Hi
I have a dataset with 22 health variables, (N=400) trying to make prediction models for Outcome, which is 3 categories. So, I applied stepwsie multinomial regression, and also Predictor Screening to select the optimal variables.
I am NOT a statistician! So, reaching out to this community for advise.
A statistician reviewed the analysis, and told me that stepwise regression should not be used in clinical settings...? Rather, variables should be selected based on clinical experience. My argument was "how can we then find new methods, or treatments, if we use old knowledge?"
He is referring to a paper from 1991... Reading newer publications, I get a different view,
https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0202344
Copas, J. B., & Long, T. (1991). Estimating the Residual Variance in Orthogonal Regression with Variable Selection. Journal of the Royal Statistical Society. Series D (The Statistician), 40(1), 51–59. https://doi.org/10.2307/2348223
In summary:
"Stepwise regression is one of the most widely used statistical techniques. In a regression of dependent variable y on independent (or explanatory) variables x₁, x₂, …, xₚ, the idea is to include only those x s which are 'significant' and omit those which make little or no contribution to the prediction of y. Although variable selection can lead to a greatly simplified model, there are dangers which are not always recognised by users of statistical packages. The choice of the variables to be included depends on estimated regression coefficients rather than their true values, and so xⱼ is more likely to be included if its regression coefficient is overestimated than if its regression coefficient is underestimated. Hence the regression coefficients of selected variables are biased upwards (regression sum of squares too large) and in consequence the residual variation, which includes the contribution of the variables which are omitted, is biased downwards (residual sum of squares too small)."
Any comments on this from anyone?
