When Clean Data Is Actually Dirty
StatHarbor Analytics
We often treat data cleaning as a neutral step.
Delete missing rows. Fill gaps with the mean. Move on.
But cleaning is not neutral. It is a modeling decision.
In this episode, we unpack the statistical consequences of deletion and simple imputation, and why what looks “clean” can fundamentally alter your estimand, distort variance, and bias inference.
We walk through:
The formal role of the missingness indicator
The difference between MCAR, MAR, and MNAR
Why complete-case analysis is rarely as safe as it seems
How mean imputation collapses variance and attenuates regression slopes
When multiple imputation and inverse probability weighting are appropriate
Why sensitivity analysis becomes essential under MNAR
If you cannot defend MCAR, deletion and mean imputation are high-risk defaults.
Cleaning is not preprocessing.
Cleaning is inference