Doubly Robust Missingness Adjustment¶
Method — instantiates Missingness-Aware Estimator Selection
Combines outcome modeling with response weighting so estimates can remain consistent if one of the two model components is correctly specified.
Doubly Robust Missingness Adjustment estimates a target quantity from incomplete data by carrying two nuisance models at once — a model for the outcome given covariates and a model for the probability that a record is observed — and combining them so the estimate stays consistent as long as at least one of the two is correctly specified. Its defining move is redundancy as insurance: neither model has to be right, only one, and the analyst does not need to know in advance which one that is. That two-chances property is what separates it from any single-model repair, whose estimate lives or dies by that lone model being correct.
Example¶
A health system compares two second-line drugs for type-2 diabetes on their effect on 12-month HbA1c change, using electronic health records. Follow-up HbA1c is absent for roughly a third of patients — those who moved care, changed insurers, or simply skipped the lab — and that absence tracks observed traits like age, baseline control, and how often a patient visits. Complete-case analysis would quietly compare only the diligent attenders. The doubly robust approach fits two models. First, a response model: a logistic regression predicting whether a patient's 12-month HbA1c is observed, from their covariates. Second, an outcome model: a regression of HbA1c change on those covariates among the patients who were observed. The estimator then weights each observed patient's residual-from-the-outcome-model by the inverse of their modeled observation probability, and adds the outcome-model prediction back for everyone. If the analyst's response model is a little wrong but the outcome model captures the trajectory, the treatment contrast is still consistent — and if the outcome model is off but the response model is sound, it is also still consistent. One misspecification is survivable; the analyst buys a second chance without having to bet on which model is stronger.
How it works¶
- Fit both nuisances. A response (observation-probability) model and an outcome model, over the same covariate set.
- Augment. Start from the inverse-probability-weighted estimate, then add a correction term built from the outcome model — the "augmentation" that cancels the bias of whichever single model is wrong.
- Bank the one-of-two guarantee. Algebraically, the bias terms multiply, so it vanishes if either factor is zero (i.e. either model is correct).
- Cross-fit when models are flexible. When the nuisances are machine-learned, fit them on held-out folds so their overfitting error does not leak into the target estimate.
Tuning parameters¶
- Which nuisance to invest in — pour modeling effort into whichever of the two you can specify with real domain support; the other is your safety net.
- Weight stabilization / trimming — cap or stabilize extreme inverse-probability weights; heavier trimming tames variance but reintroduces a little bias.
- Cross-fitting folds — more folds reduce overfitting bias at the cost of compute and slightly noisier nuisance fits.
- Nuisance model class — parametric (transparent, brittle) versus flexible/ML (adaptive, opaque, needs cross-fitting).
When it helps, and when it misleads¶
Its strength is exactly the situation the archetype worries about: you suspect MAR is defensible but you are not sure whether the response process or the outcome surface is easier to model correctly. Double robustness lets you hedge instead of guess.[1]
Its failure mode is double jeopardy: the one-of-two guarantee gives nothing when both models are misspecified, and in that case a doubly robust estimator can be less stable than a simpler one, because the augmentation amplifies noise. It is also not a cure for thin overlap — when some records have observation probabilities near zero, the weights explode no matter how good the outcome model is, and the classic misuse is to treat "doubly robust" as a license to stop looking at the weight distribution. The guarding discipline is to keep an informal eye on the spread of the estimated weights and to invest the real modeling effort in whichever nuisance has genuine subject-matter grounding, treating the other as insurance rather than an afterthought.
How it implements the components¶
estimator_identification_match— it selects the augmented estimator matched to the MAR identifying assumption while deliberately hedging against getting the identifying models wrong; the estimator is chosen for its robustness to misspecification, not just its point form.imputation_model_congruence_check— it requires the outcome (imputation-like) model and the response model to be mutually coherent, referencing the same covariates and the same estimand so the two halves reinforce rather than fight each other.
It does not run the dedicated overlap diagnostic or build the response-support model as its safeguard — positivity_or_overlap_check and observed_covariate_support_model are Inverse-Probability Weighting Model's; a doubly robust estimator's insurance is its second (outcome) model, the thing that lets it survive a misspecified weight model that would sink pure weighting.
Related¶
- Instantiates: Missingness-Aware Estimator Selection — supplies the misspecification-hedged estimator for a defensible-MAR analysis.
- Consumes: Inverse-Probability Weighting Model supplies the response-weighting half that the augmentation corrects.
- Sibling mechanisms: Full-Information Maximum Likelihood Path · Inverse-Probability Weighting Model · MCAR Diagnostic Test and Balance Review · Missingness Indicator Matrix · Multiple Imputation Workflow · Pattern-Mixture Sensitivity Model · Process-Based Missingness Audit · Selection-Model Sensitivity Analysis · Tipping-Point Analysis
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Doubly Robust Missingness Adjustment operates as a computation, comparison, model, or analytic representation used to infer, estimate, or choose because it combines outcome modeling with response weighting so estimates can remain consistent if one of the two model components is correctly specified.
Independent corroboration: The frozen evidence defines Doubly Robust Missingness Adjustment as 'Combines outcome modeling with response weighting so estimates can remain consistent if one of the two model components is correctly specified', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Specialized
Rationale: Causal-inference statistics cohered augmented inverse-probability weighting and its consistency when either the outcome model or response model is correct.
Related originating lineages:
- Medicine & Healthcare — Epidemiology and clinical research supplied major applications to confounding, dropout, and missing outcomes.
Review resolution: Causal-inference statistics cohered the estimator, while epidemiology and clinical research materially shaped its missingness applications and development context.
Review outcome: Reconciled after independent review; high confidence.
References¶
[1] Bang, H., and Robins, J. M. "Doubly Robust Estimation in Missing Data and Causal Inference Models". Biometrics 61(4), 962–973 (2005). Shows how double robustness hedges model risk by remaining consistent when either of two nuisance models is correctly specified. registry ↩