ANOVA, statistical procedure used to compare means of three or more groups. ANOVA tests compare the amount of variance between and within groups to determine whether statistically significant differences exist between their means. Many variations of ANOVA exist, including one-way ANOVA, factorial ANOVA, and repeated measures ANOVA. As an omnibus test, ANOVA can indicate whether a difference between group means exists, but further tests must be conducted to determine the nature of a potential difference or differences.
ANOVA was created by British statistician R.A. Fisher in 1918. The new statistical procedure expanded upon significance tests (e.g., Student’s t-test) developed at the start of the 20th century. While those tests were limited to comparisons between two groups, ANOVA was designed to allow for comparisons between multiple groups using a single test. The procedure gained popularity after being included in Fisher’s landmark text Statistical Methods for Research Workers (1925).
ANOVA classifies the total variability as a result of either systematic differences between treatment groups or random error. As an extension of the t-test and other simple hypothesis tests that compare only the means of two groups, ANOVA compares the means of multiple groups at the same time. Although numerous t-tests could theoretically be conducted between all possible pairs of group means in an experiment, doing so increases the likelihood of a false positive result and the incorrect identification of significant differences. An ANOVA test eliminates this risk and compares the variances across all levels of an independent variable using a single test to identify differences. Generally, ANOVA uses group means along with the error between and within each group to calculate the F-statistic, a test statistic used to find the p-value and determine the test’s significance.
Consider a hypothetical experiment testing the effect of three medications on the participants’ amount of sleep. A one-way ANOVA test, the simplest variation of ANOVA, could be used to determine whether the mean hours of sleep differed between individuals taking medications A, B, or C. The variability between medications (i.e., differences due to variation in treatment) would be compared with the variability within each of the groups (i.e., error due to random chance). The proportion of error between medications to error within each group would then be calculated as a ratio, which constitutes the F-statistic. If the corresponding p-value is smaller than the predetermined threshold of significance, it can be concluded that at least one medication results in a different mean number of hours of sleep than the others. These values are often summarized in an ANOVA table, which depicts the variance between and within groups.
ANOVA can take a variety of other forms depending on experimental design and the objective of the test. Factorial ANOVA tests are an extension of one-way ANOVA and analyze the response groups of two or more independent variables and their interactions. ANOVA tests can also involve repeated measures, which are often referred to as within-subjects ANOVA, and act as an extension of a paired t-test to compare differences between the same or matched subjects across three or more conditions. Other variations include analysis of covariance (ANCOVA), in which the effects of an independent variable are assessed after adjusting the dependent variable for a covariate, as well as multivariate analysis of variance (MANOVA), in which the effects of response groups on multiple dependent variables are analyzed.
A number of assumptions must be met in order to obtain unbiased results from ANOVA, including independence and normal distribution of the data. Additionally, ANOVA requires the response groups to have approximately the same level of variance, referred to as homogeneity of variance. Furthermore, the number of observations in each comparison group should be approximately equal, as such imbalances lower the statistical power of ANOVA tests.
Care must be taken when interpreting ANOVA results, as its status as an omnibus test limits the level of detail of a conclusion. While a statistically significant ANOVA test indicates the existence of at least one group mean that differs from others, no further conclusions can be drawn from that test result alone. A significant p-value does not specify which means differ nor the size or the number of differences. Any further details must be determined using post hoc tests and further statistical testing.