7 topics. Each one opens with its key idea, then the detail you need to use it.
1. The sampling distribution of x̄ and the Central Limit Theorem
Key idea: x̄ is centred at μ with spread σ/√n, and it is normal if the population is normal or approximately normal when n is large.
Picture taking every possible random sample of size n from a population and recording each x̄. Those values form the sampling distribution of x̄. Its mean equals the population mean μ, so x̄ is an unbiased estimator, and its standard deviation is σ/√n, which shrinks as n grows. Because of the square root, you must quadruple the sample size to cut the spread in half.
Shape depends on the population. If the population is normal, x̄ is exactly normal for any n, even n = 4. If the population is skewed or unknown, the Central Limit Theorem says x̄ becomes approximately normal as n grows, and n ≥ 30 is the usual rule of thumb. The CLT is about the distribution of sample means. It does not make the population or a single sample look normal.
This lets you compute probabilities about x̄. Suppose a population has μ = 40 and σ = 12, and you take n = 36. The standard deviation of x̄ is 12/√36 = 2. To find P(x̄ > 43), compute z = (43 − 40)/2 = 1.5, so the probability is about 0.067. Notice that you divide by σ/√n, not by σ. Using σ = 12 would answer a question about one individual, not about a mean of 36 people.
2. Standard error, the t distribution and degrees of freedom
Key idea: Use s/√n as the standard error and the t distribution with df = n − 1, because you are estimating σ with s.
The formula σ/√n needs σ, the population standard deviation, which you almost never know. So you estimate it with s, the sample standard deviation, and the result, s/√n, is called the standard error of x̄, which estimates how far x̄ typically falls from μ.
Swapping σ for s adds extra uncertainty, because s is itself random. Small samples can give an s that is too small, which would make you overconfident. The t distribution corrects for this. It is centred at 0 and bell-shaped like the standard normal, but its heavier tails mean critical values are larger, so intervals are wider and p-values are bigger than z would give.
Which t curve you use depends on the degrees of freedom. For one sample or for paired differences, df = n − 1. As df increases, s becomes a more reliable estimate of σ and the t curve moves closer to the normal curve. For df = 24, the 95% critical value is t* = 2.064, while z* is 1.96. For df = 5, t* jumps to 2.571.
This is why the exam cares about the label. If a stem gives s and asks for the test statistic, the answer is a t statistic. Calling it z is wrong even when the arithmetic matches, because z assumes σ is known.
3. Conditions: random, independent, normal enough
Key idea: State and check random, 10%, and normal/large-sample conditions in context, and apply normality to the differences when data are paired.
Every t procedure rests on three conditions, and each one protects a different part of the method. Check them before you calculate, and tie each check to the data in the stem.
Random. The data must come from a random sample or a randomized experiment. Without randomness, the sampling distribution does not exist in any useful sense, and no formula fixes the bias. A convenience sample from one university can give a significant result that still does not generalize to all students.
Independent. When sampling without replacement, the sample should be no more than 10% of the population. This keeps the observations close to independent, which the formula s/√n assumes. In an experiment, random assignment gives independence between groups.
Normal or large sample. The sampling distribution of x̄ must be approximately normal. You can claim this three ways. The stem says the population is normal. The sample size is at least 30, so the CLT applies. Or, with a small sample, a dotplot, boxplot or normal probability plot of the data shows no strong skew and no outliers. With n = 12 and a big outlier, a t procedure is not trustworthy, because the mean and s are both pulled by that one value.
For paired data, check normality on the differences, not on the two original lists. For two samples, check each group separately.
4. Building and interpreting a one-sample t interval
Key idea: Interpret the interval as plausible values for the population mean in context, and explain the confidence level as the long-run capture rate of the method.
A confidence interval has the form statistic ± margin of error. For a mean, that is x̄ ± t*·(s/√n). Here is a full calculation. A random sample of 25 commuters has x̄ = 31.4 minutes and s = 6.5 minutes. The standard error is 6.5/√25 = 1.3. With df = 24, the 95% critical value is 2.064, so the margin of error is 2.064 × 1.3 ≈ 2.68. The interval is about 28.7 to 34.1 minutes.
A good interpretation names the parameter in context. You would write that you are 95% confident the true mean commute time for all commuters in this city is between 28.7 and 34.1 minutes. Name the population mean, not the sample mean. x̄ is already known (31.4) because you computed it from the data, and the interval is built around it, so there is nothing to estimate about it.
The confidence level describes the method, not this one interval. If you took many random samples and built an interval from each, about 95% of those intervals would capture μ. This specific interval either contains μ or it does not, so you should not say there is a 95% probability that μ is in it.
Three things control the width. A higher confidence level needs a bigger t*, so the interval widens. A larger sample shrinks s/√n, so the interval narrows. More variable data means a larger s and a wider interval. The margin of error only covers random sampling error. It does nothing about bias from a bad sampling method or from nonresponse.
5. The one-sample t test, from hypotheses to conclusion
Key idea: State hypotheses about μ, compute t = (x̄ − μ₀)/(s/√n), find the p-value in the direction of Hₐ, and conclude in context in terms of evidence.
A test starts with two hypotheses about the parameter. H₀ gives a specific value, such as μ = 10 hours. Hₐ says μ is less than, greater than, or not equal to that value. The direction comes from the research question, and you must choose it before you look at the data.
Say a company claims its batteries last 10 hours on average, and a reviewer suspects they last less. A random sample of 16 batteries gives x̄ = 9.4 hours and s = 1.2 hours. The hypotheses are H₀: μ = 10 and Hₐ: μ < 10. The standard error is 1.2/√16 = 0.3, so the test statistic is t = (9.4 − 10)/0.3 = −2.0 with df = 15.
The p-value is the probability of getting a sample mean at least this far below 10, assuming the true mean really is 10. Because Hₐ is one-sided, you only use the lower tail, and p ≈ 0.032. A two-sided Hₐ would double it to about 0.064.
Compare p to α: here 0.032 < 0.05, so you reject H₀. The conclusion should say there is convincing evidence that the true mean battery life of these batteries is less than 10 hours. If p had been above α, you would fail to reject H₀ and say there is not convincing evidence. You never "accept" H₀, and rejecting it does not prove it false. A test weighs evidence, so the wording must stay at the level of evidence.
6. Paired data versus two independent samples
Key idea: If each value has a partner, analyse the differences as one sample; if the groups are separate, use the two-sample standard error √(s₁²/n₁ + s₂²/n₂).
Paired data arise when each observation has a natural partner. Examples include the same student before and after tutoring, the left and right eye of one patient, or twins split between two treatments. Pairing removes the variation between individuals, so the test is often more sensitive.
To analyse paired data, subtract within each pair to get one list of differences. Then run a one-sample t procedure on those differences, with df = number of pairs − 1. Suppose 12 students take a quiz before and after a review session, and the mean gain is 3.5 points with s_d = 4.2. The standard error is 4.2/√12 ≈ 1.21, so t ≈ 2.89 with df = 11. The one-sided p-value is about 0.007. Fifteen pairs give df = 14, not 28, because you have 15 differences, not 30 separate values.
Two independent samples arise when the groups contain different individuals, picked or assigned separately. The standard error combines both groups: SE = √(s₁²/n₁ + s₂²/n₂). The statistic is t = (x̄₁ − x̄₂)/SE when H₀ says the means are equal. Your calculator finds df with a complex formula, but a safe hand method uses the smaller of n₁ − 1 and n₂ − 1, which gives a slightly larger p-value.
For example, a researcher takes random samples of students from two large schools and compares their mean scores on the same test. School 1 has n = 20, x̄ = 78, s = 10, and school 2 has n = 25, x̄ = 72, s = 12. The hypotheses are H₀: μ₁ = μ₂ and Hₐ: μ₁ ≠ μ₂, where μ₁ and μ₂ are the true mean scores at each school. The SE is √(5 + 5.76) ≈ 3.28, so t ≈ 1.83. With conservative df = 19, the two-sided p-value is about 0.08. Because 0.08 > 0.05, there is not convincing evidence that the true mean scores at the two schools differ.
To estimate the difference instead of testing it, use the two-sample t interval (x̄₁ − x̄₂) ± t*·√(s₁²/n₁ + s₂²/n₂). Its interpretation names the difference in population means, μ₁ − μ₂, in context, such as the true mean score at school 1 minus the true mean score at school 2.
7. Duality, statistical significance and practical significance
Key idea: An interval that excludes the null value matches a two-sided rejection, and the interval's size, not the p-value, tells you whether the effect matters.
A two-sided test at level α and a confidence interval at level 1 − α give matching answers. If a 95% interval for a mean difference is 5 to 11 points, then 0 is not a plausible value. So a two-sided test of H₀: μ_d = 0 would reject at α = 0.05. If the interval had been −1 to 6, 0 would be plausible and the test would fail to reject. You do not need the p-value to make this call, but the interval must match the test: 95% with α = 0.05, 99% with α = 0.01.
The interval also tells you something the p-value cannot: how big the effect is. That is where practical significance comes in. Statistical significance only says that a difference this large would rarely happen by chance if the true effect were zero, so the data give convincing evidence that the effect is not zero. With a huge sample, even a tiny difference gives a small p-value, because the standard error shrinks with √n. A drug that lowers cholesterol by 2 mg/dL might reach p = 0.03 yet be medically pointless when levels run 100 to 200 mg/dL.
Effect size is the size of the difference in real units, or relative to the spread of the data. Judge it against the context, not against α. An interval like 0.5 to 3.5 points on a 100-point test is significant but small. An interval like 8 to 20 points is both significant and large.
The last check is scope. A significant two-sample result from a convenience sample still does not generalize to a wider population. Only random assignment supports a cause-and-effect claim.