8 topics. Each one opens with its key idea, then the detail you need to use it.
1. The sampling distribution of p-hat
Key idea: p-hat centers at p, has SD sqrt(p(1-p)/n), and is approximately Normal when np and n(1-p) are both at least 10.
Imagine 40% of a school's students walk to school, so p = 0.40. You take a random sample of 150 students and compute p-hat. Repeat that thousands of times, and the plot of every p-hat is the sampling distribution. Its mean is exactly p, 0.40, which is why p-hat is called an unbiased estimator. Its standard deviation is sqrt(p(1-p)/n) = sqrt(0.40 × 0.60 / 150) = sqrt(0.0016) = 0.04.
Judge any estimator on two separate features. Bias is about center: does the sampling distribution sit on the parameter? Variability is about spread, and a bigger random sample shrinks it. But a bigger sample cannot fix a biased method, such as a voluntary-response poll.
Two conditions let you use this model. The 10% condition says n should be no more than 10% of the population when you sample without replacement, so that one pick barely changes the next. Here that means the school needs at least 1,500 students. The Large Counts condition says np ≥ 10 and n(1-p) ≥ 10. Here np = 60 and n(1-p) = 90, so the shape is approximately Normal.
Why Normal? Code each student as 1 for walks and 0 for does not. Then p-hat is the mean of those 0s and 1s. The Central Limit Theorem says sample means from a large enough sample are approximately Normal, whatever the shape of the population. The Large Counts check is how you show the sample is large enough for a yes/no variable. With small counts, p-hat is skewed and Normal areas are wrong.
2. Probability questions about p-hat
Key idea: Find the SD of p-hat first, convert each boundary to a z-score, and subtract left-tail areas to get the probability between values.
Some questions give you the true p and ask how likely a sample result is. Use the sampling distribution, not the population. Stay with p = 0.40, n = 150 and SD = 0.04. What is the probability that p-hat falls between 0.36 and 0.46?
Convert each boundary to a z-score with z = (p-hat - p) / SD. For 0.36, z = (0.36 - 0.40) / 0.04 = -1.00. For 0.46, z = (0.46 - 0.40) / 0.04 = 1.50. The probability between two values is the area to the left of the upper z minus the area to the left of the lower z. That is 0.9332 - 0.1587 = 0.7745, so about a 77% chance.
Three slips cost points here. First, students use the population SD, sqrt(p(1-p)), and forget to divide by n. That gives a spread far too wide. Second, they use p-hat in the SD formula when the true p is given; when p is known, use p. Third, they skip the conditions. On a free-response question, you must state that np and n(1-p) are at least 10 before you are allowed to use Normal areas. On a calculator, normalcdf(0.36, 0.46, 0.40, 0.04) gives the same answer in one step, but label the mean and SD so a reader can follow you.
3. Confidence intervals and margin of error
Key idea: Higher confidence means a wider interval; a larger n means a narrower one; the confidence level describes the method, not one interval.
A confidence interval for p has the form estimate ± margin of error. The estimate is p-hat. The margin of error is the critical value z* times the standard error, sqrt(p-hat(1-p-hat)/n). You use p-hat in the standard error because you do not know p. The critical values to know are 1.645 for 90%, 1.960 for 95% and 2.576 for 99%.
Suppose 500 randomly chosen parents are surveyed and 63% are satisfied with school lunch. The standard error is sqrt(0.63 × 0.37 / 500), about 0.0216. At 95% confidence the margin of error is 1.96 × 0.0216, about 0.042, giving roughly 0.588 to 0.672.
The interpretation is where students lose credit. Say: we are 95% confident the interval from 58.8% to 67.2% captures the true proportion of all parents in the district who are satisfied. The 95% describes the method. If you repeated the sampling many times, about 95% of the intervals built this way would contain the true p. It is wrong to say there is a 95% probability that p is in this one interval; p is fixed and it either is or is not inside.
To plan a sample size, solve ME = z* sqrt(p(1-p)/n) for n. With no prior guess, use p = 0.5, which gives the largest n. For a 3% margin at 95%, n = (1.96/0.03)^2 × 0.25 ≈ 1067.1, so round up to 1,068.
4. The one-proportion z test, step by step
Key idea: In a test, use p0 in both the conditions and the standard deviation; compare the P-value to alpha and answer in context.
A school claims 30% of its students have a part-time job. You think it is higher. You survey a random sample of 200 students and 74 have jobs, so p-hat = 0.37. Work through four steps.
State. H0: p = 0.30 and Ha: p > 0.30, where p is the true proportion of all students at the school with a part-time job. Pick α = 0.05. Hypotheses are always about the parameter p, never about p-hat.
Plan. Name the test, a one-proportion z test. Check random sampling, the 10% condition, and Large Counts using the null value: np0 = 200 × 0.30 = 60 and n(1-p0) = 140, both at least 10. You use p0 because the test assumes the null is true.
Do. The standard deviation also uses p0: sqrt(0.30 × 0.70 / 200) ≈ 0.0324. The test statistic is z = (0.37 - 0.30) / 0.0324 ≈ 2.16. The P-value is P(Z > 2.16) ≈ 0.015.
Conclude. Because 0.015 < 0.05, reject H0. There is convincing evidence that more than 30% of students at this school have a part-time job. Always link the P-value to alpha and state the conclusion in context. If the P-value had been 0.20, you would fail to reject H0 and say there is not convincing evidence. You would never say you accept or prove the null.
5. Two-proportion z test and the pooled proportion
Key idea: Pool the two samples only when testing p1 = p2; intervals for a difference use separate p-hats.
To compare two groups, look at the difference in proportions, p-hat1 - p-hat2. Say 45 of 200 students who got a text reminder turned in homework late, versus 60 of 200 who got no reminder. Call the no-reminder group 1, so p-hat1 = 0.30 and p-hat2 = 0.225, a difference of 0.075.
The null is usually H0: p1 = p2, or p1 - p2 = 0. If the two proportions are equal, your best single estimate of that shared value combines both samples. This is the pooled proportion: p-hat_c = (60 + 45) / (200 + 200) = 105/400 = 0.2625. The standard error for the test is sqrt(p-hat_c(1 - p-hat_c)(1/n1 + 1/n2)) = sqrt(0.2625 × 0.7375 × 0.01) ≈ 0.0440. Then z = 0.075 / 0.0440 ≈ 1.70.
For a two-sided Ha: p1 ≠ p2, the P-value is 2 × P(Z > 1.70) ≈ 0.089. At α = 0.05 you fail to reject H0, so there is not convincing evidence the reminder changes the late rate.
Pooling happens only in the test. A confidence interval for p1 - p2 does not assume the proportions are equal, so it uses the unpooled standard error, sqrt(p-hat1(1-p-hat1)/n1 + p-hat2(1-p-hat2)/n2). For conditions, check that the samples are random or that treatments were randomly assigned. The two groups must be independent of each other. All four counts of successes and failures must be at least 10.
6. Intervals for a difference, and using intervals to judge claims
Key idea: An interval for p1 - p2 uses unpooled p-hats; values inside any interval are plausible, values outside are not, and 0 inside means no convincing difference.
When you compare two groups, the statistic is p-hat1 - p-hat2. Its sampling distribution has mean p1 - p2 and standard deviation sqrt(p1(1-p1)/n1 + p2(1-p2)/n2). The variances add because the two samples are independent. The shape is approximately Normal when all four counts, n1p1, n1(1-p1), n2p2 and n2(1-p2), are at least 10. You also need random samples or random assignment, plus the 10% condition for each sample when sampling without replacement. The 10% condition does not apply to a randomized experiment, because subjects are assigned to groups rather than sampled from a larger population.
Use the reminder data again. The unpooled standard error is sqrt(0.30 × 0.70/200 + 0.225 × 0.775/200) ≈ 0.0438. A 95% interval is 0.075 ± 1.96 × 0.0438, or 0.075 ± 0.086, which runs from about -0.011 to 0.161. We are 95% confident the true late rate without a reminder is between 1.1 points lower and 16.1 points higher than the rate with one. Because 0 is inside, no difference is a plausible value, so the interval gives no convincing evidence that reminders change the late rate. That matches the test.
An interval also lets you judge a claim about one proportion. The lunch interval, 0.588 to 0.672, lies entirely above 0.50, so it gives convincing evidence that a majority of parents are satisfied. A value inside it, like 0.60, is plausible. A value outside it, like 0.55, is not. If the district instead reported a margin of ±4.4%, it used the conservative p = 0.5, because 1.96 × sqrt(0.25/500) ≈ 0.044.
7. Type I errors, Type II errors and alpha
Key idea: Type I = false positive, with probability alpha; Type II = false negative, made less likely by a bigger n, a bigger alpha or a bigger true effect.
Your test ends with a decision, and reality is either H0 true or H0 false. That gives four outcomes. Rejecting a true H0 is a Type I error. Failing to reject a false H0 is a Type II error.
Always describe errors in context. Take a fertilizer study with H0: the fertilizer has no effect. A Type I error is concluding the fertilizer raises yield when it does nothing, so a farmer pays for a useless product. A Type II error is concluding it has no effect when it really works, so a good product gets shelved. In a vaccine trial, a Type II error means a vaccine that really prevents disease is judged useless.
Alpha is the probability of a Type I error when H0 is true. If you set α = 0.05 and H0 is true, you will wrongly reject it 5% of the time in the long run. That is why 20 tests of true nulls at α = 0.05 are expected to give 20 × 0.05 = 1 false positive. Running many tests and reporting only the significant one inflates false positives.
The chance of a Type II error, called β, falls in three ways: a larger sample, a larger alpha, or a true value further from the null value. Each one makes a real effect easier to detect. The tradeoff follows directly. Cutting alpha from 0.05 to 0.01 makes rejection harder, so Type I risk falls and Type II risk rises. The way to lower both at once is to increase n.
8. Chi-square tests of homogeneity and independence
Key idea: Several samples, one variable = homogeneity; one sample, two variables = independence; expected = row × column / total; df = (r-1)(c-1); always right-tailed.
A two-way contingency table gives you two chi-square options with the same arithmetic. How the data were collected decides between them.
A test of homogeneity compares the distribution of one categorical variable across several populations or treatment groups. You might sample 100 freshmen, 100 sophomores and 100 juniors separately and record each student's favorite lunch. H0: lunch preference has the same distribution in all three grades.
A test of independence uses a single sample and records two variables on each person. You might survey 300 random students and record both grade and lunch preference. H0: grade and lunch preference are independent, meaning not associated.
For either test, the expected count in each cell is (row total × column total) / grand total. With a row total of 120, a column total of 80 and a grand total of 300, that is 120 × 80 / 300 = 32. That is the count you would see if rows and columns were unrelated. Degrees of freedom are (rows - 1)(columns - 1), so a 3 × 4 table has df = 6. The test is always right-tailed, because only a large statistic signals a poor match.
Conditions: random sampling or random assignment, every expected count (not observed count) at least 5, and the 10% condition when sampling without replacement. Use counts, never percentages. A 2 × 2 chi-square test gives the same P-value as a two-sided two-proportion z test, since chi-square equals z squared.
The current framework for this unit covers only homogeneity and independence. You may still meet a third chi-square test, goodness-of-fit, in older practice materials. It uses one sample and one categorical variable and compares the counts to a claimed distribution. Each expected count is n × the claimed proportion, and df = number of categories - 1. The conditions are the same, and the test is still right-tailed.