Statistics › Section I · Hypothesis testing with one sample › Topic I.5
Complete one-sample tests
Every piece is built: hypotheses, error types, reference curves, p-values, decisions. This topic bolts them together and runs complete tests start to finish - the four steps, in order, every time.
The whole machine
Every piece is now built: hypotheses, error types, reference curves, p-values, decisions. This topic bolts them together and runs complete tests, start to finish - a t test for a mean, a z test for a proportion, each with hypotheses, conditions, statistic, p-value and a conclusion written in context. The skill is not new machinery but orchestration: four steps, in order, every time.
The four-step recipe
1. State the hypotheses: \(H_0\) with its value, \(H_a\) with the question's direction - before anything else. 2. Check the conditions: random, independence (the 10% rule), and the shape or counts requirements that license the chosen curve. 3. Compute the statistic and its p-value on the correct distribution. 4. Decide by \(p \le \alpha\) and conclude in context, naming the direction and the strength of evidence. Skipping a step - or running them out of order - is where complete tests most often fail, even when each individual calculation is right.
Worked example: do students sleep enough?
A sleep study of \(n = 16\) students gives \(\bar{x} = 6.6\) hours and \(s = 1.2\). Is the mean below the recommended 7? 1. \(H_0: \mu = 7\); \(H_a: \mu < 7\) - one-sided, the question names a direction. 2. Conditions: random students; \(n = 16\) is small, so the data should look roughly symmetric without outliers (the study reports they do). 3. \(t = \frac{6.6 - 7}{1.2/\sqrt{16}} = \frac{-0.4}{0.3} = -1.33\) with 15 df; the one-sided p-value is about 0.10. 4. \(p = 0.10 > 0.05\): fail to reject. The evidence is suggestive - the sample mean is below 7 - but not convincing: at \(n = 16\), a real shortfall could easily hide.
Worked example: which package design?
In a random sample of \(n = 100\) customers, 58 prefer a new package design. Is preference different from an even split? 1. \(H_0: p = 0.5\); \(H_a: p \ne 0.5\) - "different" is two-sided. 2. Conditions: random sample; \(np_0 = 50\) and \(n(1-p_0) = 50\) both clear 10. 3.
and the two-sided \(p = 2 \times P(Z > 1.6) = 2 \times 0.0548 \approx 0.11\). 4. \(p = 0.11 > 0.05\): fail to reject. A 58-42 split sounds decisive, but with 100 customers it sits within an even split's ordinary wobble.
Worked example: the advertised delivery time
A cafe advertises a 30-minute average delivery; a student times \(n = 36\) orders, finding \(\bar{x} = 28.2\) minutes and \(s = 4.5\). 1. \(H_0: \mu = 30\); \(H_a: \mu \ne 30\) - two-sided, since either direction indicts the advertisement. 2. Conditions: random orders; \(n = 36\) lets the CLT speak. 3. \(t = \frac{28.2 - 30}{4.5/\sqrt{36}} = \frac{-1.8}{0.75} = -2.4\) with 35 df; two-sided \(p \approx 0.022\). 4. \(p \le 0.05\): reject - convincing evidence the true average differs from 30, with the data pointing to faster deliveries. Check: the companion interval \(28.2 \pm 2.03 \times 0.75 \approx (26.7, 29.7)\) excludes 30, agreeing with the test.
Conditions in practice
The checks are quick once they become habits. Randomness: was the sample drawn by chance, or grabbed? Independence: is the sample under 10% of its population? Shape: \(n \ge 30\) settles means; small samples need roughly symmetric data; proportions need \(np_0\) and \(n(1-p_0)\) at least 10 - counts under the null. The sleep study's \(n = 16\) required the symmetric-data check; the poll's 100 required the 50-and-50 counts. A test that cannot show its conditions has produced a p-value-shaped number, and no more.
Reporting the whole story
A complete report says more than "significant". The sleep study: mean 6.6 hours, \(p \approx 0.10\) - suggestive but inconclusive; a larger study might settle it. The delivery test: significant, and the estimate 28.2 with interval (26.7, 29.7) says how large the difference from 30 is - about 1.8 minutes, plausibly anywhere from 0.3 to 3.3, which the cafe might call an advertising victory rather than a scandal. The effect's size and its interval carry the decision; the p-value only says the data would be surprising if the true mean were 30. The next section's effect-size topic makes this discipline formal.
Common mistakes to avoid
First, running the steps out of order - computing a statistic before writing hypotheses, or choosing the alternative after seeing the p-value. Second, skipping the conditions, especially the small-sample shape check the sleep study needed. Third, reading "not significant" as "no effect": the sleep study may be a Type II miss, underpowered at \(n = 16\). Fourth, the wrong degrees of freedom - 15, not 16, for the sleep study. Fifth, one-sided/two-sided mismatch: the poll's 0.11 is two-sided; a one-sided reading would halve it. Sixth, concluding without context - "reject \(H_0\)" is arithmetic, not a conclusion.
Key takeaways
Complete tests run four steps in order: hypotheses (\(H_0\) with its value, \(H_a\) with the question's direction), conditions (random, 10%, shape or counts), statistic with p-value on the right curve, decision with a contextual conclusion. The sleep study: \(t = -1.33\), 15 df, one-sided \(p \approx 0.10\) - fail to reject; suggestive, not convincing. The design poll: \(z = 1.6\), two-sided \(p \approx 0.11\) - fail to reject; 58 of 100 sits within an even split's wobble. The delivery test: \(t = -2.4\), \(p \approx 0.022\) - reject; the mean differs from 30, with interval (26.7, 29.7) agreeing. A non-rejection is never proof of no effect, and every rejection wants its size reported.
Flashcards
Flashcard
Hypotheses, conditions, statistic and p, conclusion
Test your understanding
Question 1 of 5
For the sleep study (n = 16, x̄ = 6.6, s = 1.2, μ0 = 7), what is the test statistic?
Question 2 of 5
The sleep study's one-sided p-value is about 0.10. At α = 0.05:
Question 3 of 5
In the design poll, 58 of 100 prefer the new design. Against H0: p = 0.5, the test statistic is:
Question 4 of 5
The design poll's two-sided p-value (z = 1.6) is about:
Question 5 of 5
The delivery test gives t = −2.4 with two-sided p ≈ 0.022 at α = 0.05. The conclusion:
