Does IPCW impute each missing value?

No. It reweights observed contributions for an aggregate analysis.

Can treatment and censoring weights be combined?

In suitable setups, yes, with attention to the joint assumptions, timing and estimator.

Does administrative end of study always require a complex model?

No. The appropriate handling depends on the design, outcome and whether the censoring mechanism is informative for the analysis.

AI Glossary / Causal inference and experimentation / Advanced

Inverse probability of censoring weighting

Definition

Inverse probability of censoring weighting, or IPCW, adjusts an analysis for incomplete follow-up by giving observed contributions weight according to the inverse probability of remaining uncensored. It aims to address selection caused by censoring under suitable assumptions. Despite its name, the relevant denominator is typically a probability of remaining observed, not a probability of being censored.

What it means

People, devices or accounts can disappear from a study before an outcome is measured. If that disappearance relates to factors that also predict the outcome, analyzing only those still observed can distort the result. IPCW increases the contribution of observed units whose histories resemble those more likely to be lost. It does not fill in each missing outcome. It instead uses a model of observation to reconstruct an aggregate comparison or outcome summary, subject to the information and assumptions available.

How it works

Define censoring and the time at which each contribution is needed. Model the probability of remaining uncensored using relevant observed history. In longitudinal settings, conditional probabilities across successive intervals are multiplied to obtain a cumulative observation probability. Uncensored contributions are weighted by its inverse, with possible stabilization depending on the analysis. Validity requires a suitable censoring model, adequate positivity and a conditional independence assumption linking censoring to outcomes given the included history. Treatment weighting may be needed separately. The exact contribution and time index depend on the outcome and estimator.

Three practical examples

A service study follows equipment until failure, but some units leave the monitoring network. Analysts investigate whether recorded operating conditions explain loss of follow-up.
A training evaluation measures six-month performance, but some employees no longer provide the planned assessment. A missingness analysis considers whether observed baseline and follow-up information supports weighting.
A digital-product study loses outcome visibility when accounts disconnect a data integration. Treating all disconnected accounts as successful or unsuccessful would answer a different question from adjusting for observation loss.

Think of it this way

Think of following a group on a long walk when some participants move out of sight. The people still visible may not represent everyone who started. Weighting asks visible participants with particular recorded histories to stand in for similar people no longer seen. That analogy depends on similarity being meaningful: if an unseen reason predicts both disappearance and the outcome, the correction may remain inadequate.

Worked example

In an original two-interval illustration, a participant has estimated probabilities of remaining observed of 0.8 in interval one and 0.75 in interval two, conditional on the relevant history and prior observation. The cumulative probability is 0.8×0.75=0.6, giving an unstabilized weight of 1/0.6≈1.67 for an eligible observed contribution at the second interval. Another participant with cumulative probability 0.9 receives weight about 1.11. The first contribution has greater influence because that observation history is less likely to remain visible. These are weights, not probabilities of the final outcome, and the calculation alone does not validate the censoring model or the full estimator.

Illustrative case study

A fictional maintenance provider compares two service schedules over a year. Equipment from smaller sites is more likely to stop transmitting logs, and those sites also differ in operating intensity. Analysts define when follow-up becomes unavailable and preserve measurements recorded before that point. They model continued observation, inspect extreme weights and compare weighted summaries with the unweighted analysis. They also investigate whether unrecorded connectivity failures could undermine their assumptions. The report separates uncertainty from missing follow-up from uncertainty in the schedule comparison itself. The scenario illustrates a possible analysis, not proof that weighting recovers a correct answer in every monitoring system.

Common misconceptions and mistakes

Censoring is not always the event of interest. If customer cancellation is the outcome, it should not automatically be treated as a nuisance disappearance. Another error is using the probability of censoring where the formula requires the probability of remaining uncensored. Analysts can also use future information in a history model, misalign interval timing or ignore time-varying predictors. Large weights deserve examination because a few observations can dominate. Finally, adjusting for observed predictors cannot guarantee correction when important reasons for loss are unmeasured.

Business use

IPCW can help teams evaluate outcomes when observation ends selectively, but it starts with a clear operational definition of missing follow-up. Businesses need to know whether an unavailable record means an unknown outcome, a genuine endpoint or a data-system problem. Reports should describe loss patterns, modeling choices, weight behavior and sensitivity to assumptions. Improving follow-up collection may provide more value than increasingly elaborate corrections. Weighting is an analytical response to incomplete evidence, not a substitute for maintaining reliable outcome records.

Check your understanding

Why were the interval probabilities multiplied?

Answer: Remaining observed through both intervals requires surviving each observation stage.

Why was the weight greater than one?

Answer: It was the inverse of a probability below one.

What should happen if cancellation is itself the studied outcome?

Answer: Define and analyze it as an event rather than casually censoring it away.

Related terms

Censoring
Selection bias
Survival analysis
Missing data
loss to follow-up
inverse probability weighting

Learn next

Draw a timeline showing observation, censoring and outcome events. Calculate cumulative observation probabilities from interval-specific probabilities. Then study conditional independent censoring and how weight diagnostics relate to the support available for the target analysis.

Censoring

Sources

Hernán and Robins, Causal Inference: What If, discussion of weighting for censoring: https://www.hsph.harvard.edu/miguel-hernan/wp-content/uploads/sites/1268/2024/04/hernanrobins_WhatIf_26apr24.pdf Diagnosing Covariate Balance Across Levels of Right-Censoring Before and After Application of Inverse-Probability-of-Censoring Weights: https://pmc.ncbi.nlm.nih.gov/articles/PMC7212402/ The interval calculation and maintenance example are original educational illustrations.

Previous: Inverse odds of sampling weighting Next: Inverse probability of treatment weighting