Determining whether a sampling distribution of a sample proportion ($\hat{p}$) is approximately normal is a foundational step in inferential statistics. Before constructing confidence intervals or running hypothesis tests for a single proportion, you must verify that the shape of the distribution meets the normality assumption. This verification relies primarily on the Large Counts Condition (often called the Success-Failure Condition), which ensures the binomial distribution underlying the proportion can be safely approximated by a continuous normal curve. Without this check, the resulting z-scores, p-values, and margin of error calculations may be inaccurate, leading to flawed conclusions.
Understanding the Sampling Distribution of a Proportion
When you take a random sample of size $n$ from a population with a true proportion of success $p$, the sample proportion $\hat{p}$ becomes a random variable. If you repeated this sampling process thousands of times, the distribution of those $\hat{p}$ values forms the sampling distribution Not complicated — just consistent..
Some disagree here. Fair enough.
The Central Limit Theorem (CLT) for proportions tells us that as the sample size increases, the shape of this sampling distribution approaches a normal distribution, regardless of the shape of the population distribution (which is binary: success/failure). Even so, "large enough" is not a fixed number; it depends on the value of the population proportion $p$ Practical, not theoretical..
People argue about this. Here's where I land on it.
The sampling distribution of $\hat{p}$ has a mean $\mu_{\hat{p}} = p$ and a standard deviation (standard error) $\sigma_{\hat{p}} = \sqrt{\frac{p(1-p)}{n}}$. So for the normal approximation to be valid, the distribution must be symmetric enough and have negligible tail probability beyond the logical boundaries of 0 and 1. Since a proportion cannot be less than 0 or greater than 1, a normal curve—which extends to infinity in both directions—is a poor fit if the mean is too close to either boundary.
The Large Counts Condition: The Golden Rule
The standard rule of thumb for verifying normality is the Large Counts Condition. It requires that both the expected number of successes and the expected number of failures in the sample are at least 10.
Mathematically, you check two inequalities:
- $n \times p \geq 10$ (Expected successes)
- $n \times (1 - p) \geq 10$ (Expected failures)
Crucial Distinction: Context Matters The value used for $p$ changes depending on the statistical procedure you are performing:
- Hypothesis Testing: You assume the null hypothesis ($H_0$) is true. Because of this, you use the hypothesized proportion $p_0$ from the null hypothesis to check the condition. You calculate $n \times p_0$ and $n \times (1 - p_0)$.
- Confidence Intervals: You do not have a hypothesized value. Instead, you use the sample proportion $\hat{p}$ as the best estimate of $p$. You calculate $n \times \hat{p}$ and $n \times (1 - \hat{p})$. Note that these are simply the observed counts of successes and failures in your sample.
Why the Threshold is 10
You might wonder why the threshold is 10 and not 5 or 30. The binomial distribution is discrete and skewed when $p$ is far from 0.5.
- If $p = 0.5$, the distribution is symmetric even for small $n$ (e.Also, g. , $n=10$ gives expected counts of 5 and 5, which is often considered borderline acceptable). In practice, * If $p = 0. 1$, the distribution is heavily right-skewed. That's why you need a much larger $n$ (at least 100) to get 10 expected successes ($100 \times 0. 1 = 10$) to pull the left tail away from 0 and create symmetry. In real terms, * If $p = 0. 01$, you would need $n = 1000$ just to expect 10 successes.
The "10" threshold is a conservative convention ensuring the normal curve does not assign significant probability to impossible values (negative proportions or proportions > 1). Some textbooks use 5 as a lower bound for "approximately normal," but 10 is the standard for AP Statistics and most college-level introductory courses to ensure the continuity correction is negligible and the approximation is strong.
Step-by-Step Verification Process
Follow these steps every time you perform a one-proportion z-test or construct a one-proportion z-interval.
Step 1: Identify the Parameter and Sample Size
Define $p$ (population proportion) and $n$ (sample size). Ensure the sample is a Simple Random Sample (SRS) or from a randomized experiment. Randomization is a separate condition (Random Condition) but is a prerequisite for the normality check to matter And it works..
Step 2: Select the Correct $p$ Value
- For a Hypothesis Test: Locate $p_0$ in $H_0: p = p_0$.
- For a Confidence Interval: Calculate $\hat{p} = \frac{x}{n}$, where $x$ is the number of successes.
Step 3: Calculate Expected Counts
Compute the two products:
- Expected Successes = $n \times p$ (or $n \times p_0$ / $n \times \hat{p}$)
- Expected Failures = $n \times (1 - p)$ (or $n \times (1-p_0)$ / $n \times (1-\hat{p})$)
Step 4: Compare to Threshold
Check if both values are $\geq 10$.
Step 5: State Your Conclusion
Write a clear verification statement.
- Example: "Since $n \times p_0 = 50 \times 0.6 = 30 \geq 10$ and $n \times (1-p_0) = 50 \times 0.4 = 20 \geq 10$, the Large Counts Condition is met. The sampling distribution of $\hat{p}$ is approximately normal."
Worked Examples
Example 1: Hypothesis Test (Using $p_0$)
Scenario: A company claims 70% of customers are satisfied ($p_0 = 0.70$). You survey a random sample of 30 customers ($n=30$) to test this claim And that's really what it comes down to..
Check:
- Expected Successes: $n \times p_0 = 30 \times 0.70 = 21$
- Expected Failures: $n \times (1 - p_0) = 30 \times 0.30 = 9$
Verdict: Condition NOT met. The expected failures (9) is less than 10. The sampling distribution will be left-skewed. You should not use a one-proportion z-test. Instead, use an exact binomial test (calculating p-value directly from the binomial distribution).
Example 2: Confidence Interval (Using $\hat{p}$)
Scenario: In a random sample of 100 voters ($n=100$), 40 support a candidate ($\hat{p} = 0.40$). You want a 95% confidence interval.
Check:
- Observed Successes: $n \times \hat{p} = 100 \times 0.40 = 40$
- Observed Failures: $n \times (1 - \hat{p}) = 100 \times 0.60 = 60$
Verdict: **Condition MET
Verdict: Condition MET. Since $n \times \hat{p} = 100 \times 0.40 = 40 \geq 10$ and $n \times (1 - \hat{p}) = 100 \times 0.60 = 60 \geq 10$, the Large Counts Condition is satisfied. The sampling distribution of $\hat{p}$ is approximately normal, and we can confidently proceed with a one-proportion z-interval using the critical value $z^* = 1.96$ for a 95% confidence level Most people skip this — try not to. Worth knowing..
Example 3: Borderline Case
Scenario: A political pollster surveys 25 randomly selected residents ($n = 25$) of a small town to estimate the proportion who support a new policy. Of those surveyed, 3 support it ($\hat{p} = 0.12$) Worth knowing..
Check:
- Observed Successes: $n \times \hat{p} = 25 \times 0.12 = 3$
- Observed Failures: $n \times (1 - \hat{p}) = 25 \times 0.88 = 22$
Verdict: Condition NOT met. Although the expected failures are well above 10, the expected successes (3) fall far short. With so few successes, the sampling distribution of $\hat{p}$ will be heavily right-skewed, and the normal approximation will produce unreliable p-values and confidence intervals. In practice, the pollster should either increase the sample size substantially or rely on an exact binomial method to draw valid conclusions.
Common Pitfalls to Avoid
-
Using the wrong $p$ value. A frequent mistake is plugging in $\hat{p}$ when conducting a hypothesis test. Remember: for hypothesis tests, always use $p_0$ from the null hypothesis. For confidence intervals, use $\hat{p}$ since $p_0$ is unknown Simple, but easy to overlook..
-
Checking only one side. The condition requires both expected successes and expected failures to meet the threshold. A large number of successes alone does not guarantee normality if the failure count is too small, and vice versa.
-
Confusing the Large Counts Condition with the Random Condition. These are two separate requirements. The Random Condition ensures that the data collection method (SRS or randomized experiment) produces unbiased results, while the Large Counts Condition ensures that the shape of the sampling distribution is close enough to normal for the z-procedure to be valid. Both must be verified independently That's the whole idea..
-
Applying the condition after collecting data. In some contexts, particularly when planning a study, you can check the condition before data collection by estimating $p$ using a pilot study or a conservative guess of 0.50 (which maximizes $n \times p(1-p)$). This prospective check helps determine the minimum sample size needed.
How the Continuity Correction Fits In
When the Large Counts Condition is satisfied but the sample size is relatively small (say, between 10 and 30 expected counts), a continuity correction can improve the accuracy of the normal approximation. The correction involves adjusting the numerator of the z-statistic by $\pm \frac{1}{2n}$:
$z_{\text{corrected}} = \frac{|\hat{p} - p_0| - \frac{1}{2n}}{\sqrt{\frac{p_0(1-p_0)}{n}}}$
This adjustment accounts for the fact that the binomial distribution is discrete while the normal distribution is continuous. As $n$ grows larger, the correction becomes negligibly small—which is precisely why a threshold of 10 (rather than 5) is preferred in introductory courses: it ensures that the continuity correction has minimal impact and the approximation remains dependable across a wide range of scenarios Nothing fancy..
