Professional Documents
Culture Documents
Pearson coefficients of correlation and other of effect size, their interpretation is at times
derivative statistics (e.g., phi, point-biserial difficult. O'Grady (1982) described several
correlations) are used with considerable fre- problems associated with measures of ex-
quency in the analysis of psychological data. plained variance, noting that measurement
One reason for computing these statistics is error, features of the research design, and the
to determine the size or magnitude of the multiplicity of behavioral causes may limit
relation between two variables. When used the magnitude of such statistics. Sechrest and
for this purpose, it is common to use the Yeaton (1981, 1982) provided a detailed ex-
square of the correlation coefficient as an amination of effect size estimation issues.
index of the size of the relation. This index, They reviewed several methods of estimating
referred to by Guilford (1936) as the coeffi- the amount of explained variance and dem-
cient of determination, is interpreted as the onstrated how these procedures yield esti-
percentage of variance in one variable that is mates that are unlikely to be generalizable to
predicted or explained by the other. Amount situations that differ from the initial research
of explained variance is one way of expressing design and measurement context.
the degree to which some relation or phe- One may also question whether the amount
nomenon is present—what Cohen (1969) re- of variance accounted for is always an appro-
fers to as effect size. Effect size estimation is priate metric for assessing effect size. D'An-
an important concern, providing a basis for
evaluating the effectiveness of an experimental drade and Dart (1983) criticized the routine
treatment or perhaps even the success of a and common use of r2 as a measure of degree
basic research program. of association. Their arguments were based
Although measures of explained variance in part upon a rejection of the variance as
appear to be intuitively meaningful measures an appropriate measure of variation in a set
of scores. Instead of variance, which is based
upon the sum of squared deviations from the
The author thanks David Ems, Ralph D'Agostino, mean, D'Andrade and Dart suggested that
Kevin Lanning, Howard Terry, and two anonymous the sum of absolute deviations from the mean
reviewers for their comments on the manuscript. is a better measure of variation. They then
Requests for reprints should be sent to Daniel J. Ozer,
Department of Psychology, Boston University, 64 Cum- demonstrated that r, not r2, provides a close
mington Street, Boston, Massachusetts 02215. fit to the proportion of variation accounted
Psychological Bulletin, 1985, Vol. 97, No, 2, 307-315
Copyright 1985 by the American Psychological Association, toe, 0033-2909/85/$00,75
307
308 DANIEL J. OZER
(the determination of y by x through c), so native interpretation has been elaborated us-
fxv2 = T^rJ- and ing the common elements interpretation of
correlation and Tryon's anticipation of the
concept of suppression. In View A and View
This last result is revealing; it parallels B of Figure 2, both the standard and alter-
results obtained in the common elements native accounts agree that r2 provides the
interpretation. The correlation between two coefficient of determination. But what of the
variables is the geometric mean of the two situation depicted in Figure 1, Figure 2 (View
coefficients of determination, as shown in C), and Figure 3? Is the region of overlap of
Equations 5 and 1 1 . If Regions A and B in x and y equal to r or r2? This area of overlap,
Figure 3 are equal (i.e., if the variances of x Region C in Figure 3, is the covariance of x
and y are equal), then r^ = rcy2 = rxy. As in' and y. Regions A, B, and C each have been
the common elements interpretation, if x and defined as independent of one another, so
y are the same size, then r and not r 2 provides
the coefficient of determination. This is ap- and a2 = aa +
parent from Equation 1 1 : If r^2 - rcy2, then
O'Grady, K. E. (1982). Measures of explained variance: Sechrest, L., & Yeaton, W. J. (1981). Estimating magni-
Cautions and limitations. Psychological Bulletin, 92, tudes of experimental effects. JSAS Catalog of Selected
766-777. Documents in Psychology, 11. (Ms. No. 2355)
Pedhazur, E. J. (1982). Multiple regression in behavioral Sechrest, L., & Yeaton, W. J. (1982). Magnitudes of
research (2nd ed.). New York: Holt, Rinehart & Win- experimental effects in social science research. Evalu-
ston. ation Review, 6, 579-600.
Rosenthal, R., & Rubin, 0. B. (1979). A note on percent
variance explained as a measure of the importance of Tryon, R. C. (1929). The interpretation of the correlation
effects. Journal of Applied Social Psychology, 9, 395- coefficient. Psychological Review, 36, 419-445.
396. Wiggins, J. S. (1973). Personality and prediction: Principles
Rosenthal, R., & Rubin, D. B. (1982). A simple, general of personality assessment. Reading, MA: Addison-
purpose display of magnitude of experimental effect. Wesley.
Journal of Educational Psychology, 74, 166-169.
Rushton, J. P., Brainerd, C. J., & Pressley, M. (1983).
Behavioral development and construct validity: The
principle of aggregation. Psychological Bulletin, 94, Received October 11, 1983
18-38. Revision received June 7, 1984