Skip to content

Base Rate Integration

Why It Matters

Humans chronically ignore base rates in favor of case-specific information, even when base rates are more diagnostic. This creates systematic errors in medical diagnosis (rare disease with positive test = patient probably doesn't have disease if disease is sufficiently rare), discrimination (judging individuals by stereotype rather than individual evidence), and prediction (spectacular individual features override mundane statistical realities).

The Key Move

Before evaluating case-specific evidence, ask: "What's the base rate of the thing I'm trying to predict in the reference class?" Then update from that base rate using the evidence. Never start from 50-50 or gut feeling. Use: P(A|B) = P(B|A) × P(A) / P(B), or at minimum think: "If most things in this category are X, this particular thing is probably X unless I have strong evidence otherwise."

Classic Application

Medical diagnosis. Disease D affects 1 in 1,000 people (base rate = 0.1%). Test T is 99% accurate (sensitivity and specificity both 99%). You test positive. What's the probability you have D? Most people say 99%. Correct answer: about 9%. Why? Of 1,000 people, 1 has disease and tests positive (99% chance), and 10 don't have disease but test positive (1% false positive rate × 999 healthy people ≈ 10). So 1 true positive, 10 false positives = 1/11 ≈ 9%.

Surprising Application

Evaluating your own insight. You have what feels like a brilliant insight that contradicts expert consensus. Base rate: most contrarian insights by non-experts are wrong (99%+). Your case-specific evidence: this feels really compelling and you can't see the flaw. How much should "feels compelling" update you from the 99%+ base rate? Usually: not much. This doesn't mean never trust yourself, but it properly calibrates confidence.

Failure Modes

Using inappropriate reference classes (base rate of disease in general population when patient is pre-selected by symptoms). Ignoring when you're selected from a filtered population (your base rate for meeting someone at a conference is not the same as meeting someone randomly). Over-updating from base rates when you have strong case-specific evidence. Not updating at all because "every case is unique" (they aren't).

Go Deeper

Kahneman & Tversky, "Judgment Under Uncertainty: Heuristics and Biases"; Meehl, "Clinical Versus Statistical Prediction"