I am trying to build a binary regression model and I will be using LASSO. The dataset is the credit card dataset where the binary response is defaulted and not defaulted. Now, I know that LASSO will shrink some of the coefficients of my variables. I have seen that some people "group" categorical variables - i.e., this makes the algorithm to either have the whole categorical variable in or completely remove it. I have a question particular for this problem and one that's generalized:

For example I have categorical variables PayStatusApril,PayStatusMay... until June, and they each have 13 categories - -2 means there was no credit used, -1 means the person paid on time, 0 means use of revolving credit, and 1 means delay for one month (and so on until 9). So, would it make sense to leave just some categories of some variables? So for example take PayStatusJune-1 to be the variable which states whether the person paid duly in june. Would it make sense to include only the -1 category and exclude other ones from PayStatusJune? Note - I am aiming not only for predictive power but also for explainability. My thinking is that if this exact category (-1) from PayStatusJune is indicative enough and LASSO algo leaves it in, then why would I want to include other categories from PayStatusJune and just make the model oversaturated with irrelevant categories. Furthermore, as for the explainability aspect of my model - leaving only one category doesn't hinder the explainability (i.e., it doesn't make the model harder to interpret)?

And the more general question:

Why would I want to group any variable, even if I want my model to be explainable?


Your Answer

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