the scheme. I don't know how to mitigate the overfitting/estimation bias that would result. Perhaps a better approach is model approximation, sometimes called pre-conditioning. Here you use traditional models or single trees (which would require a ton of nodes) to approximate the output of a black box, inheriting the shrinkage of the black box.
1d
revised Can Random Forest be used for Feature Selection in Multiple Linear Regression?
edited tags
1d
answered Can Random Forest be used for Feature Selection in Multiple Linear Regression?
1d
comment How to compare (probability) predictive ability of models developed from logistic regression?
The Hosmer-Lemeshow test was shown by le Cessie, Hosmer, et al to be arbitrary. It also lacks power. Nonparametric calibration curves have better performance and are less arbitrary. There are also powerful single d.f. tests for model fit: the le Cessie Hosmer sum of squared errors test (related to the Brier score) and the Spiegelhalter test.
1d
revised How to compare (probability) predictive ability of models developed from logistic regression?
Answered OP followup questions.
2d
comment Setting different depth in rpart, but didn't actually change
Don't use a 'decision' or 'classification'. This is arbitrary and will not (1) properly transport to other situations where the sampling is different nor (2) allow for gray 'no decision' zones. Your approach also uses a very strange loss/utility function.
2d
revised How to compare (probability) predictive ability of models developed from logistic regression?
edited tags
2d
answered How to compare (probability) predictive ability of models developed from logistic regression?
2d
comment How does upsampling rare events affect the interpretation of logistic regression?
I wouldn't find that useful unless you can point to a place in his excellent paper where he said you must do it that way. When I read the paper I didn't see that.
2d
comment Setting different depth in rpart, but didn't actually change
There is no need to sample the dataset, and sampling will destroy the ability to estimate probabilties. Don't use just dichotomous classifications. Make sure you are using an efficient splitting test such as one based on deviance (log likelihood) an efficient score statistic (sometimes this is Pearson's chi-square).
2d
comment Setting different depth in rpart, but didn't actually change
What is your sample size and distribution of class? Single tree methods require enormous sample sizes to be reliable.
2d
comment Why would parametric statistics ever be preferred over nonparametric?
The power gain from parametric tests is miniscule when compared to the power loss they suffer when their assumptions are not met.
2d
revised Equation of a fitted smooth spline and its analytical derivative
clarified knot choice
Jul
30
comment Why would parametric statistics ever be preferred over nonparametric?
I beg to differ. You can use a Cox model to estimate quantiles, mean (if highest Y value is uncensored) and all sorts of probabilities. Not being able to forecast beyond the range of the data is a problem though (as you mentioned) but you could be dangerously extrapolating.
Jul
30
revised Why would parametric statistics ever be preferred over nonparametric?
added 68 characters in body
Jul
30
revised Why would parametric statistics ever be preferred over nonparametric?
Answered a follow-up question.
Jul
30
comment Why would parametric statistics ever be preferred over nonparametric?
I'll add more description to my answer to deal with that.