Common Misreadings of Proportion Intervals

Why Careful Interpretation Matters

A confidence interval for a population proportion is easy to misread. A reported range can sound like a statement about the chance that the true proportion lies there, about the outcomes of particular individuals, or about whether the sample was representative. Those are different questions. An interval can quantify uncertainty from random sampling under suitable conditions, but it does not automatically answer all three.

This tutorial focuses on identifying these distinctions. We will use the one-proportion \(z\) interval as a concrete example and keep its assumptions visible: its usual interpretation relies on an appropriate random sampling process, independent observations, and a sufficiently accurate normal approximation.

Definition: A population proportion \(p\) is the fraction of individuals in a defined population that have a specified binary characteristic. It is a fixed, usually unknown quantity for that population. A sample proportion \(\hat{p}=x/n\) is the fraction with the characteristic among \(n\) sampled individuals, where \(x\) is the number with it. A confidence interval is a data-dependent range produced by a procedure intended to cover \(p\) at a stated long-run rate, under its assumptions.

What the Interval Procedure Does

For a sample with \(x\) successes among \(n\) observations, the familiar one-proportion Wald interval is

\hat{p}\ \pm\ z^*\sqrt{\frac{\hat{p}(1-\hat{p})}{n}}.

Here \(z^*\) is the standard normal critical value for the chosen confidence level; for a 95% interval, \(z^*\) is approximately 1.96. The square-root term estimates the standard error of \(\hat{p}\). The interval is an approximation, not a guarantee: its customary coverage interpretation depends on random sampling and independence, and on the sampling distribution being adequately approximated by a normal distribution. A common check for this Wald procedure is at least 10 observed successes and 10 observed failures.

The crucial timing distinction is that before data are collected, the endpoints of the interval are not yet known and will vary from sample to sample. After one sample has been observed and the interval calculated, its endpoints are fixed, and \(p\) is fixed. The long-run probability refers to the procedure’s coverage across repeated samples, not to a random parameter moving in and out of one already calculated interval.

Worked Example: Interpreting One Calculated Interval

Suppose a hypothetical simple random sample of 120 adults is taken from a much larger population. In the sample, 72 report meeting a stated activity target. Find a 95% Wald interval and assess the claim, “There is a 95% chance that the population proportion is in this interval.”

The sample proportion is \(\hat{p}=72/120=0.6000\). There are 72 successes and 48 failures, so both observed counts exceed 10. The sample is stipulated to be random and from a population much larger than the sample, making independence a reasonable approximation under the usual sampling assumptions.

The estimated standard error is

SE=\sqrt{\frac{0.6000(1-0.6000)}{120}} =\sqrt{0.002000} \approx 0.044721.

Using \(z^*=1.96\), the interval is

0.6000\pm1.96(0.044721) =0.6000\pm0.087653 \approx(0.5123,\ 0.6877).

The calculation is consistent with the sample: the estimated population proportion is 0.6000, or 60.00%, and the interval runs from about 51.23% to 68.77%. The quoted claim is not the correct frequentist interpretation. After the data are observed, \(p\) is fixed; the interval either contains it or does not. The 95% refers to the approximate long-run coverage of the interval procedure across repeated appropriate random samples, not a 95% probability assigned to \(p\) for this particular interval.

Misreading 1: Treating Confidence as Probability About a Fixed Parameter

The phrase “95% confident” can tempt a reader to say, “There is a 95% probability that \(p\) lies between the endpoints.” In the frequentist interpretation used here, that is not what a confidence level means. The population proportion is fixed; the random element is the sample, and therefore the interval produced from it.

A useful way to justify the standard interpretation is to consider the sampling distribution before the data are collected. When the approximation conditions hold, the standardized sample proportion is approximately standard normal. For a 95% interval, this gives

P\left(-1.96\leq\frac{\hat{p}-p}{SE}\leq1.96\right)\approx0.95.

Rearranging the inequality shows that the procedure covers \(p\) when

\hat{p}-1.96SE\leq p\leq\hat{p}+1.96SE.

Thus, before sampling, about 95% of intervals generated by this procedure would cover the fixed \(p\), if its assumptions and approximation are adequate. Once a particular interval has been computed, the repeated-sampling probability does not turn into a probability that the fixed \(p\) is in that particular interval. “We are 95% confident” is standard shorthand for this procedure-based interpretation.

The word “approximately” matters. The Wald interval does not have exactly 95% coverage in every setting. If its conditions fail or its normal approximation is poor, the nominal confidence level may not describe its actual long-run performance well.

Worked Example: What “95% Confidence” Refers To

Imagine repeatedly drawing simple random samples of 120 adults from the same large population, applying the same 95% Wald interval each time, and checking whether each interval contains the population proportion \(p\). Assume the sampling conditions and approximation are adequate.

Before any sample is selected, each sample could produce a different \(\hat{p}\) and different endpoints. The 95% confidence level describes the approximate fraction of intervals that would contain \(p\) in this repetition process. For instance, if the procedure were applied to 1,000 such samples, roughly 950 intervals would be expected to cover \(p\); the exact number would vary.

This does not mean every interval contains \(p\), nor does it say that 95% of the adults’ outcomes are inside an interval. For a single computed interval, the population proportion either lies between its endpoints or it does not. The repeated procedure, not a probability that changes the fixed parameter, is what the confidence level describes.

Misreading 2: Treating the Interval as a Range for Individuals

A confidence interval for \(p\) concerns a population fraction, not a range of individual measurements or outcomes. For a binary characteristic, an individual either has it or does not. The endpoints of an interval such as \((0.5730,0.8270)\) are possible values for the population proportion, expressed as fractions; they are not possible fractional outcomes for a person.

The interval also does not say that a specified percentage of individuals will have the characteristic. It gives an estimated range for the population fraction, with uncertainty from sampling represented through a procedure-based coverage level. Nor does it identify which individuals have the characteristic.

Worked Example: A Population Fraction Is Not an Individual Outcome

In a hypothetical random sample of 50 eligible patients, 35 report a particular symptom. Suppose the patients are sampled independently from a much larger target population. Find the 95% Wald interval and evaluate the statement, “The interval says that an individual patient’s probability of having the symptom is between 57% and 83%.”

Here \(\hat{p}=35/50=0.7000\), with 35 observed successes and 15 failures. Both counts are at least 10. Under the stated random-sampling and independence assumptions, the estimated standard error is

SE=\sqrt{\frac{0.7000(1-0.7000)}{50}} =\sqrt{0.004200} \approx0.064807.

The interval is

0.7000\pm1.96(0.064807) =0.7000\pm0.12702 \approx(0.5730,\ 0.8270).

The interval estimates the population proportion of eligible patients who have the symptom: about 57.30% to 82.70% using this procedure and its assumptions. The quoted statement confuses a population proportion with a range of individual outcomes. Each patient's recorded symptom status is binary. The interval does not say that an individual has a fractional outcome, nor does it give an individual-specific prediction. It is about the population fraction, not any one patient's status.

Misreading 3: Assuming Confidence Guarantees Representativeness

A confidence level does not certify that the sample represents its target population. The interval's sampling-based justification depends on how observations were obtained. A nonrandom sample may systematically over-represent some groups and under-represent others. Increasing the confidence level cannot, by itself, correct that selection bias, nonresponse, or a mismatch between the sample and the population named in the claim.

Random sampling and representativeness are related but not identical ideas. A well-designed probability sample supports inference to its defined target population, subject to implementation and nonresponse issues. A large sample or a high confidence level alone does not establish that the sampling design was suitable. Always identify the population and assess the data collection process before interpreting an interval.

Worked Example: A Narrow Calculation Cannot Repair Selection Bias

A hypothetical city news website asks visitors to click a poll about support for a proposed transit change. Of 300 voluntary respondents, 180 support it. The website headline says, “A 95% confidence interval proves that 54% to 66% of city residents support the change.” Assess this claim.

The poll proportion is \(\hat{p}=180/300=0.6000\). If one mechanically applies the Wald formula, the estimated standard error is

SE=\sqrt{\frac{0.6000(1-0.6000)}{300}} =\sqrt{0.0008000} \approx0.028284.

The resulting nominal interval is

0.6000\pm1.96(0.028284) =0.6000\pm0.05544 \approx(0.5446,\ 0.6554).

This arithmetic does not justify the headline’s claim about all city residents. The poll is voluntary, not a random sample of residents, so people who choose to visit the site and respond may differ from those who do not. The usual random-sampling basis for the interval's population coverage is absent. The calculation describes the poll's sample proportion and a nominal formula-based range; it does not establish that the interval covers the citywide proportion at a 95% rate. Confidence level is not a measure of sample representativeness.

A Practical Interpretation Check

When someone states a conclusion from a proportion interval, ask what quantity the interval estimates, how the data were collected, and what the confidence level refers to. These checks expose the most common errors without requiring a new calculation.

Claim	Why it is misleading	Better interpretation
“There is a 95% probability that the fixed \(p\) is in this interval.”	It assigns probability to a fixed parameter after the interval is observed.	The procedure has approximately 95% coverage over repeated suitable samples.
“The interval gives the range of outcomes for individual people.”	The interval concerns the population proportion, not individual binary outcomes.	It estimates the fraction of the defined population with the characteristic.
“A 95% interval means the sample is 95% representative.”	Confidence level does not evaluate selection or response bias.	Judge the sampling design and target population separately from the interval calculation.

A useful reporting sentence names the population and the quantity, then gives the interval and its procedure-based interpretation. For example: “The 95% Wald confidence interval estimates the proportion of eligible adults in this target population who meet the activity target as approximately 51.23% to 68.77%. Under the random-sampling, independence, and approximation conditions, the procedure would cover the population proportion in about 95% of repeated samples.”

Common Mistakes and Corrections

Mistake: Saying the probability that \(p\) is in a computed interval is 95%. Correction: Explain the approximate long-run coverage of the method.
Mistake: Reading endpoints as outcomes or probabilities for particular individuals. Correction: State that the parameter is a proportion in the defined population.
Mistake: Treating a high confidence level as evidence of a representative sample. Correction: Examine random selection, nonresponse, independence, and the intended target population.
Mistake: Treating the nominal level as a guarantee even when assumptions are implausible. Correction: Check the conditions and describe the interval's coverage as approximate when using an approximation procedure.

Key takeaway: A proportion confidence interval estimates a fixed population proportion through a procedure whose long-run coverage is controlled under stated assumptions. It is not a probability statement about the fixed parameter after observing one interval, a range for individual outcomes, or proof that a sample represents its target population.

Check Your Understanding

For each item, identify what the interval does or does not support.

A report says, “There is a 90% chance the true population proportion is in our computed interval.” What is the precise problem with this wording?
A binary-outcome interval has endpoints 0.42 and 0.58. Explain why these endpoints do not describe the possible outcomes for one individual.
A 95% interval is calculated from a large voluntary online poll. Why does its confidence level not establish coverage for the full target population?
What varies across repeated samples in the frequentist interpretation: the population proportion, the interval, or both?
State two conditions that support the usual interpretation of a one-proportion Wald interval.