Survivorship Bias

intent

survivorship bias is the mistake of drawing conclusions from a filtered sample (survivors only) and treating survivor traits as the cause of survival, when non-survivors (absent from data by definition) may have had identical traits and still failed. activate this skill when someone cites "what winners did," investment returns, fund performance, backtested strategies, or career advice from top performers. the skill forces explicit reconstruction of what non-survivors likely had, estimates the strength of the causal claim, and marks conclusions as conditional on survivor data until selection-corrected population data exists. canon: wald (1943) reversed the navy's bomber-armor recommendation by flipping the frame from "where returning planes show damage" to "where returning planes show no damage" (because those planes were shot down).

inputs

• user claim: explicit statement ("successful startups do X," "this fund returned 15% annually," "billionaires all read books")
• sample description: what data the user has (list of companies, fund prospectuses, published studies, interview transcripts of top performers)
• data source: where the sample came from (surviving companies database, published research, active fund listings, news coverage of winners)
• population estimate: optional but required for selection correction (total number of entities before the filter; e.g., "100 AI startups funded" implies ~10k unfunded attempts)
• survival rate or dropout rate: percentage removed by the filter (e.g., "92% of startups fail," "X% of mutual funds close")
• context on intent: is the user making a causal claim ("do this to succeed") or a descriptive one ("here's what winners looked like")?

external connections: none. this skill runs on logic and user-supplied data only.

edge cases to handle:

• empty or single-case survivor sample (n=1 example proves nothing about frequency)
• user conflates "survivorship bias" with "selection bias" (different; note the distinction)
• user has already corrected for selection (verify by asking what non-survivor data they used)
• user claims non-survivor data is "unknowable" or "proprietary" (it may be; skill response is then "mark as conditional" rather than "conclude")
• sample contains both survivors and non-survivors but user only cites the survivors (common; extract this explicitly)

procedure

Step 1 , State the claim and sample.

• Input: user's verbatim statement + source of data.
• Output: one-sentence restatement of the causal or descriptive claim, plus explicit description of what the sample contains.
• Example: "You say 'successful founders read 50 books/year.' Your sample is a list of 10 billionaires who mention reading in interviews."

Step 2 , Identify the survival filter.

• Input: sample description, data source, user's knowledge of population.
• Output: explicit statement of what process removed non-survivors.
• Probe: "What had to be true for someone to appear in your sample? What happened to people who didn't have that trait or weren't included?"
• Output: estimate of population size before filter, fraction removed, and what the filter selected for.
• Example: "Filter: 'founder is now a billionaire.' Population: ~7 billion humans. Fraction removed: 99.99999%. Selection mechanism: extreme wealth + access to media + willing to be interviewed."

Step 3 , Construct the non-survivor hypothesis.

• Input: the trait cited as causal (e.g., "reads 50 books/year").
• Output: plausible statement of what non-survivors likely had (e.g., "failed founders also read, or read zero books and succeeded anyway via luck or networks").
• Probe: "Did non-survivors (failed founders, non-billionaires) have this trait? What evidence would show they did?"
• Output: explicit uncertainty range ("it's plausible non-survivors had the same trait" vs. "it's unlikely they did, and here's why").

Step 4 , Re-estimate causal strength.

• Input: the claim, the non-survivor hypothesis, any available comparative data.
• Output: confidence range from "best case" (non-survivors lacked the trait, strong causal signal) to "worst case" (non-survivors had it too, trait is not causal).
• Probe: "If we could see the reading habits of 1000 failed founders, what would change your confidence?"
• Output: explicit specification of what data would resolve the ambiguity.
• Example: "Best case: 90% of billionaires read 50+/year, 5% of failed founders do. Worst case: identical distribution. Data that settles it: reading habits of a random sample of founders who failed."

Step 5 , Correct or mark conditional.

• Input: population data (if available) or explicit lack thereof.
• If population data available: re-run analysis with selection correction (use observed rates in both survivors and non-survivors to compute adjusted odds ratio or effect size).
• If population data unavailable: mark the conclusion as conditional on the survivor sample. State explicitly what assumption is required for the claim to hold.
• Output: selection-corrected inference or conditional statement.
• Example: "Conditional on survivor sample: reading is correlated with billionaire status. Selection-corrected inference: unavailable without reading data from failed founders. Claim requires the assumption that non-survivors read at a lower rate; if they read at the same rate, reading does not explain success."

decision points

if user has explicit population data (survivor count + non-survivor count, or known dropout rate):

• run selection correction: compute odds ratio, hazard ratio, or effect size conditional on both groups.
• output: selection-corrected confidence interval.

else if user has only survivor data:

• construct non-survivor hypothesis based on domain knowledge.
• mark conclusion as "conditional on survivor sample" or "unfalsifiable from available data."
• output: confidence downgraded; specify what data would upgrade it.

if sample is explicitly restricted to survivors with no population claim:

• e.g., "I am studying what successful founders do, not why they succeeded."
• decision: skill does not apply. acknowledge and stop.
• output: "this lens doesn't fit your question."

if user conflates correlation with causation:

• e.g., "billionaires read books, so reading causes billionaire status."
• decision: reframe as selection problem first, then note that even selection-corrected correlation ≠ causation.
• output: explicit statement of causal claim and what additional evidence (randomized trial, instrumental variable) would be needed.

if user claims non-survivor data is "proprietary" or "impossible to get":

• decision: accept this constraint and mark conclusion as conditional.
• output: "this claim is unfalsifiable without X data; proceeding assumes [constraint]."

output contract

output is a markdown document with this structure:

# Survivorship Bias Analysis: [Claim Title]

## Claim, Sample, Source
[One-sentence restatement of the causal or descriptive claim.]
Sample: [explicit description of what the sample contains, how many, from where].
Source: [where the data came from, who collected it, any curation or filtering in the source itself].

## Survival Filter
Population before filter: [estimate, with uncertainty if needed].
Survival mechanism: [explicit statement of what removed non-survivors].
Estimated survival rate: [percentage or fraction, with confidence].
Selection criterion: [what trait or outcome defined "survivor"].

## Non-Survivor Hypothesis
Traits cited as causal in survivors: [list, e.g., "reads 50 books/year"].
Plausible traits in non-survivors: [e.g., "also read, or read zero books"].
Evidence for/against non-survivors having the same traits: [what we know, what we don't].

## Causal Strength: Best Case vs. Worst Case
Best case (trait explains survival): [scenario + estimate, e.g., "90% of billionaires read, 5% of failed founders do"].
Worst case (trait does not explain survival): [scenario + estimate, e.g., "identical distributions"].
Data that would resolve the ambiguity: [specific population data needed].

## Selection-Corrected Inference (or Mark as Conditional)
[If population data is available: adjusted effect size, confidence interval, causal conclusion.]
[If population data is unavailable: conditional statement + required assumption.]
Confidence level: [high / medium / low].
Recommendation: [proceed with caution / mark as unfalsifiable / get [specific data]].

outcome signal

the user knows the skill worked when:

1. claim is restated with explicit sample and source , user sees "you said X, from a sample of Y survivors, drawn from Z source."
2. survival filter is named , user can identify what process removed non-survivors (e.g., "only billionaires appear in media coverage").
3. non-survivor hypothesis is constructed , user can articulate a plausible alternative ("failed founders also read books, or didn't read and still failed").
4. causal strength is bounded , user sees a confidence range ("best case: strong causal signal; worst case: trait is not causal").
5. conclusion is marked conditional or selection-corrected , user sees either "this claim is conditional on [assumption]" or "here is the selection-corrected estimate."
6. user's next decision is informed , user either decides to get population data, deprioritizes the claim, or proceeds with explicit caveats.

method in action:

• abraham wald and the statistical research group, 1943 (bomber armor): https://en.wikipedia.org/wiki/Survivorship_bias#Abraham_Wald
• mutual fund survivorship and reported returns, 1971-1996: https://doi.org/10.1111/j.1540-6261.1997.tb03808.x
• ai-startup survivorship, 2023-2026: funded unicorns vs the dead-wrapper graveyard

common survivor patterns by domain:

common rationalizations and traps:

red flags (activate skill immediately):

• sample described as "successful X" or "X who made it"
• data source is survivor-filtered (active funds, surviving companies, published studies, billionaire interviews)
• base rate of failure or dropout is not stated
• population claim is inferred from survivor subset without correction
• user conflates correlation with causation in survivor data
• backtest or historical performance is cited without disclosure of survivorship bias

verification checklist (use before signing off):

• survival filter explicitly identified
• non-survivor population size estimated or marked "unknown"
• non-survivor hypothesis constructed and plausible
• causal strength bounded (best case / worst case)
• conclusions marked conditional on survivor sample, or formally selection-corrected
• required data for resolution specified
• recommendation is actionable (proceed with caveats / get data / deprioritize)

source & attribution:

• original skill authored by clawhub.
• enriched for implexa standards with explicit decision logic, edge cases, and outcome signals.
• canonical reference: abraham wald, "a method of estimating missing values in multivariate data" (1943, statistical research group, columbia).
• survivorship bias library: https://en.wikipedia.org/wiki/Survivorship_bias