Survivorship bias

praeiudicium superstitum

About

Survivorship bias is the drawing of conclusions from what remains observable, while cases removed by failure or exclusion go uncounted. The missing are precisely those with the most to report.

Taxonomy - Selection bias, sampling-frame error. The defect is inferring from a conditioned sample as if it represented the full domain.

Informal Example - "Successful founders dropped out of school; therefore, dropping out is a good path to success," while ignoring failed dropouts.

Minimal Symbolic Form

Canonical Form:

Observe(X | Survived) ⊢ Generalize(X)

Where:

The survivor-conditioned sample is treated as if it were unconditioned; the filter is forgotten.

Sampling Analysis

Invalidity Statement:

P(X | Survived) ⇏ P(X)

A conditional rate among survivors does not entail the population rate.

Missing Necessary Condition

Generalization requires the selection process not distort the property being measured:

Survived ⊥ X

If survival and X are dependent, survivor observations are biased.

Selection Model

Biased Model:

Observed = Population

Corrective Model:

Population → Selection(Survived) → Observed

The observed set is downstream of a selection filter. Missing cases are part of the evidential structure.

Conditional Probability Formalization

The biased substitution is:

P(X) := P(X | S)

The correction requires modeling selection:

P(X) = P(X | S)P(S) + P(X | ¬S)P(¬S)

where S denotes survival or inclusion.

Epistemic Representation

𝔼(Observed(X | S)) ⇏ □(General(X))

Observation of survivors alone does not justify a population-level conclusion.

Countermodels

Survivors Differ From Failures

P(X | S) ≠ P(X | ¬S)

The observed cases can systematically differ from missing cases.

Failure-Adjusted Sample

Observe(S) ∧ Observe(¬S)

Including failures can reverse the apparent inference.

Historical Background

There is no traditional Latin name for survivorship bias. The Neo-Latin praeiudicium superstitum is descriptive.

The broader logical problem has older roots in induction: a sample cannot support a general conclusion unless the sampling conditions are understood.

Its modern formal home is selection bias and conditional probability, but the logical concern is continuous with older cautions against hasty generalization from partial cases.

© Logicae. A logic reference project.