Why does a model perform worse after reintroducing observations with missing data imputed?

Question

My dataset contains about 12% missing data, and much of the missing data is grouped along observations (not randomly scattered or along columns). I optimized a regression method after removing the observations with large amounts of missing data (and used KNN-impute to infer the remaining missing data). I am now trying knn-impute on the entire data set, without removing any data prior, and my previous optimal regression method is now performing worse. Why might this be?

Note: optimal regression method == optimizing parameters.

So is adding in this new data worse i) as it contains lots of missing data and knn-impute is the wrong method, or ii) I just need to search for new optimal parameters.

Answer 1

Presumably your model was optimised for one sample. If you add a large amount of new data it will generally perform worse. The performance could be even worse if the observations with missing data are systematically different in some way.