0

I have performed a leave-one out cross-validated prediction using a lasso regression (with both lambda min and lambda 1se). My sample size is 52 and I have a bit more than 20 predictors. While lambda min give me acceptable results, the lambda 1se gives me weird results: enter image description here

The correlation between predicted and actual data is negative, the range of predicted value is very small. On each crossvalidation iteration the lambda 1se is always higher than lambda min and the deviance from the ideal model is almways 0% with lambda 1se (while it is 59% with lambda min on average on the 52 iterations). Given I have about the same pattern of results using ridge or elastic net regression I believe this the right results but I am just unsure what is the reason of this result?

Simon
  • 1
  • I think that a plot of the cross-validation fit, like the plot on page 5 of the glmnet vignette, would help explain what's going on. Please edit the question to show that plot. – EdM Nov 14 '23 at 14:48
  • There's an example of such a plot in what I think is a similar situation on this page. – EdM Nov 14 '23 at 15:19
  • Please edit the question to limit it to a specific problem with enough detail to identify an adequate answer. – Community Nov 15 '23 at 01:31
  • I have 52 times to plots, given I was running again cv.glmnet every leave out one crossvalidation step. cv.glmnet was used to estimate the minimum lambda to use for the training set n=52-1 (minus 1 because of the leave one out procedure) and this was repeated 52 times so that every single subject was left one once. Do you need the 52 plots of the model obtained by cv.glmnet to have a better idea what is going on? – Simon Nov 15 '23 at 15:20

0 Answers0