recipes step_impute_linear() in R: Linear Model Imputation
recipes step_impute_linear() imputes missing numeric values in R by fitting a linear regression model that predicts each incomplete column from the others. The model is learned during prep() and applied during bake().
step_impute_linear(rec, Ozone) # impute one numeric column step_impute_linear(rec, Ozone, Solar.R) # impute several columns step_impute_linear(rec, all_numeric_predictors()) # all numeric predictors step_impute_linear(rec, Ozone, impute_with = imp_vars(Wind, Temp)) # set model predictors prep(rec) |> bake(new_data = NULL) # fit lm models, then apply
Need explanation? Read on for examples and pitfalls.
What step_impute_linear() does
step_impute_linear() fills missing numeric values with predictions from a linear regression model. It is part of the recipes package, the preprocessing engine of tidymodels. For each column you name, recipes fits an lm() that predicts that column from the other predictors, then uses the fitted model to estimate every missing entry from the complete values in the same row.
Like every recipe step, it runs in two phases. When you call prep(), the step fits one linear model per imputed column on the training data. When you call bake(), it scores the rows that hold an NA and writes the fitted value back into the gap.
Syntax and arguments
One argument does the real work: impute_with. Here is the call with its defaults:
The arguments you will actually set:
recipe: the recipe object you are adding the step to....: one or more selectors naming the numeric columns to impute.impute_with: animp_vars()selector listing the predictors fed into each linear model. Defaults to all predictors.skip: leave itFALSEfor imputation so the step also runs on new data.
The models argument is filled automatically during prep() and you should not set it by hand.
step_impute_linear() examples
The airquality dataset has gaps in two numeric columns, which makes a clean test case. Start by counting the missing values.
Ozone has 37 missing values. Build a recipe that predicts Ozone from two columns that are themselves complete.
Every gap is filled. Because Wind and Temp have no missing values, the linear model can score all 37 incomplete rows. The value of a model-based fill shows up against a flat mean.
Mean imputation would write 42.13 into every gap. The linear model instead gives each row its own estimate from that day's wind and temperature, so the imputed column keeps realistic variation.
You can impute several columns in one step, and each gets its own model.
Ozone and Solar.R are each predicted from Wind, Temp, and Month. Those three predictors are complete, so no gaps survive in either imputed column.
step_normalize() before step_impute_linear() does not change the imputed values. Order normalization after imputation if you need scaled columns downstream.step_impute_linear() vs other imputation steps
Linear imputation is the middle ground between a constant fill and a full model. Each step in the recipes imputation family suits a different data shape and compute budget.
| Step | Approach | Speed | Mixed types |
|---|---|---|---|
step_impute_mean() |
Column mean | Fastest | No |
step_impute_median() |
Column median | Fastest | No |
step_impute_linear() |
Linear regression model | Medium | No |
step_impute_knn() |
Average of k neighbours | Slow | Yes |
step_impute_bag() |
Bagged tree model | Slowest | Yes |
Decision rule: use step_impute_mean() or step_impute_median() for a quick baseline. Reach for step_impute_linear() when the missing column has a roughly linear relationship with complete predictors. Switch to step_impute_knn() or step_impute_bag() when that relationship is nonlinear or the gaps sit in categorical columns.
IterativeImputer with a linear estimator. Both predict each missing column from the others with a regression model, though IterativeImputer cycles over columns iteratively while step_impute_linear() fits one model per column in a single pass.Common pitfalls
A linear model can predict values the column could never take. Ozone is non-negative, but a regression line extended to a cold, windy day can return a negative estimate, as the -2.1 in the earlier example shows.
A negative ozone reading is impossible. Add a step_mutate() afterward to clamp the column, or switch to a step whose fills stay inside the observed range.
Two more traps to avoid:
- Predictors with their own NAs. Columns named in
impute_withmust be complete for a row to be scored. If a predictor still holds missing values, rows that share those gaps stayNA. Impute the predictors first, or list only complete columns. - Imputing categorical columns. step_impute_linear() handles numeric columns only. For factor or character gaps, use
step_impute_mode()orstep_impute_knn()instead.
Try it yourself
Try it: Build a recipe on airquality that imputes Ozone with step_impute_linear() using Wind and Temp as the model predictors, then confirm no missing values remain. Save the baked data to ex_imputed.
Click to reveal solution
Explanation: prep() fits a linear model of Ozone on Wind and Temp from the training rows, and bake(new_data = NULL) replaces each missing Ozone value with that model's prediction.
Related recipes steps
These steps pair naturally with linear imputation in a preprocessing pipeline.
step_impute_mean(): fast mean fill for a quick numeric baseline.step_impute_median(): median fill, robust to skew and outliers.step_impute_knn(): neighbour-based fill that also handles categorical columns.step_impute_bag(): bagged-tree imputation for nonlinear relationships.step_normalize(): centre and scale numeric columns after imputing.
FAQ
What types of columns can step_impute_linear() impute?
step_impute_linear() imputes numeric columns only. The imputed column becomes the response of a linear regression model, so it must hold numbers. Factor and character columns cannot be the target of this step. For categorical gaps, use step_impute_mode() for a most-frequent-level fill or step_impute_knn(), which supports mixed data types through Gower distance. The predictors named in impute_with can include factors, since regression handles them as dummy variables.
Does step_impute_linear() cause data leakage?
Not when used correctly. prep() fits the linear models on the training data only and stores them in the recipe. When you bake() validation or test rows, recipes applies those stored models without refitting, so test outcomes never influence the imputation. Leakage happens only if you prep the recipe on the full dataset before splitting, so always split your data first and prep on the training portion.
Why does step_impute_linear() leave some values missing?
The most common cause is incomplete predictors. A linear model can only score a row when every predictor in impute_with has a value for that row. If a predictor column has its own NA entries, rows sharing those gaps cannot be predicted and stay missing. Fix this by imputing the predictor columns earlier in the recipe, or by restricting impute_with to columns you know are complete.
How is step_impute_linear() different from step_impute_knn()?
step_impute_linear() fits a global linear regression model and assumes a roughly linear relationship between the gap column and its predictors. step_impute_knn() makes no such assumption: it fills each gap from the k most similar complete rows, which captures nonlinear and local patterns. KNN also handles categorical columns, while linear imputation is numeric only. Linear imputation is usually faster and easier to interpret when the relationship really is linear.