Understanding Uncertainty · Mini-Lesson · pairs with Class 04 & Class 11

How Bad Is My Estimator?

The standard error of the mean: one number for how far a sample mean typically lands from the truth

Understanding Uncertainty · in-class group work

A standalone mini-lesson in a small, growing set on estimator properties. The running example is still the fair die, \(\theta = 3.5\), \(\sigma \approx 1.708\) — the same die from unbiased, but never consistent, now asking a sharper question: not just "does \(\bar X_n\) tighten as n grows," but "by how much, exactly, and how far off should I expect to be right now?"

A companion exploration built to accompany the sample-mean material in Class 04 and the sampling-distribution material in Class 11 — no external dataset, the widget generates its own die rolls in the browser.

New vocabulary is underlined like this — hover or tap any term for a one-line explanation that appears right below the line you are reading.

Think · Pair · Share · is 4.1 far from 3.5?

Think · 2 minYou roll a die n times and get \(\bar X_n = 4.1\), when the true θ is 3.5. Is that a bad estimate? Would your answer change between n = 5 and n = 500 — and what single number would you need to know to answer precisely, instead of by gut feel?

I've done the Think step — reveal Pair & Share

Pair · 3 minCompare what each of you would need to know before calling 0.6 "big" or "small."
Share · 2 minAgree as a table on a rule of thumb, then check it against the widget below.

The standard error of the mean answers exactly this question. It's the standard deviation of \(\bar X_n\) itself, across hypothetical repeated samples:

$$SE(\bar X_n) = \frac{\sigma}{\sqrt n}$$

Not a bound and not a guarantee — a typical scale. Being off from θ by about one SE is ordinary; being off by three or four SEs would be a surprise. Larger n shrinks SE like \(1/\sqrt n\), which is the precise version of "more data means a better estimate."

Two distributions are stacked below — keep track of which is which. The die has one, fixed distribution that never changes. Every dot in the swarm is not a die roll — it's an entire sample's worth of rolls collapsed into one number, \(\bar X_n\). The spread of those collapsed numbers is the sampling distribution of the estimator, one level removed from the die, and SE is exactly that spread's standard deviation.

The highlighted point above the axis is one specific sample's estimate, with its own ±1 SE error bar. The swarm below shows 250 more hypothetical replications against shaded ±1 SE / ±2 SE bands around the true θ.

SE turns "more data helps" into a number. Drag n up and the shaded bands visibly narrow, and the live percentages track how many of the 250 replications actually land inside them. The bands are doing real work: they are not asserting a fixed rule, they're showing what "typical" means for this population at this n, and you can watch that change as n changes.

Verify by hand

Compute SE at n = 5 using \(\sigma \approx 1.708\): \(SE = 1.708/\sqrt5\). Then find the n at which SE first drops below 0.2.

Show the arithmetic

At n = 5: \(SE = 1.708/2.236 \approx 0.764\) — check it against the widget's readout. For \(SE < 0.2\): \(1.708/\sqrt n < 0.2 \Rightarrow \sqrt n > 8.54 \Rightarrow n > 72.9\), so n = 73 is the first integer sample size where the typical error drops below 0.2 dots.

FAQ

The widget uses the true σ to compute SE. In practice we don't know σ — now what?

You plug in an estimate of σ instead — the sample standard deviation s, computed with Bessel's correction so it isn't itself biased low. \(SE \approx s/\sqrt n\) is the version you'd actually compute from real data, and it's what most software reports by default.

Does "within 2 SE" mean a 95% guarantee?

No — that's a rule of thumb from the normal approximation (the Central Limit Theorem), and it's only accurate for large enough n. At small n, especially for a bounded, non-normal population like a die, the actual percentage can differ noticeably from 95% — which is exactly what the widget's live percentages let you check, rather than assume.

More estimator-properties mini-lessons: Unbiased, but never consistent What the sample mean minimizes Why the sample variance undershoots One sample, one point One sample, one point (skewed) One sample, one point (variance)