Biometrika Trust

The Probable Error of a Mean

Author(s): Student
Reviewed work(s):
Source: Biometrika, Vol. 6, No. 1 (Mar., 1908), pp. 1-25
Published by: Biometrika Trust
Stable URL: http://www.jstor.org/stable/2331554 .
Accessed: 23/10/2012 11:13

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 support@jstor.org.

Biometrika Trust is collaborating with JSTOR to digitize, preserve and extend access to Biometrika.

http://www.jstor.org
VOLUME VI MARCH, 1908 No. 1

BIOMETRIKA.

THE PROBABLEERROR OF A MEAN.

By STUDENT.

Inttroduction.
ANY experimentmaybe regardedas formingan individualof a " population"
of experimentswhich might be performedunder the same conditions. A series
of experimentsis a sample drawnfromthis population.
Now any seriesofexperimentsis onlyof value in so faras it enables us to form
a judgment as to the statistical constantsof the populationto whichthe experi-
mentsbelong. In a great ntimberof cases the question finallyturnson the value
ofa mean,eitherdirectly,or as the mean difference betweenthe two quantities.
If the numberof experimentsbe verylarge,we may have precise information
as to the value of the mean,but if our sample be small,we have two sourcesof
uncertainty:-(1) owingto the "errorof randomsampling" the mean of our series
of experimentsdeviates moreor less widelyfromthe mean of the population,and
(2) the sanmple large to determinewhat is the law of distribution
is not sufficiently
of individuals. It is usual, however,to assume a normal distribution, because, in
a verylarge numberof cases, this gives an approximationso close that a small
sample will give no real information as to the mannerin which the population
deviates fromnormality:since some law of distributionmliustbe assumed it is
better to work with a curve whose area and ordinates are tabled, and whose
propertiesare well known. This assumptionis accordinglymade in the present
paper,so that its conclusionisare not strictlyapplicable to populationsknown not
to be normallydistributed; yet it appears probable that the deviation from
normalitymustbe veryextremeto lead to serious error. We are concernedhere
solelywith the firstof these two sourcesof uncertainty.
The usual methodof determiningthe probabilitythat the mean of the popula-
tion lies withina given distanceof the mean of the sample,is to assume a normal
distributionabout the mean of the sample with a standard deviation equal to
s/a/n, where s is the standarddeviation of the sample, and to use the tables of
the probabilityintegral.
Biometrikavi 1
2 The ProbableError of a Mean
But, as we decreasethe numberof experiments, the value of the standard
deviation foundfrom thesampleofexperiments becomesitselfsubjecttoan increas-
ingerror, untiljudgments reachedin thiswaymaybecomealtogether misleading.
In routineworkthereare two waysof dealingwith this difficulty: (1) an
experiment maybe repeatedmanytimes,untilsuicha longseriesis obtainedthat
the standarddeviationis determined once and forall withsufficient accuracy.
This valuecan thenbe usedforsubsequentshorterseriesof similarexperiments.
(2) Whereexperiments are donein duplicatein the naturalcourseof the work,
the meansquareof the difference betweencorresponding pairsis equal to the
standarddeviationof the populationmultipliedby v/2. We can thuscombine
together severalseriesof experiments forthepurposeof determining thestandard
deviation. Owinghowever to secularchange,thevalueobtainedis nearlyalways
toolow,successive experitnentsbeingpositively correlated.
Thereare otherexperiments, however, whichcannoteasilybe repeatedvery
often;in such cases it is sometimes necessaryto judge of the certainty of the
resultsfroma verysmall sample,whichitselfaffords the onlyindicationof the
variability.Some chemical,manybiological,and most agricultural and large
scaleexperiments belongto thisclass,whichhas hitherto beenalmostoutsidethe
rangeofstatistical enquiry.
Again,althoughit is wellknownthatthe methodof usingthe normalcurve
is onlytrustworthy whenthe sampleis "large,"no one has yet told us very
clearlywherethelimitbetween"llarge" and " small" samplesis to b-edrawn.
The aimof the presentpaperis to determine the pointat whichwe mayuse
thetablesoftheprobability integralinjudgingofthesig,nificance of the meanof
a seriesof experiments, and to furnishalternativetablesforutsewhenthenumber
of experiments is toofew.
The paperis dividedintothefollowing ninesections:
I. The equationis determined of the curvewhichrepresents the frequency
distributionofstandard deviations ofsamplesdrawnfroma normalpopulation.
IJ. Thereis shownto be no kindof correlation betweenthe meanand the
standarddeviationof such a sample.
III. The equationis determined of the curve representing the frequency
distributionof a quantityz, whichis obtainedbydividingthe distancebetween
themeanofa sampleand the meanof the population by the standarddeviation
of the sample.
IV. The curvefoundin I. is discussed.
V. The curvefoundin III. is discussed.
VI. The two curvesare comparedwithsomeactual distributions.
VII. Tablesofthecurvesfoundin III. are givenforsamplesofdifferent size.
VIII and IX. The tablesare explainedand someinstances are givenof their
use.
X. Conclusions.
 BY STUDENT 3

SECTIONI.

Samples of n individualsare drawhout of a populationdistributednormally,

the frequencyofthe standarddeviations
to findan equation whichshall represenit
ofthesesamiiples.
If s be the standarddeviation foundfroma sample x1x2....x, (all these being
measuredfromthe mean of the population),then
2 S(X12) S (1I) = S (X12) _ (X12) 2i (x1x2)

Summing forall samples and dividing by the numberof samples we get the
mean value of S2 whichwe will writes2,
n2 n_ /X2 (n-1)
82-= n
U2
n
where , is the secondmomentcoefficient in the originalnormaldistributionofx:
since x1, x'2, etc., are not correlatedand the distributionis normal,productsin-
volvingodd powersof x, vanishon summing,so that 2S(xx1a'2)
is equal to 0.

If MI; representtheRth momentcoefficient

of the distributionof S2 about the
end of the range wherese= 0,
M =2 (n-1)

Again 84= ($1) (S ($1))2}

=(iS (a2)j2 - 2S (X,2) (S (XI)y2 + (S (aXI))4

S (X'4) + 2S (x,2x22) 2S (X14) 4S (x,2x22) +S (14)

?z2 n2 'It3 n3 T4

68_(X,2a'2
2)
+ .- ?+other terms involving odd powers of x, etc.,
14
whioh will vanish on summation.
Now S(X14) has n terms but S(x,2x22) has jn(n-1), hence summing for all
samples and dividing by the number of samiipleswe get
- 1) - + 2 (-1
2. (n 2A_ 1) +3?
*4
M2'=M2 = + 2 (?1
2 ~~~~~2
-
2

{2n
fig
- 2 - 1} (i -
1) {ni -2n + 3}.

Now since the distributionofa is normal,,. hence

= 3,u22,

. (n
M2.' =P -
W3n {388 + 2 - 2n + 31.=A?2 (n - 1) (n +41-
1-2
4 The Probable Error of a Mean
In a similar tedious way I find:

M3,=3 ,S (n~~n
-1)(n+1)(n+3)

and4 M4'= A4 ( 1) (n+ 1) (n +3) (n +5)

2 4

The law of formnation appears to be a simple one,

of these momentcoefficients
but I have not seen my way to a generalproof.
ofe2 about its mean,we have
If now MR be the Rth momentcoefficient

(n-1) n
M2= n +1)-(n-1)}= 22(n1)
~~~~~~n

M3 (l(n+l)(n+ ) _ (n-i)

3
2
(n {n2+ 4n+3-6t+ 6-n2+2n-1}-83(- --I)

IA4 n
- 2p2'(n -1)(n+3)
4

Hence M =2 8 A A4 = 3(2i +3)

18,M2'it-I 92M2 it-i

2/32-383,-66 -16
(n-1)n (n + 3)-24-6 = O.

Consequentlya curve of ProfessorPearson's type III. may be expected to fit

the distributionof s2.
The equation referredto an originat the zero end of the curvewill be
y= GxPe-X,

whiere ry= 2 M2 4bL22(n- )n3 n

M3 8n2,u23(n- 1) 2p'
4 n-i n-3
and p=~-l 2
4 n n
and
= - 1=-2
1 = I- - 1 -- 3_

Consequently the equation becomes

*n-3 nx
y= CX 2 e 2t,,

which will give the distributionof 82.

Ti end-(3 n)
The area ofthiscurveis C| x e--,
eydx I (say).
 BY STUDENT 5

about the end of the range will therefore

The firstmomentcoefficient be
Xn-l flX -2
Cf x2e2b2dx C -
n-i
f0--1flX
2 e, 252
=
n n-8 no
2 e_A2dx
I = I $ ? + J I
The firstpart vanishesat each limitand the second is equal to

n-1
I = n

and we see that the higher momentcoefficients

will be formedby multiplying

successivelyby n +n 1 /A22 +
n
3 J etc.,just as appeared to be the law of formation
of M2', Ms', M4', etc.
Hence it is probable that the curve foundrepresentsthe theoreticaldistribu-
tion of 82; so that althoughwe have no actual proofwe shall assume it to do so in
what follows.
The distributionof s may be found fromthis,since the frequencyof s is equal
to thiatof 82 and all that we must do is to compressthe base line suitably.
Now if y - .q (82) be the frequencycurveof S2
and Y2 (s) ,, ,, ,, ,, ,,s

then y1d(82) = y2ds,

or y2ds = 2y,sds,
Y2 =2syi.
n-3 *82

Hence y =2Cs(S2)2 e '2

is the distributionof s.
ns2
This reducesto y 2Csn-2e 2A2 .

nZ2

Hence y= AXn-2e will give the frequiency

20 distributionof standarddevia-
tionsofsamplesofn, taken out of a populationdistributednormallywithstandard
deviationa. The constantA nmay be found by equating the area of the curve as
follows:-
Area = A fx2e 2X2d. (Let 4prepresent xPe

Then IP = I X P- d (-e To) dox

Onx1nx
e2a2_ n,nX2. =00 2 go _
= - - Xpe 2a2 + (p -1) XP-2e 2oT2dX
n L x=O i

=- 1) 1p2,
21L(p-
since the firstpart vanishesat bothlimits.
6 The ProbableError of a Mean
thisprocesswe find
By continuing

In-2 = 2 (n-3) (n -5) ***3.1 Io

or = 2 (n-3)('n-5). 4. 2 I

as n is evenor odd.
according

But l0 is e
je 4
=x2V
J0 n2 [ 2 #X21-o

anid is xe -e--dx
dx= [ n
e
x=O
Henceifn be even,
Area
A=
(n- 3)(n-5)... 3. 1I 2

and ifitbe odd

Area
(n -3) (n-) ...4 2(_ 2

Hencetheequationmaybe written

Y X-)n5~.31Tf2 Z"2-?(even)

or y _ -2eN
N 2 (f~l)f2
2 ln 2(n odd)
whereN as usualrepresents
thetotalfrequency.

SECTION II.
between(a) the distanceof the meanof
To showthatthereis no correlation
a samplefromthe meanof the populationand (b) the standarddeviation of a
samplewithnormaldistribution.
of themeanof the sampleare
(1) Clearlypositiveand negativepositions
equallylikely,and hencetherecannotbe correlationbetweenthe absolutevalue
of the distanceof the meanfromthe meanof the populationand the standard
deviation,but (2) theremightbe correlation
betweenthe squareof the distance
and thesquare ofthe standard
deviation.
Let U2 = (S()) and 32= S__'32)
_ ( (X))2
n n n1
Then if mnl',
M1'be the mean valuiesof u2 and 82, we have bythe preceding
partM1'=p
#(n-i) and n1'= 2
 BY STUDENT 7

Now 282s S (X12)(S

-
(XI))2
(S (Xi))4

(S (X12))2 + 2S (X1IX2).S (X12) S(X14)

n n3~~~l n4

12 _ -other termsofoddorderwhichwillvanishon summation.

n4

Summingforall valuesand dividingby thenumberofcaseswe get

BRu82satoagA + nzM, = 24 + 2(n -1) _s4

y-l22 (n3)

betweenu2 and e2.

whereR,,2,is thecorrelation

+ 81222(n, 1) /= 2 (n-i) 3 1 -1
Bto2 at62 qp (______
n2 2~ {3'-}~ 2 n

Hence Rp8aouq,82= 0 or thereis no correlationbetweenu2 and s8.

SECTION III.

To findthe equationrepresenting
the frequency of the mneans
distribution
of samplesof n drawnfroma normalpopulation,the Ineanbeingexpressedin
termsofthestandarddeviationofthesample.

We have y= 0-22422e
e as the equationrepresenting of 8,
the distribution
the standarddeviationof a sampleof n, whenthe samplesare drawnfroma
withstandarddeviationa.
normalpopulation
Nowthemeansofthesesamplesofn are distributed to theequation
according
VnGN -n?2

=/XNe~ 2as

and we have shownthat thereis no correlation

betweenx, the distanceof the
meanofthesample,and s, thestandarddeviationofthesample.
Now let us supposex measuredin termsof 8, i.e. let us findthedistribution
ofz=-.

If wehavey,= + (x) and y2 = f(z) as theequationsrepresenting

thefrequency
ofx and ofz respectively,
then
d$
yldX=y2dz=Y2 ,a

.a. y2 = sY1.

* Airy,Theoryof Errorsof Observations,

Part II. ? 6.
8 The ProbableError of a Meant

Hence NVnse 12
V/27ro
is the equiationrepresentingthe distributionof z forsamples of ' with standard
deviations.
Now the chance that s lies betweens and s + ds is:
rs+ds 1182

JS+dC
a
-Sn-2 e .92 d
00 ns$3

sn-2 e 2a2ds

whichrepresentsthe N in the above equation.

Hence the distributionofz (lue to values of s whichlie betweens and s + ds is
n s2(+z2) s2(1+z2)
+sds n s+ds

Ysr e
/

srx
tv28n-i 2o2 ds \2J
.- e ds
YXa
31= fl0082 I' 1l82
2o7
| -
0_sX-2 e 2.72dS 41 SUl-2e *2o'ds

and summingforall values of s we have as an equation givingthe distribution

of z
/- 00 ns2(1+z2)

2 8(sn-1e ds
ff
f00e
2(r2

Ja-
=~ ~ - - 22
ci
y

By what we have alreadyprovedthis reducesto

n-2 n-4
1 772n4 5 3
3(1 + ifn be odd,
-
y Z2)2

andto l~1n- 2 n- 4 4 2
andto Y= n1 ( if 2 if be even.
-
7n-3 z2)

of the
Since this equation is independentof a- it will give thjedistribuition
distance of the mean of a sample fromthe mean of the populationexpressedin
termsof the standarddeviationof the sample forany normalpopulation.

SECTION IV.

SomePropertiesof theStandardDeviationFrequencyCurve.
By a similar methodto that adopted forfindingthe constantwe may findthe
mean and moments: thus the mean is at 119-1,

whichis equal to (1% -2) (n-4) 7 (if

nn),
(n - 3) (n - 5) v2
1 ;'n

or, -
(ifunbe odd).
 BY STUDENT 9

The secondmomentabout the end-of the range is

l2 (n-1) o2
'fl-2n

about the end of the range is equal to

The thirdmomnent
In+_ In+1 In-i
'fl-2 In-I 'Il-2

= a2 x the mean.
about the end of the range is equal to
The fourthmomeint
'fl+2 (n - 1)(n1 + 1) 4
q?2
Do
If we write the distaniceof the mean fromthe end of the range Do- and the
'lt-2

nomentsabout the end of the rangeVI, v2,etc.

- 1 2
te
then I = Do2
n) n-n 3=Do3 4
it Vn v
n2
aboutthemean
Fromthiswegetthemoments

2 = n(n -1-D2),
n

/33 {nD-3 (z - 1) D + 2D3}= {2D2 - 2n+ 3},

{4n2
= 1- 4D2n+ 6 (n-1) D2- 3D4}= {n
2 1- D2(3D2 - 2n+ 6)}.

out what these become whenit is large.

It is of interestto firnd
findout what is the value ofD.
In orderto do this we nmust
Now Wallis's expressionfor7rderivedfromthe infiniteproductvalue of sin x is

2 (2+ 1)=i- 32 as52.. (2n_-1)2

If we assume a quantity 0 = ao + a' + etc.) whichwe may add to the 2n + 1

in orderto make the expressionapproximatemorerapidlyto the truth,it is easy
to show that 6--+ -I--etc. atid we get
2 1-6n
w+ 1 1 _ 2'.*42*6 ... (2)2 *
2V 2 16n. =12 -.aa ... (2n -1)
i. 33%2..( 1)2

From thiswe findthatwhethern be even or odd D2 approximatesto n 2+

whenn is large.
* towrthanWallis's.
willbe foundto givea muchcloserapproximation
Thisexpression
Biometrikavi 2
10 ThtePrcobableError of a Mean
Substitutingthis value of D we get

Tit 16n12
3a4 1 1
11-2
2ii 4)1 ' X =4as(1+2
IA4~
~~~~~~ 162)'
1

( )
1X 4n2
2m1+
Consequentlythe value of the standarddeviationof a standarddeviationwhich
'
wehavefound - thesameas thatfoundforthenormal
becomes
t^/n/ -4n)
whenn is large enough to neglectthe 1/4nin
curveby ProfessorPearson (o-/V/2n)
comparisonwith 1.
Neglectingtermsof lowerorderthan 1 we find

2n-3 32=3I- +
-2n (1+*2n)
n (471- )
=
In3'=
Consequentlyas n increases 2 verysoon approachesthe value 3 of the normal
curve,but ,8 vanishesmoreslowly,so that the curve remainsslightlyskew.
DIGRAM I. Frequencycurvegivingthe distributionof Standard Deviationsof samples of 10 taken
from.a normal population.
I _ 1002

Equation y N 10 ,2 8e 22

1-62/

fT [ ,o t\ <
5 _ _ __ _ I,i* 245 a1 515

Diagram I shows the theoreticaldistributionof the S.D. foundfromsamples

of 10.
N 102 /2x -z2

3/=7.5.3Vtq<9e
SECTION V.

Someproperties
of thecu.rvey=n-2 n (142( rifnb ee) J _

Writingz= tan 9 the equation becomes y - ... etc. x cosiB9, which

an easy way ofdrawinlgthe curve. Also dz = dO/cos2

affiords 9.
 BY STUDENT

He-nceto findthe area of the curvebetweenany limitswe must find

-2 n-4 etc x fcos'-29d9 d
n -3 n -sa et.. xCO
3-

n-2 n-4 1.
fet, 'cosn-4 OdO+[cs'si

n-4 n-6 1 n-4

n - .5 n - .7.. etc. Cosn-Ode+ . ... etc. [coso3 9 sit) 9],

and by continuingthe processthe integralmaybe evaluated.

For example,if we wishto findthe area between0 and 9 forn = 8 we have
6 4 2 1 6
oxd
area= 5-.2c- -.4'9dr
s-dOS 14
3'

=-8+!.Cos9sin4 d +! cos39sii9+i .-cos9sin9i

3r 7r 3'7r a37r
and it will be noticed that forn =10 we shall merelyhave to add to this same

expressionthe term A - - - cos79 sin 9.

The tables at the end of the paper give the area between - oo and z

(or =-- and =tan-l z).

This is the same as *5+ the area between 9=0, and 9= tan- z, and as the
whole area of the curve is equial to 1, the tables give the probabilitythat the
mean of the sample does not differby more than z times the standarddeviation
of the sample fromthe mean of the population.
The wholearea of the curve is equal to

9
2. n-4 ... etc. x | cos-2 OdO,
2 - 3'nn- .5

and since all the parts betweenthe limitsvanishat both limitsthis reducesto 1.
Similarlythe secondmomentcoefficient
is equal to

nt-2 n-4
22-8 qz-5 2d
_~~~~~+7r
= -n82 n_
. ... etc. x tcos 9tal2
an )d
2

-3
n-2 5-31
-3- - _8
it 2
2--2
12 The Probable Error of a Mean
Hence the standarddeviationof the curve is 1//n-3. The fourthmionment
coefficielnt
is equialto
it-2 .n 4 x
oetc.
2
2 9 tan4OdO

- l 3*f-5 J'i -

.....etc.x |
+
= -_2 . n (cosqO- 2 cosql4O cosn9-2)
dO
nt-2 n -4 2(- _
2 _ _ _ _ _ _

n-2 n-4 2 (it -2) 13

n-3'n-5 n-3 (n-3)(n-5)
The odd momentsare of cotursezero as the curveis symmetrical,
so

11= ?, 18 = 3 (n- 3)= 3 + 2-

Hence as n increases the cutrveapproaches the normalcurve whose standald
deviationis 1/Vn--3.
,82 howeveris alwaysgreaterthan 3, indicatingthat large deviationsare more
commonthan in the normalcurve.
N 8 64 2
DIAGBAM Il. Solid curve Y= Sx - . -, , 2 cos0? 0, x/s'= tan 0.
7x'
Broken line curve Y= ,, N e is, the normal curve wvith
the same S.D.

3~~~~~~~~~~
N /
SN

.N ________
1-S 1/S*SOsjj S15

N~~~~~~~~~~~~~~~~~

1.53 -Os 50
3I .5 *58
Os 5S
Distance of meanfionm
mean of population
I have tabled the area forthe normalcurve withstandarddeviation1/V7so as
to comparewith mycurve forn =10*. It will be seen that odds laid according
to either table would not seriouslydiffertill we reach z =-8, where the odds are
about 50 to 1 that the mean is withinthat limit: beyond that the normalcurve
gives a. false feelingof secturity,
forexample,accordingto the normalcurve it is
99,986 to 14 (say 7000 to 1) that the mean of the populationlies between-o
and + 1 3s whereasthe real odds are only99,819 to 181 (about 5.50to 1).
* See p. 19.
 BY STUDENT 13

Now 50 to 1 correspondsto threetimesthe probableerrorin the normalcurve

and formost purposeswould be consideredsignificant;forthis reasonI have only
tabled my curves forvalues of n not greater tharn10, but have given the n = 9
and n = 10 tables to one further
place of decimals. They can be use(das foundations
forfindingvalues forlargersamples*.
The table forn =2 can be readilyconstructedby looking out 9 = tan-' z in
Chambers'Tables and then*5+ 9/7rgives the corresponding value.
Similarlyi sin 9 +*5 gives the values whenn = 3.
There are two points of interestin the n = 2 curve. Here s is equal to half
the distance between the two observations. tan-'5 8sothat between + s and
.7r
-s lies 2 x x 1 or half the probability,i.e. if two observationshave been made
4 wr
and we have no other informationi, it is an even chance that the mean of the
(normal)populationwill lie betweenthema. On the otherhand thesecondmoment
coefficient
is
?r
2

1f 9d9
tan" [tan - a j =?,
17r_7r W
r 7r
2

or the stan(darddeviationis infinitewhilethe probableerroris finite.

SECTION VI. Practical Test of theforegoingEquations.

BeforeI had succeededin solvingmyproblemanalyticallv,I had endeavoured
to do so empirically. The material used was a correlationtable containingthe
height and left middle fingermeasurementsof 3000 criminals,fromn a paper by
W. R. Macdonell (Biometrika,Vol. i. p. 219). The measurementswere written
out on 3000 pieces of cardboard,whichwere then very thoroughlyshuffledand
drawnat random. As each card was drawn its numberswere writtendown in a
book whichthus containsthe measurementsof 3000 criminalsin a randomn order.
Finally each consecutiveset of 4 was taken as a sample-750 in all-and the
mean, standard deviation, and correlation t of each sample determined. The
differencebetweenthe mean of each sample and the mean of the populationwas
then dividedby the standarddeviationof the sample,givingus the z of SectionIII.
This provides us with two sets of 750 standard deviations and two sets of
750 z's on which to test the theoreticalresuiltsarrivedat. The height and left
middle fingercorrelationtable was chosen because the distributionof both was
approximatelynormaland the correlationwas fairlyhigh. Both frequencycurves,
however, deviate slightlyfromnormality,the constantsbeing forheight 8,1= '0026,
192=3-175,and forleftmiddlefingerlengths19,=0030, ,/2=3-140,and in consequence
there is a tendencyfor a certain numberof larger standarddeviationsto occur
than if the distributionswere normal. This, however,appears to make verylittle
differenceto the distributionofz.
* E.g. if n= 11, to the correspondilng O
value for it= 9, we add X x x xixi x cos8 sill : if it= 13
we add as well 9 X x X xI x X cosi0 sin0 and so on.
X x
t I hope to publislhthe resultsof the correlationworkslhortly.
14 The ProbableError of a Mean
Anotherthingwhichinterferes withthe comparisonis the comnparatively large
groupsin whichthe observationsoccur. The heightsare arrangedin 1 inchgroups,
the standard deviation being only 2-54 inches: while the fingerlengths were
originallygroupedin millimetres,
but unfortunatelyI did not at the time see the
importanceof having a smaller unit, and condensedthem into two millimetre
groups,in termsof whichthe standarddeviationis 2-74.
Several curiousresultsfollowfromtakiiigsamplesof 4 fiom material disposed
in such wide groups. The followingpointsmay be noticed:
(1) The means onlyoccuiras multiplesof *25.
(2) The standarddeviationsoccuras the square rootsof the following
types
of numbersn, n + -19,n + *25,A + 50, n + *69,2n + *75.
(3) A standarddeviation belonging to one of these groups can only be
associated witha meau of a particularkind; thus a standarddeviationof /2 can
only occur if the mean differsby a whole number fromthe group we take as
origin,while V1A69will onlyoccurwhen the mean is at n + *25.
(4) All the fourindividualsof the sample will occasionallycome fromthe
same group,givinga zero value forthe standarddeviation. Now this leads to an
infinitevalue ofz and is clearlydue to too wide a grouping,foralthoughtwo men
may have the same heightwhen measured by inches,yet the finerthe measure-
mentsthe moreseldomwill theybe identical,till finallythe chance that fourmen
will have exactlythe same heightis infinitely small. If we had smallergrouping
the zero values of the standarddeviation might be expected to increase,and a
similar considerationwill show that the smallervalues of the standarddeviation
would also be likely to itncrease,such as -436, when 3 fall in one group and 1
in an adjacent group,or -50 when 2 fall in two adjacent groups. On the other
hand whenthe individualsof the sample lie farapart,the argumentof Sheppard's
correctionwill apply,the real value of the standarddeviationbeing morelikelyto
be smallerthan that foundowingto the frequencyin any group being greateron
the side nearerthe mode.
These two effectsof groupingwill tend to neutraliseeach otherin theireffect
on the mean value of the standarddeviation,but bothwill increasethe variability.
Accordinglywe findthat the mean value of the standard deviation is quite
close to that calculated,whilein each case the variabilityis sensiblygreater. The
fitof the curveis not good,both forthis reason and because the frequencyis not
evenly distributedowing to effects(2) and (3) of grouping. On the other hand
the fitof the curve givingthe frequencyof z is verygood and as that is the only
practicalpointthe comparisonmaybe consideredsatisfactory.
The followingare the figuresforheight -
Mean value of standarddeviations; calculated 2-027+ *021
observed 2-026
Difference= -001
 BY STUDENT 15

Standarddeviationof standarddeviations:
Calculated *8556+ 015
Observed *9066
= + *0510
Difference

Equation: y= 16X 75
Comparisonof Fit. Theoretical e

Sckleinterms!
of standard
'^ _; l I
l [ | o
o c
deviationofI 0
|OI4O 0 0 |a
population x 9 z %*.

Calculated
frequenicy 1 1O 27 451
f 641 781 87 88 S1 71 58 45 33 23 15 9i 5i 7
Observed
frequency 3 141 241 137 107 67 73 77 77i 64 52* 49| 35 28 12A 9 Ilj 7

+1j
Differenice +4 -2| -8 +421 -Ili -14 -11 -4 -7 -51+4j| +2 +5 -2|-2 +6 0

wheInCe P= 000,06(about).
X2=48O06,

In tabling the observed frequency,values between *0125 and *0875 were

includedin one group,while between*0875and *0125 theywere divided over the
two groups. As an instanceof the irregularitydue to groupingI may mention
that there were 31 cases of standarddeviations 1 30 (in terms of the grouping)
whichis *5117in termsof the standarddeviationof the population,and they were
therefore dividedover the groups*4to *5and *5to *6. Had theyall been counted
in groups *5 to *6 x2would have fallen to 29-85 and P would have risen to 03.
The x2test presupposesrandom sampling froma frequencyfollowingthe given
law, but this we have not got owingto the interferenceof the grouping.
When,however,we test the z's wherethe groupinghas not had so much effect
we finda close correspondence betweenthe theoryand the actual resuilt.
There were three cases of infinitevaluiesof z which,for the reasons given
above, were given the next largest values which occurred,namely + 6 or - 6.
The restwere divided into groupsof 1; 04, 05 and *06,being divided between
the two groups on either side.
The calculated value for the standard deviation of the frequencycurve was
1 (? 017) while the observedwas-1P039. The value of the standarddeviation is
reallyinfinite,as the fourthmomentcoefficient is infinite,but as we have arbi-
trarilylimited the infinitecases we may take as an approximation 1 from -

V%1500
which the value of the probable errorgiven above is obtained. The fitof the
curve is as follows:
16 The ProbableError of a Meant

Comparisonof Fit. TheoreticalEquiationt:y = 2 cos4 6, z =tan 0.

-
Scaleofzal ~~~~~ j~~- ~ 0
I
0
I~ +
0
+ + + + + +
Scale of z o
c) o ? o o
0 o
0 0
E

~~~'~~iIi
I~ + + 0
I Ij~~~~~~~~
~~~~i ~+ + +

Calculated l l l l l l l- l
frequen3cy5 9i 13i 34j 44* 78 1119 141 119 78i| 441 34.1 13.1 9i- 5
Observedt
frequency 9 14j 1lj 33 43i 70 119i 151k 122 67I 49 26i 16 110 6

Difference+4 -2 I I 8 1 +i| + 1J + 3 -11 +4 -8 K

+221 + s+
whence X2=12a44, P=-56.

This is verysatisfactory,especiallywhen we considerthat as a rule observa-

tions are tested against curves fitted fromthe mean and one or more other
momentsof the observations,so that considerablecorrespondenceis only to be
expected; while this curve is exposed to the full errorsof random sampling,its
constantshavingbeen calculatedquite apart fromthe observations.
III.
DIAGRRAM Comparison Curvewith750 actual
of CalculatedStandardDeviationFrequency
StandardDeviations.

loo_c__- --I I- - I- I--'-- 1111

.1 *2 3 *4 *5 *f 7 * -
*9 i-n 1.1 1.2 1-3 1-4 1 r 1-6 1,7 1-8 1-9 2-0 2-1 2-2 2,3 2-4 2-5
Scale of Standard Deviation of the Population

The leftmiddle fingersamples show much the same featuresas those of the
height,but as the groupingis not so large comparedto the variabilitythe curves
fitthe observationsmoreclosely. Diagrams III.* and IV. give the standarddevia-
tionsand the z's forthis set of samples. The resultsare as follows:
* Therearethreesmallmistakes valuesin DiagramIII., whichmakethefit
theobserved
in plotting
appearworsethanit reallyis.
 BY STUDENT 17

0 0 0 0
_ c -:f cN co

@3~~~~~~~~~~c

0
@3~~~~~~~~~~c

Cs
04~~~~~~~~~

II~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

o 00

BiometrikaVI 3
 18 The ProbableError of a Mean
Mean value of standarddeviations; calculated 2 186 + 023
observed 2-179
Difference= - *007
Standarddeviationof standarddeviations:
Calculated 9224 + 016
Observed 9802
Difference= + 0578
16 x 750 2xR
Comparison
of Fit. Theoretical
Equation: Y= V2 a2ee0.

Scaleinterms||| |0 b
of standard o "
deviationof 0 | 0 0 0 0 00 0 0 0 0
populatton| | O | j . 0 o |

Calculated
frequeney 1j 10k 27 45j 64i 78i 87 88 81| 71 58 45 33 23 15 9k 5i 7
Observe
frequency 2 14 27k 51 64j 91 94k 68k 65j 73. 48j 40j 42k 20 22i 12 5 7j

+k +3k
Difference + +5 - + 12| + 7 -19i -16 + 2 -9k -4k + 9k 3 + 7k +2| -| +k
whence
X2=2180,P=K19.
Calculatedvalueofstandarddeviation1 (+ 017)
Observed *982
Difference =-018

Comparison
of Fit. Theoretical
Equation: y=2 cos", z =tan 6.

f17
10 1

I ScaleofzIaI ?I. '7

I?I?z?
10 I '~~~~~~~x~
+ + I + Q +I + + +
____ I~~~~~~~~~~+ __ + +
a I 8 I8 IE I8 Ia I9, I 4+ <+ I a j 8 k B I E 1'< 1

ICalculatedl
9k 13
Scaeiec78z 53440 3k44 78k 119 141 119 1 3 | 13k| |
|frequency?4 15ki18 13k144 75 1122138 120k| 71 |46k| 36, 11 |9 |6|

Differenlce
|-1 +6i +4k -1 |-k| -3k ? 3 | 3I+1k-7k+2Klk+1i-2k| - +1|
whencex2=739, P=92.
A veryclose fit.
We see thenthatif thedistribution
is approximately
normalour theory
gives
us a satisfactory
measureof the certaintyto be derivedfroma smallsamplein
boththecaseswe havetested; but we havean indicationthata finegroupingis
 BY STUDENT 19

of advantage. If the distributionis not normal,the mean atid the standard

deviationof a sample will be positivelycorrelated,so that althoughboth will have
greatervariability,yet theywill tend to counteracteach other,a mean deviating
largelyfromthe generalmean tendingto be divided by a largerstandarddeviation
ConsequentlyI believe that the tables at the end of the present paper may be
used in estimatingthe degree of certaintyarrived at by the mean of a few
experiments, in the case of most laboratoryor biologicalworkwhere the distribu-
tionsare as a rule of a ' cocked hat' type and so sufficiently
nearlynormal.

/3. innodd
od tan'z
n
SECTION VII. Tables of n-2 . 1t e n) con2 6 dO
1 .- n even
for values of n from 4 to 10 inclusive.

Togetherwith e 2dx for comparisonwhen n = 10.

For comparison
( 8|n='4 n=5 n=6 n_
n=7 =8 n=9 L=1O (1f| e 2da)

*1 5633 '5745 5841 5928 6006 '60787 61462 '60411

.2 6241 '6458 *6634 *6798 *6936 70705 *71846 70159
.3 6804 *7096 7340 7549 *7733 *78961 *80423 *78641
*4 7309 '7657 7939 '8175 '8376 '85465 '86970 '85520
.5 -7749 '8131 '8428 '8667 '8863 '90251 '91609 '90691
.6 '8125 '8518 '8813 '9040 '9218 '93600 '94732 '94375
.7 '8440 '8830 '9109 '9314 '9468 '95851 '96747 '96799
'8 '8701 '9076 9332 '9512 9640 97328 '98007 '98253
.9 '8915 '9269 '9498 '9652 '9756 '98279 '98780 '99137
1P0 '9092 '9419 '9622 '9751 '9834 '98890 '99252 '99820

1.1 '9236 '9537 *9714 *9821 *9887 *99280 99539 *99926

12 9354 '9628 '9782 '9870 '9922 '99528 '99713 '99971
1'3 '9451 '9700 '9832 '9905 '9946 '99688 '99819 '99986
1'4 '9531 '9756 '9870 9930 '9962 '99791 '99885 '99989
1 '5 9598 '9800 '9899 '9948 9973 '99859 '99926 '99999
16 9653 '9836 '9920 '9961 '9981 '99903 '99951
1'7 '9699 '9864 9937 9970 '9986 '99933 '99968
18 9737 '9886 9950 9977 9990 -99953 '99978
1.9 '9770 9904 '9959 '9983 9992 '99967 '99985
2'0 '9797 '9919 '9967 '9986 9994 '99976 99990

2'1 '9821 '9931 9973 *9989 '9996 '99983 '99993

2'2
2 9841 '9941 '9978 9992 9997 '99987 99995
2'3 9858 '9950 9982 9993 '9998 '99991 '99996
2'4 9873 -9957 '9985 '9995 9998 -99993 99997
2'5 9886 9963 9987 '9996 9998 '99995 99998
2'6 9898 9967 '9989 '9996 '9999 '99996 *99999
2'7 '9908 '9972 .9991 -9997 9999 .99997 99999
2'8 '9916 '9975 9992 '9998 .9999 '99998 '99999
2'9 9924 '9978 9993 '9998 '9999 '99998 '99999
3'0 9931 '9981 9994 '9998 '99999

8-2
20 The Probable Error of a Mean

SECTION VIII. Explanationof Tables.

The tables give the probabilitythat the value of the mean,measuredfromthe
mean of the population,in termsof the standarddeviationof the sample,will lie
between- oo and z. Thus, to take the table for samnplesof six, the probability
of the mean of the population lying between - oo and once the standard
deviationof the sample is *9622 or the odds are about 24 to 1 that the mean of
the population lies between these limits.
The probabilityis therefore*0378 that it is greater than once the standard
deviation and *0756 that it lies outside + 10 times the standard deviation.

SECTIONIX. of Method.
Illustrations
IllustrationI. As an inistanceof the kind of use whichmay be made of the
tables,I take the followingfiguresfroma table by A. R. Cushnyand A. R. Peebles
in the Journalof Physiology for1904, showingthe different effectsof the optical
isomersof hyoseyaminehydrobromide in producingsleep. The sleep of 10 patients
was measured without hypnoticand aftertreatment(1) WithD. hyoseyarnine
hydrobromide, (2) with L. hyoseyaminehydrobromide.The average number of
hours'sleep gained by the use of the drug is tabuilatedbelow.
The conclusionarrivedat was that in the usual dose 2 was, but 1 was not,of
value as a soporific.

hydrobromide.
hours'sleepgainedbytheuseofhyoseyamine
Additional
Patient 1 (Dextro-) 2 (Laevo-) Difference(2-1)
1. + .7 +1-9 +12
2. - 16 + *8 +2-4
3. - *2 + 1*1 + 1-3
4. -1-2 + 1 +1-3
5. -1 - 1 0
6. +3-4 +4-4 + 10
7. + 3-7 + 5-5 + 1-8
8. + *8 + 1-6 + *8
9. 0 +4-6 +4-6
10. +2-0 +3A4 + 1-4
Mean + *75 Mean + 2-33 Mean + 1 58
S. D. 1.70 S. D. 1.90 S. D. 1.17

First let lls see what is the probabilitythat 1 will on the average give increase
of sleep; i.e. what is the chance that the mean of the populationof whichthese
experimentsare a sample is positive. + .75 = 44 and looking out z = 44 in the
1P70
 BY STUDENT 21

table forten experimentwe findby interpolatingbetween*8697and *9161that *44

correspondsto '8873, or the odds are *887to '113 that the mean is positive.
That is about 8 to 1 and would correspondin the normal curve to about
18 times the probable error. It is then very likely that 1 gives an increase of
sleep, but would occasion no surprise if the results were reversed by further
experiments.
If now we considerthe chance that 2 is actuallya soporificwe have the mean
2-33
increase of sleep = 1.9O or 1'23 times the S.D. From the table the probability
correspondingto this is -9974, i.e. the odds are nearly400 to 1 that such is the
case. This correspondsto about 4-15 times the probable error in the normal
curve. But I take it the real point of the authorswas that 2 is betterthan 1.
This we must test by making a new series, subtracting1 from2. The nmean
value of this series is + 1C58while the S.D. iS 1-17,the mean value being + 1-35
timnesthe S.D. From the table the probabilityis -9985 or the odds are about 666
to 1 that 2 is the better soporific. The low value of the S.D. is probablydue to
the differentdrugs reactingsimilarlyon the same patient,so that thereis corre-
lation betweenthe results.
Of courseodds of tlhiskind make it alrnostcertainthat 2 is the bettersoporific,
and in practical life such a high probabilityis in most matters considered as
a certainty.
IllustrationII. Cases where the tables will be useful are not uncommonin
agriculturalwork,and they would be more numerousif the advantages of being
able to apply statisticalreasoningwerebornein mind when planning the experi-
ments. I take the followinginstancesfromthe accountsof the Woburnfarming
experimentspublished yearlyby Dr Voelcker in the Journalof theAgricultural
Society.
A short series of pot culture experimentswere conductedin order to deter-
mine the causes which lead to the productionof Hard (gltutinous)wheat or Soft
(starchy)wheat. In threesuiccessiveyears a bulk of seed cornof one variety vas
picked over by hand and two samples were selected,one consistingof "hard"
grainsand the otherof " soft." Some of each of these wereplanted in both heavy
and light soil and the resultingcrops were weighed and examined forhard and
soft corn.
The conclusiondrawnwas that the effectof selectingthe seed was negligible
comparedwiththe influenceof the soil.
This conclusionwas thoroughly justified,the heavysoil producingin each case
nearly100 per cent. of hard corn,but still the effectof selecting the seed could
just be traced in each year.
But a curious point,to which Dr Voelcker draws attentionin the 2nd year's
report,is that the softseeds producedthe higheryield ofboth cornand straw. In
 22 The ProbableError of a Mean
viewof the well-known factthat the varieties
whichhave a highyieldtendto
producesoftcorn,it is interesting to see how muchevidencethe experiments
affordas to thecorrelation betweensoftness and fertility
in thesamevariety.
Further, MrHooker*has shownthattheyieldof wheatin oneyearis largely
determined bythe weatherdurinig the preceding harvest. Dr Voelcker's
results
mayafford a clue as to the wayin whichthe seed is affected, and wouldalmost
justifytheselectionofparticularsoilsforgrowing seedwheatt.
The figuresare as follows,
theyieldsbeingexpressed in grammes perpot.

Year 1899 1900 1901

-Average
_____ ___ _____ __ Standard
Average Deviation
Soil Light Heavy. Light Heavy Light Heavy

Yield of corn fromsoftseed 7'85 8-89 14-81 13,55 7-48 15-39 11'328
b,, ard ,, 7,27 8,32 13-81 13,36 7V97 13-13 10,643

Difference ... +58 +*57 +P100 +*19 - 49 + 2-26 + *685 *778 *88

Yield of straw fromsoftseed 12-81 12,87 22-22 2021 13-97 22'57 17A442
,, hard ,, 10-71 12A48 21P64 20,26 11,71 18-96 15-927

Difference. ... ... +2lo0 + 39 + *78 -*05 +2'66 +3-61 +515 1-261 1P20

If we wishto findthe odds that softseed will give a betteryieldof cornon the
average,we dividetheaveragedifference by thestandarddeviation,givingus
z = *88.
Lookingthis up in the table forn =6 we findp= 9465 or the odds are
*946.5:535, about 18: 1.
Similarly forstraws = 1-20,p = '9782,and theoddsabout45:1.
In orderto see whethersuch odds are sufficient fora practicalmanto drawa
definite I take anotherset of experiments
conclusion, in whichDr Voelckercom-
paresthe effects of different
artificialmanuresusedwithpotatoeson the large
scale.
The figures represent the difference betweenthe cropsgrownwiththeuse of
sulphateofpotashand kainitrespectively in both1904and 1905.
awt. qr. lb. ton owt. qr. lb.
1904 + 10 3 20: + 1 10 1 26 (two experimentsin each year).
1905 + 6 0 3: + 13 2 8(
* JournalofRoyal StatisticalSociety,1907.
t And perhapsa fewexperimentsto see whetherthereis a correlationbetweenyield and ' mellow-
ness' in barley.
 BY STUDENT 23

The average gain by the use of sulphate of potash was 1525 cwt. and the
9 cwt.,whence,if we want the odds that the conclusiongiven below is right,
S.D.
z= 117corresponding, when n = 4, to p = *9698 or odds of 32: 1; this is midway
betweenthe odds in the formerexample. Dr Voelckersays ' It nmay now fairlybe
concluded that for the potato crop on light land 1 cwt.per acre of sulphate of
potash is a better dressing than kainit.
As an example of how the tables should be used with caution, I take the
followving whether
pot culture experimentsto test whetherit made any difference
large or small seeds weresown.

IllustrationIII. In 1899 and in 1903 " head corn" and " tail corn" weretaken
fromthe same bulks of barley and sown in pots. The yields in grarnmeswere
as follows:
1899 1903
Large seed ..... 13-9 7.3
Small seed ..... 14-4 8_7
+,5 + -6
The average gain is thus *55 and the S.D. '05, giving z = 11. Now the table
for n = 2 is not given, but if we look up the angle whose tangent is 11 in
Chambers' tables,
tan-' 11 + 5 = 840 47 +5 = 97 1
p
i800 + 1800 5~9l
so that the odds are about 33: 1 that small corngives a betteryield than large.
These odds are those whichwould be laid, and laid rightly,by a man whoseonly
knowledgeof the matterwas contained in the two experitnents. Anyone cbn-
versantwithpot culturewould howeverknowthat the difference betweenthe two
results would generally be greater and would correspondinglymnoderatethe
certaintyof his conclusion. In point of fact a large scale experimentconfirmed
the result,the small cornyieldingabout 15 per cent. morethan the large.
I will conclude withan example whichcomesbeyondthe range of the tables,
therebeing eleven experiments.
To test whetherit is of advantage to kiln-drybarleyseed beforesowing,seven
varietiesof barleyweresown(both kiln-driedand not kiln-dried)in 1899 and four
in 1900; the resultsare given in the table.
It will be noticed that the kiln-driedseed gave on an average the largeryield
of corn and straw,but that the quality was almostalways inferior. At firstsight
this mightbe supposedto be due to superiorgerminatingpower in the kiln-dried
seed, but my farmingfriendstell me that the effectof this would be that the
kiln-driedseed would produce the betterquality barley. Dr Voelckerdrawsthe
conclusion"In such seasons as 1899 and 1900 thereis no particularadvantagein
kiln-dryingbeforesowing." Our examinationcompletelyjtustifies this and adds
 24 The Probable E1rrorqf a Mean
"and the quality of the resultingbarley is inferiorthough the yield may be
greater."

lbs. head corn per acre Price of head corn in Value of cropper acre
shillingsper quarter cwts. strawper acre in shillings*

N. K. D. K. D. Diff. N. K. D. K. D. Diff. N. K. D. K. D. Diff. N. K. D. K. D. Diff

1903 2009 +106 261 261 ) 191 25 +5 140j 152 +11]

1935 1915 - 20 28 261 -1' 221 24 + 1 152 145 -71
1910 2011 +101 29i 281 -1 23 24 + 1 158 161 +21
1899.! 2496 2463 1 - 33 30 29 --1 23 28 +5 204k 199k1 -5
2108 2180 + 72 271 27 -1 221 221 0 162 164 +2
1961 1925 - 36 26 26 26
26 00292 19~
194 122 -1
142 1391 -2j
2060 2122 + 62 29 26 -3 2412 227 - 168 155 -13
1444 1482 + 38 29 281 -1 151 16 +1 118 l171 -j
1542 - 70 28 28 1 18 171 128k 121 -71
1900 1612 - -_
1316 1443 +127 30 29 -1 141 15 +l 109* 1161 +7
1511 1535 + 24 28i 28 |_ 17 174 +1 120 12Cg + 2

Average 184165 18752 +337 2845 2755 -'91 1995 2105 +110 14582 14468 +114

Standard -
j 63 1 - -
.79 - -
225 - - | 667
Standard)
}
Deviation* 223 - -- 28 - - 80 - - 240

* Strawbeingvaltiedat 15s. per ton.

In this case I propose to use the approximationgiven by the normal curve

S -
with standarddeviation V( and thereforeuse Sheppard'stables,lookingup

divided by s8
the difference The probabilityin the case of yield of coruiper

acre is given by lookingup 3237= 1 51 in Sheppard's tables. This givesp = 934,

or the odds are about 14 :1 that kiln-driedcorngives the higheryield.

Similarly-9 = 325, correspondingto p =9994,* so that the odds are very

28
great that kiln-driedseed gives barleyof a worse quality than seed whichhas not
been kiln-dried.
Similarlyit is about 11 to 1 that kiln-driedseed gives more strawand about
2 :1 that the total value of the crop is less withkiln-driedseed.
* As pointedout in SectionV. the normal curvegives too large a value forp when the probability
is large. I findthe truevalue in this case to be p = 9976. It matterslittlehoweverto a conclusionof
this kind whetherthe odds in its favourare 1,660: 1 or merely416: 1.
 BY STUDENT 26

SECTION X.

Conolutsion8.
I. A curvehas been foundrepresentingthe frequencydistributionofstandard
deviationsofsamples drawnfroma normalpopulation.
II. A curvehas been foundrepresentingthe frequencydistributionof values
of the means of such samples,when these values are measuredfromthe mean of
the populationin termsof the standarddeviationof the sample.
III. It has been shown that this curve representsthe factsfairlywell even
whenthe distributionof the populationis not strictlynormal.
IV. Tables are givenbywhichit can be judged whethera seriesofexperiments,
howevershort,have given a result which conformsto any required standard of
accuracyor whetherit is necessaryto continuethe investigation.

Finally I should like to expressiny thanksto ProfessorKarl Pearson,without

whoseconstantadvice and criticismthis paper could not have been written.

Biometrikavi 4