Академический Документы
Профессиональный Документы
Культура Документы
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
Taylor & Francis, Ltd. and American Statistical Association are collaborating with JSTOR to digitize, preserve
and extend access to The American Statistician.
http://www.jstor.org
In the two-sample means test, the need for a preliminary Table 1. For each brand of heater, the procedure lists
variance test and the special emphasis given to the equal the sample size, means, standarddeviations, standarder-
variance assumption are questioned. The conclusions can rors, minimum and maximum values. A preliminary
be easily used in the classroom to teach the two sample equality of variance test follows, where the calculated F
means test. ratio, the degrees of freedom, and the P value are pre-
sented. The SWS and t statistics are listed in the rows
KEY WORDS: Size; Power; Preliminary variance test.
titled Unequal and Equal Variances, respectively. Thus,
the user checks the equality of variance test, if equality
The problem of testing the equality of the means from is accepted, he proceeds to the t statistic, if equality is
two independent normally distributedpopulations is cov- not accepted, he uses the SWS statistic.
ered in every elementary statistics textbook. Under the It is apparentfrom this discussion that a great deal of
assumption that the two population variances o-I and emphasis is placed on the homogeneity of variance as-
2are equal, most authors recommend the t test. If the sumption in the two-sample means test. Authors of el-
variances are unequal, then an alternativeprocedure, such ementary statistics textbooks stress the importance of the
as the method suggested by Smith (1936), by Welch assumption, to the extent that textbooks are divided into
(1937) and later by Satterthwaite (1946), is often rec- sections where the t test is presented if variance equality
ommended. To apply such an approach, the variance ra- is assumed and the SWS test is presented if variance
tio, 0 = o- 2/o2, must be known. In most applications, equality is violated. Likewise, popular statistical pack-
however, the variance ratio is unknown. In such cases, ages have incorporatedpreliminaryvariance tests in their
it is current practice to apply a formal preliminary test two-sample means test routines. The following questions
for -2 = -2. If c2 = o2 is accepted in the preliminary arise: Is the currentpractice of preliminary variance tests
test, then a t test is performed on the means; otherwise appropriate and is this emphasis on variance homoge-
a Smith/Welch/Satterthwaite (SWS) test on the means neity warranted?
is performed. The objective of this article is to address these ques-
Preliminaryvariance tests have been incorporatedinto tions. To this end, the relationship between the t test and
the two-sample means tests of popular statistical pack- SWS test is examined when the choice between the two
ages, such as SAS, BMDP and SPSS. All of these pack- procedures is based on a preliminary variance test. In
ages have two-sample means tests that include a prelim- the following sections, the problem is outlined, the sizes
inary equality of variance test. The preliminary variance and powers of the means tests are summarized, and the
test is calculated first, followed by the choice between two questions are answered.
a t test and an SWS test. To illustrate how the statistical
routines are used, we examine an example presented by THE PROBLEM
Milton and Arnold (1986, p. 312, example 10.6.1). In
the example, two brands of kerosene heaters are tested. Let xl, . . ., xn, and Yi, ..., y,,2 be independent random
The observations are the times required to raise the room samples from two normally distributedpopulations, where
temperature10?F. We analyzed the logarithm of the time xi N(,u, 1) and yj - N(4t2, o), for i 1, ...,
using SAS PROC TTEST. The SAS output is listed in and j = 1, ... , n2
The preliminary test for Ho 1 =2 versus H1: I- #
o-2 is to calculate
?) 1992 American Statistical Association The American Statistician, Febrtary 1992, Vol. 46, No. 1 19
TTEST PROCEDURE
Standard Standard
Brand N Mean deviation error Min Max Variances T DF Prob > IT
B 12 3.685 .487 .141 2.62 4.24 Unequal 1.93 16.4 .0710
A 15 3.381 .273 .071 2.63 3.73 Equal 2.05 25.0 .0508
NOTE: For Ho: Variances are equal, F = 3.18 DF = (11, 14) Prob > F = .0449. Dependent variable = logarithm of time in seconds.
is the 100(1 - a*) percentile point of an F distribution bility of rejecting Ho is the size of the test. The power
with a and b degrees of freedom, s2= (n, - 1)-i El" of the test corresponds to the probability of rejecting
(X i - X) 2 =(n2 - j)1En21( y) x = I1 l Ho for any A > 0. In either case, the probability of re-
xi/n, and y = Ej`2 yj/n2. The above rejection criterion jecting Ho is expressed as the sum of two mutually ex-
is equivalent to the P value criterion that 1 2 clusive alternatives:
rejected for all P values less than a. Pr(rejectHo) P,.(rejectHo and do not reject Ho)
If Ho is not rejected, then the t test for Ho*:i? = 12
versus HI: III 7& I-,2 (or Ho*: IiI= -u2versus H*: i,? > + P,.(rejectHo and reject Ho). (5)
1t2) is to calculate Sizes and powers for combinations of nl,- n2, 0, 8, a,
and A were calculated by Moser, Stevens, and Watts
t*
y (2) (1989). Their results are summarized as:
sp,\(1/n, + 1/jn2)
(a) If the sample sizes are equal, then the ASWS, ST
and reject Ho if t*2 > F ,(+522 (or reject Ho* if t* > and AT tests have almost identical sizes and powers for
where 8 is the prescribed means test signifi- all nl, n2, A > 0, and 1 c 0 ? 10.
t4 nI+112-2),
cance level and sp = [(n1-1)sI + (n2 - + n2
(b) If the sample sizes are unequal but the variance
21)s]/(ni
- 2).
ratio, 0, is near one, then the AT test attains the largest
If Ho is rejected, then the SWS criterion for testing power while maintaininga size near the prescribed8 level.
(c) If the sample sizes are unequal, 0 is not near one,
Ho*versus H* (or Ho*versus H*) is to calculate
and the smaller variance is associated with the larger
sample size, then the AT test still has the largest power
y-8= (3) for all A > 0, followed by the ST and ASWS tests. How-
-\(sl/n,
,\(S2/l +
+ 2 n2)
Sf2)
ever, the AT test attains this large power at the cost of
and reject Ho*if t**2 > Fb v (or reject if Ho*if t** > tv), a large size. For certain values of 0, the ST test will also
for the same prescribed significance level 8 given in (2), attain its larger power at the cost of a large size. The
with ASWS test, however, maintains a reasonable size near
8 for all 1 ? 6 ? 10.
(1/n, + u/n2)2 (d) If the sample sizes are unequal, 0 is not near one,
11/n(n -1 ) + u2/n2(n2-1) (4) and the smaller variance is associated with the smaller
sample size, then the ASWS test has the largest power,
and u = s1/s. followed by the ST and AT tests. The ASWS test con-
It should be noted that, if the prescribed variance test tinues to maintain a reasonable size near 8 for all 1 ' 0
significance level, a, is set to 0, then Ho is accepted and ?10.
the t test (2) for Ho is always performed. Likewise, if
a is set to 1, then the SWS test (3) for Ho is always DISCUSSION AND RECOMMENDATIONS
performed. Therefore, specifying an a level of 0 (or 1)
We now address the question: Is the current practice
is equivalent to performing a t test (or SWS test) without
of preliminary variance tests appropriate?The answer is
any preliminary test for Ho. For purposes of identifica-
no. We justify this response in the following paragraphs.
tion, the preliminarytest for Ho with 0 < a < 1 followed
The t test provides the largest power while maintain-
by a t test or SWS test for Ho*will be referred to as the
ing a size near the prescribed 8 level whenever the sam-
Sometimes t (ST) test. Performing a t test for H* without
ple sizes are unequal and the variance ratio is near one.
any preliminary test for Ho (i.e., a = 0), will be referred
Therefore, if the sample sizes are unequal and the vari-
to as an Always t (AT) test. Using a SWS test for Ho*
ance ratio is known to be near one, then the t test is
without any preliminary test for Ho (i.e., a = 1) will be
appropriate.
referred to as an Always SWS (ASWS) test.
The SWS, the ST and the t tests all have the same
sizes and powers whenever the sample sizes are equal.
THE SIZE AND POWER OF THE TEST
Therefore, a preliminary variance test is superfluous, as
The probability of rejecting H*: itl = ,U2,in favor of it simply adds an extra analysis step. Thus, in the equal
H* or H* , is a function of nI, n2, 0, 8, a, and A = (,U2 sample size case, either the t test or the SWS test is
- ,u1)2/[2(o-2I/n, + o-2/n2)]. When A 0, the proba- appropriate.