I have a dataset with some data missing completely at random (most < 10% missing). So far, so good. I am planning on computing regression models. However, the variables in my regression model, and also in the entire dataset, are, at best, only weakly correlated (if correlated at all). The total N of my sample is 95. I am using SPSS 27.
As far as I know, multiple imputation is a regression-based technique. So I am wondering: does it make any sense to impute data? I read that one can use 'auxiliary' variables for imputation from the entire dataset (usually people would, for instance, use items from a questionnaire with complete data to impute the missings; but I don't have that luxury). In my case, the only auxiliary variables that I could use from my dataset are correlated around .20 - .30 at best; and measure different constructs and also were often measured at different time points than the data I would need to impute.
Additional question: Is median- or mean-based imputation for covariates (control variables) okay in that case, where I really cannot/ should not impute based on regression due to lack of correlation in my dataset and missingness < 10% ?!?
I find these issues are really under-reported in publications that use imputation techniques.
"Missing completely at random" (MCAR) means that the missing values are not related to any variables in the data set. This is the most benign missing data mechanism for which multiple imputation (MI) or, equivalently, full information maximum likelihood estimation (FIML) can be used (and are widely recommended) to address missing scores.
Auxiliary variables need not be highly correlated with the target variables that contain missing scores. There is no harm in including auxiliary variables even with low correlations (see Enders, 2010, Applied missing data analysis, Guilford Press).
Mean imputation is generally a less appropriate technique to address missing data compared to MI and FIML (see again Enders, 2010, for a very clear and non-technical explanation of this issue).
You might check and see if the option being used is to delete the cases listwise. If this is so, you could change the option to pairwise deletion and this would preserve some of your cases. This would be a solution ONLY with missing at random.
Check https://academic.oup.com/bib/article/18/2/312/2562741
You need to estimate the missing data first before dong the correlation. If you cannot estimate, then try regression without the data and see how it behaves.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling