Missing-value pattern check model

Description

Recodes every input as present or missing and models the outcome from that pattern alone, to test whether missingness itself carries information before imputing.

Pyle argues that values are almost never missing at random. Before any imputation, copy the data set, replace each input with 1 if present and 0 if missing, keep the outcome, and fit a tree model (he uses a CHAID-style tree) of the outcome on these flags alone. If the tree predicts the outcome, the pattern of missingness is informative: it should be kept as features, and any replacement method must not destroy it. Pyle prefers replacement that preserves each variable's distribution and its relationships with others rather than mean substitution, which shrinks variance. The check is quick, model-agnostic and makes a hidden assumption visible. It shows that missingness matters, not why it occurs, so the reasons still need investigation.

Attributes

Loss type
Siloed
Effort
Hours

Steps

  1. Copy the data set so the original values are untouched.
  2. Recode each input variable as 1 (present) or 0 (missing), keeping the outcome variable as it is.
  3. Fit a decision tree of the outcome on the missingness flags, with cross-validation to avoid overfitting.
  4. Inspect which flags split the tree and how much predictive accuracy they give compared with a null model.
  5. Keep informative flags as features in later models, and choose an imputation method that preserves distributions and relationships.
  6. Investigate why the predictive fields are missing (skip logic, tracking gaps, user choice) with the data owners or users.

Sources

  • Handbook of Data Mining (Ye, 2003)

Also known as

  • Missing-value check model
  • MVCM
  • Missingness pattern model

Categories

  • Analytics & Behaviour
  • Statistics & Modelling