J Clin Epidemiol Vol. 49, No. 4, pp. 435-439, 1996 0895-4356/96/$15.

00
Copyright 0 1996 Elsevier Science Inc. SSDI 08954356(95)00511-2

ELSEVIER

Statistical Considerations in the

Design and Analysis of Community Intervention Trials

DEPARTMENT OF EPIDEMIOLOGY AND BIOSTATISTICS,

THE UNIVERSITY OF WESTERN ONTARIO, LONDON, ONTARIO, CANADA

ABSTRACT. Community intervention trials are often characterized by the allocation of intact social units
to different intervention groups. The assessment of adequate sample size for such trials must take into
account the statistical dependencies among responses observed within an allocated unit. However, the small
numbers of units typically involved in such trials imply that many methods of analysis that have been
proposed for analyzing correlated data, particularly in the case of a dichotomous outcome variable, are not
applicable to such designs. In this article we investigate this issue and determine the minimum number of
units required per group, for the case of both a dichotomous and a continuous outcome variable, needed to
provide adequate statistical power for detecting various levels of treatment effect. The use of significance
testing as a method of detecting intracluster correlation is also investigated, and, in general, discouraged.
J CLIN EPIDEMIOL 49;4:435-439, 1996.
KEY WORDS. Cluster randomization, power, sample size

INTRODUCTION COMPARISON OF MEANS

There has been substantial interest in community-based approaches to
disease prevention and health promotion. Koepsell et al. [l] have re-
viewed many of the special issues that arise in such studies. Two of
these issues are (1) the assignment of entire communities to interven-
tion and control groups and (2) the ability to study only a small number
of communities. These issues often complicate the task of designing a
study that has acceptable statistical power. In this article, we consider
the sample size requirements for such studies in terms of the validity
and power of the statistical methods that might be applied. We focus
on allocation methods based on random assignment of communities
to different intervention groups when outcomes are measured at a
single time point. Many of our conclusions apply equally to nonran-
domized comparisons. We deal with both the case of continuous and
dichotomous outcome variables, presenting exact sample size require-
ments for the former case and approximate requirements for the latter.
The practice of using significance tests to detect the presence of in-
tracluster correlation in community intervention trials is also dis-
cussed.
Sample size requirements for trials randomizing intact clusters of
individuals have been discussed by several authors [2-51. However,
relatively little attention has been given to this issue in the context
of community trials, where the unique design features pose special
problems.
The term community in the context of this article will be inter-
preted broadly, so that it encompasses a wide spectrum of intact social
units including, for example, schools, factories, worksites, medical
practices, and hospital wards. Examples of such intervention studies
include community-based heart disease trials [6], trials to promote
smoking cessation [7], school-based intervention studies 181,worksite
programs [9, lo], and studies randomizing medical practices [ 111.
Received in revised form 16 May 1995
We assume in this section that the primary end point for a community
intervention trial randomizing intact clusters to one of two groups is
continuous. Thus the data will consist of observed mean outcomes
computed on each of mi clusters observed in group i, i = 1, 2; with
nb denoting the size of the jth cluster in group i, j = 1, 2 . . . , m,
and x(e the cluster-specific mean. Note that n,, denotes the actual
number of individuals in a cluster that are followed up with respect to
the end point of interest; because of subsampling this number may in
fact be less than the original cluster size. We further define the overall
means in group i by x(, = Z~In&/ni where n, = C>lnl,.
If the number of clusters per group tn, is large the difference x, - &
may be tested for significance by a simple adjustment to the standard
two-sample normal deviate test for comparing two means [2]. This
adjustment is based on an empirical estimate of the within-cluster
correlation coefficient. However, if the number of clusters per group
is small relative to the average cluster size, the required adjustment
factor cannot be estimated reliably [12] and an alternative approach
must be developed. A natural approach is to perform a standard two-
sample t test comparing the cluster-specific means. If the computed
value of t exceeds the lOO( 1 - cy/2)% critical value with m, + m,
- 2 degrees of freedom, one would reject the null hypothesis of no
intervention effect at the lOOa% level of significance. To take into
account the variable cluster sizes, a weighted version of the t test may
also be developed [ 131.
In computing the power associated with this test, we restrict our-
selves to the case of an equal number of clusters per group m having
constant size n. Denote the observation taken on the kth individual
within the jth cluster assigned to group i by yjk, k = 1, 2, . . . , n;j
= 1, 2, . . . , m; i = 1, 2. The cluster-specific mean & = Z;=,X,,,/
n is assumed to be normally distributed with mean II,, i = 1, 2, and
variance Var(&) = (&/n)[l + (n - l)~], where a2 = Var(Xgk) and
p is the intracluster correlation coefficient. The parameter p is most
simply interpreted as the ordinary product moment correlation be-
tween any two observations (Xi,,, Xgl) in the same cluster. It may also
436 Donner and Klar

be regarded as the proportion of total variance in the observations that
may be attributed to differences between clusters. The exact power of
the t test may then be computed using standard results 1141 while
recognizing that the variance of the analytic unit is given by Var(X,,).
Using the Cohen [14] definition of effect size, given in the present
context as the magnitude of d = 1~~ - /.$a, the power of the t test
was computed at the following parameter values:
d = 0.20 (small), 0.50 (medium)
p = 0.001, 0.005, 0.01
m = 3, 6,9, 12
n = 100,300,500
Further details are given in the Appendix. Small effect sizes fre-
quently characterize community intervention trials, especially when
the intervention is multidimensional and the subjects own attitudes,
motivation, and commitment are critically important to its success.
Cohen [14] provides further discussion of the effect size d and its
interpretation.
The values of p listed are fairly typical of those arising in community
intervention trials. For example, the average intracluster correlation
coefficient for the number of cigarettes smoked per week was 0.02,
using results reported by Murray et al. [15] from their review of school-
based adolescent smoking prevention studies. Estimates of intracluster
correlation reported by Hannan et al. [16] using data from the Minne-
sota Heart Health Program, a community intervention study, were
generally in the range 0.002-0.012 for a number of heart disease risk
factors including measures of blood pressure, cholesterol, and body
mass index. Values of comparable size have also been reported for
binary outcomes (e.g., smoking incidence or prevalence [12,15-181).
Although these reported values appear small, their impact on Var( X,),
and thus on statistical power, can be substantial for the large cluster
sizes typical of community trials.
The results of the power calculations are shown in Table 1. They
show that trials with as few as three units per group are sufficiently
large to detect medium effect sizes, while the detection of small
effect sizes requires from six units per group (p = 0.001) to nine units
per group (p = 0.01) in order to achieve a desired power of at least
80% or more. These results also demonstrate that, for a fixed number
of subjects, it is preferable in terms of power to increase the number
of clusters relative to the average cluster size. For example, at p =
0.001, three clusters of size 300 randomized to each of two intervention
groups provide a power of approximately 80% for detecting a small
effect size. Nine clusters of size 100, on the other hand, provide a
power exceeding 95% for detecting the same size of effect.
This general result is also reflected by considering the variance of
an estimated treatment mean, given by &[l + (n - l)p]l(mn). It is
clear from this expression that indefinitely increasing n for fixed m
does not reduce this variance to zero.
COMPARISON OF PROPORTIONS
The previous section dealt with sample size requirements for the com-
parison of means in community intervention studies. However, the
primary analysis for such studies will frequently consist of a comparison
of observed event rates computed
1 for each of mi clusters observed
in group i, i = 1, 2, with Pij = Y& the corresponding event rate,
where Y, is the number of observed events for the jth cluster in group
1.
Several procedures have been developed for testing the equality of
event rates in the presence of clustering [ 191, all of which are approxi-
mate. They include the adjusted chi-square approach [20], the ratio
estimate approach [21], and the generalized estimating equations ap-
proach [22]. However, extensive simulation studies [20,23,24] have
shown that these procedures are generally unsuitable for designs in
which the number of clusters per group is small, a defining characteris-
tic of most community intervention trials.
Alternatively, exact tests of significance could be calculated using
randomization theory [25-271. Such tests are constructed by consider-
ing the number of different ways in which cluster-specific event rates
could be permuted between treatment groups while maintaining the
same number of clusters per group. The test statistic would then be
calculated for each permutation of the data. The two-tailed statistical
significance of the test is equal to the proportion of test statistics found
using the permuted data, which are at least as large as the absolute
value of the test statistic found using the observed data.
An important limitation of such tests when there are few clusters per
treatment group is that it is not always possible to obtain statistically
significant results. For example, there are only 20 unique permutations
possible when there are 3 clusters per treatment group [25]. The mini-
mum two-sided p value in this case is only 0.1 ( = 2120) because the
accompanying test statistics for pairs of mirror image permutations
are equal in absolute value. At least four clusters per treatment group
are needed to be able to obtain p values less than 0.05.
There is, however, strong evidence that the standard twosample t
test may be applied to the cluster-specific event rates arising from
such designs, even though the principal assumptions intrinsic to the
procedure, such as normality and homogeneity of variance, are satisfied
only approximately. Such evidence is available from simulation studies
[28,29] that demonstrate the robustness of the t test to moderate viola-
tions of these assumptions. Theoretical justification for the robustness
of the two-sample t test, especially when there are equal numbers of
observations in each treatment group, is also available 130,311.
To investigate further the validity and power of the t test in the
context of community intervention trials, we performed a simulation
study in which the binary outcome scores in each cluster were gener-
ated from a beta-binomial distribution with intracluster correlation
coefficient p. The beta-binomial distribution is an extension of the
binomial distribution, which has been used to model correlated binary

TABLE 1. Exact power of t test for comparing means in community intervention trials
p = 0.001 P = 0.005 p = 0.01
m m m
n d 3 6 9 12 3 6 9 12 3 6 9 12

100 0.20 0.43 0.85 0.97 0.99 0.34 0.72 0.90 0.97 0.27 0.60 0.81 0.91
0.50 0.99 1.0 1.0 1.0 0.96 1.0 1.0 1.0 0.89 1.0 1.0 1.0
300 0.20 0.79 1.0 1.0 1.0 0.53 0.93 0.99 1.0 0.37 0.77 0.93 0.98
0.50 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 0.97 1.0 1.0 1.0
500 0.20 0.91 1.0 1.0 1.0 0.60 0.96 1.0 1.0 0.40 0.81 0.95 0.99
0.50 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 0.98 1.0 1.0 1.0
 Statistical Considerations in Community Intervention Trials 437

data in a variety of experimental settings [32,33]. It reduces to the
binomial distribution as the between-cluster variability, and hence
the intracluster correlation coefficient, tends toward zero. A positive
correlation among the binary responses in a cluster is induced by as-
suming that the between-cluster variability in success rates follows a
beta distribution. The distribution of observations in each cluster was
characterized by an underlying success rate for the experimental and
control groups given by Pi and Pz, respectively.
In investigating the properties of the t test the specific objectives
were to (1) evaluate the empirical type I error rate at a nominal type
I error rate of (Y = 0.05 (two-sided) and (2) evaluate the associated
empirical power. The empirical comparisons of type I error (Pi = P2
= 0.1, 0.3) were performed at m = 3, 6, 9 and 12 for n = 100,
300, 500 over 10,000 iterations. However, to allow comparability of
the power results with that of the continuous case, the calculations
were performed at effect size levels corresponding to those presented
in Table 1. Defining the effect size 1141 as d = 2larcsin(Pr) -
arcsin(P,) the values of (PI,P2) were taken as (0.1, 0.168) and
(0.1, 0.293), corresponding, respectively, to d = 0.20 (small) and
0.50 (medium).
The empirical significance levels shown in Table 2 are in all in-
stances very close to the nominal level of 0.05. Thus it may be con-
cluded that the standard two-sample t test may be applied to the
comparison of event rates arising from community intervention trials
when there are as few as m = 3 clusters per group. This conclusion
may not hold if the variation in cluster sizes within or between inter-
vention groups is extreme. However, in this case a design involving
some prestratification or matching on cluster size would be strongly
preferred to a design involving simple randomization only.
The power results (not shown) never differed from those shown in
Table 1 by more than 0.01. This near equivalence occurs because
the combination of large clusters and small intracluster correlation
coefficients implies that the event rates in each cluster are close to
normally distributed. Furthermore, because there was no variability in
cluster size, the assumptions of homogeneity of variance are satisfied,
at least under the null hypothesis of equal intervention effects.
SIGNIFICANCE TESTING AS A METHOD
OF DETECTING INTRACLUSTER CORRELATION
It is occasionally implied in the epidemiological and clinical literature
that the existence of intracluster correlation may be detected or ruled
out on the basis of a pretrial test of significance. If the outcome variable
is continuous, this would essentially involve performing a one-way
analysis of variance among and within a sample of clusters typical of
those to be involved in the intervention study. A significant value of
the variance ratio statistic F would imply that the variability among
clusters exceeds that within clusters, that is, that the value of the
intracluster correlation coefhcient p is positive. Suppose we are inter-
ested in testing Ha: p = 0 vs. H,: p = pi in a preliminary sample of
m clusters, each of size n. Adopting a one-way random effects model,
the power of this test at the lOOa% significance level is given 1311by
1 - p = Pr{F > CF,}, where C = (1 - p,)l[l + (n - l)pi] and
F, denotes the 100a percentile point of the F distribution with m -
1 and m(n - 1) degrees of freedom.
Table 3 presents the calculated power of various values of m, n, and
p. It is clear from these results that even if p is as high as 0.01, at
least six clusters of size 500 or nine clusters of size 300 are needed to
provide reasonable power (~80%) for this test. However, for large
clusters such as communities, p is likely to be considerably smaller
than 0.01, and in this case the sample size requirements are much
larger. Note also that small values of p, combined with large average
cluster sizes, can yield design effects, which, if unaccounted for, can
seriously disturb the validity of standard statistical procedures [19,34].
The calculations presented in Table 3 refer only to the case of
a continuous outcome variable. However, corresponding power for
detecting intramuscular correlation with respect to a binary outcome
variable would be expected to be even lower [35].
Our recommendation on the basis of these results is that investiga-
tors randomly assigning intact clusters to treatment groups should in-
herently assume the existence of intracluster correlation rather than
attempt to rule it out using statistical testing procedures. This is consis-
tent with the well-accepted strategy of adapting an analytical strategy
to the study design. An additional concern is that the low power of
such procedures implies that a nonsignificant result cannot be taken as
assurance that the intracluster correlation is of negligible importance.
A related issue is the use of estimates of intracluster correlation
found by previous investigators to assesssample size requirements for
a subsequent study. Feng and Grizzle 1121 have demonstrated that
the variability associated with a single point estimate of intracluster
correlation is particularly high in the context of community trials,
where such estimates are frequently obtained from only a small number
of clusters. Their results suggest that a sensitivity analysis be per-
formed, in which the implication of using a range of values of p for
sample size estimation is explored. Feng and Grizzle (121 discuss the
use of simulation for this purpose.
EXAMPLE
To illustrate the calculation of power requited when designing a com-
munity intervention trial we consider data from a school-based, adoles-
cent smoking prevention trial [8,19] in which 12 schools were ran-
domly assigned to each of 4 conditions, including 3 intervention
conditions and a control condition (existing curriculum). One aim of

TABLE 2. Empirical significance levels associated with two-sample t test as applied to comparison of event rates in community
intervention trials

p = 0.001 P = 0.005 P = 0.01

m m m
n PI =Pr 3 6 9 12 3 6 9 12 3 6 9 12

100 0.1 0.050 0.052 0.050 0.049 0.054 0.051 0.048 0.049 0.053 0.050 0.051 0.050
0.3 0.053 0.051 0.049 0.049 0.054 0.051 0.049 0.049 0.054 0.051 0.049 0.050
300 0.1 0.055 0.050 0.050 0.049 0.053 0.051 0.050 0.049 0.053 0.049 0.050 0.050
0.3 0.053 0.051 0.048 0.048 0.054 0.051 0.049 0.049 0.054 0.052 0.049 0.048
500 0.1 0.054 0.052 0.050 0.048 0.053 0.051 0.050 0.049 0.053 0.050 0.049 0.049
0.3 0.054 0.051 0.049 0.049 0.054 0.052 0.049 0.049 0.054 0.052 0.049 0.049
The significance levels are computed at a = 0.05 (two sided) over 10,000 independent iterations (see text for further explanation).
438 Donnet and Klat

TABLE 3. Power associated with testing &: p = 0 vs. 29,: p = p1 under a one-way random effects model
P = 0.001 p = 0.005 p = 0.01
m m m
n 3 6 9 12 3 6 9 12 3 6 9 12
100 0.07 0.07 0.08 0.08 0.14 0.19 0.24 0.29 0.22 0.35 0.46 0.55
300 0.10 0.13 0.15 0.18 0.30 0.49 0.63 0.73 0.48 0.74 0.87 0.94
500 0.14 0.19 0.24 0.29 0.43 0.68 0.82 0.90 0.61 0.87 0.96 0.99

the study was to compare the effect of Smoke Free Generation intet-
vention with the existing curriculum on the proportion of children
who report using smokeless tobacco after 2 years of follow-up [19].
The mean event rates in the experimental and control groups were,
respectively, 0.039 and 0.060, with corresponding standard deviations
given by 0.026 and 0.035. The resulting value of t, with 22 degrees
of freedom, was t = 1.66 (p = 0.11).
Investigators might find such data helpful in planning future studies
designed to compare new and mote promising smoking prevention
strategies. The observed effect size is given by d = 2latcsin(P,) - to adolescent tobacco-use prevention. Prev Med 1992; 21: 449-472.
atcsin(Pz)l = 21atcsin(0.039)12 - atcsin(0.060)2~ = 0.1, with the
estimated value of p calculated as 0.01.
Following arguments given by Murray et al. [S], the new intetven-
tion, to be clinically relevant, should yield an effect size of 0.2. The
results displayed in Table 1 indicate that there would be 80% power
to detect this effect size if at least 9 clusters were randomly assigned
to each intervention group, assuming participation by 100 students
from each school and setting p = 0.01. Increasing the sample size by
lo-20% to allow for potential loss to follow-up results in a requirement
to assign 11 schools randomly to each intervention group. An SAS
[36] computer program, which can be adapted to calculate power for
parameter combinations not displayed in Table 1, is provided in the
Appendix.
DISCUSSION
The sample size and power requirements presented in this article cotte-
spond to a design that randomizes intact clusters to intervention groups
in the absence of stratification. Because of the increased probability
of imbalance on important baseline characteristics when the number
of clusters is small, some stratification or matching is often tecom-
mended [S]. Stratification on a baseline variable strongly related to
outcome will increase precision and in this sense implies that sample
size assessments that ignore stratification will be conservative. How-
ever, when the number of clusters pet group is small this gain in
precision must be weighed against the corresponding loss in degrees of
freedom [37].
Dr. Dormers research was supported by a grant from rhe Natural Sciences and
Engineeling Research Council. Dr. K&s research was supported by a postdoctoral
fellowship from the Ontario Ministry of Health. The authors thank Dr. Carol Buck
for her helpful comments.
References
1. KO~ZDS~ TD. Waenet EH, Cheadle AC, et al. Selected methodological
I
issues in evaluating community-based health promotion and disease pre-
vention programs. Annu Rev Public Health 1992; 13: 31-57.
2. Dormer A, Birkett N, Buck C. Randomization by cluster. Sample size
requirements and analysis. Am J Epidemiol 1981; 114(6): 906-914.
3. Hsieh FY. Sample size formulae for intervention studies wirh rhe cluster
as unit of randomization. Stat Med 1988; 8: 1195-1201.
4. Shipley MJ, Smith PC, Dramaix M. Calculation of power for matched
pair studies when randomization is by group. Int J Epidemiol 1989; 18(2):
457-461.
5. Donner A. Sample size requirements for stratified cluster randomization
designs. Stat Med 1992; 11: 743-750.
6. Fatquhar JW, Fortmann SP, Flora JA, et al. Effects of community wide
education on cardiovascular disease risk factors. J Am Med Assoc 1990;
264: 359-365.
7. The Commit Research Group. Community intervention trial for smoking
cessation. Am J Public Health 1995; 85: 183-200.
8. Murray DM, Perry CL, Griffin G, et al. Results from a state-wide approach
9. Orleans CS, Shipley RH. Worksite smoking cessation initiatives: Review
and recommendations. Addict Behav 1982; 7: l-16.
10. Jeffety RW, Forster JL, French SA, et al. The healthy worker project: A
work-site intervention for weight control and smoking cessation. Am J
Public Health 1993; 83(3): 395-401.
11. Addington-Hall JM, MacDonald LD, Anderson HR, et al. Randomized
controlled trial of effects of coordinating care for terminally III cancer
patients. Br Med J 1992; 305: 1317-1322.
12. Feng Z, Grizzle JE. Correlated binomial variates: Properties of estimator
intraclass correlation and its effect on sample size calculation. Stat Med
1992; 11: 1607-1614.
13. Marubini E, Correa Leite ML, Milani S. Analysis of dichotomous response
variables in teratology. Biometr J 1988; 30: 965-974.
14. Cohen J. A power primer. Psycho1 Bull 1992; 112(l): 160-164.
15. Murray DM, Rooney BL, Hannan PJ, et al. Intraclass correlation among
common measures of adolescent smoking: Estimates, correlates, and appli-
cations in smoking prevention studies. Am J Epidemiol 1994; 140(11):
1038-1050.
16. Hannan PJ, Murray DM, Jacobs DR Jr, McGovern PG. Parameters to aid
in the design and analysis of community trials: lntraclass correlations from
the Minnesota Heart Health Program. Epidemiology 1994; 5: 88-95.
17. Dormer A. An empirical study of cluster randomization. Int J Epidemiol
1982; ll(3): 283-286.
18. LaPrelle J, Bauman KE, Koch GG. High intercommunity variation in
adolescent cigarette smoking in a lo-community field experiment. Evalua-
tion Rev 1992; 16(2): 115-130.
19. Donnet A, Klar N. Methods for comparing event rates in intervention
studies when the unit of allocation is a cluster. Am J Epidemiol 1994;
140(3): 279-289.
20. Dormer A, Donald A. The statistical analysis of multiple binary measure-
ments. J Clin Epidemiol 1988; 41(9): 899-905.
21. Rao JNK, Scott AJ. A simple method for the analysis of clustered data.
Biometrics 1992; 48: 577-585.
22. Liang K-Y, Zeger SL. Longitudinal data analysis using generalized linear
models. Biometrika 1986; 73( 1): 13-22.
23. Dormer A, Eliasziw M, Klar N. A comparison of methods for testing
homogeneity of proportions in teratologic studies. Stat Med 1994; 13:
1253-1264.
24. Fung KY, Krewski D, Rao JNK, et al. Tests for trend m developmental
toxicity experiments with correlated binary data. Risk Anal 1994; 14(4):
639-648.
25. Edgington ES. Randomization Tests, 2nd Ed. Marcel Dekker, New York,
1987.
26. Conover WJ. Practical Nonparametric Statistics, 2nd Ed. John Wiley &
Sons, New York, 1980, p. 328.
27. Williams DA. Extra-binomial variation in toxicology. In: Procedures of
the 14th International Biometric Conference, Namur, Belgium, July
1988.
28. Harwell MR, Rubinstein EN, Hayes WS, er al. Summarizing Monte Carlo
Statistical Considerations in Community Intervention Trials 439

results in methodological research: The one- and two-factor fixed effects the alternative hypothesis, follows a noncentral t distribution with
ANOVA cases. J J&c Stat 1992; 17(4): 315-339. 2( m - 1) degrees of freedom and noncentrality parameter given by
29. Heeren T, DAgostino R. Robustness of the two-independent samples t-test
when au&d to ordinal scaled data. Stat Med 1987: 6(l): 79-90. d
30. Miller lk Jr. Beyond ANOVA: Basics of Applied &a&s. John Wiley
6. Sons, New York, 1986.
NC = [2(1 + ((n - l)p))/(mn)]
31. Scheffe H. The Analysis of Variance. John Wiley & Sons, New York,
1959. The power of the two-sample t test can be calculated using the
32. Paul SR, Liang KY, Self SC. On testing departures from the binomial and computer package SAS [36]. The computer code is written for a com-
multinomial assumptions. Biometrics 1989; 45: 231-236. munity intervention trial that enrolls m clusters of size n in each
33. Williams DA. Estimation bias using the beta-binomial distribution in tera- intervention group, where the intracluster correlation coefficient is
tology. Biometrics 1988; 44: 305-309.
34. Cornfield J. Randomization by group: A formal analysis. Am J Epidemiol given by p, the effect size by d, and (Y = 0.05.
1978; 108(2): 100-102.
35. Dormer A, Eliasziw M. Statistical implications of the choice between a data xl;
dichotomous or continuous trait in studies of inter-observer agreement. do m=3 to 12 by 3;
Biometrics 1994; 50(2): 550-555. do rho=O.OOl, 0.005, 0.01;
36. SAS Institute, Inc. SAS Language: Reference, Version 6. SAS Institute, do n= 100, 300, 500;
Inc., Caty, North Carolina, 1990.
do d=0.2, 0.5;
37. Martin DC, Diehr P, Perrin EB, Koepsell TD. The effects of matching on
the power of randomized community intervention studies. Stat Med 1994; df=2*(m- 1);
12: 329-338. t= tinv(0.975,df);
inf=2*( l+((n-l)*rho) )/(m*n);
nc=d/sqrt(inf);
APPENDIX power = 1 - probt(t,df,nc) + probt( - t,df,nc);
output;
The calculated power for the two-sample t test is given by end;
1 - p = P(lt( 2 t& end;
= 1 - P(t I t&J + P(t I -t&z) end;
end;
where tdz denotes the lCOct/2 percentile point of the t distribution
with 2(m - 1) degrees of freedom. The two-sample t statistic, under proc print;var m n d rho power;