search

pbreheny / ncvreg

Regularization paths for SCAD- and MCP-penalized regression models
http://pbreheny.github.io/ncvreg
41 stars 28 forks source link

Support for passing initial coefficient starting values (initial estimates/guesses) #12

Closed tomwenseleers closed 4 years ago

tomwenseleers commented 6 years ago

I have a problem where I already have initial coefficient estimates obtained using nnls, or where I fit multichannel longitudinal models and where the coefficients for a next time slice are known a priori to be very similar to those of the previous time slice. Would there by an performance advantage in these cases of being able to pass these coefficients as initial coefficient starting values perhaps (i.e. using a warm start)?

pbreheny commented 6 years ago

ncvreg is organized around the idea of fitting a regularization path, not just obtaining a fit for a single value of λ. The idea of supplying initial values doesn't really make sense in this framework, since the initial values would just come from the previous values of λ.

Now, I do agree that there are many cases where it would be useful to have a separate function that lets you fit a model at a single value of λ and supply an initial value. This function does not currently exist in ncvreg, but it isn't because I don't think it's valuable -- I just haven't had the time to implement it. Maybe this summer? But I can't make any promises.

pbreheny commented 4 years ago

This is now implemented in the function ncvfit(), available in ncvreg > 3.12.