Survivorship Bias
Survivorship Bias: You only see the winners. The ventures that failed, the planes that were shot down, the strategies that went bankrupt β they're not visible to give you data. When you study successful companies, successful people, or successful strategies without examining the failures, you systematically overestimate success rates and misidentify the causes of success.
What Is Survivorship Bias?β
The canonical example comes from World War II statistician Abraham Wald. The US military was studying damaged aircraft returning from missions to determine where to add armour β they found bullet holes concentrated in the fuselage and wings. The naive conclusion: reinforce those areas.
Wald recognised the fatal flaw: the military was only looking at planes that returned. The planes hit in the engine and cockpit didn't return. The visible bullet holes were on the least critical areas β the planes survived despite those hits. The invisible data (the shot-down planes) held the critical information. Wald's recommendation: reinforce the areas with no bullet holes on returning planes, because those are the areas that, when hit, prevented return.
This insight applies to virtually any domain where success is more visible than failure: business strategy, investment, personal development, medical treatment, academic research.
How It Worksβ
The survivorship bias mechanism:
1. A population of entities attempts something (start businesses,
invest in strategies, apply to jobs)
2. Most fail; a minority succeeds
3. Failures become invisible (companies dissolve, funds close,
people don't write memoirs about failures)
4. Observers study only the visible survivors
5. Conclusions drawn from survivors are systematically biased
Common domains:
β Business strategy: "successful companies all did X" (ignoring failed
companies that also did X)
β Investment: hedge funds with the best 5-year returns (ignoring the
funds that closed in the same period)
β Survivorship in publishing: successful authors who wrote 10 novels
before getting published (ignoring equally persistent writers who never did)
β Medical: treatments that "worked" in patients who survived
(ignoring patients who died despite identical treatment)
Three Real-World Examplesβ
Mutual Fund Performance Studiesβ
Studies of mutual fund performance systematically suffer from survivorship bias: funds that perform badly close or merge; only the survivors are in the database used for analysis. A landmark study by Burton Malkiel (1995) found that studies of mutual fund performance using only surviving funds overstated average returns by approximately 1.4% per year β enough to make mediocre active management appear market-beating over 10-year periods. This bias contributed significantly to the false belief that active fund management beats index funds.
World War II Aircraft Armourβ
The original Wald case. The military's initial analysis recommended reinforcing the areas with visible bullet damage on returning aircraft. Wald's insight: those areas were where planes could survive being hit. The truly vulnerable areas showed no damage on survivors β because planes hit there didn't return. The correct action was the counterintuitive one: armour where there are no bullet holes.
Silicon Valley Dropout Mythologyβ
The "dropout founder" mythology β drop out of college to start a startup β is a survivorship bias narrative. Bill Gates, Steve Jobs, Mark Zuckerberg, and a handful of others successfully dropped out. The many thousands who dropped out and failed are invisible. A study of startup founders found that college-educated founders had significantly higher success rates than dropouts; the dropout success stories are vivid precisely because they're rare exceptions.
When to Recognise Itβ
π¨ Survivorship Bias is likely present when:
- You're studying successful entities to learn what caused their success
- Your dataset only includes things that "made it through" a process
- The failures are invisible or uncounted in your data
- You're drawing lessons from visible examples without asking "what about the examples I can't see?"
β Countermeasures:
- Explicitly ask: "What happened to the failures? Where are they?"
- Seek data on the full population, not just the visible winners
- Before generalising from success stories, ask: "Did equally many failures share these characteristics?"
- Apply reference class forecasting: what's the success rate for the entire population?
| Pairs well with | Why |
|---|---|
| Reference Class Forecasting | RCF uses full population base rates rather than survivor samples |
| Confirmation Bias | Survivorship bias provides the biased sample; confirmation bias prevents noticing |
| Base Rate Neglect | Survivorship bias produces base rate errors by excluding non-survivors |
| Narrative Fallacy | Success narratives are mostly survivorship bias stories |
Common Misuses and Limitationsβ
Over-applying it to dismiss all success-based learning. Survivorship bias is a warning about studying only survivors without controlling for failures β it doesn't make success stories uninformative. You can learn from survivors; you just need to compare them to equivalent non-survivors to distinguish causation from coincidence.
Assuming the missing data always reverses the conclusion. Sometimes the survivors do share genuinely differentiating traits that failures lacked. The problem is assuming success-linked traits are causal without examining the distribution among failures.
Related Modelsβ
| Model | Relationship |
|---|---|
| Reference Class Forecasting | RCF uses full population data to counteract survivorship bias |
| Confirmation Bias | Both distort what evidence reaches the analysis |
| Narrative Fallacy | Success narratives are survivorship bias made emotionally compelling |
Frequently Asked Questionsβ
How do you correct for survivorship bias in research?
Three approaches: (1) broaden the dataset β explicitly include failures by searching for companies, funds, or strategies that attempted the same thing and failed; (2) use inception-to-date data β instead of studying only currently existing entities, use data from when all entities in the cohort started; (3) control studies β compare successful cases to equally matched cases that failed, looking for what differentiates them. Any dataset that only includes "things that still exist" or "things people write about" is potentially survival-biased.
Is survivorship bias the same as selection bias?
Survivorship bias is a specific type of selection bias β the selection mechanism is survival (passing through some filter). Selection bias is the broader category: any systematic non-random selection of data. Medical studies with non-representative patient recruitment have selection bias but not necessarily survivorship bias. Survivorship bias specifically involves the invisibility of failures due to their removal from the observable population.
What's the best example of survivorship bias being corrected and changing conclusions?
The mutual fund example is the best-documented. Before survivorship-bias-corrected databases (like CRSP's survivor-bias-free US mutual fund database), academic studies consistently found that active management appeared to outperform index funds over long periods. When the closed and merged funds were added back into the dataset, the advantage disappeared in most studies. The correction didn't change the data β it changed which data was included.
Further Readingβ
- Malkiel, B. (1995). "Returns from Investing in Equity Mutual Funds." Journal of Finance β the mutual fund survivorship study
- Taleb, N.N. (2001). Fooled by Randomness β survivorship bias and its effects on financial thinking
- Wald, A. (1943). "A Method of Estimating Plane Vulnerability" β the original wartime report
Apply with AIβ
π Check your analysis for survivorship bias with MindMax β
This page is part of the MindMax Mental Models Knowledge Base.