Description
Correct time-to-event estimates when people enter the data only after surviving some time, or are recorded only if their event happened before a cut-off.
Truncation is not censoring. A censored person is known to exist, with a bound on their time; a truncated person is invisible if their time lies outside the observable range. Left truncation (delayed entry) arises when people join observation only after surviving for a while, such as users who enter a dataset only once they pass onboarding, so early leavers are missing entirely. Right truncation arises when only events before an analysis date are reported, such as issues logged by users whose problems have already occurred, with no count of those still at risk. The remedy is to divide each person's likelihood contribution by the probability of having been observable given their truncation time. Turnbull's method gives a nonparametric conditional estimate beyond the smallest entry time, and parametric models can estimate the unconditional distribution, but precision drops sharply as more of the population is unseen.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Days
Steps
- For each person record the entry (truncation) time before which they could not have been observed, or the cut-off after which they would not have been reported, alongside their event or censoring time.
- Write each likelihood term as a conditional probability: the event or censoring probability divided by the probability of surviving to the entry time (left truncation) or of an event before the cut-off (right truncation).
- For a nonparametric estimate, use risk sets that include only people already under observation at each time (Kaplan-Meier with delayed entry, or Turnbull for overlapping intervals); report the curve only beyond the smallest entry time.
- For a parametric estimate, fit the conditional likelihood by maximum likelihood and compare candidate distributions.
- Where possible, recover the number of truncated people (for example from sign-up records), which converts truncation into censoring and greatly improves precision.
- Report how much of the population is likely unseen; with right truncation, treat estimates as fragile unless only a small share (around 5 percent) is truncated.
Sources
- Statistical Methods for Reliability Data (Meeker & Escobar, 1998)
Also known as
- Left truncation
- Delayed entry
- Right truncation
- Conditional likelihood for truncated data
- Analysis of left- and right-truncated life data by conditional likelihood
Categories
- Analytics & Behaviour
- Statistics & Modelling