Loading
Use interval estimates to determine whether two population proportions differ meaningfully.
Comparing two groups on a categorical outcome—whether a new drug reduces infection rates more than a placebo, or whether voter turnout differs between two districts—is one of the most fundamental questions in applied statistics. For centuries, researchers relied on intuition and raw percentages to draw such comparisons, but the lack of a formal framework meant that sampling variability could easily masquerade as a genuine difference. The development of confidence intervals for the difference of two proportions gave scientists and policymakers a principled tool for quantifying uncertainty and justifying claims about population-level differences.
The central question this lesson addresses is: once you have computed a confidence interval for the difference p₁ − p₂, how do you use it to justify—or refuse to justify—a claim about whether the two population proportions are equal, different, or ordered in a specific direction? Mastering this reasoning is essential for the AP Statistics exam, where free-response questions routinely ask you to interpret confidence intervals in context and draw defensible conclusions.
Before diving into the mechanics of justification, it is essential to understand the foundational ideas that make confidence-interval-based reasoning valid. These principles connect the mathematical machinery of interval estimation to the inferential logic that lets us make statements about populations from samples.
The diagram below illustrates the three qualitatively different outcomes when a confidence interval for p₁ − p₂ is plotted on a number line. In each case, the position of the interval relative to zero determines the conclusion you can draw.
The critical reasoning pattern is straightforward: examine whether 0 falls inside or outside the interval. When the interval is entirely positive (lower bound > 0), every plausible value for p₁ − p₂ is positive, meaning p₁ is plausibly greater than p₂ at the given confidence level. Conversely, an interval that is entirely negative (upper bound < 0) supports the claim that p₁ < p₂. An interval that contains 0 means we cannot distinguish the two proportions at that confidence level—note that this does not prove they are equal, only that the data are insufficient to rule out equality.
The confidence interval for the difference of two population proportions uses the normal approximation to the sampling distribution of p̂₁ − p̂₂. Understanding each component of the formula clarifies where the interval's boundaries come from and why its width is governed by sample sizes and variability.
Once you have constructed the confidence interval, justifying a claim requires a structured argument. The AP Statistics exam expects you to connect the interval's contents to a conclusion stated in context. Below is a decision flowchart followed by a detailed breakdown of each pathway.
| Interval Location | Contains 0? | Justified Conclusion |
|---|---|---|
| Entirely positive (L > 0) | No | We are C% confident that p₁ − p₂ > 0, i.e., p₁ is greater than p₂. There is convincing evidence of a difference. |
| Entirely negative (U < 0) | No | We are C% confident that p₁ − p₂ < 0, i.e., p₂ is greater than p₁. There is convincing evidence of a difference. |
| Spans zero (L < 0 < U) | Yes | Because the interval contains 0, there is not convincing evidence that p₁ ≠ p₂ at the C% confidence level. |
A university researcher wants to know whether the proportion of students who prefer online classes differs between freshmen and seniors. In a random sample of 200 freshmen, 118 prefer online classes. In an independent random sample of 150 seniors, 72 prefer online classes. Construct a 95% confidence interval for p₁ − p₂ (freshmen minus seniors) and use it to justify whether there is convincing evidence of a difference.
| Strengths | Limitations |
|---|---|
| Provides a range of plausible values, giving richer information than a simple reject/fail-to-reject decision. | Relies on the large-sample normal approximation; coverage can be poor when sample sizes are small or proportions are near 0 or 1. |
| Directly addresses the magnitude and direction of the difference, supporting practical significance judgments. | A wide interval (small samples) may contain 0 even when a true difference exists, leading to an inconclusive result rather than evidence of no difference. |
| Duality with hypothesis testing: a CI that excludes 0 mirrors a significant test result at the corresponding α level. | Cannot establish causation from observational data, even when the interval excludes 0. Causation requires random assignment. |
| Easy to communicate to non-statisticians: 'We are 95% confident the difference is between ___ and ___.' | The confidence level applies to the procedure, not to a specific interval. Saying 'There is a 95% probability that p₁ − p₂ is in this interval' is incorrect. |
Confidence intervals and significance tests for the difference of two proportions are deeply intertwined but differ in important ways that AP Statistics students must understand. The table below highlights the relationship and the distinctions between the two approaches, along with a look toward more advanced methods you may encounter in college-level statistics courses.
| Feature | CI for p₁ − p₂ | z-Test for p₁ − p₂ |
|---|---|---|
| Purpose | Estimate the range of plausible values for p₁ − p₂ | Test whether p₁ − p₂ = 0 (or some other hypothesized value) |
| Standard Error | Unpooled: uses p̂₁ and p̂₂ separately | Pooled: uses p̂c = (x₁ + x₂)/(n₁ + n₂) under H₀ |
| Output | An interval (L, U) | A z-statistic and p-value |
| Conclusion format | "We are C% confident that p₁ − p₂ is between L and U." | "At significance level α, we reject/fail to reject H₀." |
| Duality | If 0 ∉ CI at level C, the test rejects H₀ at α = 1 − C | If p-value < α, then 0 is outside the corresponding CI |
In more advanced coursework, you will encounter methods that refine the Wald interval presented here. The Agresti–Caffo interval adds one success and one failure to each sample before computing the interval, which improves coverage probability for small samples. Bayesian methods replace the fixed-confidence interpretation with posterior probability distributions over p₁ − p₂, allowing direct probability statements about the parameter. Logistic regression extends the comparison of proportions to multiple groups and adjusts for confounding variables—an important step beyond the two-sample inference framework. However, the fundamental logic of examining whether zero is contained in the interval of plausible values remains the backbone of all these approaches.
A confidence interval for p₁ − p₂ is constructed as (p̂₁ − p̂₂) ± z* × SE, where the standard error uses unpooled sample proportions. Before constructing the interval, verify the Random, Independence (10% rule), and Large Counts conditions. The resulting interval provides a range of plausible values for the true difference in population proportions.
To justify a claim, examine the interval's position relative to zero. If the interval is entirely positive, there is convincing evidence that p₁ > p₂. If entirely negative, there is convincing evidence that p₁ < p₂. If the interval contains 0, the data do not provide convincing evidence of a difference—but this does not prove the proportions are equal. Always state conclusions in context, reference the confidence level, and describe both the direction and plausible magnitude of the difference.
Keep learning with more lessons from the same subject.