SHORT COMMUNICATIONS 1023

LITERATURE CITED SCHMIDT-KOENIG, K. 1958. Der EinfluB experimen-

ABLE,K. P. 1982. Skylight polarization at dusk in-
fluencesmigratoryorientationin birds. Nature 299:
550-551.
ABLE, K. P. 1989. Skylight polarization patterns and
the orientation of migratory birds. J. Exp. Biol.
141:241-256.
BALDA,R., AND A. C. KAMIL. 1989. A comparative
study of cache recovery by three corvid species.
Anim. Behav. 38:486-495.
BRMES,M., AND J. GOULD. 1982. Skylight polariza-
tion patternsand animal orientation. J. Exp. Biol.
96~69-91.
HELBIG,A. J., AND W. WILTSCHKO. 1989. The sky-
light polarization at dusk affects the orientation
behavior of Blackcaps,Sylvia atricapilla. Natur-
wissenschaften761227-229.
KAMIL,A. C., ANDR. P. BALDA. 1985. Cacherecovery
and spatial memory in Clark’s Nutcracker (Nu-
cifiazzacolumbianal J. Exn. Psvchol.: Anim. Be-
h&.-Processes 11:95-l 1l.- _ nitive processesand spatial orientation in animal
MOORE,F. R., ANDJ. B. PHILLIPS. 1988. Sunset,sky-
light polarization and the migration of Yellow-
rumped Warblers, Dendroicacoronata.Anim. Be-
hav. 36:1770-1778.
tell verlnderter Zeitschltzung auf das Heimfin-
devermijgen von Brieftauben. Naturwissenschaf-
ten 45:47.
WALLRAFF,H. G. 1974. Das Navigationssystemder
Vijgel. Schriftenreihe “Kybemetik,” R. Olden-
bourg Verlag, Milnchen, Wien.
WEHNER,R. 1976. Polarized light navigation by in-
sects.Scient. Amer. 235:106-l 15. _
WILXCHKO. W.. AND R. P. BALDA. 1989. Sun com-
pass oiientation in seed-caching Scrub Jays
(Aphelocomacoerulescens). J. Comp. Physiol. 164:
717-721.
WILTSCHKO, W., AND R. WILTSCHKO.1982. The role
of outward journey information in the orientation
of homing pigeons,p. 239-252. In F. Papi and H.
G. Wallraff reds.],Avian navigation. Proc. in Life
Sciences,Siring&-Verlag, Berlin.
WILTSCHKO. W.. ANDR. WILTSCHKO. 1987. Cognitive
maps and navigation in homing pigeons,$ 20 l-
216. In P. Ellen and C. Thinus-Blanc [eds.], Cog-
and man. Martinus Nijhoff, Dordrecht, The Neth-
erlands.

The Condor93~1023-1025
0 The CooperOrnithologicalsociety I99 I

A PROBLEM IN STATISTICAL ANALYSIS: SIMULTANEOUS INFERENCE’

KATHLEEN G. BEAL
Department of Mathematics and Statistics, Wright State University,Dayton, OH 45435
HARRY J. KHAMIS
Department of Mathematics and Statisticsand Department of Community Health, Schoolof Medicine,
Wright State University,Dayton, OH 45435

Key words: Statistics;simultaneousinference:Bon- tually related hypotheses. These related hypotheses

ferroni method.
An analysis of variance (ANOVA) problem is often
used as a general example of simultaneous inference
since decisionsabout main and interaction effectsare
made concurrently.Furthermore, the various multiple-
comparison techniques(e.g., Fisher’s LSD Procedure,
Duncan’s New Multiple RangeMethod) are well-known
methods of simultaneousinference. we wish to bring
to the attention of ornithologistsa common statistical
error involving simultaneousinferencesor conclusions
based on two or more tests of hypotheses. Simulta-
neous inferencesare drawn from a family of concep-
I Received 1 July 1991. Final acceptance 22 July
1991.
usually emanate from groups of observationsthat are
collectedby an individual researcheror researchteam
(Miller 1981).
As an example, considera researcherwho quantifies
severalresponsevariablesin a singlestudy.Finch (199 1)
used a number of reproductive measures(e.g., mean
laying date, rates of nest failure, initial clutch size,
fledgling rates) in a study of the effect of three levels
of flooding on timing of reproductionand productivity
of House Wrens (Troglodytesaedon). BecauseFinch
reaches a set of simultaneous conclusionsabout the
relationshipsamong reproductive measuresand flood-
ing, they constitute a family of hypotheses.
We set (Y,the probability of a type I error (rejection
of the null hypothesiswhen it is true), at a conventional
value such as 0.05. Therefore, the probability of ac-
ceptingthe null hypothesiswhen true, (1 - a), is quite
high for a single test of hypotheses(see, for example,
1024 SHORT COMMUNICATIONS

Zar 1984:44 or Ott 1984:llO). When we perform P trol our experimentwisetype I error rate so that it does
tests of hypotheseson P responsevariables, the prob- not exceed0.05. Johnsonand Wichem (1988) discuss
ability of acceptingall null hypotheseswhen true is (1 theory and development of the Bonferroni method in
- a)’ if the testsare independent. For example, if P = multivariate analysis,however some of their formulas
3 then the probability of acceptingall three null hy- are in error (Johnson, pers. comm. with H.K.; corri-
pothesesif true is (1 - 0.05)’ or 0.8574 (Johnsonand genda can be obtained from H.K.).
Wichem 1988:187). The combined probability of a The use of the Bonferroni method, namely usingan
type I error for all hypotheses,called the experiment- alpha level of 0.01 for each t-test in this example, in-
wise type I error rate and denoted by (Y’, is then 00 = suresthat the probability of incorrectly rejecting any
(1 - 0.8574) or 0.1426, not 0.05. The experimentwise one or more of the five null hypothesesis bounded by
type I error rate increaseswith the number of simul- 0.05. If the Bonferroni method is not used, then each
taneoustests of hypotheses;for five tests it is 0.2262, t-test would be conducted at the 0.05 level of signifi-
for 10 tests it is 0.4013, etc. So, for 01= 0.05 and a canceleading to an experimentwisetype I error rate as
large number of tests, the chancesof rejecting a true high as 0.2262. So, null hypothesesmight be rejected
null hypothesisat least once, (Y’, becomesmuch larger in good faith when they should not be.
than 0.05. According to a formula called the Bonferroni The primary disadvantageof the Bonferroni method
inequality, the experimentwise type I error rate, OI’, is is that it may be more conservativethan a multivariate
always less than or equal to (P.,) regardlessof the procedure, that is, the actual experimentwise type I
correlation structureamong the P tests;that is error rate may be somewhat less than (Y.Rice (1989)
discussesthis problem and provides an adjustment,
00 5 P.a.
called the sequential Bonferroni technique, which in-
Consequently,in order to insure that the experiment- creasesthe power of the Bonferroni procedure. For
wise type I error rate, LY’,does not exceed, say 0.05, discrete data, Tarone (1990) has provided a modified
simply use a value of a equal to 0.05/P (see Morrison Bonferroni method.
1990:32-33). Finally, the Bonferronimethod is most effectivewhen
Researchersseem to be largely unaware of the prob- P, the number of tests, is small. As P increases,LY’=
lem regarding simultaneous conclusionsdrawn from a/P decreases.Since the Bonferroni method is conser-
multiple tests of hypothesesconcerning two or more vative when P is large, the probability of the type II
response variables. Introductory statistical texts and error (failing to reject the null hypothesis when it is
statistical references (such as Sokal and Rohlf 1981 false) becomes a concern. Generally, researchersare
and Snedecorand Cochran 1989) usedmost frequently urged to carefully consider the number of tests of hy-
by ornithologists restrict discussion of the experi- pothesesduringthe designphaseof their research.Many
mentwise type I error rate to multiple comparisontech- researchersappear to feel that the more tests of hy-
niauesfor ANOVA. An examination of the 1989 issues pothesesthe better. Eliminating redundant or unnec-
of-The Auk, The Condor, and The Wilson Bulletin essarytests of hypothesesand carefully choosing the
revealed at least forty instancesin which multiple uni- sets of tests to be Bonferroni adjusted reducesP, pre-
variate testswere used for a data set without regardto vents 01’from becoming smaller than necessary,and
inflated values of the experimentwise type I error rate. bolstersthe power of the tests.
We describe a univariate procedure that addresses The Bonferroni method provides a solution to a se-
this problem (simultaneousinference)which is simpler rious statisticalproblem, namely control of the exper-
than the multivariate procedures,but is relatively un- imentwise type I error rate for simultaneousinferences.
known among researchers.In fact, this univariate al- This solution is simple computationally and in terms
ternative, called the Bonferronimethod,is actually su- of interpretation, and it is superiorto other more com-
perior to multivariate methods when the number of plex multivariate techniqueswhen the number of hy-
hypothesesis small (see Johnson and Wichem 1988: pothesesis small. The Bonferroni method can be ap-
188-l 90). plied to any test procedure or confidenceinterval and
For a singletest of hypothesisusinga pooled sample is not restrictedto the useof t-tests (seeMorrison 1990
t-test (as described,for example, by Ott 1984, df = 10, for a variety of applications). It is a simple way of
and LY= 0.05) the null hypothesisis rejected in favor compensatingfor the multiplicity of significancetests
of a two-sided alternative if t is greaterthan the critical and requires no assumptionsother than those neces-
value of 2.228 or less than that of -2.228. If five sets sary for the validity of the individual significancetests
of hypothesesare testedsimultaneouslythe calculation (Miller 1981:8).
of the test statistics is typical of multiple t-tests but
interpretation of the observed level of significance(P
value) is adjustedaccordingto the Bonferroni method, LITERATURE CITED
a method that accountsfor the inflation of the exper-
imentwise type I error incurred by simultaneoustests. FINCH,D. M. 1991. House Wrens adjust laying dates
The Bonferroni method requiresus to divide the prob- and clutch size in relation to annual flooding. Wil-
ability of a type I error, here 0.05, by the number of son Bull. 103:25-43.
tests and to draw our conclusionsbased on that new JOHNSON, RICHARD A., ANDDEANW. WICHERN. 1988.
level of significance,namely (Y’= 0.05/5 = 0.01. The Applied multivariate statisticalanalysis.Prentice-
critical valuesoft are now + 3.169. more extreme than Hall, Englewood Cliffs, NJ.
those for the single test of hypothesis.This technique MILLER,R. 6., JR. 1981. Simultaneousstatisticalin-
allows us to use simple testsof hypothesesand to con- ference. Springer-Verlag,New York.

MORRISON,DONALDF. 1990. Multivariate statistical
methods. McGraw-Hill, New York.
Orr, LYMAN. 1984. An introduction to statistical
methods and data analysis. Duxbury Press, Bos-
ton, MA.
RICE, WILLIAM R. 1989. Analyzing tablesof statistical
tests. Evolution 43:223-224.
SNEDECOR, G. W., AND W. G. COCHRAN. 1989. Sta-
SHORT COMMUNICATIONS 1025
tistical methods. Iowa State Univ. Press, Ames,
IA.
SOKAL,R. R., AND R. J. ROHLF. 1981. Biometry. W.
H. Freeman and Co., New York.
TARONE,R. E. 1990. A modified Bonferroni method
for discrete data. Biometrics 46:5 15-522.
ZAR, JERROLD H. 1984. Biostatisticalanalysis.Pren-
tice-Hall, Englewood Cliffs, NJ.