All rights reserved.
This, however, is only the first part of its strength. An even greater advantage of the analysis of covariance is that it allows you to compensate for systematic biases among your samples. To return for the moment to the example of two methods of elementary mathematics instruction, suppose you were to begin with a subject pool of 20 third-graders, randomly sorting them into two independent groups of 10 subjects each, with the design of teaching one group by Method A and the other by Method B. The aim of the initial random sorting is to ensure that the two groups are starting out approximately equal, on average, with respect to all factors that might be pertinent to how well they are likely to respond to elementary mathematics instruction. Nonetheless it could happen through sampling erroranother name for sheer, cussed random variabilitythat the two groups do not start out on an equal footing with respect to one or another of these factors. To take the obvious examples, it could happen through mere chance coincidence in the sorting process that one of the groups starts out with a higher average level of intelligence or motivation, either one of which would surely complicate the situation. The analysis of covariance provides a way of measuring and removing the effects of such initial systematic differences between the samples.
In the correlated-samples design, it is typically assumed that these potential sequencing effects can be obviated by systematic counter-balancing. With two experimental conditions, A and B, you would test half your subjects in the sequence A·B and the other half in the sequence B·A. With three conditions, you would test equal numbers of your subjects in the six possible sequences: A·B·C, A·C·B, B·A·C, B·C·A, C·A·B, C·B·A. And so on.
But there are certain kinds of situations where a repeated-measures design might not be feasible; and there are others where, even if it is feasible, it might not be desirable. If you were interested in determining which of two methods of elementary mathematics instruction is more effective, it would clearly make no sense to have half your subjects taught first by Method A, then by Method B, and the other half taught first by Method B, then by Method A. Analogously, suppose you were interested in determining which of two methods of hypnotic induction is the more effective. Here it would certainly be possible to test each of your subjects under each of the two conditionshalf in the sequence A·B and the other half in the sequence B·Athough I suspect you would end up with a substantial sequencing effect, even with the counter-balancing. By the time the A·B subjects get to Method B, they will have already experienced Method A; and that prior experience is likely to influence their response to Method B. Similarly for the B·A subjects: their responses to Method A are likely to be substantially influenced by their prior experience with Method B. In these two cases and in many others, the analysis of covariance provides the best of both worlds: for it allows you to examine the several experimental conditions independently, in isolation from each other, while at the same time removing from the situation the obscuring effects of pre-existing individual differences.
|X =||the measure of intelligence||Y =||the measure of learning|
Suppose, now, that we were actually to perform this experiment and end up with the results listed in the following table. For the sake of simplicity, I am showing only 4 subjects per group; in real-life research you would rarely want to have such tiny samples. I am also color-coding the values of X and Y to help you keep this cast of characters straight.
|Method A||Method B||Subject
I have designed this set of data to make two things obvious at a glance: First, that there is a considerable range of individual differences within the values of X and Y for each of the two groups; and second, that the values of X and Y within each group have a high positive correlation. In brief, and not surprisingly: the higher the measure of intelligence (X), the higher the measure of learning (Y). Not to put too fine a point on it: the smarter a subject is at the outset, the more likely it is that he or she will learn more rather than less, irrespective of the method of instruction. So a substantial portion of the variability that occurs within each of the sets of Y measures is actually covariance with the corresponding set of X measures. Remove that covariance from Y and you thereby remove a substantial portion of the extraneous variability of individual differences.
The second thing I have designed into this set of data is an outcome in which the means of the two sets of Y measures (learning) are precisely the same:
|Method A||Method B||Xa||Ya||Xb||Yb||Means
By mere chance coincidence, the initial random sorting gave us two groups with different mean levels of intelligence:
This example is a follow-up on an item briefly mentioned above, structurally similar to the elementary-mathematics illustration, though with a somewhat more exotic setting. A team of researchers is interested in determining whether two methods of hypnotic induction, A and B, differ with respect to their effectiveness. They begin by randomly sorting 20 volunteer subjects into two independent groups of 10 subjects each, with the aim of administering Method A to one group and Method B to the other. But then, before either of the induction methods is administered, each subject is pre-measured on a standard index of "primary suggestibility," which is a variable known to be correlated with receptivity to hypnotic induction. The dependent variable, measured during the administration of Method A or Method B, is the subject's score on a standard index of hypnotic induction. Thus
|X =||the score on the index of primary suggestibility||Y =||the score on the index of hypnotic induction|
|Method A||Method B||Sub-|
|Means||13.1 ||29.2 ||18.0 ||28.1 ||
X = suggestibility scoreT|
Y = induction score
I'll ask you to begin by focusing on the two sets of induction
Figure 17.1. Variability of Induction Scores within Groups A and BT
MYa = 29.2|
MYb = 28.1
The bottom line here is that the induction scores, considered in and of themselves, would suggest that there is no particular difference between Method A and Method B, one way or the other.
But now see what happens when we look concurrently at the covariate X, "primary suggestibility." As shown in Figure 17.2, a substantial portion of the variability among the induction scores (both groups combined) is associated with pre-existing individual differences in suggestibility. In brief: the greater the level of pre-existing receptivity to hypnotic induction, the greater the response to the induction method.
Figure 17.2. Correlation between X and Y: Both Groups CombinedT
r = +.803T
r2 = .645
The upshot of the correlation between X and Y is simply this:
Of all theTvariability that exists among the induction scores (Y), 65.4% can be traced to pre-existing individual differences in primary suggestibility (X).
And then there is also this to fold into the mix. The aim of the initial random sorting was to ensure that the two groups were starting out approximately equal, on average, with respect to all factors that might be relevant to how well they were likely to respond to any particular method of hypnotic induction. But random variability being the tricky thing that it is, they in fact end up with substantially different mean levels of the key covariate "primary suggestibility." The following abbreviated version of the data table will give you an overview of this portion of the picture.
|Method A||Method B||Xa||Ya||Xb||Yb||Means
Group A started out with the lower of the two mean levels of suggestibility (13.1 versus 18.0), yet ended up with the higher of the two mean levels of hypnotic induction (29.2 versus 28.1). Our analysis of covariance on this set of data will allow us to answer a
|Home||Click this link only if the present page does not appear in a frameset headed by the logo Concepts and Applications of Inferential Statistics|