and random sampling of treatment levels. As Koele phrased it, “Not only should there be
many observations per level, but also many levels per treatment (independent variable).
Experiments that have random factors with only two or three levels must be considered as
absurd as ttests on samples with two or three observations” (p. 516). This is important ad-
vice to keep in mind when you are considering random models. We will say more about this
in Chapter 13.
One final point should be made about power and design considerations. McClelland
(1997) has argued persuasively that with fixed variables we often use far more levels of
our independent variable than we need. For example, if he were running the Eysenck
(1974) experiment on recall as a function of levels of processing, I suspect that he would
run only the two extreme groups (Counting and Imagery), or perhaps three groups, adding
the Adjective condition. He would argue that to use five groups dilutes the effect across
four degrees of freedom. Similarly, he would probably use only the 0, 0.5 mg, and 2 mg
groups in the Conti and Musty (1984) study, putting the same number of subjects in the
0.5 mg group as in the other two conditions combined. I recommend this paper to those
who are concerned about maximizing power and good experimental design. It is impor-
tant and very readable.11.13 Computer Analyses
Exhibit 11.3 contains printout for the SPSS analysis of Everitt’s data on the treatment of
anorexic girls. Instead of choosing the one-way procedures from Analyze/Compare
Means/One-Way Anova,I have used the Analyze/General Linear Model/Univariate354 Chapter 11 Simple Analysis of Variance
Exhibit 11.2 G*Power estimation of power for Conti and Musty experiment