The study involved mathematics students completing a survey with several scales at three time points.
1) What is a reasonable threshold for missing data to remove an observation? Does this threshold apply to any given time point or to the entire longitudinal data set?
2) Should we remove people missing entire scales (at any time point) or impute them (e.g. using maximum likelihood)?
3) Should we impute at each time point or impute all observations from the longitudinal study at once?
4) If the latter, for cross-sectional analysis at time point 1, should we impute the time point 1 data set separately or should we take the imputations from the full data set, and then impute remaining missing data for extra students at time point 1 (i.e. who dropped the course or did not complete the other surveys)?
5) Should we remove observations who would be cleaned out at a specific time point (e.g., for straight-lining) but seem okay at other time points?
6) What is the current status of the debate on rounding and/or restricting range when imputing data?
Hi Kaitlin
The link below shows how to impute longitudinal data using Stata:
https://stats.idre.ucla.edu/stata/faq/how-can-i-perform-multiple-imputation-on-longitudinal-data-using-ice/
To answer your other questions, the standard approach (if you have enough data) is to remove a few known values from your data. Then impute, then compare observed to expected. If the error in your imputation method is low, then it is safe to assume that the imputed dataset was sampled from the same underlying population as the unimputed. If not, then the data do not support imputation and instances with missing values should be discarded.
- Badges
- Science topic