News for hackers. Stuff that matters.

A comment on Ways to Address Collinearity in Regression Models. Parent comment

Re:Ways to Address Collinearity in Regression Models by n00b101 on Wednesday September 04, 2013 @12:47PM (#6328768)
> If you are doing linear regression with correlated variables. PLS is the correct choice.

False. PLS (Partial Least Squares) is an ad-hoc algorithm that was invented by people working in chemometrics. PLS was not based on any rigorous statistical theory, it is merely a heuristic. It might appear to work reasonably well on some data sets, but asserting that "PLS is the correct choice" is cargo cult science.

[ Parent | on HN ]
Re:Ways to Address Collinearity in Regression Models by tlarkworthy on Thursday September 05, 2013 @09:18AM (#6333941)
"cargo cult science" Um, no, unless you have decided a ton of empirical peer reviewed journal articles are written by cargo cult scientists?

the same general argument you make can be said of PCA ... but it works. We know it works becuase it can be used to predict the real world.

[ Parent | on HN ]