1

I want to run series of simple lm in R, with a continuous/categorical outcome and binary group membership (patients - controls) and categorical predictors. However the categorical predictor is positively skewed and none of the transformations I have tried so far (log, ^2 etc.) has had an effect on the normality assumption. Some of the outcome variables are Error responses therefore causing the same problem. Is it appropriate to use Nonparametric Regression (Lowess/Loess) in this case and which one would you recommend?

Your help would be much appreciated.

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
user1442363
  • 151
  • 1
  • 7
  • 8
    This is the same situation addressed at http://stats.stackexchange.com/questions/33193. Both questions derive from the same misconception, so let's be very clear: **there is no normality assumption about the distribution of _any_ of the variables in regression modeling.** The present question also covers the same ground as another (which I haven't looked up) asking for a transformation to convert a categorical variable to approximate normality: the answer is that no such transformation can possibly exist (except perhaps when there is a large number of categories). – whuber Jul 29 '12 at 21:04
  • @whuber is right, +1. Also note that in R, `lowess()` & `loess()` are essentially the same thing. [?lowess](http://stat.ethz.ch/R-manual/R-patched/library/stats/html/lowess.html) is the older & [?loess](http://stat.ethz.ch/R-manual/R-patched/library/stats/html/loess.html) is the augmented version (although, for simple things, I find `lowess()` easier to use). – gung - Reinstate Monica Jul 30 '12 at 00:39
  • the normality assumption - when it is relevant (you assume it for testing/CI's and particularly for PIs) - relates to the distribution, not the form of the model for the conditional mean. Nonparametric regression of the kind you mention an example of is related to not making a parametric assumption about the form of relationship between y and x rather than not making a parametric assumption about the form of the conditional distribution about the mean. One can test a linear fit avoiding specific distributional assumptions, but that's not what local polynomial and spline regression does. – Glen_b Aug 18 '18 at 15:51

0 Answers0