This question is more or less a theoretical question. So I am not sure whether it is suitable to post this here. Anyway it is great if I can get the opinion of the subject experts regarding this matter.
I have fitted a LASSO logistic regression model using classical approach and Bayesian approach and the results are different. I am curious about this because, the Bayesian estimates using double exponential priors similar to the classical LASSO estimates.
Here are the results based on
caret package (which internally use
glmnet package) in R. I used 10 fold cross validation to find the optimal lambda.
Here are the results based on Bayesian lasso by taking double exponential priors using my own code.
In fact, there seems to be no bug in my code as I obtained similar results based on
I also tried the horseshoe prior in
brms package. But I got divergent transitions.
So is it possible to explain the difference in results ?
Thank you in advance.