Lesson 19 · Yesterday's posterior is today's prior (order of evidence)

Same Data, Either Order

First, a quick retrieval check — different concept, new setting.

Two servers, in different data centers, each independently have a 3% chance of failing on any given day (Lesson 4’s independence: learning about one tells you nothing about the other). What’s the probability at least one of the two fails today? (Hint: it’s easier to find P(neither fails) first.)

(Worked out in the solution.)


Lesson 15 gave you the beta-binomial update rule: start with Beta(α, β), observe k successes out of n trials, and the posterior is Beta(α+k, β+(n−k)) — addition of counts. Every worked example so far updated on one clean batch of data at a time. Real evidence rarely arrives that tidily — it trickles in across days, sources, and reports.

Setup. You start with prior Beta(2, 3) (a mild pull toward lower rates — 2 virtual successes, 3 virtual failures). Two batches of real evidence exist, and you don’t get to choose which one shows up first:

  • Batch A: 5 successes, 3 failures (8 trials)
  • Batch B: 2 successes, 4 failures (6 trials)

Part 1 — Process A, then B. Starting from Beta(2,3), update on Batch A to get an intermediate posterior. Treat that as your new prior, and update on Batch B. Write down the final Beta(α, β) and its mean.

Part 2 — Process B, then A. Start over from Beta(2,3) again, but update on Batch B first, then treat that result as your prior and update on Batch A. Write down this final Beta(α, β) and its mean.

Part 3 — Compare (the numeric answer above). Are the two final posteriors identical? Give the shared posterior mean as a decimal. In one sentence, say why — what property of “just add counts” guarantees the order genuinely can’t matter here?

Part 4 — Break it. Order-independence quietly assumes something: that both batches are evidence about the same, fixed underlying rate p. Describe a realistic scenario where Batch A and Batch B come from a system whose true rate actually changed between the two batches — and explain why “just add up all the counts, order doesn’t matter” would then be the wrong thing to do, even though the arithmetic would run exactly the same way and produce a number that looks equally confident.

Final posterior mean of p, as a decimal, regardless of which order the two batches are processed in?