Construction of an index by maximization of the sum of its absolute correlation coefficients with the constituent variables

Transcription

1 Construction of an index by axiization of the su of its absolute correlation coefficients with the constituent variables SK Mishra Departent of Econoics North-Eastern Hill University Shillong (India) I. Introduction: On any occasions we need to construct an index that represents a nuber of variables (indicators). Cost of living index, general price index, huan developent index, index of level of developent, etc are soe of the exaples that are constructed by a weighted (linear) aggregation of a host of variables. The general forula of construction of such an index (OECD, 003) ay be given as I = w x w x + w x w x ; i = 1,,..., n i i 1 i1 i i = 1 where w is the weight assigned to the observations of x = ( x1, x,..., xn ). The weights, w ( w1, w,..., w ) the variables x ; 1,,..., th variable, x and reains constant over all =, are deterined by the iportance assigned to =. The criterion on which iportance of a variable (vis-à-vis other variables) is deterined ay be varied and usually has its own logic (Munda and Nardo, 005). For exaple, in constructing a cost of living index iportance of a coodity is deterined by the proportion of consuption expenditure allocated to that particular coodity and in constructing the huan developent index variables such as literacy, life expectancy or incoe are weighted according to the iportance assigned to the in accordance with their perceived roles in deterining huan developent status. In any cases, however, the analyst does not have any preferred eans or logic to deterine the relative iportance of different variables. In such cases, weights are assigned atheatically. One of the ethods to deterine such atheatical weights is the Principal Coponents analysis (McCracken, 000). In the Principal Coponents analysis (Kendall & Stuart, 1968, pp ) weights are deterined such that the su of the squared correlation coefficients of the index with the constituent variables (used to construct the index) is axiized. In other words, weights in I = w x are deterined such that r ( I, x ) is axiized. Here = 1 r( I, x ) is the coefficient of correlation between the index I and the variable x. The Principal Coponents analysis is a very well established statistical ethod that has excellent atheatical properties. Fro x = ( x1, x,..., x ) one ay obtain (or fewer) indices that are orthogonal with each other. These indices together explain cent percent variation in the original variables x = ( x1, x,..., x ). Moreover, the first Principal Coponent (often used to ake a single index) explains the largest proportion of variation in the variables x = ( x1, x,..., x ).

2 II. Soe Practical Probles with the Principal Coponents Analysis: Although the Principal Coponents analysis has excellent atheatical properties, one ay face soe difficulties in using it if one desires to construct a single index of the variables that are not very highly correlated aong theselves. The ethod has a tendency to pick up the subset of highly correlated variables to ake the first coponent, assign arginal weights to relatively poorly correlated subset of variables and/or relegate the latter subset to construction of the subsequent principal coponents. Now if one has to construct a single index, such an index underines the poorly correlated set of variables. As a result, practically speaking, the index so constructed is the weighted aggregation of only the preferred (highly correlated) set of variables. In this sense, the index so constructed is elitist in nature that has a preference to the highly correlated subset over the poorly correlated subset of variables. Further, since there is no dependable ethod available to obtain a coposite index by erging two or ore principal coponents, the deferred set of variables never finds its representation in the further analysis. III. A Wider View of Constructing an Index: possibilities of axiizing = 1 r( I, x ) L 1 / L Let us now investigate into the to obtain weights to construct I = wx. This is only a Minkowsky generalization of axiization of r ( I, x ) or (equivalently) 1 / r( I, x ). It can be shown that as ( ) = 1 = 1 L, the index becoes ore and ore egalitarian with an ever-stronger tendency to assign weights such that all or ost of the variables are equally correlated with the index. In so doing, it axiizes the inial correlation of the index with its constituent variables or in other words it gives us the axiin index. However, for L = 1 the index is inclusive in nature that assigns reasonable (although saller) weights to the ebers of less correlated subset of variables, but has no tendency to underine the less correlated variables and their representation. This property of the index obtained by axiizing r( I, x ) or axiizing the inial correlation is attractive and useful. The obective of this paper is to illustrate this fact. IV. An Experient: We have conducted (liited) experients on constructing indices by axiizing (a) su of squared correlation, which is the standard Principal Coponents analysis, (b) axiin correlation, and (c) axiizing the su of absolute correlations. For sake of identification, we would call the I-, I-M and I-1 respectively. The experients have been conducted for (i) highly correlated variables and (ii) poorly correlated variables. V. The Method of Optiization: The ethod of constructing indices by the Principal Coponents is available in any software packages such as STATISTICA or SPSS. However, the ethod to construct indices by axiin correlation or axiization of the su of absolute correlation is not available. We have obtained all indices (I-, I-M and I-1) by solving ax = 1 r( I, x ) L 1 / L such that I = 1 = wx where w are the decision variables. It

3 is an intricate non-linear optiization proble. Any powerful non-linear prograing software ay possibly be used for optiization (see Kuester and Mize, 1973 for classical ethods and FORTRAN progras). However, we have used the Differential Evolution (DE) ethod of Global Optiization (which is in the broader faily of the Genetic algoriths). The optiization ay also be done by the Particle Swar ethod often used in Artificial Intelligence (see Mishra, 006). We have found that the Repulsive Particle Swar (RPS) ethod perfors as effectively as the Differential Evolution ethod. We have not presented the results of the RPS optiization to avoid duplication of results. The FORTRAN codes of DE or RPS ay be obtained fro the author on request. VI. Findings: The results of our experients are presented in tables 1 through -c. It is evident fro the correlation atrices associated with tables 1 through -c that in case of highly correlated variables [Table-1(i)], all the three ethods have a tendency to yield indices that represent all the constituent variables. However, when the variables are poorly correlated [Tables -a(i) onwards], the principal coponent index (I-) has a tendency to underine soe variables by poorly correlating with the (and thus not representing the, or relegating the to be represented by the subsequent principal coponents). On the contrary, I-M and I-1 assign reasonable weights to those variables and thus includes the. Nevertheless, it ay be noted that I-1 and I-M pay the cost in ters of the explained variance [su of squared r(i, x )] in the constituent variables. VII. Concluding Rearks: In this exercise we have shown that the principal coponent indices are elitist and they have a tendency to underine the iportance of poorly correlated variables. On the other hand, I-1 is ore inclusive, and has a tendency to represent even the poorly correlated variables. The I-M indices are egalitarian in nature. It would depend on the analyst whether he is interested in egalitarian, inclusive or elitist ethod of constructing indices when the constituent variables are not very highly correlated aong theselves. This paper has opened up the option to choose the ethod of constructing a desired type of index. References Kendall, MG and Stuart, A (1968): The Advanced Theory of Statistics, Charles Griffin & Co. London, vol. 3. Kuester, J.L. and Mize, J.H. (1973): Optiization Techniques with Fortran, McGraw-Hill Book Co. New York. McCracken, K (000): Soe Coents on the Seifa96 Indexes, Paper presented in the 10 th Biennial Conference of the Australian Population Association, Melbourne, Australia. Mishra, SK (006): Global Optiization by Differential Evolution and Particle Swar Methods: Evaluation on Soe Benchark Functions. SSRN Munda, G. and Nardo, M (005) : Constructing Consistent Coposite Indicators: The Issue of Weights, EUR 1834 EN, Institute for the Protection and Security of the citizen, European Coission, Luxebourg. OECD (003) Coposite Indicators of Country Perforance: A Critical Assessent, DST/IND(003)5, Paris. 3

4 Table-1(i): Construction of Indices with Highly Correlated Variables x 1 x x 3 x 4 x 5 I- I-M I Coefficient of correlation of x with the Index SAR SSR Index I I-M I-1 SAR=Su of absolute correlation coefficients; SSR=Su of squared correlation coefficients Table-1(ii): Correlation aong Variables and Indices [Ref. Table-1(i)] 4

5 Table--a(i): Construction of Indices with Poorly Correlated Variables x 1 x x 3 x 4 x 5 I- I-M I Coefficient of correlation of x with the Index SAR SSR Index I I-M I-1 SAR=Su of absolute correlation coefficients; SSR=Su of squared correlation coefficients Table--a(ii): Correlation aong Variables and Indices [Ref. Table--a(i)] 5

6 Table--b(i): Construction of Indices with Poorly Correlated Variables x 1 x x 3 x 4 x 5 I- I-M I Coefficient of correlation of x with the Index SAR SSR Index I I-M I-1 SAR=Su of absolute correlation coefficients; SSR=Su of squared correlation coefficients Table--b(ii): Correlation aong Variables and Indices [Ref. Table--b(i)] 6

7 Table--c(i): Construction of Indices with Poorly Correlated Variables x 1 x x 3 x 4 x 5 I- I-M I Coefficient of correlation of x with the Index SAR SSR Index I I-M I-1 SAR=Su of absolute correlation coefficients; SSR=Su of squared correlation coefficients Table--c(ii): Correlation aong Variables and Indices [Ref. Table--c(i)] 7