Estimation, confidence intervals and tests

S3 Topic 3, spec point by spec point (Estimation, confidence intervals and tests). Each spec point has the method, worked examples written the way the mark scheme wants them, how the marks go and the mistakes that cost the most marks.

Spec3.1Estimators, bias and standard error

Concepts of standard error, estimator and bias. The sample mean \(\bar{x}\) and the sample variance \(s^2 = \frac{1}{n - 1}\sum(x_i - \bar{x})^2\) as unbiased estimates of the corresponding population parameters.

The vocabulary

Unbiased estimates of mean and variance

\[\bar{x} = \frac{\sum x}{n} \qquad s^2 = \frac{\sum x^2 - n\bar{x}^2}{n - 1}\]

Why divide by \(n - 1\): deviations are measured from \(\bar{x}\), which is closer to the data than \(\mu\) is, so dividing by \(n\) would underestimate \(\sigma^2\) on average.

Showing bias, and choosing between estimators

  1. Find \(E(T)\) using \(E(aX + b) = aE(X) + b\) and the known \(E(X)\).
  2. Compare with the parameter. If not equal, it's biased; the difference is the bias.
  3. Of two unbiased estimators, the better one has the smaller variance.

bias

\(X \sim U[\alpha - 3, 2\alpha + 3]\). Show that \(\bar{X}\) is a biased estimator of \(\alpha\), state the bias and find an unbiased estimator.

\(E(\bar{X}) = E(X) = \frac{(\alpha - 3) + (2\alpha + 3)}{2} = \frac{3\alpha}{2} \neq \alpha\), so biased. Bias \(= \frac{3\alpha}{2} - \alpha = \frac{\alpha}{2}\).

\(E\left(\frac23\bar{X}\right) = \alpha\), so \(\frac23\bar{X}\) is unbiased.

unbiased variance estimate

\(n = 10\), \(\sum x = 52\), \(\sum x^2 = 300\). \(\bar{x} = 5.2\), \(s^2 = \frac{300 - 10(5.2)^2}{9} = \frac{29.6}{9} = 3.29\).

the better estimator

\(Y = aX_1 + (1 - a)X_2\) for independent \(X_1, X_2\) with mean \(\mu\) and variance \(\sigma^2\). \(E(Y) = a\mu + (1 - a)\mu = \mu\), so unbiased for any \(a\).

\(\text{Var}(Y) = a^2\sigma^2 + (1 - a)^2\sigma^2 = (2a^2 - 2a + 1)\sigma^2\), which is least when \(a = \frac12\). So the ordinary mean is the best of these.

  • M1 for using \(E(X) = \frac{a + b}{2}\), A1 for \(E(\bar{X}) = \frac{3\alpha}{2} \neq \alpha\) with a conclusion, B1 for the bias \(\frac{\alpha}{2}\) (WST03A Jan 2026 Q2a).
  • Making an estimator unbiased: choose \(c\) and \(d\) in \(c\bar{X} + d\) so its expectation equals the parameter (WST03 Jan 2025 Q6c).
  • Variance estimates: M1 for \(\frac{1}{n - 1}(\sum x^2 - n\bar{x}^2)\); dividing by \(n\) loses the marks (WST03 Oct 2020 Q5a).

Often asked as: very common (42 questions). Show an estimator is (un)biased; find the bias; compare estimators by their variances; compute \(\bar{x}\) and \(s^2\).

Saying an estimator is "better" because it's unbiased, without comparing variances when both are unbiased.

Spec3.2The distribution of the sample mean

\(\bar{X}\) has mean \(\mu\) and variance \(\frac{\sigma^2}{n}\). If \(X \sim N(\mu, \sigma^2)\) then \(\bar{X} \sim N\left(\mu, \frac{\sigma^2}{n}\right)\). No proofs required.

The idea

Averages vary less than single values: the mean of \(n\) observations has the same centre but its variance is divided by \(n\). (This follows from 1.1: \(\bar{X} = \frac1n(X_1 + \ldots + X_n)\), variance \(\frac{1}{n^2} \times n\sigma^2\).)

\[\bar{X} \sim N\left(\mu, \frac{\sigma^2}{n}\right) \qquad Z = \frac{\bar{X} - \mu}{\sigma/\sqrt{n}}\]

Finding \(n\)

Standardise with \(\frac{\sigma}{\sqrt{n}}\), set equal to the \(z\) value from the table, and solve for \(\sqrt{n}\). Round up if the condition is an inequality.

finding n from a probability

Bags weigh \(T \sim N(5, 0.3^2)\) kg. For \(n\) bags, \(P(\bar{T} < 5.07) = 0.8829\). Find \(n\).

\(\frac{0.07}{0.3/\sqrt{n}} = 1.19\), so \(\sqrt{n} = \frac{1.19 \times 0.3}{0.07} = 5.1\), \(n = 26\).

the smallest n

\(\bar{X} \sim N\left(1.96, \frac{0.003}{n}\right)\). Find the least \(n\) with \(P(\bar{X} > 2) < 0.01\).

Need \(\frac{0.04}{\sqrt{0.003/n}} > 2.3263\), so \(\sqrt{n} > 3.186\), \(n > 10.15\). Least \(n = 11\).

  • M1 for standardising with \(\frac{0.3}{\sqrt{n}}\) and equating to a \(z\) value, B1 \(z = 1.19\), dM1 rearrange, A1 \(n = 26\) (WST03A Jun 2026 Q4d).
  • B1 for the standard error, M1 standardise, A1 (WST03 Jun 2025 Q7c).
  • B1 for \(N\left(1.96, \frac{0.003}{n}\right)\), B1 \(z = 2.3263\), A1 \(n = 11\) (WST03 Oct 2021 Q7b).

Often asked as: probabilities about a sample mean; find \(n\).

Dividing the standard deviation by \(n\) instead of \(\sqrt{n}\).

Spec3.3Confidence intervals: the concept

Concept of a confidence interval and its interpretation, and its link with hypothesis tests.

What a 95% confidence interval means

If you took many samples and built a 95% interval from each, about 95% of the intervals would contain the true \(\mu\). The interval is random; \(\mu\) is fixed.

So for a given interval, don't say "there's a 95% probability that \(\mu\) is in it". Say "95% of intervals constructed this way contain \(\mu\)".

Counting intervals

Each 95% interval independently contains \(\mu\) with probability 0.95. So the number of intervals out of \(k\) that miss \(\mu\) is \(B(k, 0.05)\).

Link with hypothesis tests

A value \(\mu_0\) lies outside a 95% confidence interval exactly when a two-tailed 5% test would reject \(H_0: \mu = \mu_0\). So you can comment on a claim by checking whether it lies in the interval.

many intervals

20 independent 95% intervals are formed. Find the probability that more than 3 do not contain \(\mu\).

\(X \sim B(20, 0.05)\). \(P(X > 3) = 1 - P(X \leqslant 3) = 0.0159\).

two intervals

Two independent 99% intervals. \(P(\text{exactly one contains } \lambda) = 2(0.99)(0.01) = 0.0198\).

  • M1 for \(B(20, 0.05)\) or \(B(20, 0.95)\), A1 for the right probability statement, A1 (WST03A Jan 2026 Q4b).
  • Comment on a claim: B1ft the value lies inside the interval, dB1ft so there's no reason to doubt the claim (WST03 Jan 2023 Q6).
  • M1 for \(2p(1 - p)\), A1 (WST03 Oct 2020 Q6c).

Often asked as: interpretation; binomial counts of intervals; comment on a claim using an interval.

"There is a 95% chance \(\mu\) lies in (229.4, 231.6)."

Spec3.4Confidence intervals for a Normal mean (variance known)

Confidence limits for a Normal mean, with variance known.

The formula

\[\bar{x} \pm z \times \frac{\sigma}{\sqrt{n}}\]

Confidence \(z\)
90% 1.6449
95% 1.96
98% 2.3263
99% 2.5758

The interval is centred on \(\bar{x}\). Its width is \(2z\frac{\sigma}{\sqrt{n}}\): more confidence → wider; bigger sample → narrower.

Working backwards

an interval

5 mugs have mean mass 230.5 g; \(\sigma = 1.2\) g. Find a 95% confidence interval for \(\mu\).

\(230.5 \pm 1.96 \times \frac{1.2}{\sqrt5} = 230.5 \pm 1.05\), giving \((229.4, 231.6)\).

from one interval to another

A 95% interval is \((11.52, 13.75)\). Find the 90% interval from the same sample.

\(\bar{x} = 12.635\). Half-width \(1.115 = 1.96 \times \text{SE}\), so \(\text{SE} = 0.5689\).

90%: \(12.635 \pm 1.6449 \times 0.5689 = 12.635 \pm 0.936\), giving \((11.7, 13.6)\).

finding the confidence level

An \(r\)% interval has width 0.04 and the standard error is 0.0124. \(2z(0.0124) = 0.04 \Rightarrow z = 1.61\). \(P(Z > 1.61) = 0.0537\), so \(r = 100(1 - 2 \times 0.0537) = 89.3\).

  • M1 for \(\bar{x} \pm z\frac{\sigma}{\sqrt{n}}\), B1 \(z = 1.96\), A1 one end, A1 both ends (WST03A Jan 2026 Q4a).
  • B1 for the midpoint, M1 for the standard error from the half-width, B1 \(z = 1.6449\) (1.64 or 1.65 is B0), A1 (WST03 Jun 2021 Q3a).
  • Confidence level from a width: M1 for setting \(2z \times \text{SE}\) equal to the width, A1 \(z\), M1 M1 to get \(r\), A1 awrt 89.3 (WST03 Jan 2025 Q4c).

Often asked as: calculate an interval; widths; minimum sample size for a given width; convert between levels.

Using \(z = 1.645\) for 95%, or \(z = 1.96\) for a one-sided 5% condition.

Forgetting the \(\sqrt{n}\).

Spec3.5Hypothesis tests for a Normal mean (variance known)

Hypothesis tests for the mean of a Normal distribution with variance known, using \(\frac{\bar{X} - \mu}{\sigma/\sqrt{n}} \sim N(0, 1)\).

Method

  1. \(H_0: \mu = \mu_0\); \(H_1: \mu < \mu_0\), \(\mu > \mu_0\) or \(\mu \neq \mu_0\).
  2. Test statistic: \(z = \dfrac{\bar{x} - \mu_0}{\sigma/\sqrt{n}}\).
  3. Compare with the critical value (1.6449 for one tail at 5%, \(\pm1.96\) for two tails at 5%, and so on), or find the \(p\)-value \(P(Z < z)\) and compare with the level.
  4. Conclude statistically, then in context.

lower tail

A manager claims screws have mean length 30 mm; it's suspected they're shorter. A sample of 80 has \(\bar{x} = 29.5\); \(\sigma = 2.5\). Test at 5%.

\(H_0: \mu = 30\), \(H_1: \mu < 30\).

\(z = \frac{29.5 - 30}{2.5/\sqrt{80}} = -1.79\). Critical value \(-1.6449\) (or \(p\)-value 0.0367 < 0.05).

Reject \(H_0\): there is evidence that the mean length of the screws is less than 30 mm.

not significant

\(H_0: \mu = 280\), \(H_1: \mu > 280\), \(\sigma = 70\), \(n = 125\), \(\bar{x} = 290\). \(z = \frac{10}{70/\sqrt{125}} = 1.60 < 1.6449\). Not significant: insufficient evidence that the mean yield is higher than 280 g.

  • B1 for hypotheses in terms of \(\mu\), M1 for the test statistic, A1 awrt the \(z\) value, B1 for the critical value (or \(p\)-value), A1 for a contextual conclusion (WST03 Oct 2021 Q1).
  • M1 for a non-contradictory statement, A1 cso for the contextual conclusion (WST03 Jun 2019 Q6a).
  • Reverse questions: find the largest \(\mu\) for which a result is not significant, by solving \(\frac{\bar{x} - \mu}{\sigma/\sqrt{n}} > -1.6449\) (WST03 Jan 2021 Q4c).

Often asked as: a full test, often after a confidence interval on the same data.

Hypotheses in terms of \(\bar{x}\).

A conclusion that "proves" the claim.

Spec3.6The Central Limit Theorem

Use of the Central Limit Theorem to extend hypothesis tests and confidence intervals to samples from non-Normal distributions. Use of large sample results when the variance is unknown: \(\frac{\bar{X} - \mu}{S/\sqrt{n}}\) can be treated as \(N(0, 1)\) when \(n\) is large.

The theorem

For a large sample (\(n \geqslant 30\) as a guide) from any distribution with mean \(\mu\) and variance \(\sigma^2\):

\[\bar{X} \approx N\left(\mu, \frac{\sigma^2}{n}\right)\]

So the methods in 3.4 and 3.5 work for non-Normal populations too. If \(\sigma\) is unknown, use \(s\) from the sample (large samples make \(s\) a good estimate). The \(t\)-distribution is not needed.

Using it

  1. Find the population mean and variance (e.g. from a uniform, Poisson or binomial model).
  2. Divide the variance by \(n\).
  3. State that \(\bar{X}\) is approximately Normal by the CLT.

a uniform population

Waiting times are \(U[0, 7]\) minutes. Estimate the probability that the mean of 46 waiting times is between 3.4 and 3.6.

\(\mu = 3.5\), \(\sigma^2 = \frac{49}{12}\). By the CLT, \(\bar{T} \approx N\left(3.5, \frac{49}{12 \times 46}\right)\), standard error 0.298.

\(z = \pm\frac{0.1}{0.298} = \pm0.34\). \(P = 2(0.6331 - 0.5) = 0.266\).

a test with s

A claim says the mean customer age is 30. A sample of 75 has \(\bar{x} = 28.2\), \(s = 8.5\). Test at 10%.

\(H_0: \mu = 30\), \(H_1: \mu \neq 30\). \(z = \frac{28.2 - 30}{8.5/\sqrt{75}} = -1.83\). Critical values \(\pm1.6449\).

Reject \(H_0\): there is evidence that the mean customer age is not 30.

  • M1 for the mean, M1 for the variance divided by \(n\), A1ft for the Normal (WST03A Jun 2026 Q7a).
  • M1 for the CLT distribution, M1 standardise, A1, M1 for \(p - (1 - p)\), A1 (WST03 Jun 2022 Q4c).
  • B1 hypotheses, M1 A1 for \(z\), B1 the critical value, A1 contextual conclusion (WST03 Jun 2017 Q3a).

Often asked as: very common (19 questions). Approximate distributions of \(\bar{X}\); tests and intervals from large samples; "explain why the CLT was needed" (the population isn't Normal).

Saying the CLT makes the population Normal; it's the sample mean that's approximately Normal.

Spec3.7Testing the difference between two means (variances known)

Hypothesis test for the difference between the means of two Normal distributions with variances known, using \(\frac{(\bar{X} - \bar{Y}) - (\mu_x - \mu_y)}{\sqrt{\sigma_x^2/n_x + \sigma_y^2/n_y}} \sim N(0, 1)\).

Why the formula looks like that

From 1.1 and 3.2: \(\bar{X} - \bar{Y}\) is Normal with mean \(\mu_x - \mu_y\) and variance \(\frac{\sigma_x^2}{n_x} + \frac{\sigma_y^2}{n_y}\) (variances add).

Method

  1. \(H_0: \mu_x = \mu_y\) (or \(\mu_x - \mu_y = d\)); \(H_1\) one- or two-tailed.
  2. Test statistic:

\[z = \frac{(\bar{x} - \bar{y}) - (\mu_x - \mu_y)}{\sqrt{\frac{\sigma_x^2}{n_x} + \frac{\sigma_y^2}{n_y}}}\]

  1. Compare with the critical value; conclude in context.

two towns

Town A: \(n = 25\), \(\bar{x} = 52.4\). Town B: \(n = 30\), \(\bar{y} = 57.8\). Both have \(\sigma = 18\). Test at 5% whether town B's mean is greater.

\(H_0: \mu_A = \mu_B\), \(H_1: \mu_B > \mu_A\).

\(z = \frac{57.8 - 52.4}{18\sqrt{\frac{1}{25} + \frac{1}{30}}} = 1.11 < 1.6449\).

Not significant: insufficient evidence that the mean score in town B is greater than in town A.

  • B1 hypotheses, M1 correct denominator, dM1 correct statistic, A1 awrt 1.11, B1 critical value (or \(p\)-value 0.134), A1 conclusion (WST03 Jan 2021 Q4b).
  • B1 for \(H_1: \mu_A < \mu_B\), M1 for \(\sqrt{\frac{0.07}{50} + \frac{0.05}{40}}\), A1 for \(z = -5.83\), B1 1.6449, M1 decision, A1 context (WST03 Jan 2025 Q3b).
  • Combining two tests: interpret the two conclusions together in context (WST03A Jan 2026 Q6).

Often asked as: full tests, sometimes with a non-zero difference ("at least 2 kg heavier").

Using \(\frac{\sigma_x}{n_x}\) inside the square root instead of \(\frac{\sigma_x^2}{n_x}\).

Spec3.8Large samples with unknown population variances

Use of large sample results to extend to the case in which the population variances are unknown, using \(\frac{(\bar{X} - \bar{Y}) - (\mu_x - \mu_y)}{\sqrt{S_x^2/n_x + S_y^2/n_y}} \sim N(0, 1)\).

The idea

With large samples, replace each \(\sigma^2\) by its unbiased estimate \(s^2\) and use the same test as 3.7. The CLT also means the populations need not be Normal.

Assumptions (often asked)

two regions

Region A: \(n = 80\), \(\bar{x} = 24.3\), \(s^2 = 16.2\). Region B: \(n = 100\), \(\bar{y} = 22.9\), \(s^2 = 20.5\). Test at 5% whether the mean in A is greater.

\(H_0: \mu_A = \mu_B\), \(H_1: \mu_A > \mu_B\).

\(z = \frac{1.4}{\sqrt{\frac{16.2}{80} + \frac{20.5}{100}}} = \frac{1.4}{0.638} = 2.19 > 1.6449\).

Reject \(H_0\): there is evidence that the mean in region A is greater than in region B.

  • B1 for the assumption that the sample variances can be used as the population variances (WST03A Jun 2026 Q2c, WST03 Jun 2023 Q6d).
  • Two assumptions: independence of the samples, and \(s^2 = \sigma^2\) because the samples are large (WST03 Oct 2021 Q6d).

Often asked as: a test from summary statistics, followed by an assumption.

Stating "the populations are Normal" as the assumption: with large samples the CLT covers that.