Difficulty: Intermediate | Prerequisites: Basic hypothesis testing (STAT 211 or equivalent), familiarity with means, variances, and the normal distribution.
Analysis of Variance (ANOVA) is the main tool in STAT 212 for comparing means across more than two groups. It sits at the centre of inferential statistics for experimental and observational designs where a categorical variable defines groups and the response is continuous. If you already understand two-sample t-tests, ANOVA is the natural next step: it generalises that logic to k groups while keeping the Type I error rate under control. Everything in this section feeds into post-hoc procedures and later regression topics.
ANOVA tests whether the means of two or more populations are equal by partitioning total variability into between-group and within-group components. Despite its name, the method analyses means, not variances directly. The test relies on three key assumptions: independence, normality, and equal variances (homoscedasticity).
Analysis of Variance (ANOVA)
A statistical method for comparing the means of two or more groups by examining how much of the total variation in the data is due to differences between groups versus differences within groups. Think of it as asking: "Is the spread between my groups bigger than the spread inside them?"
Factor
The categorical independent variable that defines the groups being compared in an ANOVA. In simple terms, this is the thing you are changing or categorising by. If three classrooms each use a different textbook, the textbook is the factor.
Levels (of a factor)
The specific categories or values that a factor can take. If the factor is "textbook," the levels are Textbook A, Textbook B, and Textbook C.
Response variable
The quantitative outcome being measured across groups. In the textbook example, this is the students' final test scores.
Grand Mean (X̄..)
The overall average of every observation across all groups combined. Think of it as the single number you would get if you ignored group membership entirely and just averaged everything.
Single-factor ANOVA
An ANOVA design with exactly one factor (one categorical independent variable). Also called one-way ANOVA.
Homoscedasticity
The assumption that all k populations share the same variance σ². In plain terms, the spread of scores within each group should be roughly equal. This is sometimes called "equal variances" or "homogeneity of variance."
Independence (assumption)
The requirement that error terms for all observations are independent of one another. One student's score does not influence another's, whether they are in the same group or different groups.
Normality (assumption)
The assumption that the populations from which samples are drawn are normally distributed, or equivalently that the error terms follow a normal distribution.
Error term (εᵢⱼ)
In the ANOVA model Xᵢⱼ = µᵢ + εᵢⱼ, the error term captures random deviation of an individual observation from its group mean. These are assumed to be independent and identically distributed as N(0, σ²).
ANOVA compares means across k groups (where k ≥ 2) using a single omnibus test.
The name "Analysis of Variance" comes from the technique of splitting total variability into components, but the goal is inference about means, not variances.
ANOVA answers the question: "Is there sufficient evidence that at least two of the population means differ?"
The formal model is: Xᵢⱼ = µᵢ + εᵢⱼ
Xᵢⱼ is the jth observation in the ith group.
µᵢ is the true population mean of group i.
εᵢⱼ is the random error, assumed ~ N(0, σ²), independent across all observations.
Independence: Error terms are unrelated to one another. This is typically ensured by the study design (random sampling, random assignment). Violation makes the F-test unreliable because the probability calculations assume independent draws.
Normality: Each population is normally distributed. ANOVA is reasonably robust to moderate departures from normality, especially with larger sample sizes.
Homoscedasticity (equal variances): All k groups share a common variance σ². This matters because the pooled error estimate (MSE) assumes a single σ². If group variances differ substantially, the F-test can give misleading results.
Null hypothesis: H₀: µ₁ = µ₂ = … = µₖ (all population means are equal).
Alternative hypothesis: Hₐ: At least two of the population means are different.
Note: the alternative does not say all means differ, only that the equal-means claim fails for at least one pair.
ANOVA model: Xᵢⱼ = µᵢ + εᵢⱼ , where εᵢⱼ ~ N(0, σ²), independently
Null hypothesis: H₀: µ₁ = µ₂ = … = µₖ
ANOVA is used whenever researchers need to compare outcomes across multiple conditions or categories. A pharmaceutical company testing three drug doses on blood pressure, an agronomist comparing crop yields across four fertiliser types, or an educator comparing test scores across different teaching methods, all of these are single-factor ANOVA problems. The method prevents the error inflation that would come from running many separate two-group tests.
Students often think ANOVA tests whether variances are equal. It does not. The name refers to analysing variance components, but the conclusion is about means.
Students sometimes write the alternative hypothesis as "all means are different." The correct alternative is that at least two means differ, not necessarily all of them.
The independence assumption is about the error terms, not about whether the groups are related by topic. Two groups can study similar material and still have independent observations.
Homoscedasticity is not the same as normality. You can have normal distributions with unequal variances, or equal variances with non-normal distributions. Both assumptions must hold separately.
⚠️ Know the three assumptions by name and be able to explain each in a sentence. These appear in short-answer questions regularly.
⚠️ Be precise with hypotheses. The null is "all means equal." The alternative is "at least two differ." Writing "all means differ" for Hₐ will cost marks.
⚠️ Identifying the factor, levels, and response variable from a word problem is a common early exam question. Read the scenario carefully: the factor is what defines the groups, not the outcome.
⚠️ The ANOVA model equation Xᵢⱼ = µᵢ + εᵢⱼ and the distribution of the error term (N(0, σ²)) are frequently tested.
True or False: ANOVA is used to compare the variances of multiple populations. False. ANOVA compares means, not variances.
Fill in the blank: The assumption that all groups share the same variance is called ________. Homoscedasticity (or equal variances / homogeneity of variance).
True or False: The alternative hypothesis in ANOVA states that all population means are different from one another. False. It states that at least two means are different.
Fill in the blank: In the ANOVA model Xᵢⱼ = µᵢ + εᵢⱼ, the error terms are assumed to follow a ________ distribution. Normal distribution with mean 0 and variance σ² (i.e., N(0, σ²)).
True or False: If three classrooms use different textbooks and we compare test scores, the factor is "test scores." False. The factor is the textbook. Test scores are the response variable.
Q: What is the primary objective of ANOVA?
A: To compare the means of two or more groups by analysing variance components (between-group vs. within-group variation).
Q: In a study assigning 150 students to three classes, each using a different textbook, to compare final test scores, what is the factor?
A: The textbook. It is the categorical independent variable defining the groups.
Q: State the null hypothesis for a single-factor ANOVA with k groups.
A: H₀: µ₁ = µ₂ = … = µₖ (all population means are equal).
Q: What does the independence assumption require, and why is it important?
A: It requires that error terms are independent of one another, meaning one observation does not influence any other. This is critical because the F-test's probability calculations depend on the observations being independent draws.
Q: Name the three assumptions of the ANOVA model.
A: Independence of error terms, normality of the populations (or error terms), and homoscedasticity (equal variances across groups).
Q: In the ANOVA model Xᵢⱼ = µᵢ + εᵢⱼ, what distribution is assumed for εᵢⱼ?
A: Normal with mean 0 and variance σ², i.e., N(0, σ²), independently.
This material connects directly to the two-sample t-test from STAT 211: ANOVA with k = 2 groups is equivalent to a two-sample t-test. It also sets the foundation for post-hoc multiple comparison procedures (Tukey's, Bonferroni) covered in Part 3 of these notes, and later for factorial ANOVA and regression, which extend the same variance-partitioning logic to more complex designs.
Related Terms / Search Tags: ANOVA, analysis of variance, one-way ANOVA, single-factor ANOVA, F-test, factor, levels, treatment, response variable, grand mean, homoscedasticity, equal variances, homogeneity of variance, independence assumption, normality assumption, ANOVA model, error term, epsilon, null hypothesis ANOVA, STAT 212, Principles of Statistics II, Texas A&M