Point estimates and the standard error

Today let's work out what makes one way of estimating a number better than another, and how far you can actually trust the single number a sample hands you.

Devi runs a small bakery that is testing a new sourdough loaf. She sets out free samples by the register and tracks who comes back to actually buy a loaf. Here is what happened with today's 20 tasters: a 1 for a taster who bought, a 0 for a taster who only tried the sample.

5 of the 20 tasters bought a loaf today, a rate of 0.25. Is that number something Devi can lean on, or could a different batch of 20 tasters on a different day have handed her something else entirely?

An estimate, an estimator, and the parameter they all describe

Three words are about to do a lot of work in this lesson: estimate, estimator, and parameter. The cover's single number, 0.25, is a good place to pin all three down before going any further.

The parameter is the thing you actually want to know: Devi's true buy rate, the share of tasters who would buy a loaf if she could offer the sample to every possible taster she will ever see. For a real bakery, nobody ever gets to see that number directly. For this lesson, since it is a simulation Devi ran on purpose, the parameter is fixed at 0.30.

Hang onto that number: the whole lesson is about to measure how close different rules land to it, and how much each one wobbles while doing so.

The estimator is the rule itself: count how many of the sampled tasters bought a loaf, then divide by how many tasters were sampled. That rule, buyers divided by sample size, is written x over n and usually called phat, said "p hat". It works on any sample Devi could draw, not just today's 20.

The estimate is what that rule gives you once you plug in real numbers. Devi's rule on today's 20 tasters gives 5 over 20, which is 0.25. That is the estimate: one number, tied to one particular sample, standing in for a parameter you cannot see directly.