Informative censoring and inverse probability of censoring weights (IPCW)
Informative censoring means loss to follow-up can distort an outcome estimate. IPCW reweights observed follow-up using measured history; it does not remove every bias.
Fit
When to use it, and when not to
Use it when
- Loss to follow-up may depend on measured factors that also predict the outcome.
- You can record relevant history before each possible loss.
Not when
- Death prevents the outcome. Use a competing-risk estimand rather than treating death as ordinary loss. Read: Choosing an analysis for competing risks
- Unmeasured prognosis drives loss. Add sensitivity analysis; weighting alone cannot identify the answer. Read: Sensitivity analyses
Before you start
What you need
- The outcome, horizon and target contrasttargetsummaryMeasureArtificial censoring at treatment changes answers a different question from assignment to treatment.
- Dates and predictors of lossRecord treatment, prior health and follow-up history before the loss they predict.
- A plan for weight checksInspect predicted observation probabilities, extreme weights and sensitivity to truncation.
- MCAR excludes dependence on observed or missing values. MAR allows dependence on observed values. MNAR retains dependence on missing values after conditioning.How to check: Relate the assumption to the unobserved outcomes and the history actually recorded.
- Censoring is independent of future outcomes given the measured history.How to check: Important unmeasured causes of both loss and outcome break this assumption.
- Every relevant history has a positive chance of remaining observed.How to check: Very small probabilities create unstable weights; zero probabilities leave unsupported comparisons.
- The censoring model represents how follow-up is lost.How to check: Use appropriate time order and functional forms. Assess uncertainty from estimating weights.
- Name missing completely at random (MCAR), missing at random (MAR) or missing not at random (MNAR).How to check: MAR is plausible only if recorded history explains selective loss. If MNAR is plausible, vary unobserved event risks in a tipping-point analysis.
Plain language
How it works
Censoring means follow-up ends before the endpoint is known. Informative censoring occurs when that loss carries information about prognosis. Estimate each chance of staying observed from prior history. Over time, multiply the conditional chances along each observed path. Their inverse gives an unstabilized weight. Stabilized weights also use a numerator model. Robins and Finkelstein developed weighted survival comparisons using time-varying prognostic data. Cain and Cole applied them to treatment noncompliance.
Weighting artificial censoring at treatment changes needs causal assumptions beyond a dropout correction. It does not recover the assignment effect automatically. Report the target, weight models, diagnostics and a variance method for the weighted estimator. Weighted pseudo-counts are not a binomial sample.
Worked example
A worked example
- Setting
- A cohort has 50 lower-risk and 50 higher-risk patients. All endpoints are defined at the same horizon.
- Question
- How can selective observation change a risk estimate?
- Approach
- Observe 40 lower-risk patients with 4 events, and 20 higher-risk patients with 8 events. Observation probabilities are 0.8 and 0.4.
- Result
- Weights are 1.25 and 2.5. Weighted events total 25 over a weighted denominator of 100: 25 in 100. The unweighted risk is 12 of 60, or 20 in 100. Weighting raises this estimate by 5 percentage points.
- What it means
- Assume observation is independent of the endpoint within each risk group. These weighted counts illustrate correction under that assumption, not extra patients or proof of validity.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
IPCW cannot recover information about unmeasured causes of loss and outcome. It needs conditional independent censoring, positive chances of remaining observed and a suitable weight model. Death that prevents the outcome is a competing event, not ordinary censoring.
- Treating a larger weighted sample as new information.Instead: Weights represent people lost to follow-up; they do not replace the missing observations.
- Truncating weights and calling the result unbiased.Instead: Truncation can reduce variance while introducing bias. Report the rule and compare results.
- Using a binomial interval for the weighted risk as if the weighted count were observed.Instead: Use variance estimation for the weighted estimator, accounting for estimated weights; a suitable bootstrap must repeat weight fitting.
Inside Mallard
How Mallard uses this
The method registry links its censoring-weights sensitivity method to this article and code. Mallard has no dedicated IPCW sizing form.
inverse-probability-censoring-weights).analysis.ipcw, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Correcting for noncompliance and dependent censoring in an AIDS Clinical Trial with inverse probability of censoring weighted (IPCW) log-rank tests (opens in a new tab) · Biometrics · doi 10.1111/j.0006-341x.2000.00779.x
- Inverse probability-of-censoring weights for the correction of time-varying noncompliance in the effect of randomized highly active antiretroviral therapy on incident AIDS or death (opens in a new tab) · Stat Med · doi 10.1002/sim.3585