You are on page 1of 4

KENDALLS CONCORDANCE ANALYSIS Coefficient of Concordance The nature, degree and direction of relationship between 2 or more cardinally measured

variables may be evaluated by regression and correlation analysis. The degree of association between two or more ordinal attributes may examined by the application of Yules coefficient of association, or coefficient of contingency in case of only nominally measurable properties. The degree and direction of relationship between 2 or more ordinally measured attributes may assessed by the use of total or partial rank correlation analysis. However, as the sample si e and number of traits to be studied become larger and larger, this may become tedious costly and les efficient if many of the traits do not display strong relationship. !esides, the analyst may be interested in identifying and grouping separate traits"attributes in different categories for which only the degree of concordance or discordance may be re#uired. $endall has developed concordance analysis for such purposes. Total %ank &orrelation and partial rank correlation analysis relates to two sets of ranking among, say, n items. &oefficient of concordance pertains to the measurement of degree of relationship/agreement/disagreement of n items on the basis of multiple, say, k criteria of ranking. 'or this purpose, $endalls coefficient of concordance may be used. The ps and rs express the degree of association between the two given attributes, whereas ( shows the degree of association or agreement between the ranks based on k)2 criteria. *aturally, the concordance analysis is useful if we study a cluster of large number of variables and sample si e of the items to be ranked is also large. +n obvious procedure may be to work out K c 2 rank correlations and then work out their average to assess the overall degree of association. Two difficulties may be encountered, i- if some attribute shows positive association with some and it has negative association with others, howsoever highly significant the association may be sum of the negative coefficients will dilute the average when added to the sum of positive coefficients. This may conceal rather than reveal the overall association. .uch results may furnish misleading inference/s-0 and ii- more the number of ranking criteria, more cumbersome and tedious becomes the correlation analysis for the reason that the number of coefficients to be estimated tends to increase with the number of ranking criteria and the calculations increase with a rise in the number of items to be ranked. +s against this, estimation of ( is simple and straight forward. !esides, ( has a definite linear relationship with .pearmans rank correlation coefficient. $endall /1234, p.41- proved that
rsav = KW 1 K 1

555555555.. /1-

(here rsav is the average of all the s between K c 2 pairs of ranking, $ are the ranking criteria, and ( is the coefficient of concordance. !ut how do we determine the coefficient of concordance6

Alternative Method 7ike other methods, concordance analysis is also based on certain assumptions which are stated hereunder. 8t is assumed that i- the ranking criteria are totally independent, and therefore, have no agreement"relationship, that is, ranks are randomly distributed0 and ii- if ranks were perfectly related, we should examine how the raw data look like. 9n the twin assumptions, the coefficient of concordance will be the index of divergence of actual data from the case of perfect agreement between the ranking criteria. The above procedure may be illustrated by the following example of the rank scores awarded by : selectors to ; applicants for selection"appointment, .elector"cand + ! & 8ndividual %ank .um =%> a 1 1 ; 4 b ; < : 13 c : ; 2 11 d 2 3 < 11 e < 2 3 11 f 3 : 1 4 .um 21 21 21 ;:

where ? = 1, 2 5.* 8f all three selectors were in perfect agreement with each other, the rank scores would have been !est candidate = 1@1@1=:=$0 *umber of ranking criteria selectors in this case .econd !est candidate =2@2@2=2A:=2$0 Third best candidate =:@:@:=:A:=:$0 'ourth best candidate =3@3@3=3A:=3$0 'ifth best candidate =<@<@<=<A:=<$0 .ixth best candidate =;@;@;=;A:=;$0 .o, perfect agreement among selectors will yield the following sum :@;@2@12@1<@14=;:, though not necessarily in the same order. Benerali ation of the illustration to * items for ranking /sample- according $ criteria for summing up all the ranks for the case of perfect agreement will yield the following,
JK =$@2$@:$@3$@<$@;$@5*$=$ /1@2@:@5.@*5555.. /2 JK is the sum of the ranks of all items according to all criteria. (here

The case of total dis agreement will yield approximately e#ual sums as % >s. %1 =%2=5%> = %* Thus, the degree of agreement is shown by the variance among the * sums of ranks. The coefficient of concordance, ( is the function of this variance,

W=

S 1 K 2 /N : N 12

555555555.. /:-

(here . is the sum of s#uares of the observed deviations from the mean of %>, that is,
R S = R N
2

5555555555.. /3-

$ = *umber of ranking criteria0 * = *umber of items"ob>ects to be ranked according to every criterin0

1 K 2 / N : N - = Caximum possible sum of s#uared deviations which is expected to occur in 12


case of perfect agreement among $ ranking criteria. 8n the example cited above, the sum of ranks, .um = 4@13@11@11@11@11@4=;: Cean = ;:";=1D.<=

. = /4E1D.<-2@/13E1D.<-2@/11E1D.<-2@/11E1D.<-2@/11E1D.<-2@/4E1D.<-2 =2<.< .ubstituting in relation 2- we get


W = 2<.< 2<.< A 12 = 2 : 1 " 12/: -/; ;- 2 A ;/:; 1-

=D.1;2 The observed degree of agreement among : criteria"selectors is very low. +lternatively, we could proceed in this case as follows, /sx.y-, /sy. -, / x.y%sxy=D.:10 rsy =D.<3, r xy=ED.<30 The average of these : .pearmans coefficients is given below,

rsav =

D.:1 D.<3 D.<3 + D.: 1.D4 = = 2; : : KW 1 : A D.1;1 1 = K 1 2 D.34: 1 D.<1F = = D.2; 2 2

9r
rsav = rsav =

This is approximately the same as above. Test of Significance 'or * ranging from : to F"2D, if the calculated value of . is greater than table value, the null hypothesis, of no association or concordance may be re>ected. Gxample, .=2<.<, for $=:, *=;, the table value is 1D:.2 at D.D< probability level. Ho of no association is accepted. 'or *)F, ( is distributed as 2 with df=*E1 ,
2 =
S 1 K N / N + 112

=$/*E1-( The calculated value of 2 may be compared with the table value for the given degree of freedom. 9ne difference distinguishes the averaging of partial rank correlation coefficients from the concordance analysis and it is that the average of partial rank correlation coefficients, like the correlation coefficients is bound to fall in the range of E1 and @1 since for individual coefficients we have
1 r"y .! + 1

55555555 /<-

+s against this, the range of the calculated value is defined by


D W +1

55555555 /;-

8f .=D, (=D it shall be the case when the variance between actual and perfect consonance among ranking criteria converges towards ero. 8f there is perfect divergence between the criteria, the numerator and denominator shall be e#ual and the s#uaring of the differences ensures the positivist of the coefficient.Gxample, pp. 2:2E:: .iegels !ook

You might also like