1. Estimating the Difference Between Two Means — Variances Known

Theorem — CI for , Known

For independent random samples (normal populations, or large ):

Derivation. is (approximately) normal with mean and variance (Chapter 10); standardize and isolate as in Chapter 12.

Worked Example

Fuel economy (km/L) of engine types A and B: , , ; , , .

The interval is entirely negative, so B appears more economical by 3.5–8.5 km/L.

⚠️ Reading two-sample CIs: if the interval for contains 0, the data are consistent with no difference; if it is all positive/negative, one mean is higher.


2. Variances Unknown but Equal — Pooled

Definition — Pooled Variance


A weighted average of the two sample variances, weights = degrees of freedom.

Theorem — Pooled CI (normal populations, unknown)

Why: with a common , (sum of independent ‘s), and dividing the standardized difference by gives a variable in which cancels.

Worked Example

Exam scores with two teaching methods: method 1: , , ; method 2: , , .


⚠️ Pooling is justified only if the variances are plausibly equal. A common rule of thumb: OK if the larger is less than about twice the smaller (or check with an test, Chapter 11).


3. Variances Unknown and Unequal — Welch’s

Theorem — Welch–Satterthwaite CI


with approximate degrees of freedom

Worked Example

Phosphorus levels at two stations: , , ; , , (very different spreads).



4. Paired Observations

Definition — Paired Samples

Each observation in sample 1 is naturally matched with one in sample 2 (same unit measured twice, before/after, twins, same day). Work with the differences , which form one sample.

Theorem — CI for (paired)


where and are the mean and standard deviation of the differences.

Worked Example

Battery life (hours) of 6 laptops with the old and new power-management driver:

Laptop123456
New5.65.16.45.65.46.0
Old5.24.86.15.54.95.7
= new − old0.40.30.30.10.50.3

, , , :

Each laptop improves (lines all rise), despite big laptop-to-laptop spread Old driver New driver 5.0 h 5.5 h 6.0 h 6.5 h

⚠️ Why pairing helps: laptops differ a lot from each other (5–6.4 h), but each laptop’s improvement is consistent. Treating the data as two independent samples would mix laptop-to-laptop variation into the error and give a much wider interval. Conversely, never pair data that are not naturally matched.


5. Estimating a Proportion

Theorem — Large-Sample CI for

If is the proportion of successes in a random sample of size ( and ):

Derivation. (Chapter 10); isolate and replace the unknown in the standard error by .

Worked Example

In a survey of students, use Linux for coursework: .

Theorem — Sample Size for Estimating

For error at most with confidence :

Why the conservative formula works: is maximized at , where it equals .

Example. Error at most 0.02 with 95% confidence:

  • Using :
  • Conservative:

6. Estimating the Difference Between Two Proportions

Theorem — Large-Sample CI for

Worked Example

A process change is tested: before, 75 defectives in 1500 units (); after, 80 in 2000 ().


The interval contains 0, so the data do not clearly show that the change reduced the defect rate.

Summary Table — Which Interval?

ParameterSituationIntervaldf
known–
unknown, normal
’s known–
unknown, equal
unknown, unequalWelch
paired
large –
large ‘s–
flowchart TD
    A["Compare two means"] --> B{"Samples paired?"}
    B -- Yes --> P["Paired t on differences d"]
    B -- No --> C{"σ₁, σ₂ known?"}
    C -- Yes --> Z["Two-sample z"]
    C -- No --> D{"σ₁ ≈ σ₂?"}
    D -- Yes --> PO["Pooled t, v = n₁+n₂−2"]
    D -- No --> W["Welch t, Satterthwaite v"]