Daniel Forster

It seems you are already using some good metrics like MAE, RMSE, and density plots. In addition to these, you can explore the following standard methods for assessing data imputations:

• Log-Likelihood and Deviance Measures https://bookdown.org/mike/data_analysis/evaluation-of-imputation-methods.html
• Sensitivity Analysis
• Multiple Imputation

Hussain Nizam Thanks, it seems I'm more or less on the right track. In cross-validation these measures seem to work ok. My concern is that when I directly compare the imputed dataset with the original, all the results are 0, implying perfect fits, which seems unlikely given the slight variations shown on my density plots. My suspicion is that the R package I am using (hydroGOF) may be matching only the observed pairs in the comparison, essentially comparing only the original data points in the imputed datasets.

I'll have a look around for other packages that might perform better.

An additional question: are the variables in your data continuous or categorical?

Depending on that, you may have to change the methods.

For continuous variables, RMSE is indeed quite good, as the distance between real and imputed point can be easily calculated and evaluated

For categorical variables, you may need to look into proportions of values, i.e. what you most likely mean by density plot. Additionally, you may perform statistical tests to see whether the different proportion "means" are significantly different from each other (via T test)