Inference for Categorical Data: Proportions

What this unit covers

The topics below follow the published Statistics course framework for Unit 3. This unit is worth 15–25% of the exam, so budget your time against that rather than against how long the unit takes to teach.

Lessons in this unit

• Sampling Methods13 min · 3 objectivesDistinguish a population from a sample and a parameter from a statistic · Describe simple random, stratified, cluster, and systematic sampling · Explain why random sampling supports generalization to a population
• Experiments & Their Design15 min · 3 objectivesIdentify the treatments, experimental units, explanatory and response variables · Apply the three principles of experimental design: control, randomization, replication · Explain the purpose of a control group and a placebo
• Bias in Data Collection12 min · 3 objectivesIdentify undercoverage, nonresponse, and response bias · Explain how question wording can bias survey results · Distinguish bias from sampling variability
• Randomization, Blocking & Scope of Conclusions13 min · 3 objectivesDescribe randomized block and matched-pairs designs and their purpose · Explain how blocking reduces variability · State the scope of conclusions permitted by a study’s design
• Bias and Variability Are Different Problems14 min · 3 objectivesDistinguish bias from sampling variability · Explain why increasing sample size reduces variability but not bias · Identify which design flaws produce which problem
• Matched Pairs & Why Pairing Beats Two Independent Groups14 min · 3 objectivesRecognize a matched-pairs design and distinguish it from two independent samples · Explain how pairing removes variability between subjects · Choose the correct inference procedure for a paired design

Formulas in Unit 3

Every term in Unit 3

All 59 terms we publish for Inference for Categorical Data: Proportions, with definitions. Reading them through is the fastest way to find the ones you cannot define — then drill those in cram mode until you can produce them without the prompt.

Confidence interval structure

Statistic ± (critical value)(standard error); the margin of error covers sampling variability only, not bias.

Interpreting a confidence level

"If we repeated this sampling many times, about 95% of the intervals produced would capture the true parameter." A statement about the method, not one interval.

Interpreting a confidence interval

"We are 95% confident that the interval from a to b captures the true population proportion of …" — in context, about the parameter, not the statistic.

Conditions for a one-proportion z-interval

Random sample, 10% condition, and Large Counts using np̂ ≥ 10 and n(1 − p̂) ≥ 10.

Conditions for a one-proportion z-test

Same, but Large Counts uses the null value p₀ rather than p̂, because the test assumes the null is true.

Margin of error

The critical value times the standard error. Reducing it requires a larger sample or a lower confidence level.

Effect of sample size on the interval

Quadrupling n halves the margin of error, because standard error falls as √n.

Null and alternative hypotheses

H₀ states no effect or no difference using the parameter symbol; Hₐ states what you are testing for. Both are about parameters, never statistics.

p-value definition

The probability of getting a result at least as extreme as the one observed, ASSUMING the null hypothesis is true.

Interpreting a p-value in context

"Assuming the true proportion is p₀, there is a p% chance of observing a sample proportion this far from p₀ or farther."

Conclusion wording

Reject H₀ when p < α: "we have convincing evidence that…". Otherwise fail to reject: "we do not have convincing evidence" — never "we accept H₀".

Type I and Type II errors

Type I rejects a true null (a false positive, probability α); Type II fails to reject a false null. Lowering α raises the chance of a Type II error.

Power

The probability of correctly rejecting a false null, 1 − P(Type II). Increased by a larger sample, a larger effect size, or a larger α.

Two-proportion z-test pooling

Under the null the proportions are equal, so combine both samples into a pooled p̂ for the standard error. The interval does not pool.

Chi-square statistic

χ² = Σ(observed − expected)²/expected, summed over every cell. Always non-negative, and larger values mean worse fit to the null.

Goodness-of-fit test

Compares one categorical variable's observed counts to a claimed distribution. df = categories − 1.

Test for homogeneity

Compares the distribution of one categorical variable across several populations or treatments. df = (rows − 1)(columns − 1).

Test for independence

Tests whether two categorical variables are associated within one population. Same statistic and df as homogeneity; the difference is the sampling design.

Conditions for chi-square

Random sample or random assignment, 10% condition, and every EXPECTED count at least 5 — expected, not observed.

Calculating expected counts

For a two-way table, (row total × column total)/grand total.

Chi-square distribution shape

Right-skewed, becoming more symmetric as degrees of freedom rise. Only large values give small p-values, so the test is inherently one-sided.

Which chi-square test to use

One sample and one variable is goodness-of-fit; several samples and one variable is homogeneity; one sample and two variables is independence.

Follow-up after a significant chi-square

Identify the cells with the largest contributions to χ² and describe how observed differs from expected there, in context.

The four-step template

State the parameter and hypotheses, Plan by naming the procedure and checking conditions, Do the mechanics, Conclude in context. Each step carries points.

A p-value is not the probability the null is true

It is the probability of data this extreme GIVEN the null. The reversal is the most penalized misinterpretation in the course.

Why we never accept the null

Failing to reject means the evidence was insufficient, which is not the same as evidence of no effect.

Using a confidence interval to test

If the null value falls outside a 95% interval, a two-sided test at α = 0.05 would reject it. The two procedures agree.

Choosing sample size for a margin of error

Set the margin-of-error expression less than the target and solve for n, using p̂ = 0.5 when no estimate is available, since it maximizes the required size.

One-sided vs two-sided alternatives

Decide from the question before seeing the data. Choosing a one-sided alternative after looking at the sample is not legitimate.

Stating chi-square hypotheses

Goodness-of-fit states a claimed distribution; homogeneity and independence state no difference and no association respectively, always in context.

Why expected counts, not observed

The condition guards the approximation of the sampling distribution, which depends on what the null predicts rather than on what was seen.

Combining categories

When an expected count is under 5, adjacent categories may be merged — reducing degrees of freedom accordingly.

Interpreting a large contribution to chi-square

That cell is where observed and expected diverge most, and it is where the description of the association should focus.

Point estimate

The single best guess at a parameter — p̂ for p, x̄ for μ. It carries no information about precision, which is why it is always reported with a margin of error or an interval.

Critical value z*

The multiplier setting the confidence level: 1.645 for 90%, 1.96 for 95%, 2.576 for 99%. Higher confidence means a larger z* and therefore a wider interval.

Standard error of p̂

For an interval, √(p̂(1 − p̂)/n), using the sample proportion because p is unknown. For a TEST it is √(p₀(1 − p₀)/n), using the null value — a distinction worth memorizing.

Pooled proportion

For a two-proportion test, combine both samples: p̂_c = (x₁ + x₂)/(n₁ + n₂). It is used only in the test, because the null hypothesis asserts the proportions are equal.

Why the interval does not pool

A confidence interval makes no claim that the two proportions are equal, so there is no justification for combining them. Pooling in an interval is a standard error.

A two-proportion interval containing zero

Zero is a plausible value for the difference, so the data do not provide convincing evidence that the two proportions differ. This is the interval equivalent of failing to reject.

Significance level

The threshold α chosen before looking at the data, and the probability of a Type I error if the null is true. Choosing α after seeing the p-value invalidates the procedure.

Type I error rate

Equal to α by construction: rejecting a true null. Lowering α reduces false positives and, holding everything else fixed, reduces power.

Type II error

Failing to reject a false null; its probability is β and power is 1 − β. Which error is worse depends entirely on the consequences in context.

Four ways to increase power

Increase the sample size, increase α, reduce variability, or study a larger true effect. Only the first two are under the investigator's direct control in most studies.

Effect size and power

The further the truth is from the null value, the easier it is to detect. A study can have excellent power for a large effect and almost none for a small one, at the same sample size.

The alpha-power tradeoff

Making α smaller makes a Type I error less likely and a Type II error more likely. There is no setting that reduces both; only more data does that.

Choosing the alternative

The direction must come from the question asked before the data are seen. Choosing a one-sided alternative after noticing which way the data went doubles the true Type I error rate.

Structure of every test statistic

(statistic − hypothesized parameter)/(standard error). Recognizing the pattern means one formula covers z-tests for proportions, t-tests for means and t-tests for slopes.

Reject versus fail to reject

Rejecting means the data are inconsistent with the null; failing to reject means they are consistent with it. Neither proves anything, and "accept the null" is never correct wording.

Conclusion template

Compare p to α, state the decision, then state the conclusion in context: "there is convincing evidence that the proportion of ... exceeds ...". Context is required for the point.

Interval-test duality

For a two-sided test at level α, the null value falls outside the (1 − α) confidence interval exactly when the test rejects. This is a useful check on your own work.

Why an interval is more informative than a test

A test reports only whether a single value is plausible; an interval reports every plausible value and thereby conveys the size of the effect as well as its existence.

Statistical versus practical significance

With a large enough n, any nonzero difference becomes significant. Whether it matters is a judgment about the subject and cannot be read off the p-value.

Caution with many tests

Running many tests at α = 0.05 produces false positives at roughly 5% of them by chance alone. A significant result found after searching many comparisons requires confirmation on new data.

Sample size for a given margin of error

n = (z*/ME)²p*(1 − p*), using p* = 0.5 when no estimate exists because that maximizes p(1 − p). Always round up, since rounding down misses the target margin.

Degrees of freedom, goodness of fit

df = (number of categories) − 1. It does not depend on the sample size, which surprises students used to t procedures.

Degrees of freedom, two-way tables

df = (rows − 1)(columns − 1). A 3 × 4 table gives df = 6 regardless of how many individuals were counted.

Expected count formula

For a two-way table, (row total × column total)/grand total. Expected counts need not be whole numbers, and rounding them before computing chi-square introduces error.

Why chi-square is right-skewed

It is a sum of squared quantities, so it cannot be negative and has a long right tail. Only large values count as evidence, which is why chi-square tests are always one-sided in the upper tail.

Reading the components

Each cell contributes (observed − expected)²/expected. Identifying the largest contributors, and their direction, is how a significant chi-square result is described in context.

What examiners penalize here

• Keep two "randoms" distinct: random **selection** (sampling) lets you *generalize* to a population; random **assignment** (experiments) lets you conclude *causation*. A study can have one, both, or neither — and that determines exactly what conclusions are allowed.
• When asked to identify bias, name the *type* (undercoverage, nonresponse, response, or wording) **and** state its likely *direction* — will the estimate be too high or too low? AP rubrics reward explaining how the flaw pushes the result.
• A classic free-response ending: "Can we conclude the treatment *caused* the difference, and can we generalize to all ___?" Answer both parts using the design — cite random *assignment* for causation and random *selection* for generalization. Missing either random feature limits the claim.
• When a question asks whether a larger sample would fix a described problem, identify the problem first. If it is bias, the answer is no, and you must say **why** — the systematic tilt is present in every observation, so collecting more of them collects more of the same tilt.
• The diagnostic question is simple: **is there a natural reason to link one observation in the first group with one specific observation in the second?** Before-and-after on the same person, left hand and right hand, twins, plots of land split in two — all paired. Two separate randomly assigned groups of different people — not paired.

Practice Statistics

Our practice bank is drawn from across the whole course rather than filtered to one unit, which is closer to how the exam asks anyway — it will not tell you which unit a question is testing.

Questions about this unit

All 5 units of AP Statistics

1. Unit 1 · Exploring One-Variable Data and Collecting Data
2. Unit 2 · Probability, Random Variables, and Probability Distributions
3. Unit 3 · Inference for Categorical Data: Proportions
4. Unit 4 · Inference for Quantitative Data: Means
5. Unit 5 · Regression Analysis