Statistics › Section J · Hypothesis testing with two samples › Topic J.3
Two means with known standard deviations
When both populations' spreads are known from specifications or long calibration - rare, but real - the two-mean comparison runs on the standard normal instead of t. The structure is the two-sample t test's, with every estimated s replaced by a known σ.
The rare clean case
The two-sample t test estimated both spreads from the samples, and the t distribution priced the extra uncertainty. Sometimes the spreads are genuinely known in advance: calibrated instruments carry manufacturer specifications, industrial processes run with decades of recorded variance, and standardised tests publish their score spreads. When both \(\sigma\) values are known, the two-mean comparison trades t for z - the same test with the cleaner curve, and no degrees of freedom to worry about.
The known-sigma z statistic
With both spreads known, the test statistic is
- the difference in sample means over the standard error of a difference, exactly as before, but built from the known \(\sigma\)'s instead of the estimated \(s\)'s. The reference curve is the standard normal: nothing is being estimated from the data except the means themselves, so there is no extra uncertainty for \(t\)'s heavier tails to price. Variances still add - the same independence rule, unchanged.
Worked example: two laboratories
Two labs measure the same chemical with different instruments, whose specifications give \(\sigma_A = 0.8\) and \(\sigma_B = 1.0\) mg/L. Lab A measures \(n_1 = 40\) samples, averaging \(\bar{x}_A = 12.4\); lab B measures \(n_2 = 50\), averaging \(\bar{x}_B = 11.9\). Do the labs disagree systematically? 1. \(H_0: \mu_A = \mu_B\); \(H_a: \mu_A \ne \mu_B\) - two-sided. 2. \(\text{SE} = \sqrt{\frac{0.64}{40} + \frac{1}{50}} = \sqrt{0.036} \approx 0.19\). 3. \(z = \frac{0.5}{0.19} \approx 2.64\); two-sided \(p = 2 \times 0.0042 \approx 0.008\). 4. \(p \le 0.05\): reject - convincing evidence of a systematic difference between the labs, with A reading higher.
Worked example: the companion interval
The 95% interval for the difference uses the normal's 1.96 (no degrees of freedom to look up): \((\bar{x}_A - \bar{x}_B) \pm 1.96 \times \text{SE}\). 1. Estimate: \(12.4 - 11.9 = 0.5\) mg/L. 2. Margin: \(1.96 \times 0.19 \approx 0.37\). 3. Interval: \((0.13, 0.87)\) mg/L - excluding zero, agreeing with the test. 4. Interpretation: we are 95% confident lab A reads between about 0.13 and 0.87 mg/L higher than lab B on this chemical. Check: the lower bound sits above zero exactly because \(z = 2.64\) exceeds 1.96, so interval and test must agree.
Conditions: the same, plus honesty about σ
The conditions carry over: both samples random, observations independent within and between groups, and shape handled by \(n \ge 30\) or symmetric data in small samples. One condition is new and demanding: the \(\sigma\)'s must genuinely be known - not estimated from these data, not borrowed from a similar study, not "roughly remembered". Estimating \(s\) from the same sample and then calling it \(\sigma\) quietly removes the uncertainty that \(t\) exists to price, producing overconfident p-values and too-narrow intervals. If the spread is estimated, use the two-sample t; this test earns its z only from real specifications.
Why this case matters less every year
The known-\(\sigma\) two-mean test is mostly a teaching case now: instruments drift, processes change, and honest analysts prefer estimating the spread and paying t's small premium. The test remains valuable where it genuinely applies - calibration chains, industrial standards, published test norms - and as the cleanest illustration of the general recipe: difference over the standard error of the difference, on the curve the amount of estimation licenses. Understanding why z is legitimate here and not elsewhere is half of understanding the t machinery.
Choosing the comparison test
Two topics of comparison machinery now exist, and the choice between them is one question: is the spread estimated? One mean versus a claim, spread estimated: one-sample t. Two groups, spreads estimated: two-sample t - the ordinary workhorse. Spreads genuinely known, from specifications or long calibration: z, the clean case. And a design where the same units appear in both groups - before-and-after, matched pairs, split vials - is not a two-independent-groups design at all: the matched-samples test, two topics on, handles it by working on the differences. Naming the design and what is known selects the test before any arithmetic begins.
Common mistakes to avoid
First, using z with spreads estimated from the samples - the t test is the honest choice whenever \(s\) stands in for \(\sigma\). Second, subtracting variances in the SE, the permanent temptation for a difference. Third, quoting an interval with 1.96 when \(s\) was used - that margin belongs to the known-\(\sigma\) case only. Fourth, treating published spreads as eternal: a specification from a decade-old model may not describe today's instruments. Fifth, checking the labs' data for the shape condition but never questioning whether \(\sigma\) still holds. Sixth, forgetting that the labs' samples must be independent - the same vials split between labs is a paired design, not this one.
Key takeaways
When both population spreads are known, two-mean comparisons run on the standard normal: \(z = (\bar{x}_1 - \bar{x}_2)/\sqrt{\sigma_1^2/n_1 + \sigma_2^2/n_2}\), variances still added, no degrees of freedom. The labs: \(\text{SE} = \sqrt{0.036} \approx 0.19\), \(z \approx 2.64\), \(p \approx 0.008\) - a systematic difference; interval \((0.13, 0.87)\) mg/L. The same conditions as the t version apply, plus the demanding one: the \(\sigma\)'s must genuinely be known, or the test quietly overreaches. Estimated spreads mean t - always - which is why this clean case matters mostly for calibration work and for understanding what t is paying for.
Flashcards
Flashcard
(x̄1 − x̄2)/√(σ1²/n1 + σ2²/n2)
Test your understanding
Question 1 of 5
With σ1 = 0.8, n1 = 40, σ2 = 1.0, n2 = 50, the standard error of the difference is about:
Question 2 of 5
With x̄A = 12.4 and x̄B = 11.9, and SE ≈ 0.19, the test statistic is about:
Question 3 of 5
The two-sided p-value for z = 2.64 is about:
Question 4 of 5
When both spreads are estimated from the samples (s1, s2), the honest test uses:
Question 5 of 5
In the known-σ case, the 95% margin of error for the difference uses:
