Confidence Intervals
On election night a pollster announces: "43% support the measure." It sounds
precise. But they only asked a thousand people, not the whole country — ask a different thousand
tomorrow and you might get 41%, or 45%. So the single number 43\% is a
little bit of a lie: it hides how shaky it is.
The honest version says a range instead: "43% \pm
3%, and we're 95% confident the true figure lies between 40% and 46%." That is a
confidence interval. It takes one fragile guess and turns it into an interval with
a stated reliability — and it is exactly how every poll, drug trial, and careful scientific
measurement reports its uncertainty. A number without an interval is a rumour; a number
with one is a measurement.
The anatomy of an interval
A point estimate
is a single dot with no sense of its own error. A confidence interval fixes that by wrapping the
dot in a margin:
\underbrace{\hat{\theta}}_{\text{estimate}} \;\pm\; \underbrace{z^{\*}\cdot \mathrm{SE}}_{\text{margin of error}}
Three pieces do all the work:
- the estimate \hat{\theta} — the centre of the
interval, your best single guess (the 43\%);
- the
standard error
\mathrm{SE} — how much the estimate wobbles from sample to sample;
- the critical value z^{\*} — how many standard
errors of hedging your chosen confidence level demands (about 2 for
95\%).
The half-width z^{\*}\cdot\mathrm{SE} is the margin of
error — the \pm 3\% you always hear quoted. Why "about two
standard errors"? The
empirical rule:
about 95\% of a bell curve sits within two standard deviations of its
centre, and the sampling distribution of the estimate is (near enough) a bell curve.
Worked example: build a 95% interval
A survey estimates support at \hat{p} = 0.43 with a standard error of
\mathrm{SE} = 0.015. Build the 95\% interval
with the quick 2-SE rule.
- Margin of error: 2 \times 0.015 = 0.03, i.e. three percentage points.
- Reach out both ways from the estimate:
0.43 - 0.03 = 0.40 and 0.43 + 0.03 = 0.46.
- State it: the 95\% confidence interval is
[0.40,\ 0.46] — "43% \pm 3%".
In plain English: the method we used produces intervals that capture the true level of support
about 95% of the time, and for this sample that interval runs from 40% to 46%.
Same data, but now we want to miss less often. The critical value climbs from
z^{\*}\approx 1.96 to z^{\*}\approx 2.58,
so the margin grows to 2.58 \times 0.015 \approx 0.039 and the interval
widens to about [0.391,\ 0.469]. More certainty costs
width: to be surer of catching the truth, you have to cast a wider net. There is no free lunch —
you buy confidence with vagueness.
Worked example: a bigger sample tightens the net
Suppose we quadruple the sample. The standard error shrinks like
1/\sqrt{n}, so four times the data halves the SE — from
0.015 down to 0.0075. Redo the
95\% interval:
- New margin: 2 \times 0.0075 = 0.015, just 1.5 points.
- New interval: 0.43 \pm 0.015 = [0.415,\ 0.445].
Same confidence level, same centre — but the interval is half as wide. This is the
whole reason big polls cost money: precision is bought with sample size, and because of the
\sqrt{n}, buying it gets expensive fast (four times the people for twice
the precision).
What the 95% really describes
Here is the subtle part. A 95\% confidence interval is built by a
procedure with this property: if you repeated the whole experiment many times —
new sample, new interval each time — then about 95\% of those
intervals would capture the true \mu. The confidence lives in
the recipe, not in any one interval it produces.
So it is tempting but wrong to say "there is a 95\% chance
\mu is in this interval". Once computed, your interval either
contains the fixed \mu or it does not — there is no probability left in
it. The 95\% is the long-run hit rate of the method.
A stack of intervals, one true μ
The vertical line is the fixed true \mu. Each horizontal bar is one
confidence interval from one sample, centred on that sample's estimate. Most bars cross the line —
they captured \mu. The odd one out (drawn in a warning colour) sits
entirely to one side: it missed. Over many intervals, the fraction that catch
\mu hovers near the confidence level.
Asking for more confidence means missing less often — and the only way to miss
less often with the same data is to make every interval wider. Confidence and
width pull in the same direction.
- A confidence interval reports a range plus a confidence level, not a single point.
- Its shape is estimate \pm (critical value) \times (standard error).
- A 95\% interval comes from a procedure that, over repeated samples, captures \mu about 95\% of the time.
- The confidence is in the method: a single computed interval either contains \mu or it does not.
- Higher confidence requires a wider interval; a bigger sample gives a narrower one.
It is desperately tempting to read "95% confidence" as "there is a 95% probability the true
value lies in this interval." That sentence is wrong, and it is wrong in a
deep way, not a nit-picking one.
The true value \mu is a fixed number, not a random one.
Your interval [0.40, 0.46] is also fixed once you've computed it. A fixed
number is either inside a fixed interval or it isn't — the probability is already 0 or 1, we just
don't know which. There is no "95% chance" left to talk about.
The correct reading puts the randomness where it belongs, in the sampling procedure:
"if we repeated this whole sampling-and-interval-building process many times, about 95% of the
intervals we'd build would contain the true value." The confidence is a property of the
method, not of your one particular interval. Say "95% of intervals like this one
capture \mu", never "this interval has a 95% chance."
There is nothing sacred about 95\%. Ronald Fisher, sorting out modern
statistics in the 1920s, casually suggested that landing within two standard deviations was a
convenient line to draw — and two standard deviations happens to fence off about
95\% of a bell curve. The convention stuck out of sheer habit, not
because the universe cares about the number 95.
What the convention really encodes is a bargain. Push the confidence all the way to
100\% and watch what happens: to be absolutely certain of
catching the true value, your interval would have to stretch from -\infty
to +\infty — guaranteed correct, and utterly useless. So every confidence
interval is a deliberate compromise: narrow enough to be informative, wide enough to
usually be right. Choosing 95% is just choosing where on that trade-off to stand.
See it explained