0

Sometimes I hear about people talk about how this or that covariate or feature is "weighted" in a predictive model, but I can't find an easy mechanism to do that in glm. The weights there refer to weighting individuals, not covariates.

Seems like regressions are designed to be fit and left alone. You either put a feature in the model or you don't.

The purpose of these weights may be to reflect an existing prior or simply to explore changing the relative contribution of the covariates in a model. I can imagine several reasons why you would apply weights but no evidence anyone actually does it in a regression.

Am I missing something? Is it normal to change the coefficients of a regression to match some preconceived notion regardless of fit? How do I do it?

jermdemo
  • 263
  • 2
  • 12
  • It wouldn't be unreasonable to think of the coefficients themselves as covariate weights in a model. Could that be what they mean? – TPM Apr 11 '18 at 18:21
  • 1
    @TPM I think that is the convention that is used in some type(s) of models. [Here is an example](https://web.stanford.edu/class/cs221/lectures/learning1.pdf) where they describe the *parameters* of a neural network in such a fashion. – AdamO Apr 11 '18 at 18:27
  • There are many different ways to weight variables in models. Could you therefore edit this post to explain the function and/or purpose of the weights? – whuber Apr 11 '18 at 19:01
  • Yes the coefficients are the weights. – jermdemo Apr 11 '18 at 21:06
  • Thank you, but it is still difficult to determine what you are asking. One interpretation is that you seek some version of a [DFBETAS](https://stats.stackexchange.com/questions/141243). Another is that you want to constrain some coefficients in the regression in some manner, either through equalities or perhaps inequalities. "Match some preconceived notion" is just too vague to permit an objective interpretation. Perhaps you could make this more specific? – whuber Apr 11 '18 at 21:16
  • An insurance company wants to determine the risk of having an accident but is not convinced that a drivers zip code should be as heavily weighted in this prediction because some drivers are lying about where they live. How to they temper the weight of this covariate without eliminating it entirely? – jermdemo Apr 11 '18 at 22:21

0 Answers0