|
rvid L Otis, f~ 4;
<br />e
<br />!~......
<br />.Johnson 763
<br />1. Marcluff 773
<br />Halmata,
<br />:o Restani
<br />8. Isaacs
<br />RE
<br />781
<br />~fi ~0 ~ h C} n~ ,
<br />THE INSIGNIFICANCE OF STATISTICAL SIGNIFICANCE TESTING
<br />DOUGLAS H. JOHNSON,' U.S. Geological Survey, Biological Resources Division, Northern Prairie Wildlife Research Center,
<br />Jamestown, ND SB401, U6A
<br />~~
<br />794 - Key words: $avesian approaches, rnnfidence interval, null hypothesis, P-value, power analysis, scientific hy-
<br />pothesis -test. statistical hypothesis test. ~ •
<br />3. Christie
<br />MODEL
<br />~ A. Klaus
<br />l K. Carrie
<br />-JAY .. .
<br />H. Below
<br />q. Wagers
<br />ERN
<br />~. Mitchell
<br />:orquodale
<br />fE-
<br />~. Hartless
<br />ITRAL
<br />S. Wilson
<br />EXAS...
<br />K. Fritzell
<br />~,
<br />an Cassini
<br />803
<br />815
<br />824
<br />833
<br />843
<br />853
<br />861
<br />872
<br />880
<br />889
<br />895
<br />ioeiery. 5410 Grosvenor line.
<br />e~. and -includes all WiiDUFE
<br />,ues can tx obtained from the
<br />~. rho Wildlire Society. solo
<br />~ersi(y, Stillwater, OK 7407A.
<br />ct Biology. SWY College of
<br />w Hampshire Street. Lawrcncc.
<br />•nsvennr hoe, Bethesda. ~iD
<br />Statistical testing of hypotheses in the wildlife
<br />field has increased dramatically in recent years.
<br />Even more recent is an emphasis on power
<br />analysis associated with hypothesis testing (The
<br />Wildlife Society 1990. While this trend was oc-
<br />curring, statistical hypothesis testing was being
<br />deemphasized in some other disciplines. As an
<br />example, the American Psychological Assoc%a-
<br />tion seriously debated a ban on presenting re-
<br />sults of such tests in the Association's scientific
<br />journals. That proposal yvas rejected, not be-
<br />cause it lacked merit, but due to its appearance
<br />of censorship ~ Vleehl 199; ).
<br />The issue yeas lughlighted at the 1998 annual
<br />conference of The Wildlife Socieh-. in $uffalo.
<br />~ieyv fork. inhere the Biumetric•s ~Vo.rkin;
<br />Croup sponsored a half-day symposium on
<br />Evaluatin_ the Rule of Hypothesis Testin~-
<br />Poyver _~nalysis in Wildlife Science. Sp?akers at
<br />that session a~ho addressed statistical hypothesis
<br />testinr, yvr,re yimlally unanimous in their opin-
<br />ion that the tool wars oyenlsed. Misused. and
<br />ulten inappropriate.
<br />~[v ohjecrives are to hrieHv describe statisti-
<br />cal h~-pothesis testing, discuss common b~rt in-
<br />correct interpretations of resulting P-values.
<br />Mention some shortcolnin~s of hypothesis test-
<br />ier„ indicate. yylw hypothesis testim~ is conduct-
<br />ed, and outline some alternatives.
<br />l E-mail: dougl:l.~lLjohnsonC~nbs.~uy
<br />pence of Paper).
<br />WHAT 1S STATISTICAL HYPOTHESIS
<br />TESTING?
<br />Abstrnet: llespite their wide use in scientific journals such as The jorrnral of [Yikllife ~larut~rrnenf, statistical
<br />-' hypothesis tests add ven• little value to the products of research. Indeed, they frequently conftue the inter-
<br />'` pretation of data. This paper describes how statistical 1>,ypothrsis tests are often viewed, and then contracts
<br />that interpretation yyith the correct one. I discuss the arbitrariness of P-values, conclusions that the null hv-
<br />pothesis is tore, power analysis, and distinctions behveen statistical and biological significance. Statistical hy-
<br />- pothesis testing, itl yvhiclt the null hypothesis about the properties of a population is almost alyvavs )mown a
<br />:~ priori to be false, is contrasted with scientific hypothesis testing, which ecamines a credible null hypothesis
<br />~`_' about phenomena in nature. More meaningful alternatives are briefly outlined, including estimation and mn-
<br />fidence intervals for determining the importance of factors, decision theory for guiding actions in the face of
<br />uncertainh~, and Bavesian approaches to hypothesis testing and other statistical prlctices.
<br />JOURNAL OF WILDLIFE MANAGEMENT 63(3):763-772
<br />Four basic steps constitute statistical hypoth-
<br />esis testing. First, one develops a null hypothesis
<br />about some phenomenon or parameter. This null
<br />hppothe;is is generally the opposite of the re-
<br />search h~gothesis, which is what the investigator
<br />truly believes and wants to demonstrate. Re-
<br />search hypotheses may- be venerated either in-
<br />duc~vely'; from a study of obsen•ations already
<br />made. or deductiyeh•, Jeri`ing from theon•. Nest,
<br />data are collected that bear on the issue. typically
<br />by an experiment or by samnliu;. (dull ltypoth-
<br />eses often are developed after the data are in
<br />hand and have been nnnma~ed through. but
<br />th:lt~s another topic.) .~ statistic:i test of the cull
<br />hypothesis then is conducted, yy'nich generates a
<br />P-y:uue. Fin:tllr. the cluestiun oC •a•hat that valve
<br />mr:ms re!atiyr to the u~lll hypothesis is consicl-
<br />ere~l. Sey era! intrrj~ret:ltious of P ~,iten are rnacle.
<br />Sometimes P is yieyyecl as the pml,ahilih• that
<br />the results obtained yyere chle to chance. Small
<br />values are taken to indicate that the results yvere
<br />not just a happenstance. A lar~~e yallle ~f P. say
<br />for a test that µ = Q, yyonld suvuest that the
<br />mean r actualh~ rea,rclyd was due to chance,
<br />and µ cnuld be assuMed to hr zero iSchmidt
<br />anti I [enter 199 ~ ).
<br />Other times, 1-P is cuusidered the reliabilih•
<br />of the result: that is. the probahilih• of ~ettina-
<br />c:;
<br />RECEIVE
<br />
|