AP Statistics: Hypothesis Testing for Proportions

Hypothesis testing is the statistical engine that drives decision-making from clinical trials to quality control. When you want to test a claim about a population proportion—like whether a new drug's success rate exceeds a placebo's, or if a new manufacturing process reduces the defect rate—you use the procedures in this guide. Mastering this topic transforms you from a passive data observer into an active statistical investigator, capable of using sample evidence to draw conclusions about the wider world.

The Foundation: Stating Hypotheses

Every hypothesis test begins with a clear, contradictory pair of statements about a population parameter. Here, that parameter is the population proportion, denoted by $p$.

• The null hypothesis ($H_0$) is a statement of "no effect," "no difference," or the status quo. It always contains an equality ($=$, $\leq$, or $\geq$). For a proportion, it is stated as $H_0: p = p_0$, where $p_0$ is the specific numerical value being tested.
• The alternative hypothesis ($H_a$ or $H_1$) is what you seek evidence for. It is a statement of change, difference, or effect.

You must correctly identify whether the test is one-sided or two-sided. A one-sided test (or one-tailed test) looks for evidence of a change in one specific direction ($p > p_0$ or $p < p_0$). A two-sided test (or two-tailed test) looks for evidence of a change in either direction ($p \neq p_0$). The research question dictates the choice. For example, testing if a new teaching method increases pass rates requires a one-sided test ($H_a: p > p_0$). Testing if the proportion of defective parts is different from the standard requires a two-sided test ($H_a: p \neq p_0$).

Conditions for a Valid Test

Before any calculations, you must verify that the mathematical model is appropriate. For a one-sample z-test for a proportion, three conditions must be met:

1. Random: The sample data must come from a well-designed random sample or randomized experiment. This is essential for generalizing to the population.
2. 10% Condition: The sample size $n$ must be no more than 10% of the population size when sampling without replacement. This ensures independence between sample observations.
3. Large Counts Condition: This checks the normality of the sampling distribution. You must expect at least 10 successes and 10 failures if the null hypothesis is true. That is, verify that both $np_0 \geq 10$ and $n(1-p_0) \geq 10$.

Failing to check these conditions is a critical error. If they are not met, the resulting p-value and conclusion may be invalid.

The Mechanics: Test Statistic and P-Value

With conditions satisfied, you calculate a test statistic, which measures how far your sample result is from the null hypothesis value, in units of standard error. For a proportion, this is the z-test statistic:

$$z = \frac{\hat{p} - p_0}{\sqrt{\frac{p_0(1-p_0)}{n}}}$$

Here, $\hat{p}$ is the sample proportion (your observed statistic), $p_0$ is the hypothesized population proportion from $H_0$, and $n$ is the sample size. The denominator is the standard error of $\hat{p}$ assuming the null hypothesis is true.

This z-score tells you how many standard errors your sample proportion lies from the null value. A large absolute value of $z$ provides evidence against $H_0$.

The p-value is the probability, computed assuming $H_0$ is true, of obtaining a sample statistic at least as extreme as the one you actually observed, in the direction specified by $H_a$.

• For $H_a: p > p_0$, the p-value is $P(Z \geq z)$.
• For $H_a: p < p_0$, the p-value is $P(Z \leq z)$.
• For $H_a: p \neq p_0$, the p-value is $2 \times P(Z \geq |z|)$.

You find this probability using the standard Normal distribution (Table A, or technology). A small p-value means your sample result would be very unlikely to occur if the null hypothesis were true, thus casting doubt on $H_0$.

Making a Decision and Stating a Conclusion

You compare the p-value to a predetermined significance level, $\alpha$ (commonly 0.05).

• If p-value $\leq \alpha$, you reject the null hypothesis ($H_0$).
• If p-value $> \alpha$, you fail to reject the null hypothesis ($H_0$). You never "accept" $H_0$; you simply lack sufficient evidence against it.

Your conclusion must always be stated in context, using non-technical language related to the original research question. It should seamlessly integrate the decision, the alternative hypothesis, and the context.

Example: A campaign manager claims 60% of voters support her candidate. A poll of 400 random voters finds 220 supporters ($\hat{p}=0.55$). Test the claim at the $\alpha=0.05$ level.

1. $H_0: p = 0.60$, $H_a: p \neq 0.60$ (A "claim" test is typically two-sided unless specified).
2. Conditions: Random sample, 400 < 10% of all voters, $np_0=400(0.6)=240$ and $n(1-p_0)=400(0.4)=160$ are both $\geq 10$. ✔
3. $z = \frac{0.55 - 0.60}{\sqrt{\frac{0.60(0.40)}{400}}} = \frac{-0.05}{0.0245} \approx -2.04$
4. P-value (two-tailed): $2 \times P(Z \leq -2.04) = 2(0.0207) = 0.0414$
5. Decision: Since $0.0414 < 0.05$, we reject $H_0$.
6. Conclusion: There is statistically significant evidence at the 0.05 level that the true proportion of voters who support the candidate is different from 0.60.

Connection to Confidence Intervals

There is a beautiful duality between a two-sided hypothesis test at significance level $\alpha$ and a confidence interval with confidence level $C = 1 - \alpha$. Specifically:

• If the null value $p_0$ is contained within a $C = 1 - \alpha$ confidence interval for $p$, then you will fail to reject $H_0: p = p_0$ at level $\alpha$.
• If the null value $p_0$ is not contained within the interval, you will reject $H_0$.

In our voter example, a 95% CI for $p$ is $0.55 \pm 1.96\sqrt{(0.55)(0.45)/400} \approx (0.501, 0.599)$. The null value of 0.60 is not inside this interval (0.599 is just below it), which is consistent with our decision to reject $H_0$. The interval provides the additional information of which plausible values for $p$ are supported by the data.

Common Pitfalls

1. Misstating the Alternative Hypothesis: Let the research question, not the sample data, dictate $H_a$. If you want to know if a new method is better, it's one-sided ($>$). If you want to know if it's different, it's two-sided ($\neq$). Seeing that $\hat{p} < p_0$ in your sample does not mean you should use $<$ for $H_a$.
2. Using the Wrong Standard Error: In the formula for the z-test statistic, the standard error in the denominator uses the null hypothesis proportion $p_0$, not the sample proportion $\hat{p}$. Using $\hat{p}$ here is a common calculation error. (Note: $\hat{p}$ is used when constructing a confidence interval, but not in the test statistic under $H_0$).
3. Forgetting the "Double" for Two-Tailed P-values: When performing a two-sided test, the p-value is the probability in both tails. A frequent mistake is to report only the area in one tail, effectively cutting the p-value in half and making results appear more significant than they are.
4. Conclusion Omissions: A conclusion that states only "reject $H_0$" or gives a p-value without context is incomplete. You must state that there is/is not significant evidence for the alternative hypothesis and relate it back to the topic (e.g., "for the candidate's support level," "for the defect rate").

Summary

• Hypothesis testing for proportions is a formal process to evaluate a claim about a population proportion $p$ using sample data. It begins by stating a null hypothesis ($H_0: p = p_0$) and an alternative hypothesis ($H_a: p >, <, or \neq p_0$).
• The validity of the test depends on three conditions: Random sampling, the 10% Condition, and the Large Counts Condition ($np_0 \geq 10$ and $n(1-p_0) \geq 10$).
• The z-test statistic $z = \frac{\hat{p} - p_0}{\sqrt{p_0(1-p_0)/n}}$ measures the discrepancy in standard errors. The p-value quantifies the strength of the evidence against $H_0$ based on this statistic.
• A decision is made by comparing the p-value to $\alpha$. The conclusion must be stated clearly in the context of the original problem.
• A two-sided hypothesis test at significance level $\alpha$ will reach the same conclusion as checking whether the null value $p_0$ falls inside a $C = 1 - \alpha$ confidence interval for $p$.