Analysis of Covariance (ANCOVA)

Use in Communication Research

Communication researchers, like other scientific researchers, assume that the universe is not random. Stuff is associated with other stuff. Rainfall is associated with humidity. Earthquakes are associated with fault lines between tectonic plates. Income is associated with education. In communication, stuff just doesn’t happen; the world is not a random place. Communication apprehension is associated with poor public speaking performance. Interpersonal intimacy is associated with relational satisfaction. Affectionate behavior is associated with health. Source credibility is associated with increased persuasion. Fear appeals are associated with behavior change. The goal of all scientists, including communication scientists, is to explain and understand the seemingly random universe. All inferential statistics are based on removing unexplained or error variance from the denominator of a statistical test and moving it to the numerator, where it is explained variation or systematic variation. Understanding communication, like understanding the subject material of any other science, requires explaining the variation in stuff, showing that relations among variables are nonrandom. This is the essence of scientific understanding.

This is clearly the case with ANOVA or F ratios, where the numerator of an F ratio is the variance that can be accounted for and the denominator is unexplained variation, called error variance. For example, we might want to know what makes a good public speaker. Many things are involved in public speaking skill, but we want to test whether public speaking skill increases after a public speaking class. We randomly assigned college freshmen to take either a public speaking or some other class (say chemistry) and we compute an ANOVA to see whether there is a statistically significant difference in public speaking ability (the outcome variable) at the end of the semester between the students who took the chemistry class and the students who took the public speaking class. We could use an ANOVA to do that.

ANOVA examines whether there are statistically significant differences among groups. These can be preexisting groups such as males and females or Democrats and Republicans, or they can be experimental groups such as people receiving three kinds of public speaking training or three different types of persuasive health communication messages.

[Page 29]ANCOVA is the same as ANOVA except it uses an extra variable or variables to control for distracting, interfering, or confounding variables that may distort the real relationship between an independent variable and an outcome variable.

Applications and Benefits

Let’s discuss a couple of examples of important applications of ANCOVA. First, let’s go back to a version of the public speaking class example discussed earlier. We want to see if several kinds of public speaking training increase public speaking ability. The gold standard is an experiment, so we create three classes: one using relaxation therapy, one using traditional public speaking training, and one unrelated to communication, say chemistry. We randomly assign students to each of these kinds of classes, so that all types of students of different races, genders, backgrounds, home cities, etc., have an equal chance of getting into one of the three kinds of classes. We now assume, with at least some justification, that the students in the three kinds of classes are very similar to one another—at least not systematically biased. At the end of the semester, we test each student for her or his public speaking ability (the dependent or outcome variable) and use an ANOVA to test whether students in one group are significantly better speakers than students in the other group. But we can do better! We know that all the students were really not the same (though we are assuming and expecting that students in all three types of classes were very similar). We know that extraverts tend to be better speakers than introverts and people with better vocabularies tend to be better speakers than people with more limited vocabularies. We could use ANCOVA to statistically readjust people’s public speaking scores by controlling for extraversion and vocabulary. We statistically adjust each person’s public speaking (outcome) scores as though each student had the same extraversion/introversion level and extent of vocabulary. Now we have removed variance associated with two extraneous variables that impacted the students’ public speaking scores and have a produced a better test of the association between these three types of classes with public speaking ability.

...