ANOVA Overview

What is ANOVA?

Analysis of variance (ANOVA) assesses the differences between group means. It is a statistical hypothesis test that determines whether the means of at least two populations are different. At a minimum, you need a continuous dependent variable and a categorical independent variable that divides your data into comparison groups to perform ANOVA.

Researchers commonly use ANOVA to analyze designed experiments. In these experiments, researchers use randomization and control the experimental factors in the treatment and control groups. For example, a product manufacturer sets the time and temperature settings in its process and records the product’s strength. ANOVA analyzes the differences in mean outcomes stemming from these experimental settings to estimate their effects and statistical significance. These designed experiments are typically orthogonal, which provides important benefits. Learn more about orthogonality.

Additionally, ANOVA is useful for observational studies. For example, a researcher can observe outcomes for several education methods and use ANOVA to analyze the group differences. However, as with any observational study, you must be careful about the conclusions you draw.

The term "analysis of variance" originates from how the analysis uses variances to determine whether the means are different. ANOVA works by comparing the variance of group means to the variance within groups. This process determines if the groups are part of one larger population or separate populations with different means. Consequently, even though it analyzes variances, it actually tests means! To learn more about this process, read my post, The F-test in ANOVA.

The fundamentals of ANOVA tests are relatively straightforward because it involves comparing means between groups. However, there are an array of elaborations. In this post, I start with the simpler forms and explain the essential jargon. Then I’ll broadly cover the more complex methods.

Simplest Form and Basic Terms of ANOVA Tests

The simplest type of ANOVA test is one-way ANOVA. This method is a generalization of t-tests that can assess the difference between more than two group means.

Statisticians consider ANOVA to be a special case of least squares regression, which is a specialization of the general linear model. All these models minimize the sum of the squared errors.

Related post: The Mean in Statistics

Factors and Factor Levels

To perform the most basic ANOVA, you need a continuous dependent variable and a categorical independent variable. In ANOVA lingo, analysts refer to “categorical independent variables” as factors. The categorical values of a factor are “levels.” These factor levels create the groups in the data. Factor level means are the means of the dependent variable associated with each factor level.

For example, your factor might have the following three levels in an experiment: Control, Treatment 1, and Treatment 2. The ANOVA test will determine whether the mean outcomes for these three conditions (i.e., factor levels) are different.

Related posts: Independent and Dependent Variables and Control Groups in Experiments

Factor Level Combinations

ANOVA allows you to include more than one factor. For example, you might want to determine whether gender and college major correspond to differences in income.

Gender is one factor with two levels: Male and Female. College major is another factor with three levels in our fictional study: Statistics, Psychology, and Political Science.

The combination of these two factors (2 genders X 3 majors) creates the following six groups:

• Male / Statistics
• Female / Statistics
• Male / Psychology
• Female Psychology
• Male / Political Science
• Female / Political Science

 

These groups are the factor level combinations. ANOVA determines whether the mean incomes for these groups are different.

Factorial ANOVA

ANOVA allows you to assess multiple factors simultaneously. Factorial ANOVA are cases where your data includes observations for all the factor level combinations that your model specifies. For example, using the gender and college major model, you are performing factorial ANOVA if your dataset includes income observations for all six groups.

By assessing multiple factors together, factorial ANOVA allows your model to detect interaction effects. This ability makes multiple factor ANOVA much more efficient. Additionally, evaluating a single factor at a time conceals interaction effects. Single-factor analyses tend to produce inconsistent results when interaction effects exist in an experimental area because they cannot model the interaction effects.

Analysts frequently use factorial ANOVA in experiments because they efficiently test the main and interaction effects for all experimental factors.

Related post: Understanding Interaction Effects

Interpreting the ANOVA Test

ANOVA assesses differences between group means.

Suppose you compare two new teaching methods to the standard practice and want to know if the average test scores for the methods are different. Your factor is Teaching Method, and it contains the following three levels: Standard, Method A, and Method B.

The factor level means are the mean test score associated with each group.

ANOVAs evaluate the differences between the means of the dependent variable for the factor level combinations. The hypotheses for the ANOVA test are the following:

• Null Hypothesis: The group means are all equal.
• Alternative Hypothesis: At least one mean is different.

 

When the p-value is below your significance level, reject the null hypothesis. Your data favor the position that at least one group mean is different from the others.

While a significant ANOVA result indicates that at least one mean differs, it does not specify which one. To identify which differences between pairs of means are statistically significant, you’ll need to perform a post hoc analysis.

Related post: How to Interpret P Values

General ANOVA Assumptions

ANOVA tests have the same assumptions as other linear models other than requiring a factor. Specifically:

• The dependent variable is continuous.
• You have at least one categorical independent variable (factor).
• The observations are independent.
• The groups should have roughly equal variances (scatter).
• The data in the groups should follow a normal distribution.
• The residuals satisfy the ordinary least squares assumptions.

 

While ANOVA assumes your data follow the normal distribution, it is robust to violations of this assumption when your groups have at least 15 observations.

ANOVA Designs and Types of Models

The simpler types of ANOVA test are relatively straightforward. However, ANOVA is a flexible analysis, and many elaborations are possible. I’ll start with the simple forms and move to the more complex designs. Click the links to learn more about each type and see examples of them in action!

One-way ANOVA

One-way ANOVA tests one factor that divides the data into at least two independent groups.

Learn about one-way ANOVA and how to perform and interpret an example using Excel.

Two-way ANOVA

Two-way ANOVA tests include two factors that divide the data into at least four factor level combinations. In addition to identifying the factors’ main effects, these models evaluate interaction effects between the factors.

Learn about two-way ANOVA and how to perform and interpret an example using Excel.

Analysis of Covariance (ANCOVA)

ANCOVA models include factors and covariates. Covariates are continuous independent variables that have a relationship with the dependent variable. Typically, covariates are nuisance variables that researchers cannot control during an experiment. Consequently, analysts include covariates in the model to control them statistically.

Learn more about Covariates: Definition and Uses.

Repeated measures ANOVA

Repeated measures designs allow researchers to assess participants multiple times in a study. Frequently, the subjects serve as their own controls and experience several treatment conditions.

Learn about repeated measures ANOVA tests and see an example.

Multivariate analysis of variance (MANOVA)

MANOVA extends the capabilities of ANOVA by assessing multiple dependent variables simultaneously. The factors in MANOVA can influence the relationship between dependent variables instead of influencing a single dependent variable.

There’s even a MANCOVA, which is MANOVA plus ANCOVA! It allows you to include covariates when modeling multiple independent variables.

Learn about MANOVA and see an example.

Crossed and Nested Models

ANOVA can model both crossed and nested factors.

Crossed factors are the more familiar type. Two factors are crossed when each level of a factor occurs with each level of the other factor. The gender and college major factors in the earlier example are crossed because we have all combinations of the factor levels. All levels of gender occur in all majors and vice versa.

A factor is nested in another factor when all its levels occur within only one level of the other factor. Consequently, the data do not contain all possible factor level combinations. For example, suppose you are testing bug spray effectiveness, and your factors are Brand and Product.

In the illustration, Product is nested within Brand because each product occurs within only one level of Brand. There can be no combination that represents Brand 2 and Product A.

The combinations of crossed and nested factors within an ANOVA design can become quite complex!

General Linear Model

The most general form of ANOVA allows you to include all the above and more! In your model, you can have as many factors and covariates as you need, interaction terms, crossed and nested factors, along with specifying fixed, random, or mixed effects, which I describe below.

Types of Effects in ANOVA

Fixed effects

When a researcher can set the factor levels in an experiment, it is a fixed factor. Correspondingly, the model estimates fixed effects for fixed factors. Fixed effects are probably the type with which you’re most familiar. One-way and two-way ANOVA procedures typically use fixed-effects models. ANOVA models usually assess fixed using ordinary least squares.

For example, in a cake recipe experiment, the researcher sets the three oven temperatures of 350, 400, and 450 degrees for the study. Oven temperature is a fixed factor.

Random effects

When a researcher samples the factor levels from a population rather than setting them, it is a random factor. The model estimates random effects for them. Because random factors sample data from a population, the model must change how it evaluates their effects by calculating variance components.

For example, in studies involving human subjects, Subject is typically a random factor because researchers sample participants from a population.

Mixed effects

Mixed-effects models contain both fixed and random effects. Frequently, mixed-effects models use restricted maximum likelihood (REML) to estimate effects.

For example, a store chain wants to assess sales. The chain chooses five states for its study. State is a fixed factor. Within these states, the chain randomly selects stores. Store is a random factor. It is also nested within the State factor.

Scroll down to find more of my articles about ANOVA!