Base Rate Neglect: The Taxi Cab Problem Nobody Gets Right

A hit-and-run accident occurs at night. A single witness identifies the taxi as blue. The city's cab fleet is 85% green and 15% blue. The witness, tested under similar visibility conditions, correctly identifies cab colour 80% of the time. Question: what is the probability that the cab involved was actually blue? Most people answer something around 80%. The correct answer is 41%. The gap between intuition and reality in this problem is not a rounding error — it is a window into one of the most consequential cognitive biases we know.

The Taxi Cab Problem

This scenario was developed by Amos Tversky and Daniel Kahneman in 1982 to illustrate base rate neglect: the tendency to ignore general statistical information (the prior probability, or base rate) when evaluating specific case evidence. The witness testimony, vivid and concrete, dominates our reasoning. The background statistic — 85% of cabs are green — feels abstract and almost irrelevant. But it is crucial.

The correct calculation uses Bayes' theorem. Of all blue cabs, the witness correctly identifies 80% as blue: that's 0.15 × 0.80 = 12% of all cabs. Of all green cabs, the witness mistakenly calls 20% blue: that's 0.85 × 0.20 = 17% of all cabs. So when the witness says "blue," the actual blue cabs account for only 12/(12+17) = 41% of such cases. Because green cabs are so common, the 20% misidentification rate generates more false "blue" identifications than the genuine blue sightings.

Tversky and Kahneman found that when participants were given this problem, the median answer clustered around 80% — essentially ignoring the base rate entirely and treating the witness report as a direct probability. When the base rate was framed causally (the cab company had a bad safety record, causing 15% of accidents) rather than statistically, performance improved somewhat. But the purely statistical framing — the one most relevant to real-world reasoning — was consistently underweighted.

Why We Ignore Base Rates

The failure is not stupidity. It is a feature of how the mind processes different types of information:

• Vividness: The witness's report is a concrete event. The 85/15 fleet distribution is an abstract statistic. Our brain gives more evidential weight to vivid, specific, emotionally resonant information than to dry statistics — a pattern Kahneman describes as the dominance of System 1 (fast, intuitive) over System 2 (slow, deliberate) thinking.
• Representativeness heuristic: We ask "does the description match my mental image of the case?" rather than "given the full population of cases, what is the probability?" The witness report matches "blue cab" — end of reasoning.
• Causal versus statistical information: People integrate base rates more readily when they feel causally relevant (this company causes accidents) rather than merely descriptive (this company has 15% of the fleet). Statistics feel like background noise; causes feel like explanations.

Medical Testing: A Life-or-Death Application

Base rate neglect is most consequential in medical diagnosis. Consider a disease that affects 1 in 1,000 people. A test for it is 99% sensitive (correctly detects the disease when present) and 99% specific (correctly excludes it when absent). You test positive. What's the probability you actually have the disease?

Intuition says: 99%. Reality says: roughly 9%.

Here's why: Test 100,000 people. About 100 have the disease; the test catches 99 of them (99% sensitivity). Of the 99,900 who don't have it, the test incorrectly flags 999 as positive (1% false positive rate). So out of 1,098 positive results, only 99 are true positives. That's a positive predictive value of about 9%.

This calculation — known as the positive predictive value — depends critically on the base rate (prevalence) of the condition. A test with the same accuracy will have a very different positive predictive value depending on whether you're screening the general population (low prevalence → many false positives) or testing a high-risk group (high prevalence → far fewer false positives relative to true positives).

Studies have repeatedly shown that physicians routinely overestimate positive predictive values, failing to update adequately for the base rate of the condition they're testing. One influential 1978 study by Casscells, Schoenberger, and Graboys asked 60 Harvard Medical School students and physicians a version of this problem: only 18% gave the correct answer. This isn't a problem unique to lay people — it affects trained professionals making consequential decisions daily.

Security Screening and Profiling

The same mathematics governs mass security screening. Suppose a screening algorithm correctly identifies a genuine terrorist 95% of the time, and incorrectly flags an innocent person only 0.1% of the time. Impressive accuracy. Now screen a population where 1 in 1,000,000 people is a genuine threat. The algorithm will flag approximately 1 terrorist and 1,000 innocent travellers. For every real threat identified, a thousand innocents are flagged. The base rate makes almost any screening system, regardless of accuracy, a false-alarm machine when the population prevalence is very low.

This is not an argument against screening. It is an argument for understanding what screening can and cannot tell you — and for recognising that the intuitive 95% accuracy figure is almost meaningless without the base rate. Security systems designed by people ignoring base rates will systematically harass innocent people while providing marginal safety benefit, precisely because the targeted population is rare.

Legal Evidence and the Prosecutor's Fallacy

Base rate neglect takes a specific, dangerous form in courtrooms, known as the prosecutor's fallacy. Suppose a DNA match has a random match probability of 1 in a million. The prosecutor argues: "The probability that this DNA match is random is one in a million — therefore the defendant is almost certainly guilty." But this inverts the correct reasoning.

The relevant question is: given the DNA match, what is the probability of guilt? To answer this, you need to know how many people could have been the source. In a large city of 5 million people, roughly 5 people would share that DNA profile by chance. The DNA match narrows the suspect pool to 5 people — it tells you the defendant is one of 5, not that he's the one. The base rate (the size of the plausible population) is essential.

People have been wrongfully convicted due to this error. Expert witnesses present match probabilities; juries hear them as probabilities of guilt; judges fail to correct the inference. The psychological mechanism — ignoring background rates in favour of vivid, specific evidence — is identical to the taxi cab problem, with higher stakes.

The Connection to Other Biases

Base rate neglect connects directly to the representativeness heuristic — judging probability by how well something matches a prototype rather than by statistical frequency. It also feeds into neglect of probability, where risks are evaluated categorically ("dangerous or not") rather than probabilistically. And it compounds hot hand thinking, where the vividness of a recent streak overrides the statistical reality of a player's baseline.

The underlying problem in all these cases is the same: the human mind is not naturally Bayesian. We struggle to weight abstract statistical information against concrete case evidence, even when the statistics are decisive. This is not irrational in every context — vivid, specific information often is more reliable than generic statistics — but it becomes a systematic error when the statistics are large-sample, well-established, and directly relevant.

Improving Base Rate Reasoning

Several techniques help, though none fully eliminates the bias:

• Natural frequencies: Reformulating problems in terms of natural frequencies ("10 out of 1,000 people") rather than percentages ("1%") dramatically improves performance. The frequency format makes the base rate visible and concrete.
• Reference class forecasting: Deliberately identifying the reference class (what population does this case belong to?) before evaluating specific evidence forces base rate consideration.
• Explicit Bayesian training: Brief training in Bayes' theorem, while not creating perfect reasoners, significantly improves performance on diagnostic probability problems.
• Decision support tools: In medicine, calculators and clinical decision support systems that automatically compute positive predictive values from test characteristics and prevalence can bypass the cognitive failure entirely.

Base rate neglect isn't a failure of intelligence — it's a failure of the default cognitive process to handle the relationship between the general and the specific. The taxi may well have been blue. But your confidence should be 41%, not 80%, and the difference between those numbers is exactly the difference between evidence-based reasoning and intuition.

Sources & Further Reading

• Tversky, A., & Kahneman, D. "Evidential impact of base rates." In Judgment Under Uncertainty: Heuristics and Biases, ed. Kahneman, Slovic, & Tversky. Cambridge University Press, 1982.
• Casscells, W., Schoenberger, A., & Graboys, T. "Interpretation by Physicians of Clinical Laboratory Results." New England Journal of Medicine 299, no. 18 (1978): 999–1001.
• Kahneman, D. Thinking, Fast and Slow. Farrar, Straus and Giroux, 2011. Chapters 14–16.
• Gigerenzer, G., & Hoffrage, U. "How to Improve Bayesian Reasoning Without Instruction." Psychological Review 102, no. 4 (1995): 684–704.
• Thompson, W. C., & Schumann, E. L. "Interpretation of Statistical Evidence in Criminal Trials." Law and Human Behavior 11, no. 3 (1987): 167–187.
• Wikipedia: Base rate fallacy