Nonparametric multiple linear regression (with interaction term and all !)

[mksp]

Experts !

Facing the issue of “compulsory” need to perform a multiple linear regression. I need to put interaction terms as well. the problem is that all assumptions are violated. So my question is: Should I relax assumptions (and I mean REALLY relax assumptions) and report everything “as is”. or there is a non-parametric alternative to multiple linear regression with interaction terms?

I also noticed that Kendall–Theil regression is a nonparametric approach to linear regression where there is only one independent and one dependent variable

Also, will a “robust regression” solve my issue?

[f2harrell]

I almost never use linear models any more, opting instead for semiparametric models such as the proportional odds model. Semiparametric models allow for arbitrary distributions of Y (including bimodal ones and clumping at zero) and interactions. Interactions are on a scale we are not used to: the link function scale, e.g., log odds of the probability that Y exceeds some value. For help on semiparametric models add a reply to RMS Ordinal Logistic Regression - models - Datamethods Discussion Forum for discrete Y or RMS Ordinal Regression for Continuous Y - modeling strategy - Datamethods Discussion Forum for continuous Y.

[f2harrell]

Post to the appropriate link above and I’ll remove this post

[R_cubed]

Minor addition: This question reminded me of my reply in this thread.

Provisions for outilers and strategies

Good question! This is going to depend on how you intend to analyze the data. Bayesian methods handle this issue by specifying probability distributions for all parameters. There is a good example on how Bayesian methods handle outliers in BBR Session 5 at around the 25:00 mark. This is an attractive solution if a credible prior is used. Frequentist procedures are numerous: Nonparametric methods: If you were going to use a 2 sample t-test, the 2 sample Wilcoxon-Mann-Whitney is the prefe…

It is hard to beat the simplicity of Dr. Harrell’s recommendation of semi-parametric modelling as a general rule. I’m pretty surprised at how much of applied stats can be compressed into proportional odds models. This strikes me as something only someone with decades of applied stats experience would figure out.

If there is a strong reason to prefer a frequentist parametric approach, I’d look at the adaptive methods of O’Gorman before considering textbook robust methods (despite their theoretical and historical importance).

Addendum: Links to O’Gorman’s texts are broken and not easily available from archive.org. This is an early paper on adaptive regression worth study.

O’Gorman, T.W. (2002), An adaptive test of significance for a subset of regression coefficients. Statist. Med., 21: 3527-3542. https://doi.org/10.1002/sim.1278