It is recommended to use Multiple Imputation (MI) for missing values. In the study for my master's thesis I have 3 measurement points I'm interested in (pre-intervention, post-intervention, follow-up). There is only 1 group. However, I have high dropout rates (post: around 35%, follow-up: around 50%). These missing values I am now trying to replace, but I am struggling with SPSS's "replace missing values" function.
If I do not define any constraints it does compute an imputed data set. Problem is, that now it gives me unrealistic values (for example, I used the PHQ-9, so it would give me a decimal number for the sum, which is technically not possible - which I don't even mind as much. More problematic: it also gives me negative values or values above the possible maximum in the Questionnaire). If I put in all my constraints (how many decimal numbers, the possible range of values) it tells me that after xy draws it could not find n imputed value in the constraints for a variable (it seems to happen with all my variables, not just one specific one). This keeps happening, even when raising the number of possible draws (even up to around 3000! If I put it too high, the program crashes - so this does not seem to be the solution, either).
Does anyone know a solution to this?
How the h* could you replace missing values??? The only thing you can do is make a model with the values you have, and then use that model to guess what the missing values could be, but NEVER EVER use those 'replacement values' to create a model!!!
Thanks for taking the time to read my question and answer to it - however, I'm not sure whether there might be a misunderstanding, Koen Van de Moortel . It is possible I have not explained my plan clearly, since I do not have much experience with conducting & describing statistical analyses by myself. I do indeed want to make a model with data estimated by the data I have, hence I am using Multiple Imputation.
Unfortunaltely I cannot delete my question, only edit it. But as an update: apparently the problem was to set variables on "impute and use as a predictor" - which isn't acutally possible for data which is missing "listwise" (as data is missing for the post & follow up measure). Now the Multiple Imputation seems to work fine (I will only be using the imputed data set for a t-Test to not overstimate the effects by only using my Complete Case Analysis, but not for my predictor analysis by multiple regression).
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling