What does a one-way ANOVA test, and what are its null and alternative hypotheses?
answer
- compares more than two group means
- one categorical factor, several levels
- omnibus, not pairwise
- alternative is 'at least one'
basics
~20 sOne-way ANOVA tests whether the population means of three or more groups defined by a single factor are all equal. The null says every group mean is the same; the alternative says at least one differs.
solid answer
~40 sOne-way ANOVA compares the means of several groups formed by one categorical factor — for example mean crop yield under four different fertilizers. The null hypothesis is that all four population means are equal; the alternative is that **at least one** mean differs from the rest, which is deliberately vague about which one. That makes it an *omnibus* test: a small p-value tells you the factor matters somewhere, not which fertilizer beat which. It does this by comparing how far the group means spread apart against how much the observations scatter inside each group, and it assumes the observations are independent, roughly normal within each group, and have similar spread across groups. If you need to name the winning group you follow the significant result with a post-hoc pairwise procedure.
go deeper
Be ready to say in one breath what ANOVA compares and to state both hypotheses correctly, especially the 'at least one differs' alternative. Know that one factor with four levels is still a one-way design.
Explain why a test about means is built out of variances: between-group spread against within-group spread, comparable in size when the null holds. Also state the independence, normality and equal-variance assumptions.
Show you know what a significant omnibus result does and does not license. Interviewers listen for whether you jump straight to naming the winning group, or correctly stop and reach for a post-hoc pairwise procedure.
Own the framing decision: whether the question in front of the business is really an omnibus one at all. Often a small set of pre-planned contrasts against a control answers the real question with more power than an all-groups sweep.
## The setup You have one **categorical factor** with `k` levels — four fertilizers applied to plots of land, three onboarding flows, five machine settings — and one **continuous outcome** measured on each unit, such as crop yield in tonnes per hectare. You want to know whether the factor moves the outcome at all. One-way ANOVA ("analysis of variance", one-way because there is a single factor) is the standard test for exactly this shape of data. "One-way" refers to the number of factors, not the number of groups. Four fertilizers is still a one-way design, because fertilizer is one factor with four levels. ## The hypotheses, stated precisely - **Null (H0):** mu_1 = mu_2 = ... = mu_k. Every group's *population* mean is the same. Note these are population parameters, not the sample means you computed — the sample means will always differ a little by chance. - **Alternative (H1):** at least one mu_i differs from at least one other. Equivalently, "not all the mu_i are equal." The alternative is the part candidates most often state wrongly. It is **not** "all group means differ from each other." A single rogue group among five is enough to make the null false. The alternative is also not directional — there is no "one-tailed ANOVA" in the sense of a t-test, because with three or more means there is no single direction to point in. ## Why it is called an omnibus test An omnibus test answers one broad question with one p-value. It has no machinery for saying *which* means differ. If your four-fertilizer ANOVA returns p = 0.004, the honest sentence is "fertilizer affects yield," not "fertilizer C is the best." Ranking the sample means and declaring the top one the winner is a classic mistake: sampling noise reorders sample means all the time, and the omnibus test never evaluated that specific comparison. To name groups you run a post-hoc pairwise procedure that accounts for the fact that you are now making many comparisons at once. ## How the name makes sense The test is about means, but it is called *analysis of variance* because it works by splitting variability. The total scatter of every observation around the grand mean is decomposed into: - **Between-group** variability — how far each group mean sits from the grand mean, weighted by group size. This grows when the factor really does shift the outcome. - **Within-group** variability — how far individual observations sit from their own group mean. This is the background noise that exists regardless of the factor. If the null is true, the group means are only wandering around the grand mean by chance, so the between-group piece is just another view of the same background noise, and the two pieces come out comparable in size. If some group really is different, the between-group piece inflates. The test statistic is the ratio of the two, and its sampling distribution under the null is an F distribution. ## What it assumes One-way ANOVA assumes the observations are **independent** (one measurement per plot, no plot counted twice, no hidden clustering), the outcome is approximately **normally distributed within each group**, and the groups have roughly **equal variance**. The test is reasonably robust to mild departures from normality, especially with decent and balanced group sizes, and much less forgiving of dependence between observations — dependence quietly shrinks the effective sample size and inflates false positives. ## The two-group special case Nothing breaks if you run a one-way ANOVA on exactly two groups. It is simply equivalent to the pooled two-sample test of equal means: the F statistic comes out as the square of the t statistic, and the two p-values are identical. ANOVA earns its keep from three groups upward, where a single test replaces a pile of pairwise ones. ## What a good answer sounds like "One-way ANOVA asks whether one categorical factor shifts the mean of a continuous outcome. The null is that all group means are equal; the alternative is that at least one differs. It compares between-group spread to within-group spread, and a significant result tells you the factor matters without telling you which groups are responsible — that needs a post-hoc comparison."
- A four-group ANOVA comes back significant — what can you say about which group is best?Nothing yet. The omnibus F only says at least one population mean differs from the others. The group with the highest sample mean may simply have drawn a lucky sample, and the test never evaluated that specific pair. Naming a winner requires a post-hoc pairwise procedure that adjusts for making all the comparisons at once.
- Can you run a one-way ANOVA on exactly two groups, and what do you get?Yes, and it reduces to the pooled two-sample test of equal means. The F statistic equals the square of the corresponding t statistic and the p-values are identical, since F with 1 numerator degree of freedom is a squared t. ANOVA only becomes useful from three groups upward, where one test replaces many pairwise ones.
- Why is the ANOVA F test evaluated only in the upper tail?Because only a large F is evidence against equal means. F is between-group variability over within-group variability; the alternative pushes the numerator up, so evidence lives in the right tail. A very small F means the group means are closer together than chance would usually produce — odd, sometimes a sign of a data problem, but not evidence that the means differ.
It is a smoke alarm for group differences: it tells you something is burning somewhere in the house, but not which room.
saying these in an interview costs you the question
- States the alternative as 'all group means differ'
- Claims a significant ANOVA identifies which group is best
- Thinks one-way ANOVA requires exactly three groups
- Says ANOVA tests whether the group variances are equal
- Confuses sample means with the population means being tested