0

The problem is following. Suppose, we have a binary logit model with a set of variables. Let's call it the old model. It is estimated over a set of observations and covariates.

Now, we wan to build a new model, which is also binary logit. New model is built over completely different set of observations, and the target variable is a set in a bit different way, similar to the old model but not the same. However, each observation for the new model has log-odds calculated according to the old model. And, the old log-odds variable serves as one of the covariates for the new model. After estimating the new model, the p-value for old log-odds shows statistical significance of the variable

Does it make any sense? What are the shortcomings/advantages of such approach? I do not know the right name for it, but I feel that it is not completely correct. It seems like it is some kind of models superposition, but I cannot find any academical work for similar matters.

kjetil b halvorsen
  • 77,844
Oleksandr
  • 113
  • 1
    Could you tell us why you do this, what you want to achieve? – kjetil b halvorsen Jun 25 '15 at 10:29
  • Hi! It is actually not me who is doing this. I just spotted it in someone's work, and I cannot simply stand it as a statistician. I just want to find some proofs about such aproach being right or wrong and why. I'm trying to google it, but I do not even know the name for such approach. – Oleksandr Jun 25 '15 at 11:49
  • Maybe what they are trying to do is the same as in this answer https://stats.stackexchange.com/a/272646/11887 or https://stats.stackexchange.com/questions/272121/recalibration-by-regressing-on-intercept-only-with-offset – kjetil b halvorsen Jan 04 '24 at 18:55

0 Answers0