The bombers came back from Europe shot full of holes, and the holes were not evenly spread. Wings and fuselage were pocked and torn. Engines, cockpit, fuel tanks, almost pristine. To the planners studying the damage maps, the reading looked obvious: the parts that keep getting hit are the parts that need armor. A statistician named Abraham Wald looked at the same maps and said the opposite — the empty spaces were the dangerous ones. His reasoning had nothing to do with aircraft engineering and everything to do with a question most people never think to ask about their own data: who is missing from this picture, and why are they missing?
🧠 The Maps That Were Only Half a Census
Abraham Wald was a Hungarian mathematician who had fled Europe in 1938 and ended up at Columbia University, where during the war he joined the Statistical Research Group, a small outfit of statisticians working on military problems. His own biography is a small lesson in the arbitrary nature of obstacles: he was officially classified as an enemy alien and therefore barred from reading the restricted material he was hired to analyze, until a federal court naturalized him early to fix the problem.
Among the group’s outputs were eight memoranda on how to analyze damage data from combat aircraft, and Wald’s 1943 technical report was titled, plainly enough, A Method of Estimating Plane Vulnerability Based on Damage of Survivors. That phrase in the middle — damage of survivors — is the whole story, and Wald knew it. He was working from damage recorded on aircraft that had returned. He used those patterns to estimate how vulnerable each part of a plane was, and noted that the conclusions could be used to decide where to put protective armor.
The version of the story most people know is sharper than that. In the telling, the planners wanted to reinforce the areas with the densest clusters of holes, and Wald stopped them: reinforce the areas with no holes at all. A plane shot in the wing limped home and got counted. A plane shot in the engine never made it into the dataset, because it never made it back to be examined.
⚙️ How It Works
Survivorship bias is not the same as having too little data. It is worse than that: your data was trimmed before you got it, and the trimming was not random. Five steps.
One. Everything you can observe has already passed a filter. The returning bombers passed the filter “hit but still flyable.” Funds that exist today passed the filter “not yet liquidated.” Writers still publishing passed the filter “did not quit in chapter three.” Observing anything at all quietly imposes a condition. Usually the condition is invisible, because it is the reason you have data in the first place.
Two. The entities that got filtered out did not vanish at random — they vanished for a reason. This is the hinge of the whole bias. If losses were decided by a coin flip, a shrunken sample would just be blurrier. But the loss is mechanized: hit something vital, the plane is gone; underperform, the fund closes; get too hard to write, the manuscript stops. The thing deciding whether you can see a case is the same thing you are trying to measure.
Three. So certain features get over-represented and others get erased. Wing damage is over-represented in the maps. Long-run winners are over-represented in the performance tables. What you are looking at is not the distribution of the world. It is the distribution of the part of the world that survived to be counted.
Four. Treat that as the whole population and your conclusion flips. Notice the direction here, because it is the punchline: the areas with the most holes look like the vulnerable areas precisely because they are the survivable ones. The reverse inference is just as quiet. You read a hole-free surface as “never hit” when it may mean “hit fatally, every time.”
Five. Diligence makes it worse, not better. This is why the bias is so hard to feel. Better sampling, more careful accounting, a bigger dataset — all of it sharpens your estimate of a trimmed population, which means you get more confident in the wrong answer. The fix is not analytical. It is one question asked out loud: who is not in this sample, and what made them leave?
🤔 The Twist
The obvious intuition is that this bias is a small-sample problem — gather enough data and it washes out. It does not. It scales.
The cleanest measurement comes from finance, where the missing cases can be counted. Elton, Gruber, and Blake showed in 1996 that US mutual fund data carries a survivorship bias of roughly 0.9% of return per year, defined as the average alpha of surviving funds minus the average alpha of all funds. The bias is bigger in the small-fund sector, where funds are likelier to fold. And the same paper produced a number worth keeping: if the peer group included the funds that had closed, around 70% of today’s surviving funds could honestly claim to be in the top quartile of their peers.
There is a sneakier version that barely looks like a statistical problem. Backtest a strategy using today’s S&P 500 constituents and you have quietly pre-selected companies that survived long enough to qualify and got promoted for doing well. You did not set out to invent a bias. You just picked a list that already had the answer baked into it.
It shows up where you would not expect it too. A 1987 study reported that cats falling from under six stories were injured more severely than cats falling from higher up, which invited an elegant physical explanation about the righting reflex. A plainer explanation also holds: cats that die in the fall are less likely to be carried to a vet, so they never enter the study. The dead cats were not under-represented in the analysis. They were entirely absent from it.
🔗 Why It Matters To You
The finance version is not a curiosity. Any statement of the form “this fund compounded at X% over twenty years” is a statement about funds that were still alive to be measured. The germane question is never about the number. It is about the denominator — what closed, what got quietly merged away, what never got to the exit. In portfolio work, the same question points at your own track record: the deals in the table are the ones that either succeeded or are still standing. Where did the ones that stalled go?
For any companion product — something built to be present for a person — the bias has a specific shape. The people who tell you how it feels are the ones who stayed. The person who opened it twice and drifted away never writes you a paragraph about why. Building from retained users only is studying the returning aircraft. Your fuel tank is empty, and nobody in the room has ever seen it, because nobody who saw it came back.
And it applies to writing, including the advice about writing. Every method you have read — a thousand words a day, outline first, write drunk edit sober — comes from someone who finished. The invisible majority of people who did exactly that and stopped anyway do not publish retrospectives. This is worth remembering both to protect yourself from the advice and as material: a character who has only ever met survivors would misread the world in a specific, interesting way, and would be completely sincere about it.
🎲 A Small Trick
Next time someone assures you that everyone they know who tried this succeeded, add the sentence silently: the ones it failed for do not come around to tell you. They are not lying about their sample. They are just reading a map with the empty spaces still blank.
The one move that helps is not more effort on the data you already hold. It is naming the missing group. Who is not in this room, and what would have had to happen for them to be here?