Diagnostic film
Survivorship Bias
Survivorship bias is the error of drawing conclusions from a sample restricted to the cases that survived a selection process, while the cases that were eliminated and have disappeared from view are silently excluded.
hover to expose the x-ray · the absence is the finding
As a statistical/econometric phenomenon, survivorship bias is not a fragile psychology-lab effect that needs replicating — it is a deductively necessary consequence of non-random sample attrition and is repeatedly and independently quantified across finance datasets. Multiple groups (Brown et al. 1992; Brown, Goetzmann & Ibbotson 1999; Fung & Hsieh 2000; Liang 2001; Malkiel & Saha 2005; Ibbotson, Chen & Zhu) measure hedge-fund survivorship bias in the same direction at roughly 240-440 bps/year, and CRSP equity comparisons show ~1.6% annualized differences. The magnitude varies by database and period, but the existence and sign of the effect are not in scholarly dispute. The cognitive-bias overlay (its tie to availability and how strongly it distorts everyday human judgment) is less rigorously tested.
Reading the lesion
Survivorship bias is fundamentally a sample-construction (selection) error, not a quirk of memory: an attrition process non-randomly removes cases (failed funds, downed planes, bankrupt firms, patients who died or dropped out), so the surviving sample is unrepresentative of the original population. Because the eliminated cases are literally absent from the data, naive estimates computed on survivors are systematically shifted — usually optimistic — and the missingness is invisible to the analyst. Cognitively, the error is reinforced by the availability heuristic: survivors are the only examples one can see and recall, so their salience makes the inference feel complete.
There is a genuine distinction between the statistical phenomenon (a special case of selection/sampling bias with non-random attrition, fully formalizable with censoring/truncation models) and the cognitive phenomenon (a reasoning failure driven by attention to visible cases). Most scholars treat 'survivorship bias' as the data/selection-bias concept; the cognitive-bias framing is a popularization that links it to availability. The two are mechanistically linked but not identical — one is about how the sample was built, the other about how a mind fills the gap.
Exposures on file
Areas with few hits among survivors are the vulnerable areas; armor should be added where returning planes show the least damage, because planes hit there did not survive to be measured. — Abraham Wald (Statistical Research Group), 1943 · Damage records of U.S. bombers that returned from missions in WWII
Survival-truncated samples manufacture an apparent relation between volatility and return that looks like return predictability/persistence, even when none exists; survivorship can be strong enough to account for evidence cited as predictability. — Stephen J. Brown, William N. Goetzmann, Roger G. Ibbotson, Stephen A. Ross, 1992 · Theoretical/simulation framework calibrated to fund-performance settings
Using surviving funds only produces a strong upward bias in average returns; backfill/instant-history bias compounds the problem. — Burton G. Malkiel, Atanu Saha, 2005 · Hedge funds in the TASS database, 1990s-early 2000s
Cases that reached the lightbox
- Aviation / military operations research
WWII bomber armor (Statistical Research Group). Abraham Wald advised the U.S. military to reinforce aircraft regions that showed the least damage among returning bombers, recognizing that planes hit in those regions (e.g., engines) did not return and so were absent from the sample. His 1943 memoranda were used through the Korean and Vietnam wars.
- Finance / asset management
Mutual fund performance overstatement. Funds that close due to poor performance ('defunct fund problem') drop out of commonly used databases, so studies of surviving funds overstate average performance and can manufacture apparent persistence. Brown, Goetzmann, Ibbotson & Ross (1992) showed truncation by survival can mimic return predictability; Carhart, Carpenter, Lynch & Musto (2002) showed survivorship bias weakens measured persistence (bias ~17 bps/yr for short samples to ~1%/yr for samples of 15+ years).
- Finance / hedge funds
Hedge fund database bias (TASS/Tremont). Surviving-fund-only returns overstate hedge fund performance by ~240-440 bps/year; Malkiel & Saha (2005) found bias up to 442 bps, and Ibbotson, Chen & Zhu reported ~2.74%/yr survivorship plus additional backfill bias raising total overstatement to ~5.68%/yr over 1995-2006.
- Finance / quantitative trading
Equity index/backtest survivorship (CRSP). Backtests run on indices that drop delisted, merged, or bankrupt stocks overstate strategy returns. Comparisons using the CRSP US Stock Database show ~7.4% annualized returns in survivorship-free data vs ~9.0% in survivorship-biased data (1926-2001); Bessembinder (2018) found only ~42% of US common stocks beat one-month Treasuries over their lifetime, with most aggregate gains from a tiny minority.
Recommended correction
Explicitly enumerate and, where possible, recover the eliminated cases before generalizing: use survivorship-free datasets (e.g., CRSP/Compustat with delisting returns), intention-to-treat rather than per-protocol analysis, and pre-registration to surface 'dead' results. Ask what selection process produced the visible sample and model the missingness rather than ignoring it.
Carhart, Carpenter, Lynch & Musto (2002) showed that including defunct funds reveals survivorship bias of ~17 bps/yr for short samples up to ~1%/yr for samples of 15+ years and weakens measured performance persistence; finance moved to survivorship-free databases precisely because surviving-only samples overstate returns by roughly 1-4%+/year.
Adjacent films