Hi, I have a case where data is missing not at random... in some cases entire chunks are missing at a time. However, after doing some digging I have found the proper values of some of these cases? is it wrong to go in and manually impute??? This one variable is also highly positively correlated with another variable in the data set?
Antonio -
Regression is often good for imputation, but you may find that you need to break your population into subpopulations for which different models apply. In each case you can use "graphical residual analyses" to check for model fit, and "cross-validation," even if just trying your model on a few known values, as if they were unknown, to check on overfitting to a particular sample.
Heteroscedasticity is often to be considered. See https://www.researchgate.net/project/OLS-Regression-Should-Not-Be-a-Default-for-WLS-Regression, and updates.
Cheers - Jim
- Badges
- Science topic