Guinness, Gosset, Fisher, and Small Samples
Transcription
Guinness, Gosset, Fisher, and Small Samples
Guinness, Gosset, Fisher, and Small Samples Author(s): Joan Fisher Box Source: Statistical Science, Vol. 2, No. 1 (Feb., 1987), pp. 45-52 Published by: Institute of Mathematical Statistics Stable URL: http://www.jstor.org/stable/2245613 . Accessed: 09/10/2011 12:33 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 [email protected]. Institute of Mathematical Statistics is collaborating with JSTOR to digitize, preserve and extend access to Statistical Science. http://www.jstor.org StatisticalScience 1987,Vol. 2, No. 1, 45-52 Guinness, Gosset, Fisher,and Small Samples Joan Fisher Box Abstract.The environment in whichW. S. Gosset(Student)workedas a brewerat Guinness'Breweryat the turnof the centuryis describedfully enoughto showhow it forcedhimn to confront problemsof smallsample statistics,usingthe techniqueshe pickedup fromKarl Pearson. R. A. Fisher'sinterestin humangeneticspromptedbiometrical applicationsof his mathematical trainingevenas an undergraduate. As soonas he consideredStudent'swork,he perceivedits importance and beganto extendits applications.Consequently, whenhe startedworkat RothamstedExperimentalStation in 1919, he was readyto respondto the experimental problemsby developingstatisticaltheoryalongwithappropriate methods ofexperimental analysisand design. Keywordsandphrases:Gosset,Student'st,Fisher,smallsamples,analysis ofvariance,correlation. The twomenI wanttowriteaboutareW. S. Gosset, betterknownas Student,who inventedStudent'st test,and R. A. Fisher,whotookup Gosset'sworkand extendedand generalized it so greatly. I am interested in the circumstances thatforcedthemto breaknew groundwheretheydid.Gossetwas a brewer, Fishera mathematics studentwhen he started.Why should they have inventedstatisticalmethodsfor experimenters?How did theypinpointthe problemarea wherehumanthought was dammedbackand baffled? Sometimesthe mostimportant stepin creativework is simplyto ask the rightquestion.The question nobodythinksto ask-that mayseemtoo trivial,or too difficult-butwhich,once asked, must be answered.Gossetaskedthefirstquestionand foundthe answer,despitehis lack ofmathematics. Fishermade it elegantand went on fromthere.Why did they persistwhileexpertsin the fieldignoredor belittled theirwork?Mathematicians onlyslowlyrealizedthat an intellectualrevolutionwas takingplace through the new inductiveuses of mathematics. Even now, peopleoftenthinkstatisticscan be reducedto deduction-that once you have learnedthe theoryyou understandstatistics.Gosset and Fisher believed otherwise. ofinducTheybelievedthatunderstanding tive reasoningis acquiredthroughlearningto deal Joan Fisher Box is theauthorof "R. A. Fisher,theLife of a Scientist" (Wiley,New York, 1978) and lives at 230 North Hillside Terrace, Madison, Wisconsin 53705. 45 knowingthat practical with real data inductively, actionwillbe takenon thebasis ofyourconclusions. Knowingthetheoryis nottheanswer,butit can help youfindtheanswerto statisticalproblems. One ofthe firstthingsyou learnin statisticsis to betweenthe trueparametervalue of the distinguish deviastandarddeviationoaand the sample-standard did statisticians tions. But at theturnofthecentury not.Theycalledbotha and s thestandarddeviation. Theyalwaysused such largesamplesthattheirestithe parametervalue,so matereallydid approximate to theirresults.But it did not makemuchdifference work. theirmethodswouldnot do forexperimental You cannotgetsamplesofthousandsofexperimental largeenoughwouldspreadso points.Anyexperiment farin timeor space thatextraneousvariationwould drownout any effectsof treatmentyou mightbe lookingfor. My storybeginswithGosset,forhe came firstin time.Born in 1876,the oldestof fivechildrenof a Colonelin the Royal Engineers,Gossetenteredthe Royal MilitaryAcademy,Woolwich,to become a RoyalEngineerhimselfbeforebeingrejectedon acwith countof his pooreyesight.He was verybright, highideals,and an impishsense of humor.A most appealing character-quiet, unaffectedlyfriendly, helpful,patient,loyal-everybodylikedand trusted worldof statistics,he him.In the veryquarrelsome termswitheveryone.He managedto be on friendly Afterschoolat was neveremployedas a statistician. Winchester and New College,Oxford,wherehe won in 1899,he tooka job a firstclass degreein chemistry J. F. BOX 46 W S. Gosset,1899 as brewerat Guinness'Breweryin Dublin. He was Head Brewerwhenhe died in 1937.It was the environmentat Guinness'that made him a statistician. That is whatwe lookat next. One hundredyearsago, in 1886,therewas eager buyingon the London Stock ExchangewhenGuinness' was incorporated as ArthurGuinnessSon and Company,Ltd. and sold to the publicforsix million pounds.Until then,it had been a familyfirm,run almostexclusively byGuinnessand one otherfamily. The firstArthurGuinnessand son had startedbrewofDublinin ingat St. James'Gate on thesouthwest 1759.Fortyyearslatertheyweregetting a foothold in the London market,sellingthe brownbeer called withLondonpor"porter"because of its popularity ters.Later,Guinnessoffered whattheycalled"a stouterporter"or stout.As this gainedin popularity, it came to be knownsimplyas Guinness.All through thecenturyGuinnesskeptextendingits marketand to dealwithgrowand extending thebrewery building itwas thelargest ingtrade.By theendofthecentury, in theworld,producing over1.5 millionbulk brewery barrelsa yearfordistribution in Englandand Ireland and forexportto theendsoftheearth. Irelandis a smallislandand,all through thebarleygrowingdistrictsof the south and east, Guinness commissioned maltsters to buylocal barleyand malt it exclusively forthe St. James'Gate brewery. Their commission maltsterswereprosperouscitizensof influencein theirownareas. Afterincorporation, Cecil Guinness,head of the firm, becameChairmanoftheBoardand Christopher DiggesLa Touche (one oftwobrewerstakenin from outsidethe familyin the 1870s)was madeManaging Director.AlthoughGuinnesshimselfwas apparently removedfromthe runningofthe brewery, it was he, withLa Touche'sadvice,whomadeall theimportant decisions.In particular,Guinness,by thenthe first Lord Iveagh,and La Touchetogetherbroughtabout a quiet revolutionaroundthe turnof the century. Theyhad decidedto makethebrewingscientific. Up to thattimethebrewers had learnedas apprentices to knoweveryprocessand to followmeticulously the traditional practices-all theblackmagicofbrewing, as it were.Now,Guinnessespouseda new approach and investeda lotin it. First,theybegan to hirea seriesof the brightest youngmentheycouldfind-all ofthemnewlygraduated fromOxfordor CambridgeUniversity withfirst classdegreesinchemistry-and toappointthembrewers.Otherbreweries employedtheoccasionalchemist in researchlaboratories,but Guinnessbroughtin in positions thesechemistsas theirtop management, previously occupiedby the Guinnessfamily.As soon as each new brewercompleted2 yearsas a junior, learninghis dutiesunderseniorbrewersin everydehe was put in chargeof a sectionof the partment, breweryand researchwork.Guinnessbrewerswere awfullygrand.Stella Cunliffe(1976) recallshow,on arrivalat St. James'Gateas a newly-hired statistician in 1947,she was instructed howto dressand howto behavein hernewwork.If she was luckyenoughto meeta brewerin the corridor, she was told,she was on no accountto recognizehimbutto lowerhereyes untilhe had passed. So, thenewbrewerswereappointed:1893,Thomas Case; 1895,Alan McMullen;1897,ArthurJackson; 1899,E. G. Peake; October1899,W. S. Gosset;January 1900,Geoffrey Phillpotts;and so on. Untilthey at theGuinnesshousefor married, theylivedtogether unmarried brewersat St. James'Gate. At work,they ate togetherin the brewers'diningroom.Offduty, theyseem to have been veryactive outdoors;they skied,fished,sailed,golfed,cycled,and walkedin the WicklowHills, and visitedand read and talkedtogether.In someways,theirlifewas likean extension ofcollege. As theygot involvedin research,of course,they needednewfacilities. Case and McMullenstartedwith GUINNESS, GOSSET, FISHER, AND SMALL SAMPLES chemicalanalysesto tryto identify and quantify what itwasthatgavehopsandbarleytheirbrewing quality. Untilthenthecriteriawerequalitative, e.g.,the"rub" of hops or the "texture"of barley,whichmightbe "milky"or "steely."So, in additionto the brewers' and chemists'laboratories, a newGuinnessResearch Laboratory was openedin 1900,headedby the most chemistalive,HoraceBrown. distinguished brewing The next questionsconcernedthe raw materials, barleyand hops. Wherecould theybe got best and cheapest?What varietiesof barleyand hops,what cultivation and manuring, whatconditionsof drying and storinggavethebestmaltingvalueofbarleyand valueofmaltand hops?So, in 1899, thebestbrewing Guinnessstarteda program ofbarleyplotexperiments throughHenryBennett,theirCommissionMaltster at Ballinacurra,CountyCork.He selectedthe farms and supervisedthe experiments to comparedifferent varietiesand cultivations and fertilizers. The variety trialsseemedmostpromising and theseexperiments withcooperation wereextended, fromothermaltsters and farmersin everybarley-growing districtin Ireland.The newMinistry in Irelandwas ofAgriculture officially in chargeand in 1904 appointeda barley expert,HerbertHunter,who made his headquarters withHenryBennett.Togethertheytouredthebarley severaltimesduringthegrowing season. experiments The GrandTour was shortlybeforeharvest-Grand becausebrewersfromDublinas wellas Bennettand Hunterattendedthistourwhichlasted2 weeks;they spentlongdaysin the barleyfieldswiththe farmers in styleby and local maltsters(and wereentertained whiletheysaw forthemthe maltstersafterward), selvesthe conditionofexperimental plotsand ofthe barleythat would soon be harvestedand sent to St. James'Gate. In 1901,an experimental malthousewas built at St. James'Gate, of a suitablesize formaltingeach batchof barleyseparately.In 1903,an experimental was openedthere.Now,barley experimental brewery grownon experimental plotscouldbe followedfrom seedcornto harvest,throughmaltingand brewingto seriesof thefinalbeerin a uniqueand comprehensive observations (McMullen,1908). At the same time,enquirieswere goingforward varietiesand cultivationof hops,and how regarding bestto dryand storethem.Manyyearslater,Rupert Guinness,thesecondLordIveagh,recalled1902when he had goneon a cycletourofKenthopgardenswith Case and McMullento see whattheycoulddiscover about hop culture.They had droppedin at the new themselves WyeAgricultural Collegeand introduced to theDirector,A. D. Hall. Followingup thisvisitin 1904,thefirstLordIveaghand La TouchevisitedHall and arrangedto renttwo hop farmsnearby,putting 47 R. A. Fisher,1912 one of Hall's graduatesin as manager.Later,these addedto,untilGuinfarmswereboughtand gradually ness was one of the largesthop growersin England. But it all startedfortheyoungresearchbrewers to do researchon whathops to grow,and how,to getthe bestflavorand the longestlifefortheirbeer,and to findouthowmuchit cost(Brown,1980). The lifeofthebeerwasimportant becauseGuinness is a naturally conditioned beer-it has no additivesor preservatives nor,ofcourse,is it pasteurized-andit has to remainpotablewhileit is exportedto Africaor the Far East, or storedin the barrelat varyingtembeforereachingtheconsumer. peratures Now,can you imaginehalfa dozen energeticand brightresearchers givena freehand to explorethe wholesubjectof brewingfroma conditionof almost totalignorance?And giventhe experimental laboratories,barley fields,hop gardens,malthouse,and fortheirexperiments. What happensnext? brewery Well,ofcourse,theystudiedavailableliterature, and foundsomeofituseful.Case readabouta newanalysis of hops forsoftresincontentand triedit. He found thequantityofsoftresinin thehopsparalleledGuinness' qualitativeassessmentof hop "condition." The 48 J. F. BOX conditionof hops storedat ambienttemperatures deteriorated and theirsoftresincontentfellas soft resinwas convertedto hard resin.If the hops were cooledin storage,bothhop conditionand softresin contentfellless. Thus, softresinslookedlikea good predictorof hop condition,and the brewersbegan routineanalysisof hop softresincontent.Guinness foundthis measureveryvaluableforassessinghops beforepurchase.In particular,Americanhops were cheaperthanEuropeanand it turnedouttheyhad a naturally highsoftresincontent. in 1902,studieswerepublishedin which Similarly, maltingqualityof barleywas shownto dependon beganroutineanalyses nitrogen content.The brewers ofbarleyfornitrogen, in additionto theirrecordsof and theold qualsize ofbarleycorns, yield,moisture, and texture. itativeassessmentsformellowness and measureReadingled to analysis,experiments, ments.Theybeganto accumulatedata and, at once, becausetheirmeasurements theyranintodifficulties varied.The effectstheywere lookingforwere not as theyhad expected, usuallyclearcutor consistent, and theyhad no wayofjudgingwhetherthe differoftreatment oraccident. encestheyfoundwereeffects Two difficulties wereconfounded: the variationwas werefew.As a result,for highand the observations his hop-cooling experiexample,whenCase reported mentsin 1898,he pointedout"theweaklinkbetween or analysisand thebrewing examination value,"that and effect,and said "mostof is, betweentreatment the results . . . require substantiationby irreproacha- ble figures," whichhis obviouslywerenot. He convalue in termsof lifebecluded,"The comparative tweenthecooledand uncooledhopsmaybe regarded as of the orderof about 10%" (Brown, provisionally 1980). He was not stickingout his neck,althougha is quitea lot. 10% difference on starting theirexperiments, found Otherbrewers, thesamedifficulty. Theyhadnowayoftakingaccount theirdata.Aftertrying ofthevariationin interpreting to correlatefertilizertreatmentswith barleyplot thateach yields,E. G. Peake endedup recommending farmer shouldfindoutthemanurialneedsofhis own land, because "the resultswere most irregularand varied with everyfarm."Peake tried to correlate springrainfallwithyieldofbarleyfromRothamsted data thatA. D. Hall also had analyzed.By takinga different slightly periodthanHall had fortherainfall, in theopposite Peake managedto discerncorrelations direction to thosefoundby Hall, and had to giveup, concludingthat "the whole subjectof the effectof weather ... deserves furthercareful investigation" (McMullen,1908). In the case of the barleyplots,the variationwas excessive.First,it was obviousthatalien barleywas intermixed in theseed,forperhaps10% ofplantsin a plotofa short,late-maturing varietywouldgrowtall and matureearly,and so on. Next,plantsevenfrom seed genuinelyof one varietyshowedmanysmaller variations.Old Irishwas the mostheterogeneous variety.Unliketheotherswhichhad beenselectedoriginallyfroma singleplantorsmallstandofplants,Old Irishwas an indigenousIrishvarietythathad never beenselected.It retainedthevariability accumulated throughout its ancestry, and madean extremely unevenand unsatisfactory crop. Thiscauseofvariation, at least,couldbe eliminated. Barleyis usuallyself-fertilized. In 1904, therefore, BennettandHunterbeganpropagation ofeachvariety of interestin purelinesfromsinglegrainsofbarley. In 1907theyhad enoughpure-lineseedto sowall the experimental plots,and on the GrandTour everyone marvelled at thewonderfully evengrowth ofthefields. But geneticvariability was a specialcase. One could noteliminatevariationsofrainfall, birddamage,soil chemistry, temperature (whichaffected brewingbeer as wellas growing crops),Qranyoftheunrecognized variablesaffecting theirdata. Theyneededsomeway to decidewhichdifferences to ignoreand whichto takeseriously. The youngresearchbrewers workedwelltogethersomewereveryclosefriends. Each seemedto fitinto hisownrolein brewery affairs. Andto themit seemed naturalto take theirnumericalproblemsto Gosset. He haddonesomemathematics at Oxfordand seemed lessscaredofmathematics thantheywere.(In a report on the theoryof error,he observed,"It may seem strangethat reasoningof this naturehad not been morewidelymadeuse of,butthisis due,first, to the populardreadofmathematics.") He was alwaysready to listento them,and veryquickto grasptheirconcerns.He wouldalwaysdo his bestto comeup with an answer,goingback to firstprinciplesand arguing to a solution(McMullen,1938),and he was through notsatisfieduntilhe reachedit. So, statisticalproblemscameto him.He gotholdofAiry'stextbookon the theoryof errors,and studiedand annotatedhis In 1903he couldcalculatestandcopyin themargins. ard errors.In 1904 he wrotea reporton the subject forthe brewery. This reportled directly to his being sent to consultKarl Pearson about his difficulties, Professor Pearsonbeingthegreatnamein biometrics in thosedays.In 1905,Gossetwas usinga homemade measureof correlation, based on examinationof the difference betweenE (A + B)2 and E (A - B)2, at the timehevisitedPearsonandlearnedabouttheproductmomentcorrelation coefficient (Pearson,1938). Onegreatdifficulty in interpreting theirexperimental data was thatthe sampleswerealwayssmall.For instance,the barleyexperimentsstartedwith four farmseach growingone plot of each variety.The estimatem ofthe meanbased on a sampleoffouris GUINNESS, GOSSET, FISHER, AND SMALL SAMPLES obviously notexact,and theerrorin theestimates of This Gosset thestandarddeviationcannotbe ignored. For him,the main questionwas exactly recognized. how muchwidershouldthe errorlimitsbe to make by usingthe estiallowanceforthe errorintroduced of the parameters instead m and s mates gt and a. PearsoncouldnotanswerthatquestionforGossetin whichwas:whatlevel 1905,northeonethatfollowed, significant? be called ofprobability should tookplaceinJuly1905duringGosset's The meeting summerholidayin Englandwhenhe cycledabout20 milesfromhis parents'hometo Pearson's summer home,and theyhad a longtalk,whichGossetfound veryhelpful.Later,he recalledthatPearson"wasable in abouthalfan hourto putme in thewayoflearning the practiceof nearlyall the methodsthenin use." He wentbackto Guinnessand practicedthosemethof ods forthe nextyearas actingBrewer-in-Charge whereanalysisofthecoltheExperimental Brewery, was lecteddatabecamea majorconcern.The meeting in fact,thatGuinnessarrangedforGosso successful, at Uniset to spenda yearin Pearson'sdepartment versityCollege,London,"to takeup the studyofthe law oferror,the workingof whichwe have foundof Gossethad foundhis greatservicein the brewery." vocation. Duringthe academicyear 1906-1907at Pearson's Gossetworkedouttheexactanswerto his laboratory, questionabout the probableerrorof the mean and values of his criterion z= tabulatedthe probability (m - ,)/s for samples of N = 2, 3, ..., 10. He tried of the correlation also to calculatethe distribution coefficient by the same methodbut managedto get theansweronlyforthecase whenthetruecorrelation is zero.It is almosta miraclethathe got so far,for his mathematicshad to be helped out by inspired guesses. Gosset quietlyreturnedto Guinnessin 1907, to BrewoftheExperimental becomeBrewer-in-Charge 7 years. eryforthefollowing He was just in time to providethe answer to McMullen'sprayers.The firstseriesofbarleyexperimentswas comingto a close.The resultsof7 yearsof to Guinness. workwereto be analyzedand repoTted McMullenpreparedthe text,with assistancefrom of barJacksonon the taxonomyand identification from leys,fromPeake on mattersofsoil and weather, whileGosset Hunteron geneticsand barleybreeding, took overthe entirestatisticalanalysis:therewere barley yields, measurementsand assessmentsfor and farmsand districts, different varieties,different different seasons;thereweremaltingassessmentsand resultsforthesamebarleylots.Gossetfinally brewing in termsof analyzedbarleyyieldand.qualitytogether he quotedthe valueperacre,and at everyopportunity odds (fromhis newlycalculatedt table) it was the 49 especially naturalwayforhimto expressprobabilities, in thenativelandofhorseracing!(McMullen,1908). Now,theyknewthat AndGuinnessweredelighted. Ireland andtheywanted barley for was the best Archer 1,000 island. They discovered the to growit all over barrelsofpure-lineDanish Archerseedwas available theirComto buy.Theyboughtit all. Then,through itas seedtochosen theydistributed missionMaltsters, togetherwith300 barrelsof Danish Archer farmers a miniseedtheyhad grownin Ireland,guaranteeing mumyield,all ofit to be boughtbackat harvest.The nextyeartheyhad some 10,000barrelsto distribute as seed,whichwas grownand boughtback as before. Afterthat, they had enoughto distributeto any farmerwho wantedto growit in the ordinaryway. mostbarleyseedin Ireland Formanyyearsthereafter, was selectedbybrewersat the GrandTour reserving and stored thecleanestfieldsforseed,to be harvested by theirmaltsters.Maltsters apart and distributed continuedas seedsmenafterWorld War I when thehybridSprattArchergavewayto a newvariety, whichHunter,at his firstattemptat making Archer, a hybrid cross,hadcreatedin 1908.Betweenthewars, 90% ofthebarleygrownin Irelandwas Spratt-Archer, as was most of what was grownin England too (Hoctor,1971;Hunter,1882-1959). In hissparetimein 1908,Gossetpreparedhispapers on theprobableerrorofthemeanand ofthecorrelaforpublication(Student,1908a and tion coefficient 1908b).Guinnesshad earlieragreedto permitpubli"Itwasdecided cation.As theBoardminutesrecorded, by La Touche that such publicationmightbe made Theywouldbe without thebrewers'namesappearing. merelydesignated"Pupil" or "Student."So Gosset Student.He continuedstatistical tookthepseudonym researchesin his sparetime,because theywerenot business.And,in his sparetimealso, brewery strictly he made himselfthe trustedfriendof E. S. Beaven. Beavenwas a self-mademan and a ruggedindividualist who had no patienceat all withstatisticsand But Gossetshowedhimselfso genuinely statisticians. in Beaven'sworkthatsoonhe had Beaven interested went(Beaven eatingoutofhishandso faras statistics papers). Beavenwas a maltsterat Warminster, Wilts,celewhichwas his passion.In bratedforbarleybreeding, inthesubject, 1904,whenGuinnessbecameinterested MaltsterforGuinBeavenhadbecomea Commission ness and, in 1919,he becameGuinness'buyerof all importedbarley,but he alwaysregardedthe barley breedingas his privateaffair.At that time,he had alreadyselectedandbredpurelinesofseveralvarieties and, withhis friendsfromthe new CambridgeUniRowlandBiffinand Department, versityAgriculture T. B. Wood,he had startedto makecrossesbetween varietiesin the hope of creatingnew and better 50 J. F. BOX barleys.Theyhadbecomekeengeneticists and breederssincetherediscovery in 1900ofMendel'spaperon hybridization. In 1904,Guinnesssoughttheiradvice onthebestdirection forthebarleyresearchin Ireland. In 1907,GossetcameintoBeaven'slifejust in timeto helphimdevelopthe nursery experiments neededto test severalof Beaven's hybridsagainsteach other and against standardvarieties.In 1912 and 1913, Gossetevenmanagedto runcooperative experiments withHunterand Bennettat Ballinacurra,Biffinat all testing Cambridge,and Beaven at Warminster, theirownagainsttheothers'selectedseed.He didthe analyseshimselfand, analyzingalwaysin termsof differences betweenpairsof neighboring plots,manfortheseexperiagedto evolvea testof significance ments-an extensionof Student'stest, producing muchthesamenumerical resultsas an F test(Beaven papers). So wearriveat Cambridge in 1912whenR. A. Fisher was introduced to Gosset.Fisherwas an undergraduate at Gonvilleand Caius College,Cambridge, and his tutorwas F. J. M. Stratton,the astronomer. Biffin was a Fellowofthe same Collegeand had interested Strattonin analyzinga fieldexperiment forT. B. Wood.Gossethad methimin theircompany.In April 1912, Fisher's firstpaper was published,"On an AbsoluteCriterion forFittingFrequencyCurves,"in whichheintroduced function theideaofthelikelihood and the methodofmaximumlikelihoodbut,without the word "likelihood,"his presentationwas rather Gossetread it and thoughtit "A neatbut confusing. as faras I couldunderstand it,quiteunpracticaland unserviceableway of lookingat things."In June, Fishersat hisfinalexaminations forthemathematics degree.He mustthen-havetalkedto Strattonabouta he had foundbetweenthe formulafor discrepancy standarddeviationin hisownpaperand in Student's. Maximumlikelihoodgavea denominator n insteadof n - 1 in thisequation.Strattontoldhimto writeto Gossetabout it. He wroteto introduceFisher,and Fishersenthisproof.The restofthestoryis givenas Gossetwroteof it to Pearsonon 12 September1912 (Pearson,1968): This [proof],Stratton,the tutor,made him send me and withsomeexertionI masteredit, spotted the fallacy(as I believe) and wrotehim a letter inthematter I hope,an intelligent interest showing, and incidentally makinga blunder.To thishe repliedwithtwofoolscappagescoveredwithmathematicsof the deepestdye in whichhe proved,by thattheformulawas,afterall, usingn-dimensions ,/ (x - m) 2/(n 1) and, of course,exposingmy his stuffand wrote mistake.I couldn'tunderstand and said I was goingto studyit whenI had time.I actuallytookit up to the lakes withme-and lost it!Nowhe sendsthisto me [themathematical proof - of Student'sdistribution]. It seemedto me thatif it'sall rightperhapsyoumightliketo puttheproof in a note. It's so nice and mathematicalthat it mightappealto somepeople. Andhe said,"Wouldyoumindlookingat it forme.I don'tfeelat homein morethanthreedimensions even ifI couldunderstand it otherwise." That's Gossetforyou-his humorousaccountofa chapterof accidents,his frankinabilityto deal with stiffmathematics,and his generousappreciation nevertheless of the "nice and mathematical" proof that deservedpublicationif it was "all right."In contrast, Pearson,whomighthaveunderstood it,did notchooseto publishthe note.(Pearsondid publish it, however,in 1915,whenFisherincludedit in his ofthecorrelation paperon the samplingdistribution the sample coefficient, againderivedby representing in n-dimensional space.) Let us turnnowto Fisherand ask whyhe should be involvedwithappliedmathematicswhenit was Afterall, as Stratdespisedbypuremathematicians. tonwroteon Fisher'sbehalfwhenhe was lookingfor a job in 1919,Fisher"couldhave been a firstclass mathematician hadhe stucktotheropes,buthewould not."Why? Fisherwastheseventhchildofa fineartsauctioneer in theWestEnd ofLondon.Alwaysbrilliantat mathin mathematics to ematics,he had wona scholarship Harrowand laterto Gonvilleand Caius College,Cambridge,wherehe won a firstclass honorsdegreein mathematics. UnlikeGosset,he had no contactwith orexperimenters experiments uptothattime.Instead, he had a consuming desireto makehimselfuseful,to servehis countryand humankindin a significant, practicalway.Pure mathematics, althoughit maybe a sourceofthegreatestintellectual to its gratification professor,is not a subjectof immediatepractical utility. Fisherhad nearlydecidedto takebiologyinsteadof mathematics at Cambridge, because he could see its and he was seriouslyinterested in evoluimportance, tion and genetics.He was especiallyinterestedin humangeneticsand that broughthim at once to statisticalconsiderations becausefewhumancharactersaredichotomous, beingcontrolled bya singlepair in of geneslike the charactersMendelinvestigated gardenpeas-for example,tall or shortplants,green or yellowpeas. Variablessuchas humanheight,skin have continuousdistributions color,and intelligence and musttherefore be controlled by multiplepairsof ofgenesin human genes.To dealwiththepopulations One can one mustdeal withprobabilities. heredity, statewhathappensonlyin statisticalterms. It seemsthat as soon as he felthe could choose, Fisherveeredto biologicalapplicationsofmathematics. Having succeeded in formingthe Cambridge GUINNESS, GOSSET, FISHER, AND SMALL SAMPLES EugenicsSocietyduringhis secondunderUniversity graduateyear,he was chairmanoftheundergraduate in 1911-1912.AttheNovember1911meetcommittee inghe was the speakeron the subjectof Mendelism (i.e.,Pearson'sstatistical (i.e.,genetics)and Biometry methodsofdealingwithcontinuousvariables),which he explained,sayingbothwerenecessaryforhuman geneticresearch.He mentionedMaxwell'stheoryof gases to illustratethe necessityof speakingof gene We havenoticed in termsofprobabilities. populations frequency curves. that his firstpaperwas on fitting We have seen how quick he was not onlyto prove in 1912butto realizetheimporGosset'sdistribution of the samplingdistributions tance of determining otherstatisticsin commonuse and to go on to find ofthecorrelation coefficient thesamplingdistribution thereafter. in 1915and a seriesofotherdistributions in 1912-1913,he As a graduatestudentat Cambridge choseto studythetheoryoferrorswithStrattonand JamesJeans. quantumtheorywiththeastronomer, A veryimportant earlypaper of Fishertookup a questionthat had been foughtover for a decade: was compatable in fact,humaninheritance whether, with Mendelian principles. Biometricianswere divided:Pearson said not, and rejectedgenetics; G. Udny Yule thoughtit mightbe compatablebut couldnotproveit; geneticists just assumedall inheritancewas Mendelian.Fisherbroughthis mathematical geniusto bear on the problemand provedthat the inheritance of continuousvariableswas entirely consistentwithMendelianprinciples-in fact,that Mendelianprinciplesmust result in preciselythe numericalrelationshipsthat existed.The way he as the resultbecause,in provedit was as important orderto deal withintractablesample correlations, he formulatedthe problem in a differentway and introducedthe conceptof analysisof variance (Fisher,1918). AlthoughFisher's biologicalinterestsstimulated researches himto do theseimportantmathematical even while he was a school teacherduringWorld War I, it was afterhe became the statisticianat RothamstedExperimentalStation in 1919 that he really,as he put it,foundhis feetin research.Papers funpouredfromhis hand,manyofthemcontaining damentalnewwork.Between1921and 1926thenumrosefrom13 to 55 papers,and berofhispublications he wroteStatisticalMethodsforResearchWorkers research is theoldestagricultural (1925).Rothamsted with institution inGreatBritain.Fisherwaspresented with experiments data-lots of data fromlong-term wheatand rootsand grassplots and farmrotations. His firstjob was to analyzethe manurialexperiment withwheaton Broadbalkfieldwherethesame fertilizershad beenappliedto the sameplotsfor67 years. morebe donewiththe He was asked,couldanything 51 data than A. D. Hall had done some yearsbefore? Besides,therewereabouta dozenstaffwhenFisher wentthere,all busywithresearchprojects. The activitiesat Rothamsted,the interestsand problemsof the staff,the discussionsovera cup of tea, the data, all werea greatstimulusto Fisher's His paperswereall a mix andinventiveness. ingenuity theory andpractice. ofstatistical ofnewdevelopments FounTherewasonebigpaper"On theMathematical dationsof TheoreticalStatistics"(1922) but even in dethathe insertedthe firstnonlinearexperimental sign,whichhe had createdforone ofhiscolleaguesat The seriesof studiesin cropvariation Rothamsted. wereostensibly appliedpapers."Studiesin CropVaroftheyieldofdressed iationI" was "Anexamination grainon Broadbalk"(1921).Thatcontainsorthogonal polynomialsand the analysisof varianceprocedure "Studies and one or twootherstatisticalinnovations. in CropVariationII: The manurialresponseofdiffertheanalysisof entpotatovarieties"(1923)introduces variancetablewiththe newz testand its validityis oftheplot statedto be conditionalon randomization treatments. Theseideasare all just slippedin without ofrainfallon the A paperon "The influence fanfare. (1924)includesa short yieldofwheatat Rothamsted" data essayon problemsof analysisof meteorological of the and the derivationof the null distribution in additionto its multiplecorrelationcoefficient advertisedcontent,not to mentionthe use of what in the analfunctions wouldtodaybe calledtransfer ysis.Thiswasconceivedas "Studiesin CropVariation III" buttheprefixwasdroppedwhenthepaperbecame journal unsuitableforpublicationin an agricultural merely. all thisnewmaterial WhenFisherbrought together in the book StatisticalMethodsforResearch Workers oftheorybut in (1925),it was not as a development could of that terms methods experimenters useintheir evenif theycouldunderstand, work,whoseprinciples were them. Fisher themathematical beyond workings in thepreface: setforthhisphilosophy For severalyearsthe authorhas been workingin somewhatintimatecooperationwitha numberof the presentbook biologicalresearchdepartments; is in everysensetheproductof thiscircumstance. Daily contactwiththe statisticalproblemswhich presentthemselvesto the laboratoryworkerhas stimulatedthe purely mathematicalresearches uponwhichare based the methodsherepresented. to showthatthetraLittleexperienceis sufficient ofstatistical ditionalmachinery processesis wholly Notonly unsuitedtotheneedsofpracticalresearch. does it take a cannonto shoota sparrow,but it built The elaboratemechanism missesthesparrow! on the theoryof infinitely large samplesis not data.Onlyby accurateenoughforsimplelaboratory J. F. BOX 52 systematically tacklingsmall sampleproblemson theirown meritsdoes it seem possibleto apply accurateteststo practicaldata. Such at least has beentheaim ofthisbook. REFERENCES BEAVEN, E. S. PapersdepositedwithGuinnessBarleyResearch Station,Warminster, Wilts. BENNETT, J. H., ed. (1971-1974).CollectedPapers of R. A. Fisher 1-5. Univ.AdelaidePress,Adelaide. GuinnessSon and BROWN, J.F. (1980).Guinnessand Hops. Arthur Co. (ParkRoyal),Ltd.,London. CUNLIFFE, S. (1976).Interaction. J. Roy. Statist. Soc. Ser. A 139 1-19. R. A. (1912)On an absolutecriterion forfitting frequency curves.Messeng.Math. 41 155-160.(See Bennett1 No. 1, 1971.) FISHER, R. A. (1918). The correlation betweenrelativeson the of Mendelianinheritance. supposition Trans.Roy.Soc. Edinburgh52 399-433.(See Bennett1 No. 9, 1971.) I. An examination FISHER, R. A. (1921).Studiesin cropvariation. oftheyieldofdressedgrainfromBroadbalk.J. Agric.Sci. 1 1 107-135.(See Bennett1 No. 15,1971.) FISHER, R. A. (1924).The influence ofrainfall ontheyieldofwheat FISHER, at Rothamsted. Philos. Trans. Roy. Soc. London Ser. B 213 89-142.(See Bennett1 No. 37,1971.) FISHER, R. A. and MACKENZIE, W. A. (1923). Studiesin crop potato variation.II. The manurialresponseof different varieties.J. Agric.Sci. 13 311-320.(See Bennett1 No. 32, 1971.) HOCTOR, D. (1971). The Department's Story: A History of the DepartmentofAgriculture.Instituteof Public Administration, Dublin. H. (1926).The Barley Crop. ErnestBenn,London. H. (1882-1959).Unpublished papersdepositedwiththe History,Univ.of Reading,Reading, Instituteof Agricultural Berks. MCMULLEN, A. (1908). Barley Cultivation in Ireland. Arthur HUNTER, HUNTER, Guinness Son and Co., Ltd., Dublin. MCMULLEN,L. (1938). William Sealy Gosset 1876-1937: "Student" as a man. Biometrika30 205-220. PEARSON, E. S. (1938). William Sealy Gosset 1876-1937: "Student" as statistician.Biometrika30 220-250. and E. S. (1968). Studiesin the historyof probability betweenW. S. GosXX. Someearlycorrespondence statistics. PEARSON, set, R. A. Fisher,and K. Pearson, withnotes and commentary. Biometrika55 445-457. STUDENT (1908a). The probableerrorof a mean. Biometrika 6 1-25. (1908b). Probable errorof a correlationcoefficient. Biometrika6 302-310. STUDENT