Why add a binary missing-indicator column next to a feature you have imputed?

## What imputation throws away Filling a gap replaces an unknown with a number. Before the fill, a row said two things: this patient's creatinine is unknown, and here are the other measurements. After the fill it says only: this patient's creatinine is 1.0, same as many other patients. The model has no way to distinguish the substituted 1.0 from a lab result that genuinely read 1.0. That lost bit is often not noise. Consider an emergency-department lab panel where creatinine is absent for 40% of patients. The reason is not equipment failure or a corrupt file — it is that the test was never ordered. Ordering is a clinical decision, so the absence carries a compressed judgement: this patient did not look like someone who needed a renal panel. If your target is anything correlated with severity, that judgement is one of the more informative facts in the row, and a plain median fill deletes it. ## What the indicator is A **missing indicator** (also called a missingness flag or a shadow feature) is a binary column added alongside the imputed one, taking value 1 where the original entry was absent and 0 where it was observed. You keep the filled feature and you keep the flag; they are used together. How different model families exploit it: - **Linear models.** With the pair (filled value, flag), the coefficient on the flag acts as a separate intercept shift for the filled rows. If you fill with the column mean, that shift lets the model correct for the fact that the imputed group's true level differs from the observed group's average. Without the flag, the model has to fit one line through both groups and the fill silently biases the slope. - **Trees and tree ensembles.** A tree can split directly on the flag, which lets it isolate the missing subgroup and fit it separately from the observed subgroup — including learning a completely different relationship for them. - **Distance and kernel methods.** The flag participates in the distance like any other binary feature, so rows that share a missingness pattern become nearer to each other. ## A second, quieter benefit The indicator makes the choice of fill value less consequential. Once the model can see which rows were filled, the exact constant matters mainly as a numerical convenience: the model can absorb a systematic offset through the flag's coefficient or through a split. This is why "median plus an indicator" is such a durable default — it is cheap, it is honest about what was unknown, and it is fairly forgiving about the constant you chose. ## The costs and the cautions **Width.** One flag per imputed column. If forty columns have gaps, naive application doubles your feature count with columns that are mostly zeros. Restrict the flags to columns where absence is plausibly informative, or where the missing fraction is substantial. A flag on a column that is missing in three rows out of a million is a near-constant column that contributes nothing and adds a little variance. **Redundancy between flags.** When gaps arrive in blocks — an entire lab panel is absent together — the flags for those columns are near-duplicates of each other. One flag for the panel is usually better than eight identical ones. **The flag encodes a process, not a fact about the world.** This is the important operational caution. The creatinine flag means "this hospital's clinicians, under this year's ordering policy, chose not to test". If the policy changes — a new protocol orders the panel for everyone at triage — the flag's prevalence collapses and its meaning inverts, while the model still applies the coefficient it learned under the old regime. Any feature built on missingness is therefore tied to the pipeline and the process that produced the gaps, and it needs monitoring: track the imputation rate per column over time, and treat a sudden change as a model-health event, not just a data curiosity. **It must exist on both paths.** The flag is a feature like any other, so the serving code must compute it exactly as training did. A pipeline that adds indicators during training but forgets them at inference will feed the model a column of zeros and get systematically wrong predictions for precisely the rows where it matters. ## How to check whether it earned its place Do not argue about it — measure it. Train with the filled column alone, then with the filled column plus its flag, and compare on the same validation split. Where absence is informative the flag typically shows up clearly; where it is not, you have learned you can drop a column and simplify the pipeline. On the ED panel, the flag frequently ranks among the stronger features precisely because it is a proxy for a decision a human already made about the patient.