13 views

Uploaded by Fanny Sylvia C.

- Chapter 12
- ReviewChaps3-4
- Chapter 10
- Chapter 14
- Chapter 20
- ReviewChaps1-2
- Chapter 13
- Chapter 21
- Charles Taylor
- Chapter 8
- Chapter 11
- Non%26ParaBoot
- SampleSizeCalcRevisited
- Model- vs. design-based sampling and variance estimation
- Hypo%26PowerLecture
- R Matrix Tutor
- Intro Bootstrap
- Clustering in the Linear Model
- Good Article on Standard Error vs Standard Deviation
- Chapter 7

You are on page 1of 5

12.14 Power of Analysiv-of Variance Tests 550
Table 13.18: Expected Mean Squares for a Latin Square Design
Source of Degrees of | Mean _ Expected moan squs
squares Modell Model IT
variation
Rows a
Columns at
Treatments or a
Bor Gyr) #
Total Ses
and 97, respectively. The derivation of the expected mean squares for a model II
Latin equare design i straightforward and, for comparison, we presont them along
with those for a model I experiment in Table 13.18
“Tests of hypotheses concerning the various variance comporents are made by
computing the ratios of appropriate imean squares as indicated in Table 13.18, and
comparing with corresponcling fvalics from Table A.6.
13.14 Power of Analysis-of-Variance Tests
As we indicated earlier, the research worker Is often plagued by the problem of
not knowing how large @ sermple to choose. In planning a one-factor completely
randomized design with n observations per treatment, the main objective is to test
‘the hypothesis of equality of treatment means,
; Hy: 04 = 09 =---04 =,
Hy: At least one of the ay’s is not equal to zero.
i (Gib ebieas, lnowerar; the expe aionball error rectoacs Gao Inns thot ts tek
E procedure will be insensitive to actual differences among the b treatment means.
Tn Sectan 12.3 the expected values of the mean squares for the one-way model are
given by
‘ac SSA 2 BY $2) 5 (SSE
B53) -2 (4) -o+ 2, a0) - 2 (2% ,)
‘Thus, for a given deviation from the null hypothesis Ho, as measured by
Ba
f)
large values of ¢? decrease the chance of obtsining a value f = of/s? that is in
‘the critical region for the test. ‘The sensitivity of the test describes the ability of
‘the procedure to detect differences in the population means and is measured by
the power of the test (sce Section 10.2), which is merely 1, where 9 Is theaz ae —
560 Chapter 18 One-Facter Experiments: General
probability of accepting a false hypothesis. We can interpret the power of our
analysisof-variance tests, then, as the probability that the Ftatistic is in the
critical region when, in fact, the mull hypothesis ic falso and the treatment moans
do differ. For the one-way analysis-of-variance test, the power, 1 — 8, is
-9=P [E> fin ea wee is tu]
oe ig > false) when S08 ,
The eran fa, 2) of coure, the uppor-ialed etial point of the Pasteibution
with vy and vp degrees of freedom. For given values of Z 2/(k 1) and o?, the
power can be increased by using # larger sample size 'n The problem becomes
one of designing the experiment with a value of so that the power requirements
ture met, For example, we might require that for specific values of 2 o} #0 and
22, the hypothesis be rejected with probability 0.9. When the power of the test
is low, it severely limits the scope of the inferences that can be drown from the |
experimental data,
Fixed Effects Case
Jn the analysis of variance the power depends on the distribution of the Fratio
uncer the alternative kypothesis that the treatment means difler. ‘Therefore, in
the case of the one-way fixed effects model, we require the distribution of $3/5*
when, in fact,
Of course, when the null hypothesis is true, a; = 0 for i = 1,2,...,k, and the
statistic follows the F.distribution with k —1 and N — ke degrost of freedom. If
¥ a 40, the ratio follows a noncentral Feds
ution.
‘The basic random variable of the noncentral Fis denoted by F’, Let f.(vs,v2,2)
be a value of F' with parameters v1, t2, and . The parameters v1 and 1s of the
istribution are the degrees of freedom associated with S? and S?, respectively
and ) is called the noncentrality parameter. When \= 0, the nonceutral F
simply reduces to the ordinary Fidistribution with v and v, degrees of freedom.
For the fixed offects, one-way analysis of variance with sample sizes m,, 72...»
‘we define -
aa Doma?
a18.14 Power of Analysis-of Variance Tests 561
If we have tables of the noncentral F at our disposal, the power for detecting a
particular alternative is obtained by evaluating the following probability:
1s Pf» s-13-» om
= PUP! > fal 1,.N 8).
Although the noncentral F is normally defined in terme of 2, its more convenient,
for purposes of tabulation, to work with
20.
Table A.16 shows graphs of the power of the analysis of variance es a function of
@ for various values of v1, v2, aud the significance level a. ‘These power charts
can be used not only for the fixed effects models discussed in this chapter, but
also for the multifactor models of Chapter 14. It remains now to give a procedure
whereby the noncentrality parameter , and thus ¢, can be found for these fixed
effects cases.
‘The noaceatrality parameter \ can be writien in terms of the expected values
of the numerator moan square of the F-ratio in the analysis of variance. We
have
and thus
Expressions for \ and ¢ for the one-way model, the randomized complete block
design, and the Latin square design are shown in Table 13.19.
‘Table 15.19: Noncentrality Parameter A and ¢* for Fixed Bifects Model
‘One-way Randomized Is
Classification Complete Block Square
EB bhet hed od
Og Lod be aL?
Note from Table A.16 that for given values of vy and t2, the power of the test
increases with increasing valnes of ¢. ‘The value of A depends, of course, on 02,
and in a practical problem one may often need to substitute the mean square error
as en estimate in determining @.

- Chapter 12Uploaded byFanny Sylvia C.
- ReviewChaps3-4Uploaded byFanny Sylvia C.
- Chapter 10Uploaded byFanny Sylvia C.
- Chapter 14Uploaded byFanny Sylvia C.
- Chapter 20Uploaded byFanny Sylvia C.
- ReviewChaps1-2Uploaded byFanny Sylvia C.
- Chapter 13Uploaded byFanny Sylvia C.
- Chapter 21Uploaded byFanny Sylvia C.
- Charles TaylorUploaded byFanny Sylvia C.
- Chapter 8Uploaded byFanny Sylvia C.
- Chapter 11Uploaded byFanny Sylvia C.
- Non%26ParaBootUploaded byFanny Sylvia C.
- SampleSizeCalcRevisitedUploaded byFanny Sylvia C.
- Model- vs. design-based sampling and variance estimationUploaded byFanny Sylvia C.
- Hypo%26PowerLectureUploaded byFanny Sylvia C.
- R Matrix TutorUploaded byFanny Sylvia C.
- Intro BootstrapUploaded byMichalaki Xrisoula
- Clustering in the Linear ModelUploaded byFanny Sylvia C.
- Good Article on Standard Error vs Standard DeviationUploaded byAshok Kumar Bharathidasan
- Chapter 7Uploaded byFanny Sylvia C.
- Chapter 7Uploaded byFanny Sylvia C.
- Bio Math 94 CLUSTERING POPULATIONS BY MIXED LINEAR MODELSUploaded byFanny Sylvia C.
- Chapter 5Uploaded byFanny Sylvia C.
- Chapter 9Uploaded byFanny Sylvia C.
- GRM: Generalized Regression Model for Clustering Linear SequencesUploaded byFanny Sylvia C.
- Chapter5p2LectureUploaded byFanny Sylvia C.
- Data Modeling: General Linear Model &Statistical InferenceUploaded byFanny Sylvia C.
- Chapter 6Uploaded byFanny Sylvia C.
- The not so Short Introduction to LaTeXUploaded byoetiker