Description
Estimate a time-to-event distribution without assuming its shape when each event is known only to fall inside an interval, and intervals from different people overlap.
Kaplan-Meier needs exact event times or non-overlapping intervals so that the risk set can be counted. When people are checked on irregular schedules, some events are only known to have happened before a check (left censored), some after the last check (right censored) and some between two checks, and these intervals overlap. Peto and Turnbull's estimator writes the likelihood with each person contributing the probability mass inside their interval, attaches unknown probabilities to the elementary intervals formed by all distinct check times, and maximises it, usually with Turnbull's EM (self-consistency) algorithm. The result is a nondecreasing estimate of the fraction who have experienced the event, defined at interval endpoints, with standard errors and logit-transformed intervals. With simple right censoring it reduces to Kaplan-Meier. Crude proportions per check age can go down by chance; this estimator cannot.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Evaluation, Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- For every person record the interval (last check without the event, first check with it]; use 0 as the lower end for left-censored cases and infinity as the upper end for right-censored ones.
- List all distinct interval endpoints and form the elementary intervals between them, attaching an unknown probability to each.
- Write the likelihood as the product, over people, of the summed probabilities inside each person's interval.
- Maximise it subject to the probabilities summing to one, for example with Turnbull's EM algorithm iterated to convergence.
- Accumulate the probabilities to get the cumulative fraction with the event at each endpoint; leave it undefined inside intervals that carry mass.
- Compute standard errors and logit-transformed pointwise intervals (likelihood-ratio intervals are more accurate in small samples) and plot the step estimate.
Sources
- Statistical Methods for Reliability Data (Meeker & Escobar, 1998)
Also known as
- Turnbull estimator
- Peto-Turnbull estimator
- Nonparametric MLE for interval-censored data
- Self-consistent (EM) survival estimator
- Peto-Turnbull nonparametric maximum likelihood estimate for arbitrarily censored data
Categories
- Analytics & Behaviour
- Statistics & Modelling