Module 3 · The Standard Error
How wrong could your average be?
Last lesson, you built a bell out of people — a thousand patients, their blood pressures spreading into a familiar curve.
But step back. You never measure everyone. You measure a sample, take the average, and report one number: "mean systolic BP = 121 mmHg."
Here's the uncomfortable question: if you'd grabbed a different sample of the same patients, you'd have got a different number.
So how far off could your 121 be? That "how far off" has its own spread — and its own name.
If you repeated the study with a different sample, would you get exactly the same average?
Two spreads that look alike
Two things are easy to blur, because both ask "how much does it vary?" — but they vary about completely different things.
Spread of people — how different patients are from each other. You already met this: the standard deviation (SD).
Spread of your answer — how different your average would be if you ran the whole study again. This one is new.
The first is about patients. The second is about your estimate. Mixing them up is the single most common mistake in applied statistics — so let's pull them apart and watch.
Repeat the study, watch the answers pile up
Each tap of "Draw a sample" pulls a fresh sample of patients, measures their average BP, and drops that one average onto the lower track. Keep tapping.
Samples drawn: 0
Two things to notice. The averages don't land on one spot — they scatter. But they scatter far less than the patients do. The wide bell on top is people. The narrow bell forming below is your possible answers.
That narrow bell has a name
The pile of possible answers is a distribution too — the sampling distribution. And like any distribution, it has a spread.
That spread is the standard error (SE): how much your estimate wobbles from one repeat of the study to the next.
Put simply — SE is the SD of your answer, not the SD of your patients.
SD — how spread out the patients are. Stays roughly the same however big your study gets.
SE — how spread out your answer is. Shrinks as your study gets bigger.
Same-shaped bell. Opposite question.
What makes your answer wobble?
Two sliders now. Drag "Patient spread (SD)" and "Sample size (n)", and watch the lower bell.
SE = 3.0 mmHg
Two forces, pulling opposite ways: messier patients (bigger SD) → wider answer-bell → less sure. More patients (bigger n) → narrower answer-bell → more sure. Your certainty isn't fixed. You buy it with sample size and spend it on variability.
The rule you just discovered
Those two levers are the whole formula:
Read it plainly: the messiness of your data on top, the amount of data underneath. More spread pushes SE up; more patients pull it down.
Nothing mysterious — you just watched both happen.
Why bigger studies cost so much
Look again at that denominator — it's √n, not n. Your study has SE = 4. You want to halve it, to SE = 2.
How many times more patients do you need?
Now do the arithmetic
Let's put numbers on it. A study of systolic BP: SD = 15 mmHg, n = 25 patients.
Step 1 — take the square root of n: √25 = 5
Step 2 — divide SD by it: SE = 15 / 5 = 3 mmHg
So your average is good to about ±3 mmHg of wobble.
Your turn. Same patients, same spread (SD = 15) — but you recruit n = 100.
√100 = ?
SE = 15 ÷ 10 = ? mmHg
You quadrupled the patients — 25 → 100 — and SE halved — 3 → 1.5. There's the square root, in plain numbers. Four times the work for twice the precision; that's the deal, every time.
Why the bell keeps showing up
One quietly remarkable thing: the answer-bell came out bell-shaped — and it would have, even if the patients hadn't been. Average enough things together and the average behaves like a normal curve — statisticians call this the central limit theorem — almost regardless of the raw data's shape.
That's why the bell is everywhere in statistics — and why last lesson's 68 / 95 / 99.7 rule is about to earn its keep: it works on the answer-bell too.
Why this matters for HTA
Here's where it lands on your desk. A manufacturer's submission never hands you a bare number — every estimate of effect, cost, or survival arrives with a "±", and that "±" is standard error wearing a disguise. Reading a dossier well means reading the SE behind every confident-looking point estimate.
- A treatment effect with a wide confidence interval? That width is SE talking — a small or noisy trial, and the true benefit could sit far below the headline figure.
- A tiny orphan-drug trial promising a large effect? The √n is working against it — the estimate is honest but unavoidably imprecise, and that imprecision is the decision risk, not a detail to wave away.
- A meta-analysis where one trial seems to dominate? Studies are weighted by precision — smaller SE, bigger say — so a single large, tight trial can quietly outvote a pile of small ones.
A point estimate tells you what was found. Its standard error tells you how hard to lean on it — and in HTA, how hard you lean is the whole decision.
Standard error, in one breath
- You measure a sample, not everyone — so your answer is one draw from many possible answers.
- Those possible answers form their own bell: the sampling distribution.
- Its spread is the standard error (SE) — how much your answer would wobble on a repeat.
- SE is not SD: SD describes patients and barely moves with study size; SE describes your answer and shrinks as the study grows.
- SE = SD / √n — and the √ means halving SE costs four times the data.
Every "±" you will ever read is a standard error — the distance between a number and how much you should trust it.
You can now read — and compute — the wobble behind any estimate. The next two lessons cash it in: wrap roughly ±2 SE around the estimate and you get a confidence interval, the honest range; measure how far two arms sit apart, counted in SEs, and you get a p-value. The whole of statistical inference is built on the number you just learned to find.