Statistics Glossary → Interval-Censored Survival
Interval-Censored Survivalanalysis
A method for data in which the exact time of the event is unknown and it is known only to have happened somewhere between two check-ups, which estimates the survival curve and compares groups or factors.
Concept
In many studies an event is only noticed at a periodic check: tooth decay at a six-monthly dental visit, an infection in a routine blood test, a crack in a machine at its annual service. The event time is then not a single number but an interval, 'between the last negative and the first positive check'. Taking the middle or the end of the interval as the event time and applying Kaplan-Meier systematically shifts the event times and gives biased results.
Turnbull's method is the generalisation of Kaplan-Meier to interval-censored data. 'Innermost intervals' are formed from the intersections of people's observation intervals, and the probability mass is spread over these intervals in the way that makes the data most likely (nonparametric maximum likelihood, NPMLE). Because exactly when the event happened inside an interval is unknown, the curve is defined only at the ends of these intervals; how it falls within an interval is uncertain. Groups are compared with the generalised log-rank test; the effects of factors are estimated with parametric AFT models (time ratio) or a semi-parametric proportional hazards model (hazard ratio).
The method assumes that check-up times are unrelated to the event risk (e.g. that people having the event do not come for check-ups more often). With infrequent check-ups the intervals are wide and the estimates become uncertain. Left-censored observations (the event happened before the first check) and right-censored ones (no event yet at the last check) are handled in the same framework.
When to use it
- When the event is detected at periodic checks and its exact time is unknown (tooth decay, tumour recurrence on imaging, HIV seroconversion, machine wear).
- If the checks are very frequent and the intervals narrow, the result is almost the same as with ordinary Kaplan-Meier and Cox; with wide intervals, interval-censored methods are needed.
- If the effects of factors are wanted on the time scale, use parametric AFT models (Weibull, log-normal, log-logistic); if a hazard ratio is wanted without a distributional assumption, use the semi-parametric proportional hazards model.
An everyday example
An oral health centre follows 200 children until decay appears in their first permanent molar; the children are examined every six months. If a child had no decay at the 18-month visit but did at 24 months, the decay is known to have formed somewhere between months 18 and 24. Turnbull's method uses these intervals to draw the curve of 'the probability of staying decay-free over time', and the generalised log-rank test says whether the curves of children who did and did not receive fluoride varnish differ.
Parameters in MerQur
Group (optional)
Concept: When a group is chosen, a separate Turnbull curve is estimated for each group, and the generalised log-rank test checks whether the groups' survival curves are the same. The test is the log-rank test adapted to event times known only as intervals.
Everyday example: Children who did and did not receive fluoride varnish are given as two groups. If the test is significant, the two groups' decay-free curves differ more than chance would explain; for example, decay appears later in children who received varnish.
What it does: A separate Turnbull NPMLE curve for each level and a generalised log-rank test for equality of groups (Finkelstein 1986 scores, under the pooled NPMLE; permutation variance, Fay and Shaw 2010).
Effect on the result: Group curves, group medians and the χ² test are added; the whole-sample curve does not change. Group percentiles are searched on each group's own grid.
On the validation data: BCS trt groups: generalised log-rank χ² = 7.2033; the median of the RT group is 40. HIV Stage: χ² = 7.7379.
Regression model (None / Weibull / Log-normal / Log-logistic / Semi-parametric PH (Cox-type))
Concept: Chooses the model used to estimate the effects of factors. Parametric AFT models (Weibull, log-normal, log-logistic) give the effect as a time ratio; the semi-parametric proportional hazards model gives it as a hazard ratio without assuming a distributional shape.
Everyday example: If a Weibull AFT model finds a time ratio of 1.4 for varnish, the time to decay is about 40% longer in children who received it. The semi-parametric model expresses the same finding as 'the risk of decay forming at any moment is this much lower'.
What it does: Covariate model: parametric AFT with the interval-censored likelihood, or semi-parametric PH (S(t|x) = S0(t)^exp(xβ), S0 NPMLE on Turnbull intervals).
Effect on the result: An additional regression table (AFT: coefficient, time ratio, scale, logL, AIC; PH: HR, logL).
On the validation data: BCS trt time ratio: Weibull 0.567551, log-normal 0.421, log-logistic 0.487307. IR_diabetes semi-parametric PH gender_male coefficient -0.140236. BCS semi-parametric PH logL -133.034249.
→ Accelerated failure time (AFT) model and survival distributions
Regression covariates
Concept: These are the explanatory variables that enter the regression model. A variable coded as text is expanded into dummy variables and each level is read relative to the base level.
Everyday example: If daily sugar intake and tooth-brushing frequency are added as covariates alongside varnish, the effect of varnish is estimated with these habits held fixed.
What it does: Columns to enter the regression model; a text column becomes K-1 dummies (the alphabetically first level is the baseline level).
Effect on the result: Changes no output when the model is "None"; with no column selected only the intercept model is fitted. If semi-parametric PH is chosen with no covariate, a warning is given and only the NPMLE is produced.
On the validation data: BCS Weibull with no covariate: scale 0.666668. mice sex_k + age (Weibull): scale 0.588644, logL = -85.5467, age coefficient 0.042371.
Additional output: Turnbull intervals and masses
Concept: Lists the innermost intervals in which Turnbull's method places the probability mass and the probability falling in each. The curve is undefined inside these intervals; the table shows which parts of the data the estimate rests on.
Everyday example: A row '18-24 months: 0.12' in the table says decay is estimated to have formed in this interval for about 12% of the children. Exactly what happened inside the interval, for example at month 20, cannot be known from the data.
What it does: Table of Turnbull innermost intervals and NPMLE probability masses (only intervals with positive mass, rounded limits).
Effect on the result: Only an additional table; the curve and tests do not change.
On the validation data: The BCS table consists of 12 intervals with positive mass.
Confidence Level (%90 / %95 / %99 / Other, custom value)
Concept: Sets the level of the interval estimates for survival probabilities, percentile times and regression coefficients. Because there is no simple standard-error formula for the Turnbull estimate, the intervals are built by the bootstrap, drawing many resamples from the data.
Everyday example: The median time to decay might be 30 months with a 95% bootstrap interval of 24-36 months. The interval is this wide because the visits were six months apart; with visits every three months it could be narrower.
What it does: Level of the percentile bootstrap (B = 500, fixed seed) interval for S(t) and percentiles, and of the regression interval.
Effect on the result: Only the interval limits.
On the validation data: BCS, S(25): %95 [0.416156, 0.656948]; %99 [0.376138, 0.689]; %80 [0.45127, 0.603081].
Example on real data
Data: Breast cosmesis (Finkelstein and Wolfe 1985), SAS ICLIFETEST bcs
Variables: lower bound lTime, upper bound rTime (months; an empty upper bound is right-censored)
- Question: If the onset time of cosmetic deterioration after breast cancer treatment is known only as an interval between check-up visits (n = 94; 51 interval, 5 left, 38 right censored), how is the survival curve estimated?
- Method: Turnbull NPMLE: the probability mass is distributed over the innermost intervals formed by intersections of the observation intervals; since the time within an interval is unknown, the curve is defined only at the ends of these intervals.
- Result: The probability of remaining free of deterioration is S(8) = 0.8764, S(17) = 0.7368, S(31) = 0.4304, S(45) = 0.3039, S(48) = 0.1170.
- Percentiles: The median time to deterioration is 31 months (95% bootstrap CI [20, 39]); 25th percentile 17 months [12, 20], 75th percentile 48 months [38, 60].
- Limitation: The CIs are built by percentile bootstrap (B = 500) and depend on random resamples; treating the interval midpoint as the event time and using KM gives biased results.
- APA: According to the Turnbull estimate, the median time to deterioration was 31 months, 95% CI [20, 39] (N = 94).
Related concepts
Censoring (survival analysis)Turnbull NPMLE (interval-censored data)Log-rank test (and Breslow, Tarone-Ware)Accelerated failure time (AFT) model and survival distributionsKaplan-Meier estimator (survival curve)CovariateMaximum likelihood (ML) estimation
The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.