Why Conditions Matter
A one-way F test compares the means of several groups by relating variation between group means to variation within groups. The calculation can always produce an F statistic, but its usual reference distribution—and therefore its p-value—depends on assumptions about how the observations arose. Checking those assumptions is part of interpreting the test, not an optional step after the arithmetic.
The conditions concern three different features: whether observations are independent, whether outcomes within each group are approximately normal, and whether the groups have a common population variance. The first is primarily supported by the study design. The other two concern the outcome distributions within groups. None is established merely because an ANOVA table can be calculated.
The Model and Its Conditions
Here \(i\) indexes groups, \(j\) indexes observations within a group, and \(N\) denotes the total number of observations across the groups. The outcome must be quantitative, and the groups must be the categories of one factor. The model does not require the group means \(\mu_i\) to be equal in order to describe the data. Equal means are the null hypothesis tested by the classical one-way F test; the conditions describe the model used to calibrate that test.
| Condition | What it concerns | Study feature or evidence to consider |
|---|---|---|
| Independence | How observations relate to one another | Sampling or assignment procedure; repeated measurements, pairs, families, or clusters |
| Approximate within-group normality | The outcome distribution within each group | Group-level distributions, especially their skewness and unusual observations |
| Common variance | The population spread in each group | Group standard deviations and whether the group spreads are reasonably similar |
Independence Comes From the Study Design
Independence means that the outcome for one observational unit does not provide information about the random variation in another unit’s outcome, once the model’s group means are accounted for. In the classical model, the errors \(\varepsilon_{ij}\) are mutually independent. This is not the same as saying that group means are equal, or that observations in different groups must have different values.
Study design is the main guide. A simple random sample of people, with one outcome recorded per person, can support independence if the sampling process does not create important dependence. Random assignment to treatment groups can support independence when each unit is separately assigned and one unit’s treatment does not affect another unit’s outcome. These design features support an assumption; they do not make it automatically true. For example, roommates may influence each other, and people treated by the same clinician may have related outcomes.
Repeated measurements on the same person are not independent observations just because they were recorded at different times. Nor are several measurements from one classroom necessarily independent if students in the same classroom share influences. In these settings, the observational unit and the dependence structure need careful attention; simply entering every recorded measurement into a classical one-way ANOVA does not remove the dependence.
Independence within and across groups is also distinct from how participants were recruited. A random sample can support generalizing to a population, while random assignment can support comparisons of treatments. Neither feature by itself establishes all the conditions of the one-way F test, and independence alone does not justify a causal conclusion.
Worked Example: Repeated Measurements Are Not Independent
Hypothetical study. A team records a symptom score for each of 18 patients before treatment and again at two later visits. They propose to place the three visit measurements into three groups and use a classical one-way F test.
The number of recorded scores is \(18\times3=54\). But those are not 54 independent patients: each patient contributes three measurements. A patient’s earlier score is likely related to that same patient’s later scores, so the independent-error condition is not supported by the design.
The proposed analysis treats the visits as groups, but the repeated observations within each patient remain linked. The fact that there are 54 recorded numbers does not change that link. A classical one-way F test applied as though all 54 scores were independent would use an independence assumption that this study design contradicts. The comparison requires an approach that accounts for the repeated measurements; the classical one-way test is not justified simply by arranging the data into three columns.
Normality Is Assessed Within Each Group
The normality condition applies to the distribution of outcomes around each group’s own mean, or equivalently to the errors within each group. It does not say that all observations pooled across groups must form one normal distribution. If group means differ, the pooled data can combine several distributions with different centers even when each group’s outcomes are normal.
The condition is often described as approximate because real data rarely follow an exact normal distribution. The degree of concern depends on the group sample sizes and on the nature of departures. Strong skewness or a few extreme observations can be particularly troublesome in small groups, where they may have a large influence on the sample mean and within-group spread. Larger samples often make the F procedure more tolerant of moderate departures, but they do not make severe outliers or dependence harmless.
Normality cannot be confirmed from a small sample. A set of observations that looks fairly symmetric is compatible with approximate normality, but it does not prove a normal population distribution. Conversely, a small sample may show irregularities simply because it contains little information about its population distribution. The question is whether the group data and the process that produced them make the approximation reasonable enough for the intended analysis.
Worked Example: Check the Distributions Group by Group
Hypothetical data. Three independently assigned groups each have four outcome measurements. The values are:
Group A: 8, 9, 10, 9. Group B: 10, 11, 9, 10. Group C: 11, 12, 10, 11. The group means are 9, 10, and 11, respectively. For Group A, for example, the sample variance is \(s_A^2=\frac{\sum (x_j-\bar{x}_A)^2}{n_A-1}=\frac{(8-9)^2+(9-9)^2+(10-9)^2+(9-9)^2}{4-1}=\frac{2}{3}\approx0.667\).
Each group’s four values are arranged symmetrically around its own mean, with no value far from the other three. That is consistent with approximately normal within-group distributions, but four observations per group are far too few to establish normality. The assessment must be made separately for each group: comparing all 12 observations as one pooled distribution would mix three different centers.
The story says the measurements come from independently assigned units, which supports independence if there is no interference between units. This example illustrates a plausible check, not proof that every model condition holds. In particular, the small group sizes leave substantial uncertainty about the shape of each population distribution.
Common Variance Means Comparable Population Spreads
The common-variance condition says that the population variance is the same in every group: \(\sigma_1^2=\sigma_2^2=\cdots=\sigma_k^2=\sigma^2\). It does not require the group sample variances to be identical. Sample variances vary from sample to sample even when the population variances are equal, so the practical question is whether the observed differences in spread seem small enough to be compatible with a common population variance.
Compare spreads within each group, rather than using the pooled spread to hide group differences. Group sample standard deviations are in the same units as the outcome and are often easier to compare than variances. Large differences deserve more attention when group sizes are unequal: the combination of unequal spreads and unequal sample sizes can make the classical F test less reliable. Balanced group sizes can make the test more tolerant of some variance differences, but balance does not guarantee equal variances or make severe differences irrelevant.
Worked Example: A Large Difference in Group Spreads
Hypothetical summary. Three groups have sample sizes \(n_1=8\), \(n_2=20\), and \(n_3=8\). Their sample standard deviations are \(s_1=2.0\), \(s_2=2.2\), and \(s_3=7.5\) outcome units. The third group’s sample spread is much larger than the other two.
A descriptive comparison of the largest and smallest sample variances is \(\frac{\max(s_i^2)}{\min(s_i^2)}=\frac{7.5^2}{2.0^2}=\frac{56.25}{4.00}=14.1\).
This ratio is a warning that a common population variance may be implausible; it is not a universal cutoff that decides the assumption by itself. The group sizes are also unequal, with the largest sample in the group whose observed spread is small. That combination warrants particular caution about the classical F test. The summary cannot prove the population variances differ, since sample spreads are uncertain, but it gives a concrete reason not to proceed as if equal spreads were unquestioned.
Why These Conditions Give an F Reference Distribution
A short derivation explains the role of the assumptions. Let \(n_i\) be the size of group \(i\), so \(N=\sum_{i=1}^{k}n_i\). The between-group sum of squares is \(SS_{\text{between}}=\sum_{i=1}^{k}n_i(\bar{x}_i-\bar{x})^2\), and the within-group sum of squares is \(SS_{\text{within}}=\sum_{i=1}^{k}\sum_{j=1}^{n_i}(x_{ij}-\bar{x}_i)^2\). Their degrees of freedom are \(k-1\) and \(N-k\), respectively. The corresponding mean squares divide each sum of squares by its degrees of freedom.
Under the null hypothesis \(H_0:\mu_1=\cdots=\mu_k\), and with independent normal observations sharing variance \(\sigma^2\), the centered observations form a normal vector. The between-group and within-group sums of squares are squared lengths of its projections onto perpendicular subspaces: the between-group subspace has dimension \(k-1\), and the within-group subspace has dimension \(N-k\). For a normal vector with variance \(\sigma^2\), its squared projection length divided by \(\sigma^2\) has a chi-square distribution with degrees of freedom equal to that subspace’s dimension. Perpendicular projections of this normal vector are independent. Therefore,
Dividing each sum of squares by its degrees of freedom gives two independent chi-square quantities divided by their degrees of freedom. Their ratio is the classical statistic \(F=MS_{\text{between}}/MS_{\text{within}}\), which under \(H_0\) has an \(F\) distribution with \(k-1\) and \(N-k\) degrees of freedom. Independence supports the separation of the two sources of variation; within-group normality supplies the normal-vector result; and a common variance gives both sums of squares the same \(\sigma^2\) scale. If a condition fails, this derivation no longer establishes the usual F reference distribution.
Practical Assessment and a Common Mistake
Conditions are judgments about a model and the data-generation process, not mechanical boxes that sample data can certify. First ask how units were sampled or assigned, how many measurements came from each unit, and whether units share clusters or influences. Then examine the outcome within each group for conspicuous skewness or extreme values, and compare group spreads in context. Record limitations rather than claiming that an assumption has been proven.
- Mistake: checking normality only after pooling all groups. A pooled distribution may be non-normal because group means differ. Assess the outcome distribution within each group.
- Mistake: assuming random assignment guarantees independence. Assignment does not prevent repeated measurements, clustering, or interference between units. Consider how outcomes are linked.
- Mistake: treating similar sample standard deviations as proof of equal population variances. Sample spreads are estimates with uncertainty. Similarity supports the condition but cannot prove it.
- Mistake: thinking a large total sample automatically fixes every problem. Large samples do not remove dependence, and severe skewness or influential observations can still matter, particularly when group sizes differ.
Check Your Understanding
Answer in terms of the model and the study features that bear on each condition.
- A study records one outcome for each randomly selected person. What additional design features would you consider before treating those outcomes as independent?
- Why can a pooled distribution look non-normal even when the within-group normality condition is plausible?
- A clinic records outcomes for the same patients at three visits. Which condition is most directly challenged if all visits are treated as independent observations?
- What does a large ratio of the largest to smallest group sample variance suggest, and why is it not a definitive test of equal population variances?
- Under the null hypothesis of equal group means, explain how independence, normality, and common variance support the usual F reference distribution.