The Representativeness Heuristic and the Law of Small Numbers: Why We See Patterns That Aren't There

Kahneman and Tversky's research on cognitive heuristics gave us three dominant shortcuts: availability, anchoring, and representativeness. Of these, representativeness is the hardest to explain cleanly and the most consequential in practice.

The representativeness heuristic: when assessing the probability that an event belongs to a particular category or that a process produced a particular outcome, people judge by how much the event resembles the prototype of that category or process — not by base rates, sample sizes, or other statistical properties.

The Mechanism

The core problem: resemblance is not probability. An outcome can look exactly like the output of a process and still be unlikely — if the base rate of that process is low. An outcome can look nothing like what you'd expect and still be probable — if the base rate is high.

Example: the Linda problem (Tversky & Kahneman, 1983): Linda is described as politically active, concerned with discrimination, and a participant in anti-nuclear demonstrations. Which is more probable: (A) Linda is a bank teller, or (B) Linda is a bank teller and active in the feminist movement?

Most respondents choose (B). This is the conjunction fallacy — P(A and B) cannot exceed P(A). But (B) resembles the description more closely, so resemblance gets substituted for probabilistic reasoning.

The Law of Small Numbers

The intuitive belief that small samples are representative of the populations they came from — Kahneman's "law of small numbers" — is a direct application of the representativeness heuristic to sampling.

> 📌 Tversky & Kahneman (1971) asked statisticians whether a study with 10 subjects finding an effect at p = .05 should replicate in the same direction at similar power. Most said yes — the small sample had already provided evidence. In reality, replication probability at n=10 is far lower than intuition suggests. Even professional statisticians held the belief that small samples should reflect population effects. [1]

The consequences:

• Streak perception in sports ("hot hands"): a basketball player making 3 consecutive shots is seen as "on a hot streak." Research shows the actual correlation between consecutive shots is small to zero — but the streak resembles what people expect from a genuinely hot player, so it gets attributed to skill or state.
• Trading on short-term performance: fund managers with 2–3 years of outperformance get credited with investment skill. The base rate for sustained market outperformance over 10+ years among active managers is roughly 2–5% — consistent with random variation. Short-term performance looks like skill because it resembles the prototype of skilled performance.
• Medical pattern recognition: a cancer cluster in a neighborhood triggers investigation for an environmental cause. Random clusters in geographic distributions look identical to clusters produced by environmental causes.

The Statistical Antidote

Ask about the base rate: Before concluding that an observation resembles a category or pattern, establish the prior probability. How frequent is the category? What rate of this outcome is expected by chance alone?

Demand adequate sample: How many observations are needed to distinguish a real effect from chance variation? For effects of typical magnitude in human performance domains, the answer is usually dozens to hundreds — not 3 to 5.

---