Inference for Categorical Data: Proportions

A parameter describes a population; a statistic describes a sample. A proportion measures the fraction in a specified category. Define the success category and population before choosing a procedure.

Define p as the true proportion of a stated population with a stated characteristic. A sample proportion estimates it. Random sampling or a relevant randomized design supports the probability model; large counts support a normal approximation. For sampling without replacement, a sample no more than about 10% of the population is a common independence check.

For a one-proportion interval, check at least 10 observed successes and 10 failures. For a one-proportion null test, check expected successes and failures using the null proportion. State the actual values so the reader can see whether the condition holds.

A one-proportion z interval uses the sample proportion plus or minus a critical value times its estimated standard error. Higher confidence uses a wider critical-value multiplier; larger sample size generally narrows the interval. Neither change repairs a biased sampling method.

A 95% confidence interpretation concerns the long-run success rate of the method. For one realized interval, say you are 95% confident it captures the population proportion. Do not say 95% of individuals lie in the interval, and do not describe the fixed parameter as randomly moving between interval endpoints.

State a null hypothesis with equality and an alternative matching the question before using the data to choose a direction. A one-proportion z test standardizes the observed difference from p₀ using the null standard error. The P-value is the probability, assuming the null model, of a statistic at least as extreme in the alternative’s direction.

A small P-value provides evidence against the null, not the probability that the null is true. Compare it with a prespecified significance level. “Fail to reject” means insufficient evidence for the alternative, not proof of no effect. Statistical significance does not establish practical importance.

Define p₁ − p₂ in a clear group order. For an interval, estimate variability separately in the two groups. For a test of equality, pool successes and sample sizes because the null assumes a shared population proportion. The pooled estimate is total successes divided by total observations, not the simple average of proportions when group sizes differ.

The groups must be independent for the ordinary two-proportion procedures. Paired before/after categorical outcomes are not automatically two independent samples. Check each group’s relevant success/failure counts and the study design. Interpret an interval in percentage points, not as a relative percent change unless that is explicitly calculated.

For a test of independence, one sample is classified by two categorical variables. For a test of homogeneity, independent samples or randomized groups are compared across a categorical response. Both use expected counts based on row and column totals under the null and a chi-square statistic measuring departures.

Expected counts are row total × column total / grand total. Check that all expected counts meet the usual threshold of at least 5. Degrees of freedom are (rows − 1)(columns − 1). A significant result indicates association or differing distributions, not which cell alone caused it or that the relation is causal. Follow up with the pattern of proportions and the design.

A Type I error rejects a true null; a Type II error fails to reject a false null for a particular alternative. Power is the probability of rejecting the null at a specified true alternative. Larger samples generally increase power to detect a fixed effect.

Raising the significance level can increase power but also increases the allowed Type I error rate. Context determines the consequences of each error. Report the actual effect estimate and uncertainty rather than using a significance label as the whole conclusion.

A confidence interval estimates an unknown population proportion with a sample proportion plus or minus a margin of error. Check that the data come from a suitable random process, observations are independent, and the sample has enough observed successes and failures for the Normal approximation. When sampling without replacement, check the 10% condition against the population size.

The confidence level describes the long-run success rate of the interval-producing method. After an interval is computed, the fixed population parameter is either inside it or outside it. An AP-level interpretation identifies the population and parameter and states confidence in the interval; it does not say that the specified fraction of observations falls between its endpoints.

State the null and alternative in population parameters before analyzing results. The null gives a reference value; the alternative specifies greater than, less than, or different from. For a one-proportion z test, the standard error and large-count check use the null proportion, because the p-value describes results assuming that null model is true.

A small p-value indicates that the observed statistic, or one more extreme in the direction of the alternative, would be unusual under the null model. It is not the probability that the null is true. Compare the p-value with a previously chosen significance level, then state a contextual conclusion. Failing to reject is not proof that the null is exactly correct.

For two independent samples, define p₁ and p₂ and keep the order of subtraction consistent through hypotheses, interval, and conclusion. An interval for p₁−p₂ estimates the difference using separate sample proportions in its standard error. A test of p₁=p₂ uses a pooled estimate under the equality null. Do not substitute one standard error automatically into the other procedure.

A statistically significant difference can be too small to matter practically. Examine the estimate, interval, and context alongside the p-value. A large sample can detect a tiny effect. Also distinguish a Type I error, rejecting a true null, from a Type II error, failing to reject a false null; describe the real-world mistaken decision in each case.

A population proportion describes a categorical outcome. State the population and success category. For two populations, define the order p₁−p₂ and keep it consistent in hypotheses, calculations, and conclusions. A negative estimated difference is meaningful only relative to that order.

A confidence interval estimates a parameter; a significance test evaluates evidence against a specific null model. For a one-proportion interval, estimated standard error uses the observed proportion. A one-proportion test uses the null proportion in its null standard error and success–failure check. Do not move formulas between purposes without checking their assumptions.

For a two-proportion test of equality, the pooled estimate reflects the null assumption of a common proportion. A two-proportion confidence interval generally uses separate sample estimates in its standard error. Independent groups are required for that procedure; paired categorical outcomes are not automatically two independent samples.

Random sampling or a defensible randomized design supports the intended inference. When sampling without replacement, check that the sample is at most about 10% of the population for the usual independence approximation. Check success and failure counts using values appropriate to the procedure, and name what those counts represent.

A 95% confidence level describes the long-run capture rate of the interval method under its assumptions. For one computed interval, say that you are 95% confident the population parameter lies between its endpoints. Do not say 95% of individual people have proportions in that interval, or that the parameter changes from sample to sample.

A P-value is the probability, assuming the null model and conditions, of a statistic at least as extreme in the direction of the alternative. It is not the probability that the null hypothesis is true. Failure to reject does not prove equality; sample size and variability may limit the evidence.

For a two-way table, a chi-square procedure can assess association between categorical variables or compare distributions across populations, depending on the design. Expected counts come from row total × column total / grand total under the null. Degrees of freedom are (rows−1)(columns−1).

The statistic adds (observed−expected)²expected over cells, so large departures increase it regardless of sign. Examine counts and the sampling design before using the reference distribution. The test identifies evidence of a pattern, but does not by itself quantify which category is responsible or establish a causal mechanism.

An association test can be followed by contextual comparison of conditional proportions to describe the pattern. This unit follows the revised course scope; do not substitute a one-variable goodness-of-fit procedure for a two-way-table question. A correct numerical result still needs a conclusion about the stated populations and variables.