Вы находитесь на странице: 1из 6

Gosset,Fisher,and thet Distribution

JOAN FISHER BOX*

Mean" and "Probable Error of a CorrelationCoef-


Letters fromW.S. Gosset to R.A. Fisher are used to ficient.'"
describe Gosset's firstcontactsand growingfriendship In the summerof 1912Gosset was 36; Fisher,gradu-
withFisher, theircollaborationover the tabulationof ating from Cambridge University,was 22. Fisher's
Student's t, and the wider applications of Gosset's firstpaper, in whichthemethodofmaximumlikelihood
work that were made possible by Fisher's extension was introduced,had appeared in April. Deriving the
and generalizationofthedistribution theoryintheearly maximumlikelihoodestimateforthevarianceofa nor-
1920's. These lettersrevealGosset's practicalapproach mal sample, Fisher had come up withthe divisorn in-
to the statisticalproblemsof his brewerywork that stead of the n - 1 of Student's formula.His college
''naturallyrequired" the solutionsforwhichthe name tutortoldhimto writeto Gosset about thediscrepancy,
of Student is famous; and they exhibithis appealing which he did. Several letterspassed between them,
character,modest yet self-reliant, in relationshipnot in the course of which Fisher conceived the notionof
only to Karl Pearson and R.A. Fisher but to statistical representinggeometricallythe configurationof the
assistants stillin training. sample in n-dimensionalspace. This representation
immediatelygave him the concept of degrees of free-
KEY WORDS: Geometric representation;Sampling
dom and the correct divisor n - 1 for the variance
distributions;t distribution;Randomization;Analysis
formula,and moreoveryieldedthemathematicalproof
of variance; Degrees of freedom.
of Student'sdistribution,whichFisher sentto Gosset.
Gosset (Pearson 1968) immediatelysent the proof to
Pearson, suggestingpublication. "I couldn't under-
standhis stuff,"Gosset wrote."I don't feelat home in
1. INTRODUCTION morethanthreedimensionseven ifI could understand
it otherwise." He concluded, however, "It's so nice
and mathematicalthatit mightappeal to some people"
W.S. Gosset, betterknown under the pseudonym (p. 446).
Student,was an extraordinarily appealing individual, The proof was not published, the correspondence
generousto a fault,humble,enthusiasticin thepursuit ceased; but a seed had been sown that would grow.
of his varied interests,and helpful.He had a flairfor As Fisher wrote in Statistical Methodsfor Research
graspingtheheartof a problemthatmade his contribu- Workers(1925, p. 23), "The study of the exact dis-
tionsto statisticsimportantout ofall proportionto their tributionsof statisticscommences in 1908 with 'Stu-
number.It is a tributeto his personal qualities thathe dent's' paper The Probable Errorof a Mean. Once the
was a friendofbothKarl Pearson and R.A. Fisherfrom truenatureof the problemwas indicated,a largenum-
the timehe made theiracquaintance in the earlyyears ber of samplingproblemswere withinreach of mathe-
of the centuryuntilhis death in 1937. It is a pleasure matical solution." Having grasped the true nature of
to reviewthegrowthofhisfriendship and collaboration the problem fromStudent's work, and the means of
withFisher in the early 1920s; and, because the story mathematicalsolutionthroughthe geometricalrepre-
comes alive in thelettersfromGosset thatFishersaved sentation,Fisher proceeded nextto solve the problem
and latersummarized(Gosset 1970)-about 80 ofthem of Student's second paper by deriving the general
in the period of eight years here considered-to sampling distributionof the correlation coefficient.
allow Gosset's own words to reveal his personality When Gosset received the publishedpaper in 1915,he
and his approach to his statisticalresearch and col- wrote, " When I firstsaw it, I nearly wrote to thank
leagues. (For a more technicaltreatmentof the same you forthe kindway in whichyou referredto my un-
period, see ChurchillEisenhart's (1979) article, "On scientificefforts";he suggestedhow he would tackle
the Transitionfrom'Student's' z to 'Student's' t.") the remainingproblemof estimatingthe errorof cor-
From 1899on, Gosset was employedas a brewerat relationestimates, adding, however, "But of course
Guinness's Brewery, and his statistical researches anythingalmost would do if it gave an integrableex-
sprangfromthe needs of thiswork. In particular,dur- pression" (Gosset 1970; no. 1). This was the nub. In
ing a year he spent in Karl Pearson's departmentat principle,the problemcould be solved by calculating
UniversityCollege, London, in 1907,Gosset prepared the curves, and hence their integrals,for each case
two papers (Student1908a,b),"The ProbableErrorofa separately,and thiswas infactundertakenin Pearson's
department(Soper et al. 1917). Fisher showed, how-
* JoanFisherBox, 3437
EdgehillPkwy.,Madison, WI 53705,is the ever, thata verysimpleand close approximationcould
authorof R.A. Fisher, The Life of a Sctientist(1978) and of articles be obtained by a transformation, which renderedthe
on relatedtopics. massive tabulationof ordinatesalmost superfluous.

?UThe AmericanStatistician,May 1981, Vol. 35, No. 2 61


This content downloaded from 66.195.118.21 on Mon, 23 Mar 2015 13:23:02 UTC
All use subject to JSTOR Terms and Conditions
2. FELLOW PRACTITIONERS to get a statisticiansoon, when he gets the money I think,
and it mightbe worthwhile to keep yourears open to news
fromHarpenden. (Gosset 1970, no. 3)
Three years later, in December 1918, Fisher wrote
again to discuss applicationof Gosset's (Student 1914) Naturally,Gosset was the firstto hear eight months
difference correlationmethodto birthrate and marriage laterthatSir JohnRussell had appointedFisher statis-
data. Gosset answered at great length,ending, tician at RothamstedExperimentalStation, and was
Well, I expect you've about had enoughofthis.I am send-
glad, for,as he said, "There should be lots of interest-
ing it via UniversityCollege as I cannot lay hands on our ing workto be done there,and theymighteasily have
formercorrespondenceand theonlyclue to yourwhereabouts got someone there who would have been worse than
is the postmarkwhich seems to be Reading. If I oughtto ad- useless" (Gosset 1970,no. 4). He was happyto answer
dress you as Major or Professorpray forgiveme, here one Fisher's requestforadvice bothabout whatcomputing
failsto keep abreast of events. (Gosset 1970, no. 2)
machine to buy and about home brewing- "less
Gosset mightnothave knownhow to address Fisher,or troubleto buy Guinness," he added, "and let us do it
where,buthe could recognizethatherewas a manwith foryou" (Gossett 1970, no. 4), which Fisher did.
a similarscientificoutlook and similarpracticalgoals, Then, forthe last time,the correspondenceflagged
thougha mathematician.In 1915Gosset had explained for nearly two years. Gosset's habitual complaintof
thathe had become interestedin the correlationcoef- his laziness was merelyan excuse forpackingthe day
ficient(like the standarderrorof the mean) because he with an unusual variety of pursuits, not always so
needed the solutionto the problemin his work. "'perfectlyuseless"' as he claimed in a letter dated
It happened thatI was mixed up witha lot of large scale
20 May 1924:
experimentspartlyagriculturebut chieflyin an Experimental My laziness takes the formof pursuingwithtremendous
Brewery. The agricultural(and indeed almost any) Experi- energy some perfectlyuseless aim while totally neglecting
mentsnaturallyrequireda solutionof themean/S.E. problem, thingswhich oughtto be done at once. Instead of writingto
and the ExperimentalBrewerywhich concerns such things you I read "Eminent Victorians;- when I oughtto be filling
as the connectionbetween analysis of malt or hops, and the in numerous income tax forms (in order to recover much
behaviourof the beer, and whichtakes a day to each unitof needed cash overpaid to the British Government) I play
the experiment,thus limitingnumbers,demanded an answer patience; when I ought to be tyingup loganberrieswhose
to such questions as, "If witha small numberof cases I get a crop will otherwisebe wasted, I spend my time in an alto-
value r, what is the probabilitythatthereis reallya positive getherprematurethinningof pears. (Gosset 1970,no. 45)
correlationof greatervalue than, say, .25?' (Gosset 1970,
no. 1) In March 1922 Fisher sent Gosset some offprints,
includingthe controversialarticle "On the Interpreta-
Now Gosset explainedhow, beforehis year withPear-
tionofx2 FromContingencyTables,' in whichdegrees
son, he had evolved the differencecorrelationmethod
offreedomwerefirstintroduced.The articlewas bound
forsimilarlypracticalreasons; and why,thoughPear-
to infuriatePearson. Yet in a covering letterFisher
son had since withdrawnhis objections to the method
seems to have suggested formationof a biometrical
and indeedhad extendeditsapplications(see reviewby
society that only Pearson could at thattime have ini-
Fisher 1916), Gosset still doubted its usefulness,ex-
tiated,an extensionoftheinformalsocietyofbiometers
cept in its originalapplication,which he described as
at UniversityCollege, to include "practitioners"using
follows.
smallsample statistics.Fisherknewhe was thelast man
One of the points to be investigatedwas the connection Pearson was likelyto listento and hoped theidea might
between the LaboratoryAnalysis of malts and the lengthof be betterreceived if it came fromGosset, who was
time the resultingbeer remained potable, as measured by
acidity. (If you have ever drunk Guinness in England you
Pearson's friendas well as the firstsuch practitioner.
will understandwhy.) But one of the chieffactorsin acidity Gosset replied
productionis the temperature(both at brewingand during When I am nextover I willsee how theland lies at Univer-
storage) and at that time our arrangementsfor stabilising sityCollege. Ofcourse ifthe'Biometers' are to be anyuse they
temperatureat the ExperimentalBrewerywere ratherprimi- shouldincludetheleadingpractitioners, butI ratherfancythat
tive. Hence I was forced to take firstdifferencesbetween Pearson's idea is thatit is a sort of UniversityCollege Club.
successive brewings . . . to eliminatethe largetemperature Besides, as you say, he is perhaps a littleintolerantof criti-
effect.(Gosset 1970, no. 2) cism, most of us tend to thatI fancyas we grow older.
In most of yourdifferenceswithPearson I am altogether
Gosset did not know that Fisher also was involved in
on your side and in some cases I have agreed to differfrom
biological work untilhe received a replyto his letter, him long ago. (Gosset 1970,no. 5)
in which Fisher admittedhe was neitherMajor nor
Professor,but had been a schoolteacher throughout Thus a curious three-wayrelationshipwas formed.
the war and was now looking for a job. Fisher went Pearson and Fisher did not communicatedirectly,but
on to discuss some statisticalwork he was doing with on occasions did throughGosset, and Gosset saved his
orcharddata fromWoburn.Gosset's interestwas im- communicationswithPearson forthe occasions when
mediatelyaroused: he visitedhim in person.
By thistimeFisher's workon samplingdistributions
Your fruittree work must be interesting:as a gardenerI
am of course familiarwith alternate seasons of apples and
was producing new results in response to Gosset's
pears and (as a brewer) of hops. I don't know whetheryou needs. On 3 April 1922 Gosset inquiredabout the dis-
are lookingfora job in thatline but I hearthatRussell intends tributionof an estimated regression coefficient,"'a

62 (? The AmericanStatistician,May 1981, Vol. 35, No. 2


This content downloaded from 66.195.118.21 on Mon, 23 Mar 2015 13:23:02 UTC
All use subject to JSTOR Terms and Conditions
problemto whichhe presumablyreceived the solution prisingthat when Gosset checked the two tables, he
by return,"as Fishernoted later(Gosset 1970,no. 5). declared them both "perfectlyrotten' (Gosset 1970,
On 12 AprilGosset asked about the distribution of the no. 36).
partial regression/correlation
coefficients:"This also Fisher proposed a new methodof attack. Using the
was probablyquicklyansweredforon May 5thhe refers value of t instead of z, he saw that the probabilities
to the solution" (Gosset 1970, no. 6). The letterof could be arrivedat by an expansion in inversepowers
5 May has several pointsof interest.First,its opening of n. The expansion is asymptoticso that as n is in-
shows Gosset's self-deprecation as a mathematician- creased, the numberof nonnegligibletermsof the ex-
his feelingsabout Fisher's use of the word evidently pansion is reduced, and as n tends to infinity,the
were to become a favoritejoke in his letters.He wrote: leading term in t2 alone remains. Thus, as Student's
Many thanksforyourletterand thePhil. Trans. paper. As cos 0 formulabecame morecomplicatedwithincreas-
to the formerin whichyou say thatit is encouragingto have ingn, Fisher's methodbecame moreand moresimple.
your work understood, I fear that I can't conscientiously Before the calculations were begun, however, Gos-
claim to understandit,but I take itforgrantedthatyou know set plannedto visitFisherin England.On 15September
what you are talkingabout and thankfully use the results!
1922 he motoredover with his sister and stayed the
It's not so muchthe mathematics,I can oftensay "Well,
of course, that's beyondme, but we'll take it as correct" but nightin Harpenden. It was the firsttimehe and Fisher
when I come to 'Evidently' I know that means two hours had met,and the meetingconfirmedall theirfavorable
hard work at least beforeI can see why. preconceptions.Althoughthetalkcenteredon thenew
tabulation and its multitudinoususes, Gosset could
The same lettershows thatFisherwas usingStudent's
not conceive thathis table could ever be so important
tables (enteredforthe rightnumberof degrees of free-
as Fisherclaimed. Afterthe visithe sent-Fishera copy
dom) to test the significanceof regressions("which,"
of Student's tables, "as you are the only man that's
Gosset remarks,"is of course very satisfactoryfor
ever likelyto use them!" (Gosset 1970, no. 11).
Student!"), of partial correlations and regressions,
The tables had been in constantuse at the brewery
and ofthedifference betweentwo means fromsamples
forthe past 14 years, but they were almost unknown
of differentsize-a novel notion to Gosset. Thus it
elsewhere.Gosset accepted theirneglectbythestatisti-
appears thatFisheralreadyhad a unifiedunderstanding
cal establishmentwith a good grace. In contrast,
of the distributiontheoryinvolved, which had been
Fisher's sense ofjustice was outraged.In a "Historical
progressivelyilluminatedfor him throughgeometric
Note" at the beginningof Statistical Methods for
representation.Moreoverthe lettershows thatFisher
Research Wotkers (14th ed., 1970, p. 23), we read:
had suggestedthecalculationofnew tablesofStudent's
distribution and a new methodof calculation,forGos- "Student's" workwas not quicklyappreciated(it had, in
fact, been totallyignoredin the journal in which it had ap-
set writes,"I could probablyget theworkof tabulating peared), and fromthe firsteditionit has been one of the chief
your integraldone easily enough, though whetherI purposes of thisbook to make betterknownthe effectof his
should be allowed to publishthe work of the man I'd researches,and ofmathematicalworkconsequentuponthem.
get to do it is anothermatter" (Gosset 1970, no. 7).
Duringthe visitwithFisher,Gosset suggestedsend-
That year, for the firsttime, Gosset had a statistical
ing his assistant, E. Somerfield,to Rothamstedfor
assistant; but he knew the brewery'srestrictivepolicy a fewmonthsto studywithFisher. The visitwas even-
on publicationsmightmake trouble.
tuallyarrangedto beginin December 1922.The method
of instruction,workingthroughfielddata, was neverin
doubt.Gosset approved: "What you suggestabout run-
3. THE TABULATION OF t ningSomerfieldover any figurescomingin fromyour
various departmentsis exactly what I want." But he
Two changes were made fromStudent's former7 cautioned Fisher: "You will, I fear,findthatyou will
table to make the new tabulationssuitable for all the have to be even moreelementarywithSomerfieldthan
new applications.First,thenumberofobservationsfor withme, but thatwill be rathergood for you as your
which the table was enteredwas changed to the num- faultis thatyou considerus all to be mathematicians'
ber of degrees of freedom.Second, the now familiar (Gosset 1970,no. 13). The visitproved highlysatisfac-
values of t = z(n - 1)1/2 were tabulated instead of:. tory to all parties. Soon after Somerfield's return,
This correctionstandardizedthe gradingof the scale Gosset confessed that his assistant now understood
that had previouslybecome progressivelycoarser for Fisher's writingsbetterthanhe did; moreover,Somer-
largervalues ofn, and it incidentallymade thegrading fieldwas permitted to publishtheworkdone at Rotham-
finerthroughoutthe table. sted, although, Fisher's annoyance, it had to be
to
In calculatingthe tabulated values in 1907, Gosset undera pseudonym,"Mathetes."
had used the device of integratingby partsin termsof Meanwhile,Gosset began tabulations.In October he
cos 6. The original tables did not extend beyond wrote, "'I haven't had time to do anythingwith the
n = 10; since the formulaincludes an extratermeach Type VII [Student's curve], apples at home and busi-
time n is increased by two, the method becomes ness at the Brewery,but I hope to get on to it soon''
cumbersome for largern. The tables had later been (Gosset 1970,no. 12). A monthlater,however(to quote
extended and, in view of the difficulties,
it is not sur- a fragmentonly), he wrote:

(? The AmericanStatistician,Ma' 1981, Vol. 35, No. 2 63


This content downloaded from 66.195.118.21 on Mon, 23 Mar 2015 13:23:02 UTC
All use subject to JSTOR Terms and Conditions
Last nightI checked your values forx = 1 (discoveringa more accurate than a randomized one. In his reply
slight slip) from your correction formulae and calculated Gosset invoked practicalexperience,butgraduallyhis
the same values to seven places. As I used the sum of at least
reasoningbecame moreexplicit.On readingthe proofs
fournumbersof 7 places they also have an errorin the 7th
place due to approximations,but the correspondenceis quite ofStatisticalMethodsforResearch Wor-ker-s in October
wonderfullyclose. (Gosset 1970, no. 13) 1924, Gosset protested: "I don't expect to convince
you butI don't agree withyourcontrolledrandomness.
ImagineFisher at Rothamstedworkinghis motorMil- You would want a large lunatic asylum foroperators
lionaire,a large electricmachineon whichone turned who are apt to make mistakesenougheven now" (Gos-
a crankto set thenumberand inserteda plungerto start set 1970, no. 50). In April 1928. criticizinga block
operation. He providedthe correctionformulasand a design because the errorwas largerthan if the same
table of values calculated fromhis expansion. Imagine treatmentshad been laid out in a Latin square, Gosset
Gosset puttinghis hand-operatedBaby Triumphator argued:
intohis rucksackand carryingit home to workon the
The fact is that thereale two principlesinvolved in the
tables, calculating, checking, recalculating. Imagine Latin Square, of whichI attachthe greaterimportanceto the
Somerfield, left without the Triumphator at the balancing of the errorsand you to the randomisation.It is
brewery,and, as Gosset reported, myopinionthatin thegreatmajorityof cases, the randomisa-
tion is supplied to any properlybalanced experimentby the
...borrowing fromall and sundry.Yesterday I foundhim
soil itself.(Gosset 1970,no. 92)
withthe machineNoah used when quantitysurveyingbefore
his voyage. The storygoes that he subsequentlybarteredit Gosset understoodperfectlythe essential role of ran-
for a barrel of porter with the originalGuinness. Anyhow domizationand himselfpointedit out in criticizingthe
he doesn't seem to have been able to keep it dryand Somer-
fieldwasn't strongenough to turnthe handle. (Gosset 1970,
LanarkshireMilk Experiment(Student 1931), but this
no. 13) did not alter his conviction. In 1936, in advocating,
beforethe Royal StatisticalSociety, Beaven's half-drill
Apart from Fisher's calculations and assistance in stripmethodfor large-scale cooperative experiments
checking, Gosset was to recalculate the whole table (Student 1936),he claimed that"since the tendencyof
himself,under the restrictionthat the calculator was deliberate randomisingis to increase the error, a
in demand at the breweryforhalfthe year. At the end balanced arrangementlike the half-drillis best."
of February1923he warnedFisher,"People are getting Fisher was angered by this, and a heated debate fol-
querulousabout the machine,and I reallycannotspare lowed. Fisherthenproduced "A Test of the Supposed
day-lightto workon [it] at the Brewery,so I fearthat Precisionof SystematicArrangements"(Barbacki and
I shan't do much more tillnext winter"(Gosset 1970, Fisher 1936),whichGosset answered in an articlepub-
no. 19). It was mid-Octoberwhen Gosset announced, lishedposthumously(Student 1938). Afterconsidering
"The calculating season having now commenced, I the usually slighteffectsof departuresfromassump-
took a calculatingmachinehome on Saturdayand be- tionsof normalvariationand of equal variance,Gosset
gan work last night" (Gosset 1970, no. 34). continued:
If, however the samplingbe not random,thereare such
possibilitiesof drawingfalse conclusions that Prof. Fisher
4. ARGUING THINGS THROUGH has introduceda systemof artificialrandomizingto ensure
thatthe thirdconditionis satisfiedand brands all other sys-
tems invalid.
Even while he was makingthe calculations for his
Nevertheless,it is possible, by balancingsources of error
new tables, Gosset was reconsideringFisher's paper which would otherwiselead to bias, to obtain arrangements
(1922) on the goodness of fitof regressionformulas, of greaterprecision which are neverthelesseffectivelyran-
in which it was shown thatthe significanceof regres- dom, by which I mean thatthe departurefromrandomness
sion formulas,linearor nonlinear,simpleor multiple, is only liable to affectour conclusions to the same sort of
extent as do departures from normalityor inequality of
may be treatedexactly by Student's test. Althoughin
variances. (p. 200)
May he had accepted Fisher's results, on rereading
the paper aftervisitingFisher, Gosset became uneasy The argumentof the rest of the article supported
and wroteFishermorethanonce beforehe was satisfied this claim.
thatthe test was legitimate.He chose to workthrough Gosset's obstinacyin debate was not withoutvalue
Fisher's problematically"obvious" argumentsand, if to science, forhe raised pointsthatneeded considera-
necessary,to raise pointsand argue themthroughwith tion or elucidationand persevereduntilhe was satis-
Fisheruntilhe was sure he understoodand agreed with fied. His difficultyin 1922 was one that has troubled
the reasoning. In 1923, even before the principleof later statisticians,it is perhaps the firstinstance in
randomizationwas in print,Gosset expressed skepti- which Fisher's concept of the "relevant subset" be-
cism about therandomizationoffieldexperiments,and comes critical.In makingthe regressionof X'on x,-and
on this subject Fisher never whollyconvinced him. testingthe significanceof deviationsabout the regres-
The subject arose during lengthycorrespondence sion line b, Fisher had proved thatthe distributionof
about the analysisof experimentsusingBeaven's half- the ratio of a regressioncoefficienth to its estimated
drillstriparrangement.Fisherhad evidentlyasked why standard error follows the tdistribution. It was not
Gosset believed a systematicdesignto be intrinsically obvious to Gosset thatthe resultingtestwas legitimate

64 ? The American Statistician,Max 1981, Vol. 35, No. 2


This content downloaded from 66.195.118.21 on Mon, 23 Mar 2015 13:23:02 UTC
All use subject to JSTOR Terms and Conditions
because the samplingdistribution ofthex's themselves place, mostlyit is trueby .0001 only, and quite a numberof
was not taken into account. Fisher, however, argued otherones. The factis thatI was even more ignorantwhen I
made the firsttable than I am now and thoughtI was going
thatthe regressionwas not made on the whole popula-
to be accurate to 4 places by taking5 in the working!and the
tionof possiblex's, butonlyon the sample; forthetest second was of course constructedon the same lines though
in question,the sample of values ofx actuallyobtained not by me. I oughtto have checked it myself,but musthave
formedwhat he was later to call the relevantsubset, been prettycasual about it. Anyhowthe old man is just about
and forthemthe mean and the variance were fixedby fed up withme as a computerand wouldn'teven let me cor-
rect my own table. I don't blame himeither.
the data themselves.
Whetherhe will have anythingto do withour table I don't
know,I ratherdoubtit,butpersonallyI feelI could hardlyput
itbeforehimunless you are preparedto do quitea lotofcheck-
5. PUBLICATION OF THE t TABLES AND ing eitheryourselfor per Miss Mackenzie. Justas well you
THEIR WIDER APPLICATIONS didn't take that table fromBiometrika!It has been rathera
miserablefortnight findingout what an ass I made of myself
FromthetimeFishergraspedtheunifieddistribution and fromthe point of view of the new table whollywasted.
However, I beginworkagain tomorrow.(Gosset 1970,no. 36)
theoryforthe generallinearmodel, events moved fast
towards new applications. His investigation,"The Fisher must have jibbed at more checking,for two
Goodness of Fit ofRegressionFormulaeand theDistri- weeks later Gosset wrote, "It seems rathera shame
bution of Regression Coefficients,"was published in to burdenyou withcheckingafterwhat you say, but I
July 1922. In it he identifiedas the test criterionthe thinkI may fairlyput your own tables up to you"
functionwe know as F, and gave the correctformula (Gosset 1970, no. 37).
for its distribution.Thus the theoreticaldistribution Fisher was busy writingStatistical Methods, and
appropriatefortestsof significanceforthe analysis of Mackenzie (his statistical assistant) her Master's
variance was known. The analysis of variance itself thesis. When Gosset was invitedto act as refereefor
appeared in 1923and FisherreintroducedtheF test in this thesis in February 1924, he responded charac-
termsofz = ?/21og F. In May 1923Fisherwas evidently teristically,
tryingto get out a table of F; but tabulationin terms I suppose theyappointedme because theThesis was about
of z had obvious advantages, and it may be this idea barley, so of course a brewer was required, otherwise it
Gosset referredto in October when he wrote, "I like seems to me ratherirregular.I fearthat some of Miss Mac-
the resultforz in the case of that horriblecurve you kenzie's mathematicsmay be too 'obvious' forme. (Gosset
are so fondof'' (Gosset 1970, no. 33). 1970, no. 40)
By the summerof 1924, Fisher (1924) was ready to But he accepted the appointment,and this time he
summarize the relationshipsof the distributionsin- stayedtheweekendat Fisher's home: "I travellightso
volved, the normal,x2, t, and z. He had completed ifyou willlet me knowyouraddress I willcome straight
the typescriptof Statistical Methods for Research up withmy pack on my back" (Gosset 1970, no. 43).
Workers,in which tables of these quantitieswere to Possibly he was a littleapprehensiveabout Pearson's
appear, but he had no tables. Gosset had felt they reaction to theirreferees' report;his excuse in May
shouldofferthet tables forpublicationfirstto Pearson, for"rousing myselffrommylethargyto writeto you"
but therewere doubts about the copyrightof this and was that he was "consumed with curiosityto know
Gosset's previoustables. The outcomewas thatFisher whatthe University[of London] thoughtof our report
calculated all the tables forthe book himself. on Miss Mackenzie" (Gosset 1970, no. 45).
What happened when Gosset approached Pearson Withthisletterof 20 May 1924Gosset sentthe com-
mustbe told in his own words,in a letterof 23 Novem- pleted tables, saying,
ber 1923.
It is now up to you to writean account of the tables, and
Dear Fisher, please don't let too much be clear or obvious, I'd like to
Your interpolationformulaworks like a charm though understandas much as possible what I have been doing for
why the Dickens you chose those particularvalues of n I the last two years....
can't think.I can only go on "Watsonin'!" P.S. Ifyou could let me have youraccount quiteearlynext
I was over in London on my way to and fromhome the monthI can take it to K.P. when I next get over.
weekend beforelast and dropped in on K.P. I broached the
subject of a joint table withan introductorynote by you and Fisher hastened to send his contribution,and Gosset
he was preparedto consider it: he would I thinkhave taken took it with him to England. "But," as he explained
it butfora mostunfortunate occurrence. Whilewe were talk-
later,"unfortunately I leftitat homewhenI wentto the
ing he mentionedthat he was bringingout a new editionof
Tables for Biometriciansand I said 'Oh thereare one or two Laboratory" (Gosset 1970,no. 60). It was May of the
mistakesin my small table in it,' referringto a discovery of followingyear beforehe deliveredtheirworkto Pear-
some in the odd numbersof .2 mentionedin the introduction son. Two weeks later, on 12 June 1925, Gosset re-
to my second table in Biometrika.'Well we'll put thatright.' ported:
So I wentaway to get the correctionsforhim. That involved
gettinga 5thplace forn = 9 and I couldn'tcheck the4thplace. K.P. is very anxious to publish your note about the use
I leftmy attemptand went home and when I came back on of the table, but doesn't like the binomial approximation
myway to Dublin I foundthathe agreed withme and thatthe whichhe consideredrequiresproofof convergence.It was in
new table was wrong. On furtherinvestigation,both tables vain thatI pointedout that,convergingor diverging,theproof
were foundto be perfectlyrotten.All .1 and .2 wrongin 4th of the puddinglies (to me doubtlessnot to you) in thefactthat.

(? The AmericanStatistician,May 1981, Vol. 35, No. 2 65


This content downloaded from 66.195.118.21 on Mon, 23 Mar 2015 13:23:02 UTC
All use subject to JSTOR Terms and Conditions
yougetaboutsevenplacesthesamewithn = 21 to t = 6. and jokinglywould remark,"I take it thatwhateverit
Anyhow,he retumsbothand I send themherewith.... is followsat once 'obviously' froma considerationof
oughtto go into
In any case [thepaperon applications]
Biometrika.(Gosset1970,no. 61) n-dimensionalgeometry" (Gosset 1970, no. 12), or
demandof Fisherwhethera certainequation came out
Soon after,itwas decidednotto publishtheworkas a of n-dimensionalspace "or what is much the same
whole,underjointauthorship, butin partsseparately, thing,your head" (Gosset 1970, no. 61).
andGosset continued tohope thepaperonapplications
withthetablesthatwereacceptableto Pear- [Received August 1980. Revised October 1980.]
together
son wouldappearinBiometrika. Butitwas notto be.
In September all thetablesand paperswereaccepted REFERENCES
forpublicationin Metron,wheretheyappearedto-
gether,just fouryearsaftertheyhad firstbeen con- BARBACKI, S., and FISHER, R.A. (1936),"A Test of theSup-
templated,and 14yearsafterthatfirst correspondence posed Precisionof SystematicArrangements," Annals of
Eugenics, 7, 189- 193,CollectedPapers ofR.A. Fisher,3, no. 139.
thathad led Fisher to thegeometric representationof BOX, JOAN FISHER (1978), R.A. Fisher, The Life of a Scientist,
the sample, whichcame to its full in
fruition his New York:JohnWiley.
remarkablepaper "Applicationsof Student'sDis- EISENHART, CHURCHILL (1979), "On the TransitionFrom
tribution." 'Student's' z to 'Student's' t," The American Statistician,
Fisherhad written to Gosseton 17July1924: 33,6-10.
FISHER, RONALD AYLMER (1971-1974),CollectedPapers of
I enclosethetwonotesI mentioned,thefirstofwhichis an of
R.A. Fisher(5 vols.),ed. J.H. Bennett,Adelaide:University
attemptto givesomeidea ofthemultitude ofuses to which Adelaide.(See references Papers.)
to individual
yourtablemaybe put,andthesecondis a formal statement (1912), "On an AbsoluteCriterion forFittingFrequency
oftheapproximationformula. is largerthanI hadin-
Thefirst Curves," Messenger of Mathematics, 41, 155-160, Collected
tended,and to makeitat all completeshouldbe largerstill, Papers, 1, no. 1.
butI shallnothave timeto makeit so, as I am sailingfor (1916), "Biometrika," Eugenics Review, 8, 62-64, Collected
Canada on the 25th,and willnot be back tillSeptember. Papers, 1, no. 7.
(Gosset1970,no. 48) ofCorrela-
(1921),"On the'ProbableError'ofa Coefficient
tionDeducedFroma SmallSample,"Metron,1,3-32, Collected
In thispaperFisher(1925)gavea proofofStudent's Papers, 1, no. 14.
results,whichin1908hadbeenpartially Then
intuitive. Formulae,and
(1922),"The Goodnessof Fitof Regression
ofRegression
theDistribution Journal
Coefficients," oftheRoyal
he presentedapplicationsofStudent'stablesappropri-
Statistical Society, 85, 597-612, Collected Papers, 1, no. 20.
of two means,and of
ate to testingthe significance YieldingtheErrorFunctionsof
(1924),"On a Distribution
linear,curvilinear,and multipleregressions.Finally, Several Well Known Statistics,"Proceedings of theInternational
he introduced thewiderclass ofF (or equivalentlyz) Congressof Mathematics,Toronto,2, 805-813, CollectedPapers,
whichhe identified
distributions, as being 1,no. 26.
(1925), Statistical Methods for Research Workers,Edin-
...relatedto Student'sdistributioninthesamemanner as burgh:Oliverand Boyd.
thatof x2 is relatedto thenormaldistribution. This wider Metron,
of 'Student's'Distribution,"
(1925),"Applications
class of distributionsappears(i) in the studyof intraclass 5, 90-104, Collected Papers, 2, no. 43.
correlations(ii) in thecomparison of estimatesof thevari- GOSSET, W.S. (1970),LettersFrom W.S. Gossetto R.A. Fisher1915-
ance(iii)intesting thegoodnessoffitofregressionlines(iv)of 1936,withsummaries byR.A. Fisheranda forewordbyL. McMul-
testingthesignificance ofa multiple or(v) ofa cor-
correlation len,Dublin:Arthur GuinnessSon and Co. (Dublin)Ltd. Issued
relationratio.(Gosset1970,no. 102) (Lettersin thisvolumeare numbered.)
forprivatecirculation.
PEARSON,EGON S. (1968),"StudiesintheHistory ofProbability
Althoughthearticlewas not,in fact,publishedfor18 and Statistics.XX. Some EarlyCorrespondence BetweenW.S.
monthsafterbeingwritten, Fisherdid notenlargeit. Gosset,R.A. Fisherand K. PearsonWithNotesand Commen-
(Apartfromanything else, forsix monthsafterhe re- tary," Biometrika,55, 445-457.
SOPER, H.E., YOUNG, A.W.,CAVE, B.Y., LEE, A., andPEAR-
turnedfromCanada Fisher was calculatingthetables SON, KARL (1917),"On theDistribution Coef-
oftheCorrelation
and,withGossetand Somerfield, working onthe proofs ficient 11,328-413.
in SmallSamples,"Biometrika,
ofStatisticalMethodsforResearch Workers.)Perhaps "STUDENT" (1908a),TheProbableErrorofa Mean,"Biometrika,
if he had extendedthe last section,he could have 6, 1.
avoidedthe repeatedmisunderstandings thatarose a Coefficient,"
(1908b),"Probable Errorof a Correlation
Biometrika,7, 302.
decade later about of
justification the test of signifi- Due to
of SpuriousCorrelation
(1914),"The Elimination
cance fortheanalysis of variance.But the argument, PositioninTimeor Space," Biometrika,10, 179.
thoughconcise,was complete.It even containedan ofObserva-
(1925),"New TablesforTestingtheSignificance
algebraicproofanalogousto thegeometric methodof tions," Metron,5, 25-32.
proof, whose as
validity, Fisherput it (p. 97), "may (1931),"The LanarkshireMilkExperiment," Biometrika,
23, 398.
notbe universally admitted." Jour-
(1936),"Cooperationin Large-ScaleExperiments,"
It is rathersad thatevenwiththisaidto understand- nal of the Royal StatisticalSociety (Supplement),3, 115.
ing,Gossetfoundthegeometric representation always (1938),"ComparisonBetweenBalancedand RandomAr-
highlymysterious; buthe acceptedthemystification, rangements 29, 363.
of FieldPlots,"Biometrika,

66 ? The AmericanStatistician,May 1981, Vol. 35, No. 2


This content downloaded from 66.195.118.21 on Mon, 23 Mar 2015 13:23:02 UTC
All use subject to JSTOR Terms and Conditions

Вам также может понравиться