Augmented inverse probability weighting (AIPW)
AIPW combines outcome predictions with weighted prediction errors; the average effect can remain consistent when either the treatment model or outcome models are correct.
Fit
When to use it, and when not to
Use it when
- You compare two baseline treatments and want the average effect for a defined eligible population.
- You can fit treatment probabilities and outcome predictions using measured baseline factors.
Not when
- Treatment affects later confounders that influence later treatment. Consider a longitudinal marginal structural model. Read: Choosing an analysis for time-varying treatment
- You cannot measure the main common causes of treatment and outcome. Reconsider identification using a causal diagram. Read: Directed acyclic graphs
Before you start
What you need
- The target, contrast and outcome horizonSpecify a mean difference or risk difference for the eligible population.
- Treatment scores and two outcome predictions per personPredict each outcome under treatment and under control from baseline factors.
- A fitting and uncertainty planFor flexible learners, plan cross-fitting: train models on other folds before predicting each held-out record.
- The measured baseline factors suffice for causal adjustment.How to check: Check clinical knowledge and the causal diagram; two models do not fix an omitted common cause.
- Either treatment is possible at each target profile.How to check: Inspect probabilities and weighted errors; extreme scores still cause instability.
- The treatment versions are well defined, without relevant interference.How to check: Align eligibility, assignment and follow-up at time zero.
- Either the treatment model or both treatment-specific outcome models are correct for consistency under regularity conditions.How to check: Check both models. Double robustness is not a promise about bias in a small sample.
- Missing at random given recorded predictors may justify imputation.How to check: Specify imputation and sensitivity analyses. AIPW for treatment does not automatically solve missing outcomes.
Plain language
How it works
First fit a treatment model and an outcome model for each arm. For each person, start with their predicted outcome under treatment. For treated people, add the observed outcome minus that prediction, divided by their treatment chance. Average these corrected predictions. Repeat for control, correcting only untreated people with one minus the treatment chance. Subtract the corrected averages.
Bang and Robins establish double robustness; Funk and colleagues give a worked example and simulation study. The consistency claim requires either a correct treatment model or correct outcome models, plus causal assumptions. For regular parametric models, bootstrap the entire fitting procedure and use a percentile interval. An influence-function Wald interval with flexible models needs cross-fitting and suitable convergence rates; consistency alone does not establish interval validity.
Worked example
A worked example
- Setting
- Four records have treatment scores of 0.5, with treated pain scores 3 and 5 and untreated pain scores 5 and 7.
- Question
- What is the average treatment-minus-control difference in pain?
- Approach
- Outcome predictions are 4 under treatment and 6 under control for every record. Add each observed arm’s prediction error divided by its treatment chance.
- Result
- The corrected means are 4 under treatment and 6 under control. Their difference is -2 pain points.
- What it means
- This small dataset illustrates the correction only. Double robustness gives no guarantee of a reliable effect or confidence interval for four records.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
Double robustness concerns consistency when either nuisance model is correct. It does not remove unmeasured confounding, repair absent treatment support or guarantee valid uncertainty.
- Calling two incorrect models doubly safe.Instead: Either nuisance model must be correct for the usual consistency claim; both can fail together.
- Assuming consistency guarantees correct confidence intervals.Instead: Use a joint sandwich variance or full refit bootstrap for regular models. Flexible learners need suitable cross-fitting and rate conditions.
- Reading a single corrected contribution as a possible patient outcome.Instead: Contributions are calculation devices and may lie outside the observed outcome range.
Inside Mallard
How Mallard uses this
Plans do not link to this method yet.
aipw-doubly-robust).analysis.aipw, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Doubly robust estimation in missing data and causal inference models (opens in a new tab) · Biometrics · doi 10.1111/j.1541-0420.2005.00377.x
- Doubly robust estimation of causal effects (opens in a new tab) · Am J Epidemiol · doi 10.1093/aje/kwq439