Scribd
Upload
39 views

R 4 Different Drugs

This document discusses power calculations for F-tests in analysis of variance (ANOVA). It explains that the power of an F-test increases as the sample size increases or as the population means become more spread out. It provides an example of calculating the necessary sample size to detect a difference of at least 10 minutes among 4 drug means with 80% power. The document also summarizes procedures for investigating differences among treatment means after an ANOVA, such as constructing confidence intervals and testing pairs of means. It discusses methods for making multiple comparisons like Tukey's procedure that control the familywise error rate.

Uploaded by

Justin Joshuva
Copyright
© © All Rights Reserved
Available Formats
Download as DOC, PDF, TXT or read online on Scribd
39 views

R 4 Different Drugs

This document discusses power calculations for F-tests in analysis of variance (ANOVA). It explains that the power of an F-test increases as the sample size increases or as the population means become more spread out. It provides an example of calculating the necessary sample size to detect a difference of at least 10 minutes among 4 drug means with 80% power. The document also summarizes procedures for investigating differences among treatment means after an ANOVA, such as constructing confidence intervals and testing pairs of means. It discusses methods for making multiple comparisons like Tukey's procedure that control the familywise error rate.

Uploaded by

Justin Joshuva
Copyright
© © All Rights Reserved
Available Formats
Download as DOC, PDF, TXT or read online on Scribd
You are on page 1/ 13

Power Calculation for F-test

If the population means 1, 2, , r differ, we would like our


F-test to reject H0 with high probability.
Power =
For any fixed , the power of a test _______________ as the
sample size increases.
For any fixed , the power of the F-test is ______________
when the 1, 2, , r are more spread out.
The distribution of F* under Ha is a ____________________

In experiments, we often wish to determine the minimum


sample size needed to have a specified power to detect certain
differences among the population means.
Example: It may be important to reject H0 when the largest
and smallest i differ by at least
Example: Y = Time until headache relief (in minutes),
r = 4 different drugs
Suppose the company wishes to detect when the mean times
for the 4 drugs differ by at least 10 minutes with probability
0.80, using = 0.05.
We also must specify (or at least guess) the value of .
Suppose

Then
Table B.12 (pg. 1342-1343) tells us we need
The R function power.anova.test does similar calculations,
but not in terms of .
Investigating Differences Among Treatment Means
If the F-test is not significant, we conclude the true mean
response is (or may be) the same for all factor levels and no
further investigation is needed.
If we conclude a difference in treatment means, we investigate
further.
Plots: We can plot the sample means for each level using a bar
graph (good with a qualitative factor) or a main effects plot
(good with a quantitative factor).
See SAS and R examples.
Inference About the Population Treatment Means
Consider a CI for i, the true mean response at level i:
Point estimate of i is
Using MSE to estimate 2,

and

100(1 )% CI for i is:

SAS gives these using the CL option to the LSMEANS


statement.
Example: 95% CI for the mean sales for package design 1 is:
Comparing Two Population Treatment Means
Estimating the Difference between i and i' :
Point estimate is

So a 100(1 )% CI for i i' is:

We also may test whether two particular treatment means are


equal:

This is done with a


If we use the PDIFF option to the LSMEANS statement, SAS
gives CIs for the difference between each pair of treatment

means and the p-value for each test comparing two treatment
means.
Note: These results are only valid if we are doing inference
about one particular pair of treatment means. If we are doing
many simultaneous comparisons, we must use different
techniques.
SAS example: 95% CI for the difference between mean sales
for package design 3 and mean sales for package design 4:

Contrasts
A contrast is a linear combination of factor level means in
which the coefficients sum to zero.
Often useful for comparing several treatment means.
Previous example:
In Kenton Foods data, package designs 1 and 3 used
cartoons; designs 2 and 4 did not. To compare mean sales for
the cartoon designs vs. the non-cartoon designs, we could
use the contrast:

An unbiased estimator of a contrast L is

Using t procedures we can obtain a 100(1 )% CI for L or


perform a t-test of
SAS will do this with an ESTIMATE statement.
Previous example: 95% CI for difference in mean sales for the
cartoon designs and mean sales of non-cartoon designs is:

Interpretation:

Testing whether cartoon designs and non-cartoon designs


have significantly different mean sales:

Suppose we had wanted to test whether the non-cartoon


designs had produced significantly better mean sales than the
cartoon designs:

Simultaneous Inference
Suppose we make 3 simultaneous comparisons, say:

If we do three separate t-tests, each with = 0.05, then:


If H0 is true in each case: The probability that we reject H0
and incorrectly conclude Ha in at least one test is:

So the family of tests has significance level


Another concern: We must avoid data snooping testing
only those comparisons that are suggested by the data. This
can wildly inflate the actual significance level!
If we look at the data initially and then use that early look
to decide which tests to perform, we are implicitly making
comparisons.
Solution: Use methods designed to make many
inferences/comparisons simultaneously.

These procedures are designed so that the family significance


level is

The family significance level is also known as the


This is as opposed to the
which is
Tukeys Multiple Comparison Procedure
This compares all possible pairs of treatment means
simultaneously:
Tukey simultaneous CIs for all differences i i' :
where

We have confidence (1 )100% that the entire set of these


CIs contain the true pairwise differences between treatment
means that they purport to contain.
Simultaneous Testing
We can test

The Tukey procedure declares i and i' significantly different


if

Results of the Tukey procedure may be found in SAS.


Example (Kenton data): SAS output reveals CIs for all
pairwise treatment mean differences, with family confidence
coefficient 90%:

Testing: We can compare all pairs of treatment means, with


family significance level = 0.10. SAS output shows:

Other Multiple Comparison Procedures


The Scheff procedure is designed for CIs for (and tests
about) all possible contrasts.
The family confidence level (significance level) is at least 1
(at most ).
Note that this class of inferences includes all possible pairwise
comparisons, plus many more inferences.

If only pairwise comparisons are needed, then Tukeys


procedure is preferable.
The Bonferroni procedure is good for testing about a few
contrasts (must be specified before looking at the data or
results no data snooping allowed!)
For each of g contrasts, L1, , Lg, the Bonferroni intervals are

For testing
against the two-sided alternatives, for each H0, we reject if

where t* is the estimated contrast divided by its standard error.


The Bonferroni method is based on the Bonferroni inequality.
Let Ai indicate that the i-th CI does not contain its parameter
(or the i-th test has a Type I error):

When testing about several contrasts, it is most useful to test


orthogonal contrasts. Two contrasts

Orthogonality implies that one contrast conveys no


information about the other (i.e., no overlapping
information).
Example:
When doing all pairwise comparisons, the Tukey method is
more efficient than Bonferroni. If only a few comparisons are
of interest, the Bonferroni method may be better.
When all (or very many) contrasts are of interest, the Scheff
procedure is best. When a few contrasts are of interest,
Bonferroni may be better.
Data snooping should not be used with the Bonferroni
procedure, but data snooping is OK when Tukey or Scheff is
used, because these procedures intrinsically involve all possible
inferences of a certain type anyway.
Other procedures: Dunnetts procedure is designed to
compare several treatments to a control group
simultaneously.
Example (Drug study):
Hsus procedure: Selects best treatment and compares all
others to the best.
Checking Model Assumptions in ANOVA
This is again done through an analysis of the
We may check for outliers by examining the (internally)
studentized residuals, which for the ANOVA model are:

Rule of Thumb:
A formal t-test based on the externally studentized residuals is
also available; see pg. 780 for details.
Graphical Tools: (1) Plots of Residuals vs. Fitted Values
Can check:

(2) Normal Q-Q plots of residuals


Could do separate Q-Q plots for each factor level, if the
sample sizes are quite large.
Or could do a single Q-Q plot of all residuals (as long as error
variances are judged roughly equal).
Shapiro-Wilk test on the residuals can formally test for nonnormality.
(3) If the data are gathered over time, we may plot the
residuals against a time index (separately for each factor level)
to check for dependence of errors across time.
The equal variance assumption can be tested formally with
the Brown-Forsythe test in SAS or R:

The B-F test is based on doing the familiar F-test on the

If the F-statistic is large and H0 is rejected, then the variances


are deemed unequal across populations.

SAS example (Kenton Foods data):

Common Remedies for Unequal Error Variances and/or


Non-normal Errors
(1) Use weighted least squares and fit the ANOVA model with
the regression approach.
(2) Transform the Response Variable:
Examples:

(3) Use a nonparametric alternative to the F-test called the


Kruskal-Wallis test, which is based on the ranks of the data.

You might also like