I want to run a penalized multinomial logit and logit regression using the glmnet package in R. I understand, that before fitting the penalized model, one should standardize the variables, to penalize each coefficient equally. There are allready some posts which adress similar topics - but they don't answer the question how the mathematical process is designed:

Coefficient value from glmnet



Skimming through the glmnet vignette, I could find that the variables are standardized per default. What I don't understand is the phase: "The coefficients are always returned on the original scale" - how is this done mathematically?

So what I understand is:

  1. standardize all variables $x_i$ : $\hat{x_i}=\frac{x_i- \bar{x_i}}{\sigma_{x_i}}$

  2. obtain $\hat{\beta}= \text{argmax } log(L) - \lambda \cdot 0.5 \cdot ||\beta||_2^2$

So my quiestion basically is:

How is the outlined process to be complemented, in order to arive at the final reported estimates?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.