Base-Rate Neglect

Base-Rate Neglect

Base-rate neglect is the systematic tendency to under-weight prior probabilities (base rates) when updating beliefs from new evidence. In the Rabin (1998) parametric template
$\frac{\pi (A\mid S)}{\pi (B\mid S)}={\left(\frac{p(S\mid A)}{p(S\mid B)}\right)}^c{\left(\frac{p(A)}{p(B)}\right)}^d,$
base-rate neglect is the case $d<1$ — the prior odds enter the posterior with less than full Bayesian weight. Bayes’ rule is $c=d=1$. Base-rate neglect is the parametric expression of the representativeness heuristic: when an investor judges probabilities by how well the evidence resembles a stereotype, the base rate falls out of her calculation, which is exactly what $d<1$ encodes.

Base-rate neglect is the cognitive mirror image of conservatism ($c<1$). The two biases pull in opposite directions, and both have empirical support. Griffin & Tversky (1992) reconcile them: conservatism dominates when the signal is weak but its generating process is well understood (urn experiments), while base-rate neglect dominates when the signal feels diagnostic but the base rate is abstract (the engineer-vs-lawyer problem). The two operate in different parts of the information environment and are not mutually exclusive.

Intuition

Bayes’ rule says posterior odds equal prior odds times the likelihood ratio. If your prior is that a disease is rare ($1$ in $10,000$) and a test with $99\%$ accuracy comes back positive, the posterior probability of having the disease is still less than $1\%$ — because the rarity of the disease swamps the test’s diagnostic power. The Bayesian answer is counterintuitive precisely because it requires the prior to dominate the likelihood. Base-rate neglect is what happens when intuition fights this calculation and loses: the diagnostic-feeling test result feels more compelling than the abstract-feeling base rate, and the prior gets discounted.

The bias is robust across populations. Trained physicians, when asked the disease-test problem with realistic numbers, produce posteriors that are too high by an order of magnitude. Trained analysts, when asked which of two firms is more likely to be a “growth firm” given a description, ignore the actual fraction of growth firms in the universe. The cognitive habit of judging by resemblance — representativeness — is the underlying source.

Properties

(Parametric Form) Base-rate neglect is $d<1$ in the Rabin template.

In the Rabin (1998) deviation from Bayes’ rule, $d$ governs the weight on the prior odds. Bayes is $d=1$. Base-rate neglect is $d<1$, with empirically estimated values typically between $0.5$ and $0.9$ depending on the domain. The parameter $c$ on the signal is left at $1$, meaning the base-rate-neglecting investor still uses the new evidence correctly — she just under-weights the prior relative to it.

(Mirror Image of Conservatism) The two biases pull in opposite directions.

Conservatism ($c<1$) under-weights the new signal. Base-rate neglect ($d<1$) under-weights the prior. A pure conservative agent moves her posterior less than Bayes prescribes; a pure base-rate-neglecting agent moves it more. The two biases were initially seen as contradictory — Edwards (1968) documented conservatism, and Kahneman & Tversky (1973) documented base-rate neglect within five years. Griffin & Tversky (1992) reconciled them by noting that they apply in different settings: conservatism when the signal is weak and the process is transparent, base-rate neglect when the signal is vivid and the base rate is abstract.

(Cognitive Source) Representativeness operationalizes base-rate neglect.

Why does $d$ shrink? Because when an investor judges probabilities by resemblance — does the description fit the engineer stereotype? — the base rate is invisible to the heuristic. Resemblance is a feature of the evidence and the hypothesis, not of the population. Base-rate neglect is therefore not a separate psychological mechanism from representativeness — it is what representativeness looks like when you express it parametrically.

(Empirical Magnitude) The engineer-vs-lawyer experiment shows near-total prior neglect.

Kahneman & Tversky (1973) gave subjects a description of “Jack” and asked them to judge whether Jack was an engineer or a lawyer. One group was told the population was $70\%$ engineers; another was told $30\%$. Bayesian subjects should have produced posteriors differing by a factor of $5.4$. Real subjects produced posteriors differing by less than $0.1$ — implied $d$ values close to $0$. The bias is not subtle.

Examples

The disease-test problem

A disease has prevalence $1/10,000$ in the population. A test for it has $99\%$ sensitivity (true-positive rate) and $99\%$ specificity (true-negative rate). A randomly selected person tests positive. What is the probability she has the disease?

Bayesian calculation: $\mathrm{Pr}(\text{disease}\mid +)=\frac{0.99\cdot 0.0001}{0.99\cdot 0.0001+0.01\cdot 0.9999}\approx 0.0098$, or about $1\%$. Most respondents — including trained physicians in survey studies — answer above $80\%$. The test’s $99\%$ accuracy feels dispositive; the $1$-in-$10,000$ base rate feels abstract and gets discarded. The error is precisely $d\to 0$ in the Rabin template.

Engineer vs. lawyer (Kahneman-Tversky 1973)

Group A is told: “We selected one description at random from a pool of $70$ engineers and $30$ lawyers.” Group B is told: “We selected one description at random from a pool of $30$ engineers and $70$ lawyers.” Both groups read the same description of “Jack” — a $45$-year-old conservative man with hobbies including carpentry and mathematical puzzles. Group A reports a posterior probability of $0.55$ that Jack is an engineer. Group B reports $0.50$. The Bayesian posteriors should be roughly $0.86$ and $0.43$ — a factor of $2$ apart. Subjects’ answers differ by $0.05$. The base rate is being almost completely ignored.

See Also

• 8.6. Belief-Based Biases — parent hub.
• 8.6.5. Representativeness Heuristic — the cognitive source of base-rate neglect.
• 8.6.1. Conservatism — the parametric mirror image ($c<1$ instead of $d<1$).