Вы находитесь на странице: 1из 4
  • 248 STATISTICS: W. A. SHEWHART

P'ROC. N. A. S.

ON THE MEASUREMENT OF A PHYSICAL QUANTITY 'WHOSE

MAGNITUDE ISINFLUENCED BY PRIMARY CAUSES BEYOND THE CONTROL OF THE OBSERVER AND ON THE METHOD OF DETERMINING THE RELATION BETWEEN TWO SUCH

QUANTITIES

By W. A. SHEWHART

RESEARCH LABORATORIES Or THE AMERICAN TZLOPHONS, AND TELEGRAPH

COMPANY AND THE WOSTURN ELCTRIC COMPANY, INC. Read before theAcademy, April 25, 1922

In order to relate the physical and chemical properties of carbon to

the microphonic properties of this material, it has been found necessary

to make measurements on physical quantities which are influenced by

primary causesbeyondthe control ofthe observer. In many problems of

thephysicaland engineeringsciencesitispossiblefortheobserverto con-

trol within narrow limits, the causes of variation of a quantity while itis

being subjected to measurement. Certain problems arise, however, in

these sciences as in the field of economics and biology wherein it is im-

possible to do this. In general let Xi and X2 represent two quantities

related to others U1, U2, followingway:

Ua; Vi, V2,

Vb; W3, W2

Xi

F1 (Ui, U2,

X2= F2 (U1, U2,

Ua, Vl, V2,

Vb)

Ua, W1, W2,

Wb)

Wb inthe

(1)

(2)

where F1 and F2representunknown functional relations. In most physi-

calexperimentsitispossibletoholdthe U's, V's,and W's,practicallycon-

stantwhile a measurement isbeingmade on either X1 or X2. In the last

analysis, however, the U's, V's, can never be held constant and in many

cases, particularly where these symbols represent molecular phenomena,

thevariations abouttheir mean values may become quitelarge. Thus,ifX1 andX2representtwo microphoniccharacteristicsofgranular carbon which are functions of the physical and chemical properties rep- resented by the U's, V's, and W's, it has been found necessary to apply

certain statistical criteria to determine the nature of the cause complex

controllinga singlequantitysuch as eitherXi or X2 and alsoto determine

quantitatively the degree of relation existing-between the two micro- phonic properties. As a result of such a study involving an analysis of thousands of measurements of the above type, 'certain conclusions which .areofgeneralinteresthave been reachedinrespect to thepractical appli- cation of statistical methods in connection with physicalmeasurements

of thischaracter.

VOL. 8, 1922

STATISTICS: W. A. SHEWHART

249

Inordertostudythenatureof thecomplexofcausescontrollingasingle

quantity such as X1, one of the firstproblems is to determine whether or not the causes represented by the U's, and V's satisfy the following con-

ditions: (1) That allof the causes, n in number, are effectiveat thetime of each observation, (2) that the probability, p,that a cause will produce a positive effectisthe same for all of the causes. (3) That the probabil-

ity, p, remains the same for allofthe observations, (4) thattheeffect, Ax,

ofasinglecauseisthesameforallofthecauses. Iftheseconditionsare fulfilled, thedistributioninXicanberepresented

by the successive terms of the expansion (p + q)n where the ordinates

are separated at intervals of 2Ax. For most of the problems ithas been

found convenient to compare the observed distribution with the theoreti-

cal distributionconsistentwith theabove random conditions. In general, thefollowingprocedurehasbeen followed: Foreach observed distribution

two factors k = -3-- -23

/2

and 12 -

2 22 have been calculated and compared

with similar factors consistent with the above mentioned binomial ex-

pansion, where the first four corrected moments of the observed distri- bution about the mean are represented bythe symbols jul, jf, p3,and &4.

Itwillbenoticedthatkand 32areindependentoftheunitsusedinmaking

themeasurements.

For a symmetrical distribution k is always 0 and 12 may vary between 1 corresponding to n = 1 and 3 corresponding to n = oo. Irrespective

of the values of p and q, itshould be noted that 12does.not increase be-

yond 4 for 10,000 causes and even for 100 causes the value of 4 is not exceeded except for conditions in which the skewness k isvery large. In general large values of 12indicate eitherhigh skewness and.few causes or that the complex of causes does not follow the random conditions con- sistent with the expansion of (p + q)n. As aresultofvariationsduetosamplingthe standard deviations, 0kand of2ofk and 2are giveninterms ofthenumber of observations, s, by the

followingexpressionsak = and ap, = 44 itbeingassumed thatthe

distribution is practically normal.

If, in practice, the values of k and

12are found to differfrom 0 and 3.respectivelyby' more than three times

the standard deviations of k and 132,itis practically certain that the dis- tribution iseitherunsymmetrical or that the causes do not act at random as defined above. The method of application of these criteria depends

in general upon whether or not the observed distribution is one involving

attributes or variables.

In general for the case of attributes the procedure is as follows: The

factors k and 12are calculated to determine whether or nottheyaresignif-

  • 250 STATISTICS: W. A. SHEWHART

PROC. N. A. S.

icantly differentfrom 0 and 3. Ifk and 132are differentfrom 0 and 3, the values ofp, q, and n may be calculated since pn = i and pqn = A where

x represents the average of the distribution, and a study of these values

will indicate the probable reason for such variations in k and 132provided the distribution follows the random laws. If the observed distribution is consistent with the calculated values of p, q, and n, the "probability of fit"l between the observed distribution and the theoreticaldistribution

representing the expansion of (p + q)nshould be high.

In the case that the U's, V's and W's are variables, we start with un- known values ofp, q and n, and in addition have theunknown value of A

X which represents the effectofa singlecause. Itisobviouslyimpossible

todetermine thevalues ofp, qand n as inthecase ofattributesforamong other reasons we cannot determine the origin of the distribution. In general, however, we may make use of the factors, k and 12, since these are independent of the unit in which the measurements are made. Fur-

thermore, we may make use ofthe criterion, establishingthe "goodness of

fit" between the observed distribution and a theoreticalone corresponding to thevarious degrees ofapproximation2 to the normal law.

To return toequations (1)and (2)itisevidentthat,ifnone ofthecauses are common, the correlation between simultaneous measurements of X1 and X2 willbepracticallyzero. Furthermore, as thenumber ofcauses

that are common increases we should expect an increasein the correlation coefficientunder certainlimitingconditions. Thoughitmay not bepossi- ble to determine the number of causes that influence a variable quantity, such as either X1 or X2, ithas been found possible, however, to determine the approximate value for the ratio, of the number of common causes to the total number of causes influencing either variable. To show this for

the general case assumed in equations (1)and (2)letus startwiththedefi- nition of the correlation coefficient r,

r :xlx

(3)

where xi and x2 represent the deviations from the mean values ofX1 and X2 and -,, and a., represent the standard deviations in these quantities and N represents the total number of observations made. Let us assume

independence between the primary causes. Ifwe expandequations (1)

and (2) by Taylor'stheorem and neglectthes econd and higherpowers of thevariationsofthe primarycauses we can obtainvaluesofthedeviations xiandx2and ofthestandarddeviations axi and os which,when substitu- ted in equation (3)reduce to the simple expression

a+b

a

a + b

(4)

VoL. 8, 1922

PHYSICS: H. E. FARNSWORTH

251

providing we add the further conditions that the standard deviations in

allofthe primarycausesareequalandthattheeffectsofalloftheprimary causes are equal. These conditions appear at firstvery limited, and in a practicalcaseitisnecessaryatleastqualitativelytocheck theassumptions underlying equation (4),butingeneral ithasbeenfound thatasignificant deviationinr may be taken toindicate a variation in the effectsofone or more groups ofprimary causes, upon oneofthevariablesand notupon the

other.

A more complete account ofthis investigation giving experimentaldata, willbe published elsewhere.

1 Pearson, K.

"Tables for Statisticians and Biometricians" pp. (26-28) and Phil.

Mag., Ser. 5, 1, 1900.

2Edgeworth,F.Y., Camb.Phil. Trans.,20, 1904 (36-65)and (113-141).

ELECTRONIC BOMBARDMENT OF NICKEL

By H. E. FARNSWORTH

DEPARTMENT OF PHYSIcs, THE UNIVZRSITY OF WIscoNsIN

Communicated June 13, 1922

Studies of the secondary (emergent) electrons produced by electronic

bombardment of metal surfaces have previously been made by various experimenters.' However, thefactthat most oftheresultswere obtained

previous to the development of modern high vacuum technique combined with the failure of the small amount of recent work to agree with these older results, leaves the important questions of this problem still un' answered. It is well known that the number of secondary electrons de-

pends upon the velocity of the primary (incident) beam, that more elec-

tronsleavethesurfacethanstrikeitiftheprimaryvelocityisgreatenough,

butthefollowingcharacteristicsofthevariousmetalsarestillnotdefinitely

determined: (1) the magnitude of the secondary electron current as a function of the primary velocity; (2) the velocity distribution of the

secondary electrons forany givenprimary velocity.

The present investigation is an attempt to obtain these characteristics for nickel. The type of apparatus employed was such that a primary electronstream, ofuniformvelocityandcross-section, couldbemadeeither

toimpingeon anickeltargetortopassintoanabsorbingFaradaycylinder. This was accomplished by moving the target, by means of a magnetic control, over a holein the Faraday cylinder into the path of the primary

stream. The secondary electrons which left the target were received by

a conducting cylinder. This arrangement made it possible to move the