Bayesian Thinking for the Rest of Us

Bayes’ theorem is a piece of mathematics that describes how to update a probability estimate in light of new evidence. The formula looks intimidating but the intuition is simple: how much you should change your belief depends on (i) how strong the new evidence is and (ii) how confident you were before the evidence arrived.

The basic idea

Suppose you believe X is 30% likely. You then encounter evidence that bears on whether X is true. How should the evidence change your estimate?

Bayes’ theorem says: it depends on how likely the evidence is under X-being-true versus X-being-false. If the evidence is much more likely under X-being-true (say, the evidence happens 90% of the time when X is true but only 10% of the time when X is false), then encountering the evidence should substantially increase your estimate. If the evidence is similarly likely either way (60% under X-true and 50% under X-false), then encountering it should barely change your estimate.

The formula formalizes this intuition. The intuition itself does not require the formula.

A useful example

You see a friend acting strangely. Your initial estimate that something is wrong is, say, 20% (most days, nothing is wrong; today, something might be).

You then learn that they didn’t respond to a message from their partner. Under “something is wrong” (say, a relationship conflict), this is fairly likely — maybe 70% of the time. Under “nothing is wrong,” this is less likely but not impossible — maybe 30% of the time (they could just be busy).

The Bayesian update increases your estimate from 20% to roughly 37%. The new evidence raises your estimate, but not to certainty. The strange behavior is more consistent with “something is wrong” than with “nothing is wrong,” but it’s not so much more consistent that you should jump to a definite conclusion.

Why it’s useful

Most people, in practice, do not update their beliefs proportionally to evidence. They either update too much (a single piece of evidence flips them to certainty) or too little (no amount of evidence shifts them off a prior view).

Bayesian thinking is the discipline of updating proportionally. The size of the update should match the size of the evidence. Strong evidence should produce a substantial shift. Weak evidence should produce a small shift. Equally consistent evidence should produce no shift.

This is not how the intuitive system works. The intuitive system is often binary — convinced or unconvinced — rather than graded. Bayesian thinking introduces the graded view that the intuitive system lacks.

Practical applications

Bayesian thinking has applications well beyond formal mathematics.

Medical decisions. A positive test for a rare condition should usually be followed by more tests, because the base rate of the condition is low. Bayesian updating tells you that a single positive test, by itself, doesn’t establish the diagnosis.

Job interviews. A strong impression in a single interview is weaker evidence than it feels. Many candidates make strong impressions and turn out to be poor fits. Bayesian updating tells you to weight the interview against base rates and other information.

Reading the news. A single sensational headline is weaker evidence than it feels. Most sensational headlines turn out to be more nuanced on close examination. Bayesian updating tells you to update from headlines with appropriate skepticism.

How to practice

Bayesian thinking is built through small repeated practice. When you encounter new information, ask: how much should this shift my belief? More than this morning’s news on the same topic? Less than the consensus of three independent experts?

Quantifying the answer (“this should shift me from 40% to 55%”) is not always necessary. Often, the directional question (“this should shift me a little, not a lot”) is enough. The discipline is in noticing that you should shift at all, and shifting by an amount appropriate to the evidence rather than to your emotional reaction.

The practitioner who internalizes this discipline ends up with beliefs that track reality better than most. The beliefs are usually less certain than the intuitive system would produce, and that uncertainty is itself a feature: it allows for further updating as more evidence arrives.