Length time bias¶
A screening-selection bias in which slowly progressing disease remains detectable for longer and is therefore overrepresented among screen-detected cases, making their observed survival appear better even without a screening benefit.
Core Idea¶
Length-time bias arises because periodic screening is more likely to encounter disease that remains asymptomatic and detectable for a long time. Slowly progressing cases receive more chances to be sampled than aggressive cases with short preclinical windows. If slower progression also predicts better survival, the screen-detected group is enriched for favorable biology. If slower progression also predicts better survival, the screen-detected group is enriched for favorable biology.
Scope of Application¶
Use it when evaluating screening cohorts, survival comparisons, and observational claims whose detection process depends on disease duration. Use it when evaluating screening cohorts, survival comparisons, and observational claims whose detection process depends on disease duration.
- Cancer screening. Examines tumor growth-rate selection.
- Diagnostic programs. Checks duration-dependent ascertainment.
- Survival analysis. Questions exchangeability of detected groups.
- Trial interpretation. Separates mortality effects from selected case mix.
- Risk communication. Explains why post-diagnosis survival can mislead.
Clarity¶
The signature requires two links: a longer detectable window raises selection probability, and window length tracks prognosis. Without both, preferential detection need not create the characteristic favorable-survival illusion. The closest near miss sets the boundary: Lead-time bias is closest: it shifts the diagnosis clock earlier for the same course, whereas length-time bias changes which disease trajectories enter the screened group. A positive case must satisfy this test: A case qualifies when detection probability increases with preclinical duration and that duration is associated with prognosis.
Manages Complexity¶
Screening changes the observed population before any outcome model is fitted. Stratification by detected mode alone cannot recover the unsampled fast cases, so more elaborate survival curves may preserve rather than solve the bias. The central early detection–selected biology tradeoff is this: Better observed survival may reflect who is found rather than what screening changes. A second simple cohort–latent cases tension matters because Observed groups omit rapidly passing states.
Abstract Reasoning¶
Use three linked moves: map the disease's preclinical detectable interval; test whether screening opportunity depends on its duration; determine whether duration or growth rate predicts outcome. As a collapse test, the case exits when sampling is independent of detectable duration or when analysis appropriately accounts for the selection mechanism. A fourth check is to separate selection from earlier diagnosis time and overdiagnosis. A final check is to prefer randomized disease-specific mortality or models of the sampling process.
Knowledge Transfer¶
The duration-biased sampling skeleton transfers to inspections that preferentially observe long-lived states. It stops where detectability duration is unrelated to inclusion or outcome; cancer biology and mortality endpoints do not transfer by analogy. The nearest stopping boundary is explicit: Lead-time bias is closest: it shifts the diagnosis clock earlier for the same course, whereas length-time bias changes which disease trajectories enter the screened group. No canonical parent prime is currently asserted; broader structural comparisons remain related-prime analogies until separately adjudicated in the DAG. Inclusion probability differs systematically by prognosis-linked duration. The distorted cohort changes the observed survival distribution.
Relationships to Other Abstractions¶
Current abstraction Length time bias Domain-specific
Parents (1) — more general patterns this builds on
-
Length time bias is a kind of Bias Prime
Length time bias is a domain-specific kind of bias under the frozen identity and differentia.
Hierarchy path (1) — routes to 1 parentless root
- Length time bias → Bias
Neighborhood in Abstraction Space¶
Length time bias sits in a moderately populated region (50th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.
Family — Empirical Measurement & Statistical Inference Methods (50 abstractions)
Nearest neighbors
- Obesity paradox — 0.88
- Assay sensitivity — 0.87
- Mill's Methods — 0.86
- Continuous Individualized Risk Index — 0.86
- False coverage rate — 0.86
Computed from structural-signature embeddings · 2026-10-08