Lesson 20 · Solution · Conjugacy: beta-binomial and normal-normal

Solution: One Update, Two Distributions

Retrieval check answer. P(alarm) = 0.98×0.01 + 0.05×0.99 = 0.0098 + 0.0495 = 0.0593. P(defective | alarmed) = 0.0098 / 0.0593 ≈ 0.165 (about 16.5%) — most alarms are false alarms, the same base-rate story as Lesson 7, on a factory line instead of a clinic.


Part 1 — Virtual sample size: n₀ = 400/100 = 4. This prior is worth only 4 virtual measurements — weak, compared to 16 real ones (4× as much real evidence as prior evidence). Expect the posterior to land much closer to the data’s 180 ms than to the prior’s 200 ms.

Part 2 — Posterior mean: 184 ms.

prior precision = 1/100 = 0.01
data precision   = 16/400 = 0.04
posterior precision = 0.01 + 0.04 = 0.05

posterior mean = (200 × 0.01 + 180 × 0.04) / 0.05
               = (2 + 7.2) / 0.05
               = 9.2 / 0.05
               = 184

184 ms is 4× closer to the data (180) than to the prior (200) — distance 4 from the data, distance 16 from the prior, a 4:1 split, exactly matching the 4:1 ratio of data precision to prior precision (0.04 : 0.01). This is the precision-weighted average doing exactly what its name says: whichever side carries more precision (equivalently, more “weight” in the beta-binomial sense) pulls the posterior more of the way toward itself.

Part 3 — With n = 100: posterior mean ≈ 180.77 ms.

data precision = 100/400 = 0.25
posterior precision = 0.01 + 0.25 = 0.26
posterior mean = (200×0.01 + 180×0.25) / 0.26 = (2 + 45)/0.26 = 47/0.26 ≈ 180.77

With 100 real measurements, data precision (0.25) dwarfs the prior’s fixed 0.01 — the posterior lands almost exactly on the data’s 180, barely nudged by the prior at all. This is washing out the prior (Lesson 14) in its precision-weighted form: any finite prior’s pull is a fixed, bounded quantity, while data precision grows without bound as n grows, so enough real data eventually swamps any reasonable prior — the same phenomenon, in a completely different distribution family.

Why this is conjugacy, concretely. Neither computation touched an integral. You added two precisions and took a weighted average of two means, and the algebra of normal distributions guarantees the result is itself normal. That’s what “conjugate” buys you: a prior family paired with a matching likelihood family such that the posterior stays in the same family, collapsing the update to arithmetic on that family’s own parameters. Beta-binomial adds counts (α, β); normal-normal adds precisions and takes a weighted mean. Different parameters, same underlying shape: combine what you believed with what you saw, weighted by how much each side is worth trusting.

Where this goes: conjugate pairs like these two are the convenient exception, not the rule — most real priors and likelihoods you’d actually want to combine don’t have a tidy closed form at all. Next lesson previews exactly what breaks conjugacy, and what you do once the closed form runs out.

How was this one? Any answer marks it complete and moves on — your rating shapes future lessons.