Begin With the Question
A one-way study has one factor that defines several groups, also called treatments. The response is measured on individuals in each group. Before choosing an interval or test, decide which population quantity answers the research question. A question about the mean response in one group is not the same as a question about the difference between two group means, and neither is the same as a question about whether all group means are equal.
The distinction matters because an inferential result answers only the question represented by its target. Evidence about one treatment group’s mean does not, by itself, answer whether that mean differs from another group’s mean. Likewise, an overall comparison of all groups does not specify which groups differ.
The One-Way Model and Its Parameters
Suppose a factor has \(k\) groups. Let \(Y_{ij}\) be the response for individual \(j\) in group \(i\), where \(i=1,\ldots,k\). The one-way model represents each response as its group population mean plus an error:
Here \(\mu_i\) is the population mean response for group \(i\), and \(\varepsilon_{ij}\) is the individual’s deviation from that group mean. The usual pooled one-way ANOVA model assumes independent errors with a common variance \(\sigma^2\), and normally distributed errors within each group:
These assumptions require independent experimental units, approximately normal within-group errors, and equal population variances across groups. The design and the data should make these assumptions reasonable before using pooled one-way inference. In particular, repeated measurements on the same person are not independent experimental units merely because they appear in separate rows.
The choice of target comes from the question, not from which sample means happen to look most different after the data are collected. It is useful to write the target in symbols and in the context of the response before considering a particular inferential procedure.
Three Distinct Inferential Targets
| Question | Population target | Typical null statement |
|---|---|---|
| What is the mean response in group \(i\), or is it a specified benchmark? | \(\mu_i\) | \(H_0:\mu_i=\mu_0\) |
| How do groups \(i\) and \(j\) compare? | \(\mu_i-\mu_j\) | \(H_0:\mu_i-\mu_j=0\) |
| Are all group population means equal? | The set of means \(\mu_1,\ldots,\mu_k\) | \(H_0:\mu_1=\cdots=\mu_k\) |
For an individual group mean, the benchmark \(\mu_0\), if there is one, must have meaning in the study context; it might be a clinical threshold or an established standard. For a pairwise difference, the order of subtraction matters. The quantity \(\mu_i-\mu_j\) is positive when group \(i\) has the larger population mean, and negative when group \(i\) has the smaller one. The overall equality target instead concerns all groups together.
Worked Example: One Group Mean Compared With a Benchmark
In a hypothetical study, patients are independently assigned to one of three recovery programs. Recovery time, in minutes, is recorded for five patients in each program. Program A’s observations are 118, 121, 119, 123, and 120 minutes; the observations for the other programs are 115, 117, 116, 119, 118 minutes for B and 124, 122, 126, 125, 123 minutes for C. The question is whether the population mean recovery time for Program A differs from a 120-minute benchmark.
The target is the mean for Program A, \(\mu_A\), not a difference between programs. Its sample estimate is
The corresponding hypotheses are \(H_0:\mu_A=120\) minutes and \(H_a:\mu_A\ne120\) minutes. These hypotheses ask whether the population mean for Program A is at the benchmark; they do not ask whether Program A differs from B or C.
For pooled one-way inference, the study must have independent patients, roughly normal errors within the programs, and a common error variance. The hypothetical design assigns each patient to one program, so each patient contributes one independent response. The listed values have no extreme observations and have broadly similar within-group spreads (ranges of 5, 4, and 4 minutes); in an actual study, the error distributions and variance assumption would still need appropriate diagnostic review. The sample mean is close to the benchmark, but that fact alone does not establish whether the population mean differs from it.
Worked Example: A Difference Between Two Group Means
A hypothetical agricultural experiment compares three fertilizer treatments on independent plots. Yield is measured in kilograms per plot. The four yields for treatment A are 42, 45, 44, and 43; for treatment B they are 47, 46, 50, and 45; for treatment C they are 41, 44, 43, and 42. The research question is whether the mean yield differs between A and B.
The target is \(\mu_A-\mu_B\). The sample means are
The observed difference, in the specified A-minus-B order, is
This sample result describes A’s mean yield as 3.5 kg/plot lower than B’s. The population hypotheses for a two-sided comparison are \(H_0:\mu_A-\mu_B=0\) and \(H_a:\mu_A-\mu_B\ne0\). A positive difference would instead indicate a higher mean for A. The target is not the mean yield of all three treatments or whether any treatment differs from any other.
The plots are the experimental units, and each plot receives one treatment, supporting independence if plots do not influence one another. The values within each treatment show no striking outlier or severe asymmetry, and the ranges are 3, 5, and 3 kg/plot. Those features make the normal-error and equal-variance assumptions plausible for this hypothetical illustration, though real data require checking rather than assuming them from the treatment labels.
Worked Example: The Overall Equality of Three Means
A hypothetical study compares three diets using the reduction in a measured health indicator, in mg/dL, for four independent participants per diet. The observations are 8, 10, 9, and 11 for Diet 1; 14, 13, 15, and 12 for Diet 2; and 9, 8, 11, and 10 for Diet 3. The research question is whether the population mean reductions are all equal.
The group sample means are
The target is the overall set of population means, and the hypotheses are \(H_0:\mu_1=\mu_2=\mu_3\) and \(H_a:\) not all three means are equal. The sample means suggest that Diet 2 has a higher observed mean reduction, but sample means alone do not determine the outcome of an inferential test or prove that the population means differ.
Each participant contributes one response and is assigned to one diet, so independence is reasonable under the hypothetical design. The values within each group are fairly balanced around their means, without extreme observations. Their spreads are also similar: the ranges are 3, 3, and 3 mg/dL. These features are consistent with the normal-error and equal-variance assumptions for the illustration; they do not replace diagnostic checks in a real analysis.
Why an Overall Equality Target Is Different
The overall null hypothesis has a precise logical meaning: every population mean is equal. For \(k\) groups, this is equivalent to saying that each of the first \(k-1\) means equals the final mean:
If the overall null is false, at least one pair of population means differs. That alternative does not say which pair differs, how many pairs differ, or which group has the largest mean. An overall result therefore answers a broad question about equality; a specific pairwise target is needed to answer a question about a named pair. Conversely, a single pairwise comparison does not settle whether every group mean is equal.
This distinction is especially important when there are more than two groups. A study with four treatments may have a scientific question about one treatment versus a control, a question about two selected treatments, or a question about any differences among all four. Those are three different targets even though they use the same response data and the same one-way model.
A Practical Way to Choose
Specify which individuals and which measured outcome the conclusion concerns, including the response units.
Use \(\mu_i\) for one group’s mean, \(\mu_i-\mu_j\) for a signed comparison, or the equality statement \(\mu_1=\cdots=\mu_k\) for the overall question.
For a mean or difference, identify the benchmark or zero difference being considered. For overall equality, the alternative is that not all means are equal.
Consider independent experimental units, normal errors within groups, and the common-variance assumption required by pooled one-way inference.
The order is deliberate. A procedure cannot repair a mismatch between the research question and the parameter being analyzed. Once the target is clear, the appropriate inferential method can be selected and its uncertainty assessed.
Common Mistake: Treating the Targets as Interchangeable
A common error is to say, “The overall test asks whether treatment A differs from treatment B.” That is not the overall null: it concerns equality of all treatment means, and a result against that null does not identify a particular pair. Another error is to interpret a sample difference, such as \(\bar{x}_A-\bar{x}_B=-3.5\), as proof that the population difference is exactly \(-3.5\). It is an estimate based on the observed plots, not the population parameter itself.
The correction is to name the target before interpreting the evidence. Write “the population mean for A,” “the population mean difference A minus B,” or “equality of all population means.” Then keep the conclusion within the scope of that target.
Check Your Understanding
For each question, identify the inferential target and state what it does not directly answer.
- A researcher wants to know whether the mean response under Treatment 2 is above a specified safety threshold. What is the target?
- A study asks whether the mean response under A is different from the mean under B. Write the parameter, preserving the order in the question.
- For four treatments, what does the null hypothesis of the overall equality question state? What does its alternative state?
- If an overall test provides evidence against equality of five population means, what can be concluded about a particular pair of means, and what cannot be identified from that result alone?
- Explain why an observed sample mean difference is not itself the population mean difference.