On this page
During the Second World War, analysts looked at aircraft that came back from missions and asked where to add armour. The obvious answer was "where the bullet holes are". The statistician Abraham Wald saw why that answer could be backwards. His reasoning now has a name, survivorship bias, and it shows up everywhere from exam advice to research papers. This post sets out what is documented about Wald's work, walks through a worked example with sample numbers, and shows how to spot the same trap in your own studies.
Who was Abraham Wald?
MacTutor's biography records that Wald was born on 31 October 1902 in Kolozsvár, then in Hungary and now Cluj in Romania. He earned a doctorate at the University of Vienna in 1931 under Karl Menger. In the summer of 1938 he moved to the United States at the invitation of the Cowles Commission, and he was on the faculty of Columbia University from 1939 until 1950. He published the monograph *Sequential Analysis* in 1947 and developed the theory of statistical decision functions. He and his wife died in a plane crash in India on 13 December 1950; MacTutor places it in Travancore.
What problem was Wald working on?
During the war Wald was a member of the Statistical Research Group (SRG) at Columbia. W. Allen Wallis's history, The Statistical Research Group, 1942-1945, describes it as an Office of Scientific Research and Development activity at Columbia and gives an account of how sequential analysis began there in 1943.
According to Marc Mangel and Francisco Samaniego, in Abraham Wald's Work on Aircraft Survivability (Journal of the American Statistical Association, 1984), Wald worked on estimating the vulnerability of aircraft using data obtained from survivors. The work was published as a series of SRG memoranda, used in the Second World War and in the wars in Korea and Vietnam, and later reissued by the Center for Naval Analyses. Mangel and Samaniego describe his ideas and methods and re-examine his main results using statistical theory, which is a reminder that Wald's contribution was a method of estimation, not just a clever remark.
What was the insight?
The returning planes are a filtered sample. A plane hit in a place that brings it down never lands, so its damage is never counted. A 2015 paper in PLOS ONE, which borrowed Wald's reasoning for cancer genetics, summarises it: Wald looked for vulnerable zones by studying the bullet-free patches on aircraft that returned, because if bullets hit those areas the planes crashed and the evidence became unobservable. The bullet-free zones on returning aircraft were the ones that needed more armour.
Worked example: where are the missing holes?
These are sample numbers for a classroom example, not wartime data. Suppose 200 planes each take one hit, and the hit is equally likely to land in any of four areas: engine, fuselage, wings or tail. That is 50 hits per area. Suppose a hit on the engine brings the plane down 60% of the time, and a hit anywhere else brings it down 10% of the time.
Engine hits that return: 50 × (1 − 0.6) = 20 planes.
Each other area: 50 × (1 − 0.1) = 45 planes, so the three areas give 135.
Planes that return: 20 + 135 = 155.
Share of returning damage on the engine: 20 ÷ 155 ≈ 12.9%, although the engine took 25% of all hits.
Looking only at survivors, the engine seems the safest place to be hit. It is the most dangerous.
Now reverse the reasoning. If hits were spread evenly, you would expect 50 engine hits but you see only 20 among survivors. So 30 of the 50 planes hit there did not return, and the estimated chance that an engine hit is fatal is 30 ÷ 50 = 0.6. The missing holes reveal the weakness. Real analysis needs more care, such as planes taking several hits, which is why Wald's memoranda were mathematical.
Where does survivorship bias show up for students?
Topper interviews. You hear from students who followed a routine and succeeded, not from those who followed the same routine and did not.
Success stories about one book or app. People for whom it did not work rarely write about it.
Your own error log. If you only record mistakes you noticed, the ones you never caught are missing, and those are the dangerous ones. Comparing your work with full solutions catches more of them.
When an error keeps surviving because you cannot see where your method breaks, the Fermi Problem Library has worked problems with a step-by-step solver that explains the reasoning rather than printing an answer, so you can compare each of your steps with a complete line of logic.
What can a student learn from how Wald worked?
Wald's question was not only "what does the data show?" but "how was this data produced, and what could never reach me?" He then turned that intuition into a calculation others could check and reuse. Before you trust a pattern in any experiment, ask what had to happen for each data point to be recorded. Our post on conditional probability and Bayes' theorem uses the same "given that it survived" logic.
Try it now: redo the worked example with a 40% loss rate for engine hits and find the share of engine damage among returning planes.
