Statistics › Section I · Hypothesis testing with one sample › Topic I.3
Picking the distribution for a test
Once the hypotheses are written, a test needs the right curve to judge the data against: normal or t for a mean, normal for a proportion. What is known about the population decides which.
The test's measuring stick
A test statistic measures how far the data fall from the null's prediction - but a raw distance means nothing without a scale. The reference distribution supplies the scale: the curve on which "this far" becomes "this unlikely". Choosing it is the first technical decision of any test, and it follows one question: what is known, and what is being tested?
The test statistic: distance in standard errors
Every test statistic has the same shape: the observed statistic minus what the null predicts, divided by the statistic's standard error. For a mean with known \(\sigma\):
The result counts standard errors between the data and the null's world: \(z = 2\) says the sample mean sits two standard errors above what \(H_0\) expected. Larger magnitudes are stronger evidence - and the p-value topic will price them.
Known sigma: the standard normal
When the population spread is known - from engineering specifications or long calibration records - the test statistic is a z-score, and the standard normal supplies the tail areas. This is the same \(N(0,1)\) machinery the confidence chapter built on, with \(\mu_0\) playing the centre's role. The conditions are the interval chapter's: a random sample, independence, and either a normal population or \(n \ge 30\) for the CLT to guarantee the bell for \(\bar{x}\).
Unknown sigma: the t distribution
When \(\sigma\) must be estimated by \(s\) - the ordinary case - the statistic becomes a t-score: \(t = \frac{\bar{x} - \mu_0}{s/\sqrt{n}}\), with \(n - 1\) degrees of freedom, exactly as in the t-interval topic. The substitution adds uncertainty, the \(t\)'s heavier tails price it, and small samples need the same symmetric-data checks. The choice between z and t here mirrors the interval chapter's exactly - same conditions, same story, opposite reading: the interval asked "where might \(\mu\) be"; the test asks "is \(\mu_0\) credible".
Proportions: normal with the null's p
Testing a proportion uses a z-score built from \(\hat{p}\):
Notice what differs from the confidence interval: the standard error uses \(p_0\), the null's value - because a test computes probabilities in the null's world, where \(p = p_0\) exactly. The CI, which assumes no value for \(p\), used \(\hat{p}\) instead. That one-letter difference between the two formulas is easy to miss and worth checking every time. Conditions: random sample, the 10% rule, and \(np_0\) and \(n(1-p_0)\) both at least 10 - counts under the null. (Some textbooks accept 5; this course keeps the safer 10 used throughout.)
The recipe, one line
Statistic minus null value, over the standard error, on the right curve:
Every one-sample test in this chapter, and the two-sample tests that follow, is that one line with different ingredients (chi-square tests, later, measure distance another way). The sampling-distribution chapter said which curves exist; this topic says which curve a given test lives on: z for known-\(\sigma\) means, t for estimated-\(\sigma\) means, z for large-count proportions. The distribution is chosen before the p-value is read, never after.
Conditions travel with the curve
Choosing the distribution is also inheriting its conditions. A random sample: every curve assumes it. Independence: the 10% rule where sampling is without replacement. Shape: \(n \ge 30\) or symmetric data for means, and the 10-and-10 counts for proportions - computed with \(p_0\) in a test, since the test lives in the null's world. A test that skips its conditions produces a p-value-shaped number with none of a p-value's meaning, which is worse than no number at all: it borrows the machinery's credibility for a conclusion the machinery never certified.
Worked example: the cereal statistic
A consumer group tests \(H_0: \mu = 500\) g against \(H_a: \mu < 500\) with a sample of \(n = 36\) boxes averaging \(\bar{x} = 496\) g; \(\sigma = 12\) g is known from the filling line's specifications. 1. Known \(\sigma\) and \(n \ge 30\): the standard normal. 2. \(z = \frac{496 - 500}{12/\sqrt{36}} = \frac{-4}{2} = -2\): the sample mean sits two standard errors below the claim. 3. The tail area beyond -2 - the p-value - is the next topic's work; here the choice is settled: z on \(N(0,1)\), conditions met.
Worked example: the poll's statistic
A senator claims 50% support: \(H_0: p = 0.5\). A random poll of 400 voters finds \(\hat{p} = 0.55\). 1. Conditions: \(np_0 = 200\) and \(n(1-p_0) = 200\) both clear 10 easily. 2. The null's standard error: \(\sqrt{0.5 \times 0.5/400} = 0.025\) - built from \(p_0\), not \(\hat{p}\). 3. \(z = \frac{0.55 - 0.5}{0.025} = 2\): the poll sits two null-world standard errors above the claim. 4. Check: the CI's \(\hat{p}\)-version would give \(\sqrt{0.55 \times 0.45/400} \approx 0.0249\) - nearly identical here, but the principle stands: tests assume the null.
Common mistakes to avoid
First, pairing \(s\) with the standard normal: unknown \(\sigma\) means \(t\), always. Second, using \(\hat{p}\) in a proportion test's standard error: the test lives in the null's world, so \(p_0\) sets the spread. Third, the wrong degrees of freedom: \(n - 1\), not \(n\). Fourth, checking the success-failure counts against \(\hat{p}\) instead of \(p_0\). Fifth, choosing the curve after seeing which gives the smaller p-value - the choice belongs to the design, not the outcome. Sixth, computing a statistic before checking the conditions that license it.
Key takeaways
The reference distribution is the test's measuring stick, chosen by what is known. Known \(\sigma\): \(z = (\bar{x} - \mu_0)/(\sigma/\sqrt{n})\) on the standard normal. Unknown \(\sigma\): \(t = (\bar{x} - \mu_0)/(s/\sqrt{n})\) with \(n - 1\) degrees of freedom. A proportion: \(z = (\hat{p} - p_0)/\sqrt{p_0(1-p_0)/n}\) - the SE built from \(p_0\), the null's value, exactly where tests differ from confidence intervals. Conditions travel with the curve: random, independent, shape or counts checks. Statistic minus null over SE, on the right curve - then the p-value topic prices the distance.
Flashcards
Flashcard
The curve that prices the test statistic
Test your understanding
Question 1 of 5
A mean test has σ known from the process specs. Which reference distribution?
Question 2 of 5
A mean test with n = 16 and σ unknown uses which distribution?
Question 3 of 5
In a proportion TEST, the standard error is built from:
Question 4 of 5
For the cereal test (x̄ = 496, μ0 = 500, σ = 12, n = 36), what is the test statistic?
Question 5 of 5
A poll of 400 finds p̂ = 0.55 against H0: p = 0.5. What is the test statistic?
