Dear statalisters,
I seek assistance in employing deviance ratios for evaluating model fit in logistic LASSO regressions within STATA. I employ this help file lassolassoexamples.pdf (stata.com), specifically on page 19, which suggests that a high deviance and a low deviance ratio may signify inadequate model fit. However, this assessment require testing on separate data. When the deviance ratio is significantly lower on the testing dataset compared to the training dataset, it implies superior performance within the training data but diminished performance on previously unseen data, which is undesirable, is that correct? I want my model to perform well on unseen data right, and that would be indicated by a higher (close to one) deviance ratio? Or maybe what I perfer is that it performs similar across data.
How should I think in the case of selection of a model in cases where the deviance ratio is low but the models (training and testing) exhibit similar performance. Should I prioritize models with similar performance despite their low deviance ratio? Additionally, is the simplified calculation of the deviance ratio as follows: deviance ratio = (the intercept model deviance) / (the fitted model deviance)? If so, is it preferable to have a low deviance for the fitted model but a high deviance for the intercept model to indicate good model fit?
I appreciate any insights and guidance you can provide on this matter.
