You are on page 1of 14

Entropy and Other Measures of Concentration Author(s): P. E. Hart Reviewed work(s): Source: Journal of the Royal Statistical Society.

Series A (General), Vol. 134, No. 1 (1971), pp. 73-85 Published by: Blackwell Publishing for the Royal Statistical Society Stable URL: http://www.jstor.org/stable/2343975 . Accessed: 31/10/2011 15:12
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.

Blackwell Publishing and Royal Statistical Society are collaborating with JSTOR to digitize, preserve and extend access to Journal of the Royal Statistical Society. Series A (General).

http://www.jstor.org

1971]

73

Entropy and OtherMeasuresof Concentration


By P. E.
HART

University Reading of December1969.RevisedApril1970] [Received

In recent havebegunto use theentropy, redundancy, economists years, or ofa sizedistribution measure extent which to the to is business concentrated in the controlof giantfirms.This paper comparesthesenew measures from information derived withthe classicalstatistical theory measures of and dispersion withtraditional measures business of concentration derived the from cumulative concentration curve.It shows that when number the of is to firms large enough usestatistical distribution the theory, classical statistical measures superior theentropy the are to or redundancy. When number the of firms small,theentropy superior theredundancy, bothare is is to but inferior the traditional to measuresof concentration derivedfromthe cumulative concentration curve. Consequently, thereis littlepoint in usingtheinformation measures measure to business theory concentration.
1. INTRODUCTION

SUMMARY

statistics, thougha long timelag has oftentendedto elapse beforeeconomists have their recognized value. This is certainly forthe studyof businessconcentration, true forthe workof the Frenchengineer Gibrat(1931) was neglected economists by for manyyears. In recent have begunto use the conceptof entropy, years,economists in formulated statistical originally and thermodynamics subsequently used in informationtheory conmmunications by engineers.In particular entropy the function has been used to measurethe extent whichthe outputor employment an industry to of is concentrated thecontrol a fewlargefirms an industry Hildenbrand in of in (cf. and Paschen, 1964; Finkelsteinand Friedberg,1967; Theil, 1967; Horowitz, 1968; Stigler, 1968; Hexterand Snow, 1970). However,it will be arguedherethat,in the contextof businessconcentration, use of entropy the is theory unjustifiedt because it adds nothing traditional to statistical measures concentration. of Whenthenumber of observations largeenoughto use a theoretical is statistical distribution, classical the statistical derived the parameters from moments superior theentropy are to measure. whenthenumber observations small,forexamplein an oligopoly, Alternatively, of is thetraditional cumulative concentration curveis superior.
2. CLASSICAL STATISTICAL MEASURES

ENGINEERShave made importantcontributions economics and to economic to

Let the variatez have a continuousdistribution with arithmetic f(z) mean u'. In thecontext the economicsof businessconcentration variatez might the of the be
this that is measure other in t Ofcourse, does notimply entropy nota useful economic contexts, and non-economic. Indeed,iftheaxiomsunderlying concept entropy regarded necesthe of are as sarythenthe entropy measuremustbe used forit is the onlyfunction theseaxioms satisfying (cf.Theil,1967,p. 6).
4

74

HART - Entropy and OtherMeasuresof Concentration

[Part1,

function write distribution the We as by sizeofa firm measured itsemployment. may ofz
F(z)=

f(z) dz.

(1)

of function z maybe written Thefirst moment distribution


Fl(z)
=

zf (z) dzff zf(z) dz

, | zf(z) dz.

1 z

(2)

first for Kendall "incomplete moment" (2) and also (1958)use theterm and Stuart in value of z. However, the the between limits oo and theparticular integrate of the better follow example to is and context variate positive it seems the present "first term zeroand z. Their (1957,p. 12) and use thelimits Aitchison Brown and is also usedfor(2). distribution" moment abscissa F(z) andordinate of with TheLorenz curve z is thelocusofpoints F&(z), by with slopeat anyvalueofz, given
dF1(z) dF(z) zf(z) dz plf(z) dz z Pi

Kendalland Stuart (1958) showthatthe Lorenzcurveis convexto the F axis. the on at curve which tangent it the from that point theLorenz (3) Moreover, follows to corresponds thevaluez = p. Theabscissa is parallel thediagonal to F(z) = F&(z), the meansizeandtheordinate gives of below ofthis the gives proportion firms point are curves Lorenz belowmeansize. Frequently, of in proportion employmentfirms F(z) = 1- F(z), so thattheabscissaof theinterabout symmetrical thediagonal the F(z) = 1- Fl(z) gives proportion and thediagonal of section theLorenzcurve of is derived from properties theunderthe offirms below meansize. Thissymmetry is thatthere a largeclassof size and lying distribution Kendall(1956)had argued Lorenz curves. symmetrical distributions generate which complete F(z) = Fl(z) reflects It is customary arguethatsincethediagonal to reflects this awayfrom diagonal of of equality size,a movement theLorenzcurve the (1958), area and by as shown Kendall Stuart Furthermore, increasing inequality. F(z) = Fl(z) is halfof Gini'scoefficient between Lorenz the and curve thediagonal since of is a ofconcentration which clearly measure dispersion A1is themean Al/2p4 pairsofvaluesofz. Butin thecase all between possible difference repetition) (with the a of symmetrical Lorenzcurves, movement awayfrom diagonalF(z) = Fl(z) of in but in not implies onlyan increase inequality, also a decrease theproportion the are abovemeansize. ThuswhenLorenzcurves symmetrical Gini frequencies of they coefficient theLorenzcurveare morethanmeasures dispersion; also and the concept and reflect skewness kurtosis comecloseto measuring economists' and in is of to the industryconcentratedthe oftheextent which employmenta particular of concerns effect changes the limitation Theirmajoreconomic companies. largest later.Their and majorstatistical in N, thenumber companies, willbe considered of and intractability the absenceof appropriate are limitations theirmathematical of to so distributions them, it is preferable use the parameters the for sampling the between size which F(z) and Fl(z) govern relationship underlying distribution

1971]

HART -

Entropy OtherMeasuresof Concentration and

75

if and the and hencedetermine Ginicoefficient Lorenzcurves.For example, the of is thenar2 (the variance the of underlying distribution firms lognormal, size itgoverns others, the measure concentration; of logarithms size)is theappropriate of the F-test maybe used to see whether and is easilydecomposed thewell-known behind changes the in If to in observed changes a2 aresignificant.itis desired probe of of thenStudent's error the regression the t-test, usingthe standard variance of the t logarithm sizeat time on sizeat t-1, willsatisfy requirementssampling of by (1964) is errorof 1(bJ1b2) Horowitz of theory.The derivation the standard of b1 unnecessary, where is theregression logzton log zt-1 and b2is theregression oflogZt-i on logzt. why classical the statistical and properties important explain are Thesestatistical ofmean difference, to variance, arepreferredGinicoefficients etc.) parameters (mean, But of to a or Lorenz of measures concentration,summarizesizedistributionfirms. of is for a while statistical analysis size distributionsappropriate themeasurement of in as becausethenumber firms of business concentrationtheeconomy a whole, within it for of is very large, maynotbe appropriate themeasurementconcentration the of in the an individual industry, because number firms-particularly nearmonowhich interest economists-may too smallto applystatistical be industries polistic level distribution theory. Indeed,at theindustry it maybe thecase thatad hoc to of such ratio measures concentration, as theconcentration arepreferable measures in are of These hocmeasures considered detail ad suchas thevariance thelogarithms. section this of in thenext paper. of size. maximum Moreover, In practice dealwith we with samples firms a finite to are it is size theobserved distributions notcontinuous. Nevertheless, helpful use
of theconcepts Section2 and we estimate by Y p'
if
1

3.

CONCENTRATION CURVE MEASURES

isfi. of the the where classmark theithclassis zi andwhere classfrequency classes, the distribution, frequency Again, cumulative
k
11

1fgZz/Ek.=lf

there k size are where

of the limit the equal of frequencies to or below upper gives proportion thetotal the to from ith class,say i', and corresponds F(z). This mustbe distinguished the of employment distribution gives proportion total which the first-moment cumulative of limit theithclass, namely belowtheupper
Efi zi/Efi zi
1 11

it

since zi represents total the in employmenttheithclass. Thatis,ifwearestudying fi of between classesof employment, relevant size first the thesize distribution firms is distribution the distribution employment of moment amongthosesize classes. has distribution parameters which can estimate. example, we For The first moment mean distribution be estimated weighting by thearithmetic ofthefirst moment may' eachclassmark byemployment, instead frequencies, then of and by dividing the zi sumofweights, namely,
E zifi ZI/ Efi i=1 Ii=1
k /k

z =

Z,.

(4)

76

HART- Entropy and OtherMeasuresof Concentration

[Part1,

moment mean data,thearithmetic z and thefirst if Of course, we haveungrouped where and by be mean may estimated EN zi/N by i zi/Ei zi respectively, N = El and firm, Ely zi = Ek.lfizi is the of number firms, is thesizeoftheith is thetotal zi 3). p. (cf. in of totalemployment all thefirms thesample Prais,1961, 87,footnote Clearly, , (5) s2 so of of the where estimates variance thesize distribution firms, 2 >2. Thisis and a between weighted an unweighted difference caseofthegeneral simply special a whenthe in for mean, described, example, Yule and Kendall(1957,pp. 332-334), are and of deviations thevariables theweights equal,and means and thestandard and the between variable itsweight. correlation positive where there perfect is level at industry is theconcentration of measure concentration Themost common sales, (or employment output, of is the ratio which simply proportion an industry's an it units. Formally, is simply upper r by etc.) controlled the largest business of distributionz, namely, moment of quantile thefirst
Cr= 1-Fl(zr-7) = 1Z{ zf(z) dz (6) + ,l = (s2/2)

data, ungrouped case, or in thediscrete with


Cr = 1'r1

zZ/2zi).

IN

(7)

N, rank is rank1 andthesmallest given ifthe firm Alternatively, largest is given


Cr=
i=l

r IN r z I z/ = 9i,
JI =1 i=l

(8)

where

industry. to it because is lesslikely be affected concentration Economists Crto measure use with firms negligible small if in For of bychanges thenumber firms. example, several the firms to then enter employment orleavetheindustry, Cr is unlikely change; small on smalleffect the in belowrankr do notaffect numerator (8) and havea very the on to are firms unlikely haveanyeffect theintensity Sincethese small denominator. than, example, of of degree concentration for measure the ofcompetition,isa better Cr influenced be by or which thevariance Ginicoefficient wouldprobably significantly becomes the to case, in changes N. To takean extreme ifN fell 5 equalfirms, variance based of zero,F(z) = F1(z) and themeasure concentration on thewholesizedistriwouldbe an oligopoly. Yet bution wouldshowzero concentration. theindustry the and wouldreflect oligopolistic The Crmeasure, with r = 4, wouldbe 0-80, say upontheconfidentiality the of In nature theindustry. practice valueof r depends r In office. the UnitedKingdom is statistical Governument rulesof the relevant it 3, the generally though BoardofTradehas increased to 5 for1958-63. indifferent industries, concentration to it Problems when is desired compare arise wouldliketo economists of different or points time.Ideally, countries, at different before of comparisons. making on upper informationthe quantiles F&(z) havecomplete of facilitates comparisons concurve concentration foreachsample The cumulative curves cross,a summary frequently concentration But centration. sincedifferent
measureis required.

= ZI zj/iN

in zi and is the share of the ith firmin total employment the

1971]

HART -

Entropy OtherMeasuresof Concentration and

77

curve concentration valueofthecumulative One possible measure theaverage is

Br

_A{9

+(9A+92)+(9A+92+Y3)+

..+(91+92+...+9r)}

rA

1(r=,
lr = -NE

i+l)9i
(r-i+ 1) (z/i). (9)

since of curve, slopesoftheLorenz Thisis a type weighted of average theestimated with 2 is an estimator p4 (3), with diminish higher weights - i+ 1)/Nwhich (r for in as in Alternatively maybe regarded (9) valuesof i, i.e. with decreases size offirm. r by themeansizeofthelargest firms, eachfirm with weighted (r- i+ 1)/Ni 1 r
Br = -N (r-i+1)zi

(10)

or to affected theentry exitof by Thissummary measure unlikely be significantly is on havelittle influence Ni = ,Zi. small firms, because they sum is the An alternative of type simply cumulative ofthe9i,given measure this
by rBr= D where

D =Ed(r- i+ 1)Zi/ Ezzi.


i=l1I i=l

with curve,with r = N, D is readilycompared For the whole concentration is that the distribution: onlydifference moment z1= Ez,2Ezj, themeanofthefirst of for in D therank zi is substituted onezi inthenumerator i1. Thedisadvantage of D: in r ofthismeasure with = N is thatan increase N increases theuse ofrankin increases that evenonemoresmall firm thenumerator means E(r- i+ 1) zEzi. the the in is The samedrawback present another namely area under measure, by concentration given curve,

9 J= 9/2+ {(N- 1) +Y2/2} {(N+


=

22)92

+93/2} +

E(N-i)4I z9i+ i=1

since 9i= 1, = +N-Ei9, (12) Rosenbluth this withN. To overcome limitation, increases whichclearly (1961) concentration of the the suggests useofthereciprocal twice areaabovethecumulative curve.The areaaboveis given by K=N-J=
,i9A

(13)

is also and Rosenbluth's (1967), by measure, proposed Hall andTideman 2K 2 9l-fs1(14) a the and Butas noted Finkelstein Friedberg (1967)K is simply areaunder Lorenz by curve also holdsfortheusualLorenz abscissa.Their an with absolute curve proof

78

HART -

and Entropy OtherMeasuresof Concentration

[Part 1,

use the mean, others a ratioofa Niehans usesa weighted Whereas concentration. in of mean. In the context concentration the whole to weighted an unweighted to is enough use distriof where number firms large the or economy, in an industry they be can because are of these measures concentration superfluous bution theory, if For size of the derived from parameters theunderlying distribution. example, the lognormal, c2),then is a sizedistribution two-parameter A(p,
=

between and 2. Niehans(1958) uses 21,Adelman(1969) to z1 be reduced relationships (1945) uses (21/N)i to measure (1950) use 21/Nzand Hirschman and Herfindahl

in which eachfirm contributes to F(z) on theabscissa thetotalareaunder 1/N and no to there is curve (1/N)(Ei9- i) = K/N. Consequently, seems reason theLorenz the providing number or rather thantheGinicoefficient Lorenzmeasure use 1/2K offirms known. is a otf otf which economists proposed serles measures concentration may Other have

= exp(2 + 3&2/2) and zJ1Nz exp(a2)/N.

(15)

few industry a relatively with In thecontext concentration an individual in of to in to are but may firms, measures be useful, they likely be sensitive changes N. these N+ in with clearly decreases increases N, foriffirm 1 with The Herfindahl measure the it the employment enters industryincreases numerator byzN+ whereas ZN+l Ezi Sometimes, terms. thedenominator by (EZ,)2is increased zN+1 andthecross-product has that the creates falseimpression there beena sucha decrease butnotalways, in For firms theindustry. example, power theleading of in reduction thecompetitive to industry are likely have they concentrated if several a smallfirms enter highly giant firms, theconcentraand of power thefew on little no effect thecompetitive or the but measure necessarily and tion curve B,measures much same, theHerfindahl stay of are and falls.On theother hand, thenewentrants important havea share the if then firms, thedecrease that significantly reduces ofthegiant output which industry's But is measure notmisleading. in sucha case,theconcentration in theHerfindahl also to prudent use Br in (9), or the and theBr measures fall. Thusit seems curve to are they lesslikely be misleading N concentration for small casebecause curve, the of an point view. from economic 4. ENTROPY AND REDUNDANCY of is A morerecent to approach themeasurement concentrationbased on the Let shareof theithcompany in theory. theobserved concept information entropy are there N companies in total employmentj\, where be 9=ziz.
N
/N

(16)

is given of by of distributioncompanies employment by Theentropy measure this


H(9)
= i=1

9ilog9,

0 H(9) < logN.

(17)

equality occurswhen9i = 1/N for all i and H(9) = logN. Complete Complete inequalityoccurswhen yi = 1, and y; = 0 for all j # i so that H(y) = 0, using the

which varies a directly that convention yilog9i= 0 when = 0. To obtain measure 9i between or of we the with degree inequality, mayuse theredundancy thedifference

1971]

HART -

and Entropy OtherMeasuresof Concentration

79

theobserved and entropy itsmaximum value:


logN- H(y)
=

logN+ E9t log9i


i

= z 9q = E
i

log(N9t), since I9
i

(zi/E zi) log(Nz1/Nf)

E zilog(zi/zi)/Ez i ' P.

(18)

Thatis, theredundancy a weighted is of of average thelogarithms theestimated of Lorenz slopes the curve eachziwhere weights employment. difference at the are The between meanofthefirst the moment distribution
ff=Ez Ezi (19)

andtheredundancy (18) is simply inthenumerator (18) theterm (zi/2), in that of log the of its the deviation z from mean, of measuring logarithm therelative replaces term zi, measuring absolute is the deviation z from of zero. A further simplification since possible, (18) maybe written

E zilogzi/
( (z

( z log / *)
) - log . (20)

logzi/

The first is meanof thelogarithms z, where weights term a weighted of the are If werethefrequencieswouldbe thelogarithm the of it employment. theweights is geometric which readily with second the the of mean, compared term, logarithm thearithmetic ofz. mean It is possible useinformation toobtain measure concentration to a which of theory is simply difference the between of and of means logarithms thearithmetic geometric thedistributionz. The expected of as informationan indirect of is message defined
I(y: x) = E yilog(yi/xi),
i=1
N

(21)

where is theposterior of probability zi,estimated (16),andwhere is theprior by yi xi of the as on probability zi. We mayexpress prior probability 1/N, theassumption that should we eachoftheN companies have1/N thetotal to of expect employment, i.e. complete Thus(21) maybe estimated: equality.
1(y: x) =
N

i=1

yq N9j, log

(22)

which theredundancy zi in(18). Similarly, maydefine expected is of the we informationcontent theindirect of which transforms prior the message probabilities into xi theposterior probabilities as yi
I(x: y) = E xi log(xi/yi),
i
N

(23)

80

HART -

and Entropy OtherMeasuresof Concentration

[Part1,

which estimated is by , (IIN) log


N
N

II

E (1/N)log(z/z*) i = logz-(1/N) E logzi


i N

= logAM-log GM.

Clearly, information this theory measure concentration be obtained of may directly from frequency the distribution z usingclassical of statistical analysis.Indeed, as Theil(1967)reminds forthelognormal us, distribution which approximates many u2 size distributions firms, y) reduces cr2/2 of I(x: to where is thevariance the of logarithms firms of sizes, since arithmetic is exp + JU2) andthegeometric the mean (/ meanis exp(p), where is themeanofthelogarithms z. There of seems little point / in replacing which beenusedextensively measure c2, has to business concentration, by an entropy measure which becomes Theil(1967,p. 124) argues cr2/2. thatthe variance thelogarithmsan inconvenient of is of measure concentration because is it thesecondmoment of extension thegeometric meanwhereas totalemployment is to directly related thearithmetic mean, namely Thisobjection Nz. losesitspoint if in the distribution known. For example, a lognormal is distribution, total the is to employmentdirectly related thegeometric mean. Total employment be may c2 the obtained multiplying geometric by meanbyNexp(Ja2), where iS thevariance ofthelogarithms. 5. NUMBERS-EQUIVALENTS and the Stigler (1968)andFinkelstein Friedberg (1967)havetransformed concenin 3 the of of tration measures Sections and4 into equivalent number firms equalsize, which yet is another ofmeasuring business concentration. Herfindahl The way index between when N firms thesamesize,and 1,when jthfirm all have the has varies 1/N, alltheemployment zi,i #j,is zerofor i. Thusthereciprocal theHerfindahl all of and of as of index, = Nf/zl, be regarded theequivalent NE number firms equal size. may and It variesbetween reflecting N, complete competition, 1, reflecting complete this is the form simply actualnumber of monopoly. However, indexin reciprocal of moment distrifirms, multiplied theratio themean to themeanofthefirst N, by z in if. that bution, Equation shows (5) 2<z2, except thecase ofcomplete monopoly on of when2 = 2f,thusN is reduced a fraction which by depends thedispersion of of firms' sizes. The equivalent number firms equal sizebasedon theHerfindahl ofvariation, this and v measure can index simply is N/(v2 1),where is thecoefficient + from be obtained directly N andS2/22.For thelognormal distribution, NE+ U2= log of logarithms of minus variance the the the of number firms logN,sothat logarithmthe is of a knowledge N (which directly relevant to giveslogNE. In anydistribution, is in measures to concentration,anycase) and of S2/22 (which inequality) sufficient from entropy the of number equal firms The equivalent mayalso be obtained in be would l/9ifirms theindustry. in (17). If all firms thesamesizeas zi there were
calculateNE.

1971]

HART -

and Entropy OtherMeasuresof Concentration

81

A measure thenumber equivalent of of firms, basedon each9fis given by Ne,


Ne
i=l

flI(1 /9s)( 0,
N

KY

(24)

or = logNe= E y log(1/9i) H(U).


i=1

(25)

Thus theredundancy (18) maybe viewed thelogarithm theratioof the in as of actualnumber firms theequivalent of to number equalfirms. of this However, can be also obtained directly theparameters theunderlying from of distributionfirms. of For example, Theil(1967,p. 97) shows thattheredundancy reduces U2/2 the to in caseofthelognormal distribution, logN- (U2/2) and gives logarithm Ne. Note the of also that thelognormal Ne/NE exp(u2/2) in = case the gives relationship between the twomeasures equivalent of number firms. of 6. COMPARISONS OFMEASURES Theentropy redundancy or measure concentration be compared the of may with more traditional measures concentration thelarge of for and group thesmallgroup cases. In theformer thenumber firms large case of is enough use statistical to distribution theory. Froman economic of point view, discussion business the of concentration thiscontext associated in is with comparison growth a of rates largeand of small firms. Generally industry boundaries ignored concentration are and ismeasured for whole the corporate sector thewhole manufacturing. large or of This group case must distinguished thecase where number firms too smallto use be from the of is statistical distribution theory, where industrial boundaries important where are and thedegree monopoly oligopoly themajor of or is interest economists. of There no is reason the why samemeasure concentration of should appropriate bothcases. be in Information sizesofcompanies thelarge on in case tends be published group to in theform frequency of distributions. (1967,pp. 99, 128-134) Theil estimates the of redundancy income distributions theassumption frequencies theith on that in classearnthearithmetic mid-point income thatclass,Zi. For opensize-classes, of he uses a Paretodistribution estimate mid-point. arithmetic to the The meanis estimated, and totalincome given NY,and equations and (18) are used 2, is by (16) A to estimate redundancy, Similar the A. techniques usedto estimate forthe were distributions companies profit theUnitedKingdom 1949and 1964.t of in by for Thefirst moment distributions given itwaspossible calculate arithmetic are and to the mean in profit eachsize-class, the 2f, posterior was probability estimated Y' = Z/Ezi as andthefrequencies theithclasswere in assumed havethesamey9.Theestimated to redundancies given Table1 innatural are in logarithms that so they be compared may with variance thelogarithmszi. Thelatter estimated graphs the the of of were from of cumulative distributions cumulative moment frequency and first distributions drawn on logarithmic A= probability paper. In thelognormal distribution=JU2. For 1949 with'2 = 8-7. Bothmeasures indicate concentration that increased 1949-64, the but is measure superior A forthefollowing to reasons.
a

2A = 5-46whichcompareswell with 2 = 5-3,through 1964 2A = 7415compared for

the because it excludesvariationof sizes within t This procedureunderestimates entropy cf. income-brackets Theil,(1967),pp. 99, 128-134.

82

HART

- Entropy OtherMeasuresof Concentration and

[Part1,

'2 First, is easier compute thanR. Graphic it to methods available, so are and aretheshort-cut methods calculation of grouping intoclassintervals data equal to in unity logarithms thebase 2. Moreover, is necessary knowEz3, which to it to is notalways available, order compute Secondly, efficiencies various in to R. the ofthe TABLE 1

Size distribution companies trading of by assessed underSchedule D, United profit Kingdom, 1949 and 1964
1949 Upper limit of Numbers Profit profit (?) ('000) (?m) 250 500 1,000 1,500 2,000 3,000 4,000 5,000 10,000 15,000 20,000 25,000 30,000 40,000 50,000 75,000 100,000 200,000 1,000,000 > 1.000,000 62.8 168 21.1 120 7.9 109 71 4.7 122 52 3.1 1.9 1P5 2.1 1P3 1.9 1.0 1*4 1.1 0.2 176 3 R (loge)
&2

1964 Numbers Profit ('000) (?m) 1705 166 193 114 85 107 6.7 49 134 62 37 25 18 24 1-6 23 13 2.1 1.9 0*4 288.2
8.7

6-0 59 143 141 136 260 24 2 208 84*1 62.1 52-1 42*4 409 725 56-5 110.6 83.8 196.4 4248 466*7 18176 2 729

4.3 5.8 13.3 13.9 147 259 23*3 21V8 956 75.2 63.7 55 7 50-6 836 69.8 1425 110 1 298.8 7730 1478.2 34198 3 576

(loge)

5.3

Sources:BoardofInlandRevenue, Annual Reports. 1949relates profits to assessed1950/1. 1964relates profits to assessed1965/6.

' of methods estimating from and data is grouped ungrouped areknown.Butlittle known aboutthedegree reliability estimates fA of of of from data. Thirdly, grouped wecan apply standard the F-test thesignificancethedifference of the of between two because observations sampled the are from which approxiare variances, populations normal after the mately transformation. of logarithmic Fourthly, decomposition r2 is elegant wellknown, that are able to study effects births, and so we the of deaths and amalgamations companies concentration. of on it to Fifthly,is possible use the modeland attribute in Galton regression changes aJ2to thedifferent average growth

1971]

HART - Entropy and OtherMeasuresof Concentration

83

of ratesof firms different (measured bl) and to thesize mobility firms of sizes by (measured r),as explained HartandPrais(1956),since by in
r/OAr-_ =

b2/r2,

(26)

coefficient. where is theregression logzi(t)on logzi(t- 1) andris thecorrelation b1 of For all these 0r2 to case. reasons, is superior A in thelarge group
TABLE

Kingdom Estimated redundancy Br-small groupcase, United and


Industry
Distilling

Date

N
7 9 6 5 13 9 24 27

mrnin

Amax

Br
96 00 94 00 88x17 90 50 88x50 89 83 70 33 73-33

Hmin
0 6861 0-9136 1x5447 1 4534 1-5073 1-4803 2x7392 2 5497

Hmax
0 6861 0 9310 1x5447 1 4534 1-5471 1-4977 2 9233 2 7616

1952 1957 Cement 1952 1957 Tobacco 1952 1957 Cocoa, chocolate 1952 and sugar confectionery 1957 Units: R and H in log2.

2-1213 2-2389 1 0403 0-8685 2.1533 1-6722 1x6617 1P9933

2x1213 2-2563 1-0403 0-8685 2x1931 1x6896 1x8458 2.2052

or Table 2 gives estimates theredundancy, and ofBr forthesmallgroup of R, of case. A range A is given when 9i oligopoly individual observations(i = 8,9, ..., N) werenot available. Individual of observations 9 (i = 1,2,..., 6) are givenin the But Appendix whenN = 7 it was possible obtain by subtraction. when and to 97 the 98 equalto give minimum the redundancy. Second, 97= I6, equalled remainthat
ing proportion assetsand 9i (i = 9, 10,..., N) equalled zero, to givethemaximum of of These estimates be compared Br,as inequation estimate redundancy. with (9). may A is between andBr,andinthedistilling, cement tobacco and There no correlation N> 7, two extreme were made. First,that all 9i (i = 7, 8, ..., N) were assumptions

industries of in 1952-57 indicated are by different directions movement concentration A and Br. An inspection thedata in theAppendix of showsthatin thedistilling the curve point, industry concentration for 1952is above thatfor 1957at every in indicates decrease concentration. Br measure a suggests decrease, a which The in measures an butbothredundancy suggest increase concentration. In thecement the curve industry concentration for1952is belowthatfor1957, indicating business that concentration increased thisperiod.The Br measure over in to a showsan increase shows decrease con88417 90 50,buttheA measure from industry, Br and A the centration. thecocoa,chocolate sugar In and confectionery are and the curve measures consistent eachother with concentration datain the with that The inthe and industries suggest Appendix. inconsistencies distilling thecement A is inferior Br as a measure concentration. mainreason this thatA for is to of The with of increases logN so thata larger tends produce higher N to a degree concenis in which contrary thestandard to ofan economic tration, interpretation increase N the Thusinthesmall group inthesmall group caseas reducing degree monopoly. of case Bris more than appropriate A.

84

HART

- Entropy OtherMeasuresof Concentration and

[Part 1,

as measure concentration of is more However, use oftheentropy an inverse the on appropriate A, becauseit is less dependent N and thusshowsthesame than of as of direction movement thedegree concentration doesBr. Thisconsistency of but to between H(y) and Br is comforting Br is nevertheless preferable H(y). First data are available, is easierto calculate it becauseevenwhenall therequired Br, it to mean. Secondly, because is unnecessary havea range is a which merely simple at of estimates Br forany one industry any one time. H(y) is indeterminate, of are it the sizesof smallfirms generally because relative unknown, though is known of a small that collectively contribute they proportion an industry's output employor in of overtime havelittle ment.Thirdly becausechanges thenumber smallfirms and for of effect theshares thetopfirms it is undesirable a measure concenon of The suffers disadvantage, this tration be influenced all thefirms. H(y) measure to by it to Br whereas doesnot. For these reasons, is preferable use Br in thesmall group case. and is the of measures business The general conclusion that redundancy entropy For use concentration inappropriate. thelarge are group case,weshould a parameter of For ofthedistribution as a2, thevariance thelogarithms. thesmallgroup such cumulative concentration orthemeanofits curve casewe should thetraditional use r first values, denoted Br by
ACKNOWLEDGEMENTS

in follow from enquiry mergers an into Theresearch results reported this paper and at National Institute Economic conducted the of business concentration isbeing which was on and SocialResearch. Thisresearch project launched a grant from Social the assisted a grant from Departthe Science Research Council was subsequently and by ment TradeandIndustry. of
REFERENCES
ADELMAN,

AITCHISON,

Rev. Econ. Statist., 99-101. 51, Distribution. Cambridge: Cambridge J. and BROWN, J. A. C. (1957). TheLognormal
UniversityPress. M. 0. and FRIEDBERG, R. M. (1967). The application of an entropy theory of concentrationto the Clayton Act. Yale Law J., 76, 677-721.

M. A. (1969). Comment on the H concentration measure as a numbers equivalent.

FINKELSTEIN,

GIBRAT,
HALL,

162-168

M. and

Paris: Sirey. R. (1931). Les Inegalites Economiques.


TIDEMAN,

N. (1967). Measures of concentration. J. Amer. Statist. Assoc., 62,

HART,

0. C. (1950). Concentration in the steel industry. Ph.D. dissertation, Columbia University. HEXTER, J. L. and SNOW, J. W. (1970). An entropymeasure of relative aggregate concentration.
HERFINDAHL,

J. R. Statist. Soc. A, 119,2, 150-181.

P. E. and PRAIS,S. J. (1956). The analysis of business concentration:a statisticalapproach.

South.Econ.J.,36, 239-243.
W. and
PASCHEN,

Nr. Statist. Inform. Europ. Wirtschafts., 3, 53-61. and Structure Foreign Trade. Berkeley: HIRSCHMAN, A. O. (1945). NationalPower the of
of California Press. HOROWITZ, I. (1964). A

HILDENBRAND,

H. (1964). Ein axiomatisch begrundetes Konzentrationsmass.


University

HOROWITZ, A. and I. (1968). Entropy,Markov processes and competitionin the brewingindustry.

127,234-237.

note

on the Hart-Prais measure of concentration. J. R. Statist. Soc. A,

J. Indust. Econ.,16, 196-211. M. G. Griffin.


and STUART,

KENDALL,

M. G. (1956). Discussion on Hart and Prais (1956), p. 185.

KENDALL,

Vol. I. London: A. (1958). The Advanced Theory Statistics, of

1971]

HART Entropy OtherMeasuresof Concentration and

85

PRAIS,S. J.(1961). A methodological afterthought business on concentration. Econ.Statist., Rev.

NIEHANS, J.(1950).

An index thesizeofindustrial of establishments. Econ.Papers, 122-132. Int. 8,

G. (1961). Round Table-Gesprach ilberMessungder industriellen Konzentration. In Die Konzentration der Wirtschaft in (Amndt, p. 391. Berlin:Duncker& Humbolt. ed.), STIGLER, J.(1968). TheOrganisation Industry. G. of Homewood, Ill.: Irwin. THEIL, H. (1967). Economics and Information Theory.Amsterdam: North-Holland. YULE, G. U. and KENDALL, M. G. (1957). An Introduction theTheory Statistics. to of London: Griffin.
ROSENBLUTH,

43, 87.

APPENDIX Industry Year 1 1952 1957 1952 Cement 1957 1952 Tobacco 1957 chocolate and 1952 Cocoa, sugarconfectionery 1957 Distilling 90 86 60 68 73 70 43 49 Cumulative concentration ratios (%) forthesix largest business units 2 93 91 83 85 82 84 60 64 3 96 94 90 93 89 92 74 75 4 98 96 97 98 94 96 79 82 5 99 98 99 99 96 98 82 84 6 100 99 100 100 97 99 84 86 N 7 9 6 5 13 9 24 27

in Sources: 1952HART, P. E. (1958). Concentration selected industries. Scot.J.Pol. Econ.,5, 185-201. 1957HART, P. E. (1961). Concentration itsmeasurement theUnited and in Kingdom. In Die Konzentration der Wirtschaft in (Arndt,ed.), pp. 653-674, Berlin: Duncker& Humbolt. rules to Note: The BoardofTradeconfidentiality makeitimpossible use CensusofProduction data to estimate cumulative concentration curves.The above data werecompiled from published accounts and zi is measured nettangible assets. Unfortunately, Board of Trade the company by the no longerstandardizes accountsof companieswithless than ?0s5 millionassets so conwith curves recent for centration yearsare notcomparable thosefor1957.

You might also like