While building a predictive model, median imputations are performed while preparing the training data. How should the imputations be addressed in the validation data?

Question

Options

• AThe imputed values are irrelevant to the validation data, and are not used.
• BThe imputed values must be applied directly to the validation data without recalculation.
• CThe imputed values must be recalculated using the validation data.
• DThe imputed values must be recalculated using both the training and the validation data.

• A
  7% (2)
• B
  18% (5)
• C
  71% (20)
• D
  4% (1)

Explanation

Recalculating the median using the validation data (C) is correct because the validation set simulates new, real-world data arriving after model training - it should be preprocessed using its own representative statistics, not the training set's. The training median reflects the training distribution, which may differ from the validation distribution; imposing it on the validation set could introduce systematic bias and distort your model evaluation.

Why the distractors fail:

• A is wrong because missing values in validation data still exist and must be handled - ignoring them would cause errors or unreliable predictions.
• B is wrong because blindly applying the training median assumes both datasets share an identical distribution, which is not guaranteed and contaminates the independence of validation.
• D is wrong because combining training and validation data to compute the median leaks validation information into preprocessing, violating the separation needed for unbiased evaluation.

Memory tip: Think "each dataset owns its own statistics." Just as the validation set tests the model on unseen data, it should also be imputed from its own unseen data - never borrowing from training.

No community discussion yet for this question.