Multivariate Behavioral Research

ISSN: 0027-3171 (Print) 1532-7906 (Online) Journal homepage: [Link]

How Many Subjects Does It Take To Do A

Regression Analysis

Samuel B. Green

To cite this article: Samuel B. Green (1991) How Many Subjects Does It Take To Do
A Regression Analysis, Multivariate Behavioral Research, 26:3, 499-510, DOI: 10.1207/
s15327906mbr2603_7

To link to this article: [Link]

Published online: 10 Jun 2010.

Submit your article to this journal

Article views: 3059

Citing articles: 853 View citing articles

Full Terms & Conditions of access and use can be found at

[Link]
 Multivariate Behavioral Research, 26 (3), 499-510
Copyright O 1991, Lawrence Erlbaum Associates, Inc.

How Many Subjects Does It Take

To Do A Regression Analysis?
Samuel B. Green
University of Kansas

Numerous rules-of-thumb have been suggested for determining the minimum number of
subjects required to conduct multiple regression analyses. These rules-of-thumb are evaluated
by comparing their results against those based on power analyses for tests of hypotheses of
multiple and partial correlations. The results did not support the use of rules-of-thumb that
simply specify some constant (e.g., 100 subjects) as the minimum number of subjects or a
minimum ratio of number of subjects (N) to number of predictors (m). Some support was
obtained for a rule-of-thumb thatN ;r 50 + 8 m for the multiple correlation and N ;r 104 + m for
the partial correlation. However, the rule-of-thumb for the multiple correlation yields values too
large for N when rn ;c 7, and both rules-of-thumb assume all studies have a medium-size
relationship between criterion and predictors. Accordingly, a slightly more complex iule-of-
thumb is introduced that estimates minimum sample size as function of effect size as well as the
number of predictors. It is argued that researchers should use methods to determine sample size
that incorporate effect size.

"How many subjects does it take to do a regression analysis?" It sounds like

a line delivered by a comedian at a nightclub for applied statisticians. In fact, the
line would probably bring jeers from such an audience in that applied statisticians
routinely are asked some variant of this question by researchers and are forced to
respond with questions of their own, some of which have no good answers. In
particular, because many researchers want a sample size that ensures a reasonable
chance of rejecting null hypotheses involving regression parameters, applied
statisticians are likely topresent their responses within apower analytic framework.
From this perspective, sample size can be determined if three values are specified:
alpha, the probability of committing a Type I error (i.e., incorrectly rejecting the
null hypothesis); power, one minus the probability of making a Type I1 error (i.e.,
not rejecting a false null hypothesis); and effect size, the degree to which the
criterion variable is related to the predictor variables in the population. Although
alpha by tradition is set at .05, the choice of values for power and effect size is less
clear and, in some cases, seems rather arbitrary.
As an alternative to determining sample size based on power analytic
techniques, some individuals have chosen to offer rules-of-thumb for regression
analyses. These rules-of-thumb come in various forms. One form indicates that
the number of subjects, N, should always be equal to or greater than some constant,

MULTIVARIATE BEHAVIORAL RESEARCH 499

S. Green

A (i.e., N >A), while a second form stipulates a recommended minimum ratio B

of subjects-to-predictors (i.e., N 2 B m where m is the number of predictors).
Finally, a third form is a more general rule that encompasses the first two (i.e.,
N zA+Bm).
Tabachnickand Fidell(1989) suggest,with some hesitancy, in their multivariate
text that the minimum number of subjects for each predictor or independent
variable (IV) in a regression analysis should be 5-to-1. They state the following:

If either standard multiple or hierarchical regression is used, one would like to have
20 times more cases than IVs. That is, if youplan to include 5 IVs, it would be lovely
to measure 100 cases. In fact, because of the width of the errors of estimating
correlation with small samples, power may be unacceptably low no matter what the
cases-to-IVs ratio if you have fewer than 100 cases. However, a bare minimum
requirement is to have at least 5 times more cases than IVs - at least 25 cases if 5 IVs
are used. (pp. 128-129)

In the first edition of a multivariate primer by Harris (1975), he recommended that

the number of subjects N > 50 + m. In the revision of his text, he states,

... not too different from the Primer's earlier suggestion that N - m be > 50. More
common is a recommendation that the ratio of N to m be some number, for example,
10. I know of no systematic study of the ratio versus the difference between Nand m
as the important determinant of sampling stability of the regressionweights. However,
ratio rules break down for small values of m ... and there are hints in the literature that
the difference rule is more appropriate. (1985, p. 64.)

Harris' difference rule is an example of the general rule-of-thumb o f N r A + B m

with A = 50 and B = 1. Harris7call for a study to compare the ratio and difference
rules can be expanded and recast within the context of the present paper as a request
for a study to evaluate whether the A constant is a necessary component of the
general rule-of-thumb. If A is unnecessary, the general rule-of-thumb simplifies
to the ratio rule-of-thumb, N 2 B m.
Nunnally (1978) makes slightly different recommendations based on an
examination of an equation for determining an unbiased estimate of the population
squared multiple correlation coefficient (A2) from the sample squared multiple
correlation coefficient (R2):

(1) k2= 1- (1 - R2)[(N - 1)l(N - m)].

He states,

If there are only 2 o r 3 independent variables and no preselection is made among them,
100 or more subjects will provide a multiple correlation with little bias. In that case,

500 MULTIVARIATE BEHAVIORAL RESEARCH

 S. Green

if the number of independent variables is as large as9 or 10, itwill benecessary to have
from 300 to 400 subjects to prevent substantial bias. (p. 180)

Other individuals, based on a variety of justifications specify somewhat

different rules-of-thumb. For example, Marks (1966), as cited in Coolley and
Lohnes (1971), recommended a minimum of 200 subjects for any regression
analysis, while Schmidt (1971) suggested a minimum subject-to-predictor ratio
ranging in value from 15-to-1 to 25-to-1. Still other individuals, for example,
Pedhazur (1982), discuss rules-of-thumb, but make no general set of
recommendations themselves.
An important question is whether researchers who use these rules-of-thumb
have designed studies with adequate power. The answer to this question is
unknown; however, it is known that many empirical studies do have insufficient
power (for a review of this literature, see Cohen, 1988). Perhaps some portion of
these studies have used rules-of-thumbfor regression analysis (e.g., 5-to-1 subject-
to-predictor ratio). The purpose of the present study is to determine if researchers
who apply such rules-of-thumb are designing studies with low power. Sample
sizesfor regression analyseswill be determinedfollowingrecommendationsmade
by Cohen (1988) in the second edition of his bookon power analysis. These results
will be compared against various rules-of-thumb to judge their adequacy. It was
anticipated that none of the reviewed rules-of-thumb would be satisfactory in that
all of them ignore effect size and, accordingly, recommendations for more
complex rules-of-thumb would be required. It is hoped that this presentation will
encourage researchers to struggle with the difficult decisions required by power
analysiswith the reaiization that simplistic rules-of-thumb ignore the idiosyncratic
characteristics of research studies. In addition, it is hoped that the discussion will
encourage applied statisticians to develop methodsfor the determinationof sample
size that researchers find less esoteric and more useful (see Harris, in press, for one
such alternative).
The methods employed in this study impose a few limitations on the
conclusions. The sample sizes were based on power tables presented in Cohen
(1988) and are slightly different from those that would be obtained using other
power tables (for a more in-depth discussion, see Gatsonis & Sampson, 3 989).
Also, the power analyses assume that the regression analyses include a11predictors
and do not allow for the preselection of predictor variables (e.g., stepwise
regression). More generally, in order to have a focused presentation, it was
necessary to develop methods which suggest that decision making with regression
analysis is simpler than what should occur in practice. For example, researchers
need to consider carefully prior to determining sample size whether they can
reduce the number of predictor variables by forming a priori linear combinations
of two or more predictors and whether the variables are reliable and are highly

MULTIVARIATE BEHAVIORAL RESEARCH 501

S. Green

interrelated. Most importantly, researchers need to attend seriously to the issue of

estimating effect size based on their knowledge of a research area and the methods
of their study rather than assume a value for an effect size offered by Cohen or
discussed in this paper (see Cohen, 1988 for a detailed discussion).

Sample Sizes Required to Evaluate Multiple Correlation

Coeficients with a Power of .SO

I began by considering what sample size is required to evaluate the hypothesis

that the multiple correlation between the predictors and the quantitative variable
01) is equal to zero with a power of 3 0 . To conduct power analyses, choices of
values for alpha, power, and effect size were made:
1. Alpha was set at .05, the traditional level of significance.
2. Power was set at .80, a value proposed by Cohen (1988) as appropriate for
a wide range of behavioral research areas. He argued that the setting of power, the
probability of not committing aType I1 error, is to some extent arbitrary, but should
be dependent on the loss associated with making this error. Inasmuch as power
is partially a function of alpha, it also should be a function of the loss associated
with a Type I error. He suggested that typically across the behavioral sciences, a
4 to 1 ratio reflects the relative seriousness of a Type I error to a Type I1 error.
Consequently, when alpha is set equal to .05, the probability of a Type I1 error
should be 4 x .05 = .20 and power would be 1 - .20 = 3 0 .
3. Cohen (1988) stresses two indexes of effect size for regression anaIysis,f
and the better known R2. The two indexes are directly related:

Although Cohen argues that the choice of values for effect size (R2or f ) should
depend on the research area, he proposes, as a convention, R2sof .02, .13, and .26
(f2 of .02, .15, and .35) to serve as operational definitions for the descriptors small,
medium, and large,respectively. Cohen discussesthese selectionsrather extensively
and indicates that they agree with his subjective judgment of small, medium, and
large effect sizes obtained in behavioral sciences. These three values are used in
the current study.
Calculations as outlined by Cohen (1988) were performed to determine
sample sizes for the selected values of alpha, power, and effect size using tables
@p. 448-455) he provided for this purpose. The tables have entries in which the
number of predictors for regression analyses may assume 23 different values,
ranging from 1predictor to 120 predictors. While sample sizes were determined
for all 23 values, for sake of simplicity, only those for 15 of the 23 values are
presented in the first three columns of Table 1. Sample sizes are not presented in

502 MULTIVARIATE BEHAVIORAL RESEARCH

 S. Green

Table 1
Sample Size Required to Test the Hypothesis that the Population h4uStipSe
Correlation Equals Zero with a Power of .80 (Alpha = .05>

Sample sizes based Sample sizes based

on power analysis on new rule-of-thumb
Number of --
predictors Effect size Effect size
-.
Small Medium Large Small Medium Large
- - - ---
1 390 53 24 400 53 23
2 481 66 30 475 63 27
3 547 76 35 545 73 31
4 599 84 39 610 81 35
5 645 91 42 670 89 38
6 686 97 46 725 97 41
7 726 102 48 775 103 44
8 757 108 51 820 109 47
9 788 113 54 860 115 49
10 844 117 56 895 119 51
15 952 138 67 1045 139 60
20 1066 156 77 1195 159 68
30 1247 187 94 1495 199 85
40 1407 213 110 1795 239 103
--

Table 1 for regression analyses with greater than 40 predictors because data
analyses with this many predictors were assumed to be relatively infrequent.
Harris' (1975) iule-of-thumb that N s 50 + m reflects to some extent the results
given in Table 1. A rule-of-thumb, by definition, should require minimal
complexity and, therefore, in the determination of sample size, it might be argued
that the rule-of-thumb should be developed to give accurate answers for typical
studies rather than for all studies. According to Cohen (1988), a typical study in
the behavioral sciences would have a medium effect size, and the minimum
number of subjects required for studies with this effect size according to Table 1
is 53 (when m = I), similar to the minimum number based on the rule-of-thumb
(50 + m = 51). Also, both the tabled sample sizes and those based on Harris' rule-
of-thumb indicate that the number of subjects increases as the number of predictors
increases. However, as the number of predictors increases, the differences in the
sample sizes from Harris' rule-of-thumb and those from power analyses become
larger. One possible revision would be that the minimum number of subjects

MULTIVARIATE BEHAVIORAL RESEARCH 503

S. Green

should be 50 + 8 m. This rule-of-thumb is fairly accurate for medium effect-size

studies with a small number of predictors (m c 7), but becomes increasingly more
conservative with additional predictors (i.e., the rule-of-thumb overestimates the
required sample size according to power analyses).
Even though an argument can be made for the rule-of-thumb that N 2 50 + 8
m, it has its drawbacks. This rule, in particular, and traditional rules-of-thumb, in
general, have two problems associated with them. First, their mathematical
simplicity does not offset their inability to take into account effect size. The
number of subjects required for conducting a regression analysis with a small
effect size is dramatically different from the number of subjects required for an
analysis with a large effect size. Accordingly, it is important that researchers
consider the magnitude of the effect they want to detect when determining the
number of subjects to include in their study rather than view their study as a typical
study in the behavioral sciences and base their sample size on amedium effect size.
Second, even if effect size is held constant, the number of subjects is not linearIy
related to the number of predictors as indicated by the general form of the rule-of-
thumb. Instead, the number of subjects required for each additional predictor is
greater with few predictors than with many predictors.
In order to better reflect the results of power analyses, a new two-step rule-of-
thumb was developedbased on Cohen's (1988) power analytic approach. Cohen's
procedure requires the determination of lambda, which is obtained by entering
tables @p. 448-455) in his book. With the new rule-of-thumb, lambda is
approximated by L in step 1. Once L is determined, N can be computed in step 2
using a simple equation given by Cohen. The two steps of the rule-of-thumb are
as follows:
1. Compute L. L is 8 with a single predictor. For regression analyses with 2
through 10predictors,L increases for each additional predictor by 1.5,1.4,1.3,1.2,
1.l, 1.O, .9, .8, and .7, respectively. Algebraically, for m c 11,

For each additional predictor past 10, L increases .6.

2. Compute required minimum sample size, N. N z L l P where P = R2 /
(1 - R2).
For example, with 7 predictors, L = 8.0 + 1.5 + 1.4 + 1.3 + 1.2 + 1.1+ 1.0 =
15.5 (or using Equation 3,6.4 + 1.65 (7) - .05 (72) = 15.5). For a medium effect
size (R2of .13 and? of .15), N z 15.5 / .15 = 103. This answer is close to thevalue
of 102, the sample size derived from conducting a power analysis (see Table 1).
If this new rule-of-thumb is to replace current rules-of-thumb, it needs to be
simple mathematically and easy to remember, as well as relatively accurate.
Although the rule is more complicated than traditional rules-of-thumb, it is not

504 MULTIVARIATE BEHAVIORAL RESEARCH

 S. Green

complex. Values for L in step 1 do progress in a systematic fashion and the

equation in step 2 is simple. Also, the results of the new rule-of-thumb agree
moderately well with the sample sizes determined by power analytic methods, as
can be observed by comparing the last three columns of Table 1with the first three
columns. It is most accurate for moderate effect sizes; the sample sizes based on
the rule-of-thumb and on power analyses never differed by more than 5% from
each other if m s 20. These results compare favorably to those based on the rule-
of-thumb thatN 250 + 8rn; the sample sizesbased on this tradjltionalmle-of-thumb
and on power analyses differed by more than 5%from each other when rn = 1and
when rn B 8. For small effect sizes, the new rule-of-thumb is reasonably accurate
when rn is small; however, it produces an N that is consistently larger than the one
based on power analyses as m increases, with the two differing by 10%or more
when m 2 20. For a large effect size, the new rule-of-thumb always underestimates
the sample sizes based on power analyses, although the degree of underestimation
is not great when rn is small.
In summary, no specific minimum number of subjects or minimum ratio of
subjects-to-predictors was supported. The general rule-of-thumb of N 2 50 + 8 rn
was seen as more accurate than these simpler rules-of-thumb. However, a new
rule-of-thumb was found to have even greater accuracy, although it requires that
researchers estimate the effect size that they wish to detect. An argument could
be made that this new two-step rule-of-thumb is sufficiently complex that
researchers could almost as easily do a power analysis to determine their sample
sizes. Researchers are encouraged to conduct such power analyses..

Sample Sizes Required to Evaluate Partial Correlation

Coefficients with a Power of .80

When conducting multiple regression analyses, researchers typically evaluate

not only hypotheses that population multiplecorrelations areequal to zero, but also
hypotheses that population semi-partial correlations are equal to zero andlor
population partial correlations are equal to zero (Cohen & Cohen, 1983). The test
statistics to evaluate these latter two hypotheses are identical; therefore, discussion
will center around a test for just one of these coefficients, the partial correlation.
Researchers try to understand their multiple regression results in part by evaluating
partial correlations between a criterion variable, y, and a set of predictor variables,
partialling out the effect of a second set of predictor variables from both y and the
first set of predictors. For example, in investigating the relationship between
college grade point average (CGPA) and the predictors of high school grade point
average (HSGPA) and Scholastic Aptitude Tests (SATs), researchers might focus
their attention on three correlations: the multiple correlation of CGPA with
HSGPA and SATs, the partial correlation between CGPA and HSGPA holding

MULTIVARIATE BEHAVIORAL RESEARCH 505

S. Green

constant (partialling out) SATs, and the partial correlation between CGPA and
SATs holding constant HSGPA. The latter two partial correlations are evaluated
to determine whether one of the predictors enhances the predictability of CGPA
if the other predictor is held constant or whether a single predictor is sufficient.
Because tests of partial correlations are frequently conducted, the power of these
tests should be considered when evaluating sample-size rules-of-thumb for
regression analyses.
How to incorporate the partial correlation into our assessment of these rules-
of-thumb was problematic. For any multiple regression analysis, a researcher
might calculate a number of partial correlations. For example, with three
predictors, twelve partial correlations could be computed: six correlations relating
y to one of the predictors holding a second predictor constant, three correlations
relating y to two of the predictors holding a third predictor constant, and three
correlations relating y to one of the predictors holding the other two predictors
constant. The number of subjects required for these analyses based on power
analysis differs.
It was decided to focus on one type of partial correlation, the correlation
between y and a predictor holding all other predictors constant, because the
significance test for this correlation is often performed by researchers. However,
the test may not be labeled as evaluating this partial correlation in that tests of other
statistics from regression analyses produce the same results as the one for this
partial correlation. Specifically, the test that the partial correlation betweeny and
a predictor x, holding constant all other predictors (x2through xm) is equal to zero
is equivalent to a test that the weight for x, in a multiple regression equation
containing all rn predictors is equal to zero or that the variance ofy accounted for
by x, over and above x2through xm(the squared semipartial or unique variance) is
equal to zero. Most computer regression programs include tests of one or more of
these statistics. For example, the General Linear Models Procedure (PROC GLM)
in the Statistical Analysis System (SAS, 1985) gives tests for the regression
weights and the unique variances (referred to as Type 111 sums of squares).
Undoubtedly, researchers report these tests in part because they are standard output
of regression packages.

Sample Sizes Based on Effect Sizes Given by Cohen

To determine the sample size required to reject the hypothesis that the partial
correlation is equal to zero, values for alpha, power, and effect size that were used
to determine sample size for the multiple correlation were initially chosen. Alpha
was set at .05; power was set at 30; and the squared partial correlation was set at
.02, .13, or .26. Cohen (1988) indicates that .02, .13, and .26 represent small,
medium, and large effect sizes for not only the squared multiple correlation, but

506 MULTIVARIATE BEHAVIORAL RESEARCH

 S . Green

also for the squared multiple partial correlation, of which the scluared partial
correIation is the special case being considered here. Using Cohen7s(1988) power
analysis approach, the sample sizes for the three effect sizes were determined. The
obtained sample sizes may be found using the following simple rule:
1. Determine the number of subjects required to test the hypothesis that the
population correlation between y and a single predictor is equal to 0 (first row of
Table 1).
2. Add m - 1 to the result of step 1.
When the squared partial correlation between y and a predictor holding
(m - 1) other predictors constant is .02, the required number of [Link] is 390 +
(m - 1); when it is .13, N 2 53 + (m - 1); and when it is .26, N 8 24 + (m - 1).
The rule-of-thumb recommended by Harris (1975) offersagood approximation
to the sample sizes obtained from power analyses when the squared partial
correlation is .13, Cohen7sdefinition of a medium effect size. In particular, Harris
suggested that N > 50 + m, while the power analysis for an effect size of .13
indicated that N 2 52 + m. However, within this perspective, the rule is rather
inflexible in that it allows for only a single effect size, a squared partial correlation
of .13. Also, as discussed below, this rule-of-thumb may not be appropriate even
for the typical behavioral science study because .13 may be too large a value for
a medium effect size.

Sample Sizes Based on Revised Eflect Sizes

Cohen (1988) suggested that effect sizes should be greater for correlational
analyses which may have multiple predictors, such as multiple: and partial
correlations, than for zero-order correlational analyses. Therefore, he suggested
small, medium and large effect sizes of .02, .13, and .26 for squared multiple and
multiple partial correlationsand .0 1,.09, and .25 for squaredzero-ordercorrelations.
However, the partial correlation of interest here is not amultiplepartialcorrelation,
but is a partial correlation between y and a single predictor, holding all other
predictors constant. Even the values for azero-order correlation of .Ol, .09 and .25
might be too large in that a partial correlation between y and x controlling for all
other predictors frequently will be somewhat smaller than a zero-order correlation
between y and x. The zero-order correlation may exceed the partial correlation if
suppression occurs among the predictors (for a discussion of suppression, see
Cohen & Cohen, 6983; Pedhazur, 1982) or if the correlation between x and the
other predictors is close to zero (particularly if the multiple correlation is large, for
example, .70 or greater). However, some researchers (e.g., Nunnally, 1978) have
suggested that suppression occurs rarely in practice. Given that suppressors are
rare and given that in many research studies predictor variables are substantially
intercorrelated, a partial correlation between two variables is likely to be smaller

MULTIVARIATE BEHAVIORAL RESEARCH 507

S. Green

than the zero-order correlation between them. Therefore, a medium effect size for
a squared partial correlation might be redefined to be .07 rather than .13, Cohen's
suggested value for a squared partial correlation, or .09, his suggested value for a
squared zero-order correlation.
If the squared partial correlation of interest is .07, the required sample size for
regression analyses using Cohen's power tables is N r 104 + m. It should be noted
that a similar value may be obtained by initially using the two-step rule-of-thumb
presented for multiple correlations: (a) L = 8 with a single predictor and (b) f2 =
.07 l ( 1 - .07) = . O X , substituting partial correlations for multiple correlations in
theF equation, and L If = (8 I .075) = 107. Based on these calculations, N r 107
+ (m - 1) = 106 + m. Regardless of the approach, the choice of values for an effect
size has a dramatic effect. Based on power calculations, the minimum sample size
is 52 + m if a medium effect size is .13 or 104 + m if it is .07.
Although the resulting equations for sample size with the partial correlation
all have the general rule-of-thumb form that N 2 A + m, no one value for A is
satisfactory in that A is a function of effect size (as well as alpha and power).
Researchers should be allowed to determine the value for the partiaI correlation and
use it to establish sample size using power analysis or, if unable to access power
tables, the rule-of-thumb that N r (8 l f ) + (m - 1).

Conclusion

Researchers who use a rule-of-thumb rather than power analyses are trading
simplicity of use for accuracy and specificity of response. Traditional rules-of-
thumb that give a minimum number of subjects or a minimum ratio of subjects-
to-predictors are the simplest to use, but sample sizes based on them are rarely
congruent with power analyses. Slightly more complex rules-of-thumb of the
form that N r A + B m show slightly better agreement with power analysis results,
but only if researchers are interested in a typical effect size and hypotheses
concerning a multiple correlation and/or one type of partial correlation (as well as
an alpha of .05 and power of .80). More specifically, the power for a test of a
multiple correlation with a medium effect size is approximately .80 or greater if
N 2 50 + 8 rn and the power for a test of a medium-sized partial correlation between
y and a predictor holding all other predictors constant is approximately .80 if N z
104 + m. A more complex rule-of-thumb is necessary in order for minimum
sample size to be a function of effect size as well as the number of predictors for
both the multiple correlation (i.e., N 2 L I f ) and the partial correlation (i.e., N 2
(8 I?) + (m -1)). Greater accuracy and flexibility can be gained beyond these ruIes-
of-thumb by researchers conducting power analyses. Because researchers should
estimate effect size based on the characteristics of their study as opposed to
assuming some conventional value presumed to be appropriate for the typical

508 MULTIVARIATE BEHAVIORAL RESEARCH

 [Link]
study in behavioral sciences, it is recommended that researchers use either the
more complex rules-of-thumb that directly incorporate effect sizle or conduct
power analyses.
Researchers who evaluate hypotheses concerning partial ca~rrelationsin
addition to hypotheses about multiple correlations should determine minimum
sample sizes for both types of tests and select the largest of these minimum sample
sizes as the N for their study. For example, if a researcher was planning to conduct
a regression analysis with 5 predictors and an estimated multiple correlatioln of .50
V2 = .33) and estimated partial correlations of .40 V2 = .19), he/she rnight use the
rule-of-thumb for the multiple correlation that N 2 L lf = 13.4 / .33 = 40.2 and the
AI-atl Pr z-S,l,a-r! &n-- 1)=81'.l9.f ( 5 - ij
rule-nf -Lhr.1& hr.t,[Link] 6ar~bthr
= 46. Consequently, this researcher should conduct the study with a minimum of
46 subjects.
In conclusion, researchers who use traditional rules-of-thumb are likely to
design studies that have insufficient power because of too few subjects or
excessive power because of too many subjects. For example, researclhers who use
the rule-of-thumb of 5 subjects for each predictor (Tabachnick and IFidell,, 1989)
are conducting studies that have a high probability of not yielding significance
unless the effect size is extremely large. In contrast, researchers who follow the
recommendationsof Nunnally (1978) and collect data on aminimum of 300 or 400
subjects have likely collected more data than necessary if the number of predictors
are few and the effect size of medium value or greater. Of course, larger sample
sizes might be justified on issues unrelated to power. These other issues must be
considered on their own merits. For example, Nunnally (1978) made his
recommendation based on the amount of bias (shrinkage) in sample multiple
correlations rather than power. However, it is unclear why researchers should
collect data from 300 or 400 subjects to minimize shrinkage rather than determine
sample size based on power analyses which may yield Ns substantially less than
300 or 400. These researchers may account for the bias that occurs with smaller
sample sizes by using a shrinkage equation such as Equation 1.

References

Cohen, J. (1988). Statisticalpower analysis for the behavioral sciences (2nd ed.). New York:
Academic.
Cohen, J., & Cohen, P. (1983). Applied multiple regressionlcorrelation analysis for the
behavioral science. Hillsdale, NJ: Lawrence Erlbaum.
Cooley, W. W., & Lohnes, P. R. (1971). Multivariate data analysis. New York: Wiley.
Gatsonis, C., & Sampson, A. R. (1989). Multiple correlation: Exactpower and sample size
calculations. Psychological Bulletin, 106,516-524.
Harris, R. J. (1975). A primer of multivariate statistics. New York: Academic.
Harris, R. J. (1985). A primer of multivariate statistics (2nd ed.). Orlando, FL: Academic.

MULTIVARIATE BEHAVIORAL RESEARCH 509

S. Green
Harris, R. J. (In press). MIDS (Minimally Important Difference Significant) criterion for sample
size. Journal of Educational Statistics.
Marks, M. R. (1966, September). Two kinds of regression weights that are better than betas in
crossedsamples. Paper presented atthe meeting of the American Psychological Association.
Nunnally, J. (1978). Psychometric theory (2nd ed.). New York: McGraw-Hill.
Pedhazur, E. J. (1982). Multiple regression in behavioral research (2nd ed.). New York: Holt
Rinehart Winston.
SAS Institute Inc. (1985). SAS user's guide: Statistics (Version 5). Cary, NC: SAS Institute
Inc.
Schmidt, F. L. (1971). The relative efficiency of regression in simple unit predictor weights in
applied differential psychology. Educational and Psychological Measurement, 31, 699-
714.
Tabachnick, B. G., & Fidell, L. S. (1989). Using multivariate statistics (2nd. ed.). Cambridge,
MA: Harper & Row.

MULTIVARIATE BEHAVIORAL RESEARCH