I have a dataset where about 15% of the values in my independent variables are missing. I'm debating between using mean imputation or something more complex like Multiple Imputation by Chained Equations (MICE). What is the standard approach in the industry to ensure the regression coefficients stay unbiased?
3 answers
Avoid mean imputation if you can; it artificially reduces the variance of your dataset and can lead to biased results. For a 15% missing rate, MICE is generally the gold standard in modern data science. It creates multiple "complete" datasets by predicting the missing values based on other variables, then pools the results. This captures the uncertainty of the missing data far better than any single imputation method. If the data is Missing Completely at Random (MCAR), you might get away with listwise deletion, but MICE is safer for Missing at Random (MAR) scenarios.
Before choosing an imputation method, have you analyzed the "missingness" pattern? Is it possible the data is Missing Not at Random (MNAR), which would require a different strategy?
MICE is great, but don't overlook simple K-Nearest Neighbors (KNN) imputation. It’s often easier to implement and can be very effective for numerical datasets.
Ryan makes a good point. KNN is quite robust, although for very large datasets, the computational cost can be a bit higher than MICE or simpler methods.
Kevin, that's a fair point. It seems people with higher incomes were less likely to report their salary in this survey, so it’s likely MNAR. How would I go about adjusting my regression model if the missing data is actually dependent on the value itself?