Selecting the Correct Tail

Why the Tail Depends on the Question

A one-proportion test asks whether a population proportion \(p\) is consistent with a benchmark \(p_0\), or whether the data provide evidence for a specified alternative. The tail identifies which sample results would count as evidence for that alternative. A claim that a proportion is higher than a benchmark points in one direction; a claim that it is lower points in the other. A claim that it differs, without specifying a direction, includes departures in either direction.

The tail is not chosen by looking at which way the observed sample proportion happens to differ from the benchmark. Instead, it follows from the research question and is set before examining the data. Choosing the direction afterward changes the question in response to the observed result and can make the evidence appear stronger than it is.

Translate the Claim Before Looking at the Data

Begin by defining success in the context of the question. Then identify the benchmark and decide what departure from that benchmark the research question is asking about. The wording “greater than,” “higher,” “increased,” or “exceeds” indicates an upper-tail alternative. The wording “less than,” “lower,” “decreased,” or “falls below” indicates a lower-tail alternative. Wording such as “different,” “changed,” or “not equal,” without a stated direction, indicates a two-sided alternative.

The null hypothesis uses equality to the benchmark in this standard test setup. The alternative carries the directional or nondirectional research claim. This does not mean the null must be the most interesting claim; it provides the reference value against which the alternative is assessed. If a practical requirement is that a rate be at least \(p_0\), for example, the concern might be that the true rate has fallen below that standard. The test then uses \(H_0:p=p_0\) and \(H_a:p<p_0\), examining departures below the boundary value.

These choices concern the alternative, not the direction of an observed difference. For instance, \(p>p_0\) remains the correct alternative for an improvement question even if a particular sample happens to produce \(\hat{p}<p_0\). That sample may provide little evidence for the proposed improvement, but it does not change the question that was asked.

Worked Examples

Why These Alternatives Correspond to These Tails

The choice can be justified by considering the one-proportion \(z\) statistic under the null. For a sample of size \(n\), it compares the observed proportion with the benchmark and measures the difference in null standard-error units:

The denominator is positive when \(0<p_0<1\), so the sign and ordering of \(z\) follow the difference \(\hat{p}-p_0\). A large positive value of \(z\) corresponds to a sample proportion far above \(p_0\); a large negative value corresponds to one far below \(p_0\). Thus, an alternative predicting \(p>p_0\) calls for attention to unusually large positive statistics, while \(p<p_0\) calls for unusually negative statistics. If the alternative is \(p\ne p_0\), departures on either side are relevant.

This direction-based reasoning assumes the usual one-proportion \(z\)-test setup: observations represent independent binary outcomes from the population of interest, and the null expected numbers of successes and failures are large enough for the normal approximation to be reasonable. The statistic and its null standard error are not needed to choose the alternative; they explain why its direction determines which tail is relevant. The test procedure and its assumptions must be appropriate for the study as well as for the wording of the claim.

Common Mistakes and Why Tail Choice Matters

• Choosing the tail from the sample. Wrong reasoning: “The sample proportion is above the benchmark, so I will use an upper-tail test.” Correction: choose the alternative from the research question before examining the sample. If the question was whether the rate would decrease, the alternative remains lower-tail even when the observed proportion is higher.
• Treating “different” as “lower.” Wrong reasoning: “The sample value is lower, so a claim that the rate differs is a lower-tail test.” Correction: “differs” includes values above or below the benchmark, so use \(p\ne p_0\) and a two-sided test.
• Switching the outcome without switching the direction. Wrong reasoning: “The on-time proportion falling below 92% means the late proportion also falls below 8%.” Correction: on-time and late are complementary outcomes. A decline below 92% on time corresponds to an increase above 8% late.
• Assuming a directional test is automatically more appropriate. A one-sided alternative should be used only when the research question genuinely specifies a direction in advance. It should not be chosen merely because one direction better matches the observed sample.

Tail choice matters because it defines what counts as evidence for the alternative. An upper-tail test, a lower-tail test, and a two-sided test answer different questions, even when they use the same benchmark and sample. Choosing the wrong one can lead to a conclusion about a claim that was never actually posed. The tail should therefore be recorded alongside the hypotheses before computing or interpreting a test result.