Survival or Time-to-Event Analysis¶

Model — instantiates Survival-Conditioned Persistence Forecasting

Fits a lifetime distribution and hazard function from durations that include still-alive (censored) cases, turning a set of survivors and exits into an estimated curve of risk over time.

Survival or Time-to-Event Analysis is the estimation engine of the archetype. Given a population where some members have exited — their lifetime known exactly — and some are still alive — their lifetime known only to exceed their current age — it estimates the survival function S(t) and the hazard h(t) without discarding the still-alive cases. Its defining idea is that censoring is information, not missing data: a unit still running at age eight tells you its lifetime is at least eight, and the method (Kaplan–Meier non-parametrically, or a parametric or Cox fit) is built to use exactly that partial fact. It produces the risk-over-time curve every downstream mechanism reads from.

Example¶

An HR analytics group wants the tenure distribution of its engineers. Using only the ones who have quit would badly understate tenure — a reverse survivorship trap, since it throws away everyone still employed. So they define the exit event precisely (voluntary resignation counts; an internal transfer is censored, not an exit), assemble the survival-age record (each engineer's tenure to date, flagged as departed or still-here), and fit a Kaplan–Meier curve. Still-employed engineers enter as right-censored — each contributing "tenure ≥ current" without pretending to have left. The output is S(t): the estimated fraction still present at one, two, and five years, with a confidence band, plus a hazard curve showing whether quit-risk spikes in the first year (the familiar early cliff) or builds later. That curve is the raw material; conditioning it on a specific engineer's age comes afterward.

How it works¶

Define the exit event precisely — what counts as death, and what is a competing event to be handled separately.
Build the survival-age record — durations with explicit censoring flags for the still-alive.
Estimate non-parametrically first — Kaplan–Meier assumes no shape, so it shows what the data actually say.
Fit a parametric or hazard model only if justified — impose a functional form only when there's reason to.
Report S(t) and h(t) — the full risk-over-time curve, not merely an average lifetime.

Tuning parameters¶

Event-definition strictness — what counts as the exit and how competing events are treated; a loose definition quietly changes what "lifetime" means.
Non-parametric vs. parametric fit — Kaplan–Meier for honesty about the data, a parametric family for extrapolation and smoothness — at the cost of an assumed shape.
Censoring assumption — whether censoring is taken as non-informative or explicitly modeled; the estimate is only as sound as this assumption.
Covariates — a pooled curve vs. a Cox model that lets covariates shift the hazard.
Time granularity — the resolution at which durations are recorded and the curve is read.

When it helps, and when it misleads¶

Its strength is that it uses the censored cases every naïve approach throws away, and it estimates risk over time rather than a single average — the foundation everything conditional is built on.

Its failures are specific and quiet. Informative censoring — censoring correlated with the risk being estimated — biases the curve, as does left-truncation, where only entities that survived long enough to enter the sample are ever observed.^[1] Too few events leave the tail unestimable, and a wrongly chosen parametric family extrapolates confidently past where the data support it. The classic misuse is running it backwards: selecting the family or the cut-off that yields the survival curve you already wanted. The discipline is to check that censoring is non-informative, to look at the non-parametric Kaplan–Meier estimate before committing to any parametric family, and to route left-truncation questions to the Censoring and Left-Truncation Audit rather than assuming them away.

How it implements the components¶

This model owns the archetype's estimation core — the components that turn raw durations into a fitted curve:

exit_event_definition — time-to-event analysis is impossible without a precise event, so it pins this down first.
observed_survival_age_record — it ingests the durations-with-censoring record as its raw material.
lifetime_distribution_and_hazard_model — its primary output: the fitted S(t) and hazard h(t).

It does not interpret the hazard's shape to classify aging vs. non-aging — that is Hazard-Shape Diagnostic — nor choose among candidate lifetime families (Lifetime Distribution Comparison), nor read off conditional remaining life, which is Age-Conditioned Remaining-Life Table.

Instantiates: Survival-Conditioned Persistence Forecasting — it is the estimation core the appraisal's conditional forecasts are built on.
Consumes: Reference-Class Forecasting — the reference-class boundary fixes which population the lifetime curve is fit on.
Sibling mechanisms: Age-Conditioned Remaining-Life Table · Hazard-Shape Diagnostic · Lifetime Distribution Comparison · Censoring and Left-Truncation Audit · Forecast Backtesting · Change-Point Detection · Reference-Class Forecasting · Rolling Window Comparison · Historical or Holdout Coverage Backtest · Periodic Durability Inspection · Stationarity Test · Lindy Decision-Horizon Review · Non-Aging Eligibility Review

Notes¶

The fitted curve describes the cohort as observed. Using it for a specific surviving member requires conditioning on that member's current age — which is the Age-Conditioned Remaining-Life Table's job. The raw S(t) is not itself a remaining-life forecast, and reading it as one for an already-aged subject is a common and costly slip.

References¶

[1] Left truncation — a form of survivorship bias — arises when only entities that have already survived long enough to enter the sample are ever observed, so the observed set is systematically longer-lived than the population it is meant to describe; informative censoring (censoring correlated with the risk being estimated) produces a related distortion. Both must be corrected for rather than assumed away. ↩