Loading
Understanding how sample proportions vary from sample to sample enables powerful inference about population parameters.
The question of how reliably a sample reflects a larger population has occupied mathematicians and scientists for centuries. When a pollster surveys 1,000 voters and reports that 54% support a candidate, the natural follow-up is: how close is that 54% to the true proportion in the entire electorate? Answering this question requires understanding the sampling distribution of the sample proportion—the probability distribution that describes how p̂ (the sample proportion) behaves across all possible random samples of a given size. The development of this idea spans several centuries and reflects a deep interplay between probability theory, mathematical analysis, and practical statistical reasoning.
These historical developments converge on a single powerful idea: even though individual samples are imperfect reflections of their populations, the pattern of variation across many samples is predictable. This predictability is precisely what allows us to quantify uncertainty through margins of error, confidence intervals, and hypothesis tests. The central question this lesson addresses is: if we know the true population proportion p, what can we say about the shape, center, and spread of p̂ across all possible samples of size n?
Before examining the mathematical machinery, it is essential to establish the foundational concepts that govern how sample proportions behave. Each principle below builds upon the preceding one: we start with the distinction between a parameter and a statistic, move to the concept of a sampling distribution, and then state the conditions under which a remarkably clean normal model applies.
The following diagram illustrates the conceptual process of constructing a sampling distribution. Imagine a large population with true proportion p = 0.40. We repeatedly draw random samples of size n = 50, compute p̂ for each sample, and then plot the distribution of all those p̂ values. The resulting histogram forms the sampling distribution, which—when the Large Counts Condition is satisfied—takes on the characteristic bell shape of a normal distribution centered at p.
Notice several important features in the diagram. First, each individual sample produces a different p̂ value—0.38, 0.44, 0.36, and so on—reflecting natural sampling variability. Second, when we aggregate thousands of these p̂ values, the histogram forms a smooth, symmetric bell curve centered on the true population proportion p = 0.40. Third, the spread of this distribution—governed by σp̂ = √(0.40 × 0.60 / 50) ≈ 0.069—quantifies how much p̂ typically deviates from p in a single sample. About 68% of all sample proportions fall within one standard deviation of the center, and about 95% fall within two standard deviations.
The sampling distribution of p̂ arises from the binomial distribution. If each observation in a sample of size n independently has probability p of being a "success," then the count X of successes follows a Binomial(n, p) distribution. Since p̂ = X/n, the properties of p̂ follow directly from those of X. The three key formulas below describe the center, spread, and approximate shape of the distribution of p̂.
The shape and spread of the sampling distribution of p̂ depend on two quantities: the sample size n and the population proportion p. Understanding their effects is critical both for designing studies and for reasoning about results on the AP exam. The diagram below shows how the sampling distribution changes as n increases, holding p constant at 0.30.
The visual makes a crucial relationship concrete: since σp̂ = √(p(1 − p)/n), the standard deviation is inversely proportional to the square root of n. This means quadrupling the sample size cuts the standard deviation in half. Going from n = 25 to n = 100 (a factor of 4) reduces σp̂ from 0.092 to 0.046—exactly half. This square-root relationship has practical implications for survey design: achieving twice the precision requires four times the sample size, which means four times the cost.
| Sample Size (n) | σ_p̂ (when p = 0.30) | Large Counts Met? | Approx. 95% Range for p̂ |
|---|---|---|---|
| 10 | 0.145 | No (np = 3) | Normal model not appropriate |
| 50 | 0.065 | Yes (np = 15, n(1−p) = 35) | (0.17, 0.43) |
| 200 | 0.032 | Yes (np = 60, n(1−p) = 140) | (0.24, 0.36) |
| 1000 | 0.014 | Yes (np = 300, n(1−p) = 700) | (0.27, 0.33) |
The value of p also affects the standard deviation and the validity of the normal approximation. The product p(1 − p) is maximized when p = 0.50 and decreases symmetrically as p approaches 0 or 1. When p is extreme (near 0 or 1), the sampling distribution is more skewed and requires a larger n to satisfy the Large Counts Condition. For instance, with p = 0.05, you would need n ≥ 200 to reach np = 10, whereas with p = 0.50, even n = 20 suffices.
A large university reports that 65% of its students graduate in four years. A researcher takes a simple random sample of 150 students. What is the probability that the sample proportion of four-year graduates exceeds 0.70?
Even students with a solid grasp of the formulas frequently lose points on the AP exam due to subtle but important errors. The table below catalogues the most common mistakes and their corrections, drawing on patterns observed in released AP scoring guidelines.
| Common Error | Why It's Wrong | Correct Approach |
|---|---|---|
| Using p̂ (from sample data) in the standard deviation formula when p is known | The standard deviation σ_p̂ requires the true population proportion p. Substituting p̂ yields the standard error, which is appropriate for confidence intervals but not when the population parameter is known (e.g., hypothesis testing). | Use σ_p̂ = √(p(1−p)/n) when p is given. Use SE = √(p̂(1−p̂)/n) only when estimating from data. |
| Confusing the distribution of X with the distribution of p̂ | X ~ Binomial(n, p) counts successes; p̂ = X/n is a proportion. Their means and standard deviations differ by a factor of n. | For p̂: mean = p, SD = √(p(1−p)/n). For X: mean = np, SD = √(np(1−p)). Be clear about which you are modeling. |
| Forgetting to check or state the conditions | Applying the normal model without verifying the Random, 10%, and Large Counts conditions loses points on FRQs. The rubric typically dedicates 1–2 points to conditions alone. | Always state all three conditions with context-specific verification: name the condition, show the numerical check, and state whether it is satisfied. |
| Claiming the sampling distribution is always normal | When np or n(1−p) is less than 10, the distribution of p̂ is noticeably skewed and the normal model provides poor approximations, especially in the tails. | State that the distribution is approximately normal only when the Large Counts Condition is met. For small samples or extreme p, use the exact binomial distribution. |
| Believing larger samples change the population distribution | Increasing n does not alter the population. It makes the sampling distribution of the statistic narrower and more normal, but the underlying population remains unchanged. | Distinguish clearly between the population distribution (fixed) and the sampling distribution of p̂ (depends on n). Larger n → narrower sampling distribution, not a different population. |
The sampling distribution of p̂ is not merely an abstract concept—it is the engine that powers all inference about proportions on the AP Statistics exam. Both confidence intervals and hypothesis tests for a single proportion rely directly on the normal model for p̂ that we have developed. Understanding how the sampling distribution connects to these two inferential procedures is essential for performing well in Units 6 and 7 of the AP curriculum.
| Feature | Confidence Interval for p | Hypothesis Test for p |
|---|---|---|
| Goal | Estimate plausible values for the unknown p | Assess evidence against a claimed value p₀ |
| Formula | p̂ ± z* × √(p̂(1−p̂)/n) | z = (p̂ − p₀) / √(p₀(1−p₀)/n) |
| Which p in the SD? | p̂ (standard error), because p is unknown | p₀ (standard deviation), because H₀ specifies p |
| Large Counts Check | np̂ ≥ 10 and n(1−p̂) ≥ 10 | np₀ ≥ 10 and n(1−p₀) ≥ 10 |
| Sampling dist. used | Distribution of p̂ centered at the true (unknown) p | Distribution of p̂ centered at p₀ (assuming H₀ is true) |
Looking forward, the same conceptual framework extends to two-proportion inference (comparing p₁ − p₂), where the sampling distribution of p̂₁ − p̂₂ is approximately normal under analogous conditions. Beyond proportions, the idea of a sampling distribution underlies inference for means (using the t-distribution), regression slopes, and chi-square tests. Mastering the single-proportion case thoroughly—especially the logic of checking conditions and computing standard deviations—provides a template that transfers directly to every other inferential procedure on the AP exam.
The sampling distribution of p̂ describes the probability distribution of the sample proportion across all possible random samples of size n from a population with true proportion p. Its center is μ_p̂ = p (making p̂ an unbiased estimator), and its spread is σ_p̂ = √(p(1 − p)/n), which decreases as sample size increases. The shape is approximately normal when the Large Counts Condition is satisfied (np ≥ 10 and n(1 − p) ≥ 10), allowing us to use z-scores to compute probabilities involving p̂.
Before applying the normal model on the AP exam, always verify three conditions: the sample must be random, the 10% Condition must hold (n < 10% of the population) to ensure approximate independence, and the Large Counts Condition must be met. This framework directly underpins confidence intervals and hypothesis tests for proportions—master the sampling distribution, and you hold the key to all proportion-based inference.
Keep learning with more lessons from the same subject.