Writing Hypotheses for One Proportion

From a Claim to a Pair of Hypotheses

A statement such as “at least 90% of orders arrive on time” sounds clear, but a statistical test needs a precise parameter and a comparison value. It also needs to say what kind of departure from that value would count as evidence against the claim: a rate that is lower, higher, or simply different. Writing the hypotheses carefully makes that question explicit before sample results are used.

This tutorial concerns a single population proportion. The goal is not to calculate a test statistic or p-value, but to express the claim in a form that defines what a test would investigate. The central convention is that the null hypothesis uses equality at a stated reference value. The alternative specifies the direction—or the two-sided difference—that the investigation seeks evidence for.

Before writing either hypothesis, specify three things: the population being described, the outcome that counts as a success, and the reference proportion. For example, “on time” needs an operational definition, such as delivery by the date promised. Without a population and outcome, a symbol such as \(p\) has no context; without a reference value, there is no precise claim to test.

The Null and the Alternative

The null hypothesis provides the reference point. For a benchmark \(p_0\), its standard point form is \(H_0:p=p_0\). The alternative says what departure from that benchmark would address the research question. There are three common forms:

The number \(p_0\) is the reference value, not an estimate calculated from the sample. The sample proportion \(\hat{p}\) describes the observed sample; it does not replace \(p\) in the hypotheses. For instance, if a question compares an advertised rate of 0.60 with the population rate, the hypotheses refer to \(p=0.60\), not to whatever value \(\hat{p}\) happens to take in a particular sample.

The choice among \(>\), \(<\), and \(\ne\) comes from the question being asked, not from which direction the sample result happens to point. If the goal is to detect a lower rate, a sample rate above the benchmark does not justify switching to an upper-direction alternative after looking at the data. That would change the question in response to the result.

Match the Alternative to the Claim

Translate the claim into an explicit comparison with the benchmark. “Higher,” “exceeds,” and “more than” point to \(H_a:p>p_0\). “Lower,” “falls below,” and “less than” point to \(H_a:p<p_0\). “Different,” “has changed,” and “not equal” point to \(H_a:p\ne p_0\). The null is written with equality at \(p_0\) in each of these standard test formulations.

Some practical claims are composite rather than exact. “At least 90%” includes every value from 0.90 upward, not only 0.90. If the investigation asks whether performance falls short of that minimum, the full status-quo claim is \(p\ge 0.90\), while the alternative is \(p<0.90\). In the usual one-proportion test setup, the null is evaluated at its boundary, so the stated test pair is conventionally written \(H_0:p=0.90\) versus \(H_a:p<0.90\). Equality is the boundary value that must be shown in the null for this setup; it is not an alternative hypothesis.

Why Equality Belongs in the Null

The equality convention gives the test a precise reference condition. For a directional question, the null side includes the benchmark, while the alternative excludes it and points in the direction of interest. In a one-sided problem, it can be useful to keep the broader composite null in mind: for a lower-direction alternative, the full null region is \(p\ge p_0\); for an upper-direction alternative, it is \(p\le p_0\). The conventional one-proportion setup states the boundary equality \(p=p_0\) as the null value used for the test.

This distinction matters for interpreting the question. Consider the lower-direction case. The possible parameter values are divided by the benchmark: values below 0.90 support the concern being investigated, while values at or above 0.90 are on the status-quo side. The equality \(p=0.90\) sits at the boundary between these regions. It is therefore not part of the alternative \(p<0.90\).

This is a justification for keeping the benchmark equality in the null formulation and for choosing the alternative from the claim. It does not determine whether the evidence will be strong; that requires the observed data and the later steps of a test.

Common Mistakes

• Putting the equality in the alternative. Writing \(H_a:p=0.12\) is incorrect when 0.12 is the claimed benchmark. To ask whether the population rate differs from 12%, write \(H_0:p=0.12\) and \(H_a:p\ne0.12\).
• Using \(\hat{p}\) in place of \(p\). A sample proportion is an observed statistic. Hypotheses describe the population parameter, so write \(p\) rather than \(\hat{p}\).
• Choosing a direction from the sample result. If the question asks whether a rate is lower, the alternative remains \(p<p_0\) even if the observed sample proportion is higher. The direction must come from the research question before assessing the data.
• Treating “at least” as “exactly.” A minimum claim such as \(p\ge0.90\) includes values greater than 0.90. In the conventional test pair for checking whether the rate is below the minimum, state the boundary equality \(H_0:p=0.90\) and the alternative \(H_a:p<0.90\); recognize that the broader status-quo region is \(p\ge0.90\).
• Leaving the success condition vague. “A successful outcome” must be defined in context. State exactly what counts as success and which population the proportion describes.