Model-Based Single Imputation (kNN, Regression, Random Forest)

Model-based single imputation fills each missing cell once with a value predicted from the other variables by k-nearest neighbours, iterative linear regression or random forests (missForest).

빈 칸마다 다른 변수로부터 k-최근접 이웃, 반복 선형회귀 또는 랜덤 포레스트(missForest)로 예측한 값을 한 번 채워 넣는 대체 방법입니다.

I need a complete data table for a clustering analysis, but 7% of cells are missing across numeric and categorical columns. What is a good way to fill them in?

군집분석에 완전한 자료가 필요한데 숫자·범주 열에 걸쳐 칸의 7%가 비어 있어요. 어떻게 채우는 게 좋을까요?

Use a model-based imputer: missForest predicts each incomplete column from the others with random forests and handles mixed numeric and categorical data, while k-nearest neighbours and iterative regression suit numeric tables. Check accuracy with the normalised root mean squared error and the proportion falsely classified, and remember that single imputation understates uncertainty for later inference.

When to use Model-Based Single Imputation (kNN, Regression, Random Forest)

• A downstream method such as clustering or prediction requires a complete table.
• Relationships among variables are non-linear or involve interactions (missForest).
• You want an estimate of imputation accuracy for each column.

When to use something else

• You need standard errors or p-values from the completed data; use multiple imputation or FIML instead.
• Almost all values of a variable are missing; drop or redesign the variable instead.

Assumptions and requirements

• Values are missing at random given the other columns.
• kNN and iterative regression use numeric columns only; categorical columns require missForest.
• Row caps: missForest 10,000, kNN 5,000, iterative regression 20,000.

What the analysis reports

• Completed data with an imputation flag for each column
• Number and percentage missing per column
• Observed versus imputed means or modes
• NRMSE and PFC accuracy (out-of-bag for missForest, masked cells for the other methods)

How to report it

Missing values ([value] cells, [value]%) in [value] variables were filled by random-forest imputation (missForest); out-of-bag NRMSE = [value]; out-of-bag proportion falsely classified = [value].

Related methods

• Multiple Imputation (MICE) with Pooled Regression
• Full Information Maximum Likelihood (FIML) Estimation
• Missing Data Pattern Analysis

Run Model-Based Single Imputation (kNN, Regression, Random Forest) on Quorant