Must the alternative be two-sided?

No. A one-sided alternative is appropriate for a justified directional question, with the test chosen accordingly.

Does rejecting the null prove the alternative?

No. It supplies evidence against the null within the test's assumptions and error framework, without logical proof.

Can a study test practical equivalence?

Yes, but equivalence requires a specified margin and an appropriate hypothesis formulation. Merely failing to reject a no-difference null is not evidence of equivalence.

AI Glossary / Statistics and probability / Beginner

Alternative hypothesis

Definition

An alternative hypothesis is the statistical claim considered in contrast to a null hypothesis. It specifies values or distributions that the test is designed to detect, such as a difference between population means or a success probability above a benchmark. It is commonly written H1 or Ha.

What it means

The alternative translates a research question into a statement about a population or data-generating process. It does not merely say that the observed sample looks different. A sample difference can occur through random variation even when the null holds.

Alternatives may be directional, such as a mean being greater than a specified value, or two-sided, such as the mean being different in either direction. The alternative should reflect the intended question and be chosen before inspecting the result used to test it. It is not a claim that automatically becomes certain when a test is significant.

How it works

Define the parameter and the null and alternative hypotheses. For example, a two-sided comparison might use H0: mean difference = 0 and H1: mean difference is not 0. A superiority question might use H0: mean difference is at most 0 and H1: mean difference is greater than 0.

Choose a test whose assumptions and rejection region match those hypotheses. Calculate the test statistic and assess how extreme it is under the relevant null model. The alternative determines which outcomes count as evidence against the null, including the direction or directions considered.

Report the estimated effect and uncertainty as well as the test result. Power is the probability of rejection under a specified alternative value, so planning requires more than saying "some difference exists." A failure to reject does not establish the null, and rejection does not establish a practical benefit or eliminate concerns about study design and model assumptions.

Three practical examples

1. A product experiment asks whether a new process increases a population completion rate above a prespecified benchmark.

2. A quality analyst uses a two-sided alternative when either an increase or a decrease in mean output would indicate a change worth investigating.

3. An evaluation team specifies a minimum meaningful improvement, distinguishing a test of any positive effect from a test of exceeding that practical threshold.

Think of it this way

Think of the alternative as the destination the test is looking toward. A two-sided question watches for movement in either direction; a one-sided question watches a specified direction. Choosing the destination after seeing where the data moved changes the rules of the assessment.

Worked example

Consider an original benchmark test with 100 independent observations from a normal population whose standard deviation is known to be 10. The benchmark mean is 50 and the observed sample mean is 52. The standard error is 10 divided by the square root of 100, or 1, so the z statistic is (52 - 50)/1 = 2.

For a prespecified alternative that the mean exceeds 50, the one-sided p-value is approximately 0.0228. For a prespecified two-sided alternative that the mean differs from 50, the p-value is approximately 0.0455. Both use the same observations but ask different questions.

At a significance threshold of 0.03, the one-sided test rejects while the two-sided test does not. This is not permission to choose the one-sided test after seeing the result. Also, if the practical requirement were a mean above 53, a sample mean of 52 would not provide evidence of exceeding that threshold. The numerical calculation assumes the stated known-variance normal model.

Illustrative case study

Fictional case study: Hypothesia evaluates a change intended to reduce task-completion time. Before collecting results, its analysts define the population, outcome, comparison and direction of improvement. They also specify a practically meaningful reduction rather than relying only on a test against zero.

After analysis, they report the estimated reduction, its uncertainty and the prespecified hypothesis-test result. They avoid describing a nonsignificant result as proof of no effect and do not switch to a more favorable alternative after inspecting the data.

This fictional case illustrates alignment between a question and a test. It does not establish that statistical significance alone is sufficient to adopt a product change.

Common misconceptions and mistakes

The alternative hypothesis is not the probability that the researcher's explanation is true. A p-value is calculated under a null model; it is not the probability of either hypothesis. Likewise, a small p-value does not reveal which causal mechanism produced the observed association.

One-sided tests are not shortcuts for obtaining smaller p-values. Their direction must be justified by the question. Testing many alternatives and reporting only favorable ones changes error behavior. Also, an alternative need not always be "different from zero": equivalence, noninferiority and threshold questions have distinct null and alternative formulations.

Business use

Clear alternatives help organizations connect statistical work to decisions. They force teams to state what would count as improvement, deterioration or an important difference, and to plan the sample size around effects that matter.

Prespecify the primary comparison, direction and outcome when feasible. Separate exploratory findings from confirmatory testing, and account for multiple comparisons when several claims are evaluated. Use effect sizes and uncertainty to assess practical consequences. In AI benchmarking, a statistically detectable change may be too small to justify deployment costs, while a poorly powered study can miss a consequential difference. The alternative defines the question; it does not replace judgment about the answer.

Check your understanding

A team wants to detect either an increase or a decrease from a population mean of 20. What alternative matches the question?

Answer: H1: the population mean is not equal to 20. A greater-than alternative would address only an increase.

Related terms

Null hypothesis
One-sided test
Two-sided test
P-value
Significance level
statistical power
effect size and equivalence testing

Learn next

write null and alternative hypotheses for a difference, a directional improvement and an equivalence question. Identify the population parameter, meaningful effect threshold and evidence required in each case before selecting a test.

Sources

Primary sources: KNIME, Data Science Glossary, Alternative hypothesis: https://www.knime.com/blog/data-science-glossary

Penn State, STAT 100, Hypothesis Testing: https://online.stat.psu.edu/stat100/Lesson10

Penn State, STAT 200, Hypothesis Testing, Part 1: https://online.stat.psu.edu/stat200/Lesson05

Previous: Alternating direction method of multipliers Next: Alveolar sound