What is Inverse Probability of Censoring Weighting?
Inverse Probability of Censoring Weighting is a method that gives each still-observed unit a weight based on the inverse of its estimated probability of remaining uncensored, so observed follow-up represents units with comparable measured histories who were lost.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Define the estimand before labeling an event as censoring
Administrative censoring can be unrelated to prognosis under a fixed study end, whereas dropout or artificial censoring at treatment deviation can be informative. A competing event may be part of a composite outcome, a terminal event in a cause-specific estimand, or an intercurrent event requiring another strategy. Treating every unavailable outcome as the same censoring process can change the question being answered.
Longitudinal weights multiply conditional survival probabilities
At each eligible time, estimate the probability of remaining uncensored given past treatment, covariates, and observation history. The denominator is the product of those probabilities up to that time; a stabilized numerator uses a reduced history to limit variance. IPCW may be multiplied by treatment weights in a Marginal Structural Model or cloning-censoring-weighting analysis. Robins and Finkelstein formalized IPCW for dependent censoring.
Weight diagnostics reveal unsupported follow-up
Inspect censoring by time and reason, predicted observation probabilities, stabilized-weight mean and tails, effective sample size, covariate balance in risk sets, and sensitivity to model form and truncation. Very small probabilities indicate that observed units must represent many censored peers. Truncation reduces variance at the cost of bias and cannot solve unmeasured causes of censoring.
Key Characteristics
- Targets selection caused by informative loss of outcome follow-up
- Weights only units still observed at the relevant analysis time
- Uses products of time-specific uncensoring probabilities
- Can be stabilized and combined with treatment weights
- Depends on conditional independent censoring and Positivity
- Requires censoring-reason, tail-weight, balance, and effective-size diagnostics
Common Use Cases
- Adjusting survival estimates for prognosis-dependent dropout
- Estimating per-protocol effects after artificial censoring at deviation
- Supporting cloning-censoring-weighting in Target Trial Emulation
- Correcting longitudinal outcomes when observation depends on measured history
- Evaluating how weight truncation changes an informative-censoring estimate
Example
Loading code...Frequently Asked Questions
What problem does Inverse Probability of Censoring Weighting solve?
IPCW addresses selection when observed follow-up differs systematically from censored follow-up because censoring depends on measured outcome predictors. It makes retained units represent similar censored units under conditional independent-censoring, Positivity, and correct-model assumptions.
How is IPCW different from inverse-propensity weighting?
Treatment weighting models the probability of the observed treatment to adjust treatment-confounder imbalance. IPCW models the probability of remaining observed to adjust censoring selection. Longitudinal causal analyses may need both, and multiplying them can amplify instability.
Does IPCW handle unmeasured reasons for dropout?
No. It requires censoring to be independent of the relevant counterfactual outcome after conditioning on measured history. If unrecorded prognosis affects dropout, the estimate can remain biased; sensitivity analyses or data collection improvements are then necessary.
Should competing events always be censored with IPCW?
No. Treatment of competing events follows the estimand. They may be part of a composite outcome, handled in a cause-specific or subdistribution framework, or addressed through another intercurrent-event strategy. Calling them censoring without defining the target effect can answer the wrong question.
How should IPCW weights be checked?
Summarize censoring reasons and timing, predicted uncensoring probabilities, weight mean, quantiles and maximum, effective sample size, risk-set balance, and results across model specifications and truncation thresholds. Investigate near-zero probabilities instead of hiding them with clipping.