• Tidak ada hasil yang ditemukan

5.2 Data analysis

5.2.5 Econometric issues

113

where ๐ถ๐ฟ๐‘‡๐ด๐‘–๐‘ก is current liabilities to total assets, ๐‘ƒ๐บ๐‘…๐‘‚๐‘Š๐‘‡๐ป๐‘–๐‘กand ๐‘๐บ๐‘…๐‘‚๐‘Š๐‘‡๐ป๐‘–๐‘ก represent positive and negative sales growth respectively, ๐‘†๐ผ๐‘๐ธ๐‘–๐‘ก is the natural logarithm of market capitalisation, a proxy for firm size, ๐ผ๐‘–๐‘ก is fixed investment during the year t deflated by total assets, ๐‘‚๐ถ๐น๐‘‡๐ด๐‘–๐‘ก is operating cash flows to total assets; ๐ฟ๐ธ๐‘‰๐ธ๐‘…๐ด๐บ๐ธ๐‘–๐‘ก is the amount of debt employed by the firm and is deflated by total assets; ๐‘…๐บ๐ท๐‘ƒ๐‘–๐‘ก is the Real GDP growth rate, ๐‘€๐พ๐‘‡๐‘ƒ๐‘‚๐‘Š๐ธ๐‘…๐‘–๐‘ก is the market power of the firm and ๐œ‚๐‘– represents unobservable heterogeneity, ๐œ†๐‘ก are the time dummy variables and ๐œ€๐‘–๐‘ก is the error term.

The study repeated the estimation of Equation (8) using the disaggregated approach the working capital finance sources, ๐ถ๐ฟ๐‘‡๐ด๐‘–๐‘ก comprising accounts payable, short-term debt and accruals.

114

the error term. This correlation does not disappear when ๐‘ in the sample gets larger or ๐‘‡ increases (Bond, 2002).

GMM in first differences was considered superior to the alternative approach of estimating Equation (5) the fixed-effects model, the least-squares dummy variables (LSDV). Although the Fixed Effects estimator, eliminates the firm-specific effects ๐œ‚๐‘– through the within transformation, it does not eliminate bias. The transformed lagged dependent variable ๐‘ฆ๐‘–,๐‘กโˆ’1โˆ’ ๐‘ฆ๐‘–.โˆ’1 will still be correlated with (๐œ€๐‘–๐‘กโˆ’ ๐œ€ฬ…๐‘–.); where ๐‘ฆ๐‘–.โˆ’1 = โˆ‘๐‘‡๐‘–=2๐‘ฆ๐‘–.โˆ’1๐‘‡ โˆ’ 1, even if the random disturbances are not serially correlated because ๐‘ฆ๐‘–,๐‘กโˆ’1 is correlated with ๐œ€ฬ…๐‘–. by construction. ๐œ€๐‘–๐‘ก is correlated with ๐‘ฆฬ…๐‘–. because the latter contains ๐‘ฆ๐‘–๐‘ก. The fixed effect estimator produces biased but consistent estimates when ๐‘‡ tends to infinity and not when ๐‘ tends to infinity. This is known as the dynamic panel bias or the Nickell bias (Nickell, 1981). While the LSDV fails to deal with the problem of ๐ธ(โˆ†๐‘ฆ๐‘–,๐‘กโˆ’1โˆ†๐œ€๐‘–,๐‘ก) โ‰  0, Generalised Method of Moments takes care of this problem by using the lagged dependent variable (๐‘ฆ๐‘–,๐‘กโˆ’๐‘  ๐‘“๐‘œ๐‘Ÿ ๐‘  โ‰ฅ 2) in level as instruments.

Using OLS regression on first differenced equations produces biased and inconsistent estimates of the parameters because (๐‘ฆ๐‘–,๐‘กโˆ’1โˆ’ ๐‘ฆ๐‘–,๐‘กโˆ’2) and (๐œ€๐‘–๐‘กโˆ’ ๐œ€๐‘–,๐‘กโˆ’1) are correlated through the terms ๐‘ฆ๐‘–,๐‘กโˆ’1 and ๐œ€๐‘–,๐‘กโˆ’1. The fixed effects estimator fails to produce consistent estimates when ๐‘ tends to infinity and ๐‘‡ is fixed. GMM in first differences produces consistent estimates because it was designed for ๐‘ tends to infinity and ๐‘‡ is fixed; that is, small-T and large-N panels.

The Instrumental Variable (IV) estimator as suggested by Anderson and Hsiao (1981), produces consistent and efficient estimates in a dynamic panel. The IV estimator takes the first differenced equation and finds a set of variables, the instruments, to apply the instrumental variable estimator. Instruments are used to eliminate the correlation between the regressors and the disturbances because they must be correlated with the regressors but uncorrelated with the disturbances. In this case, since (๐‘ฆ๐‘–,๐‘กโˆ’1โˆ’ ๐‘ฆ๐‘–,๐‘กโˆ’2) and (๐œ€๐‘–๐‘กโˆ’ ๐œ€๐‘–,๐‘กโˆ’1) are correlated, (๐‘ฆ๐‘–,๐‘กโˆ’2) or (๐‘ฆ๐‘–,๐‘กโˆ’2โˆ’ ๐‘ฆ๐‘–,๐‘กโˆ’3) are used as an instrument for (๐‘ฆ๐‘–,๐‘กโˆ’1โˆ’ ๐‘ฆ๐‘–,๐‘กโˆ’2) because they are

115

uncorrelated with (๐œ€๐‘–๐‘กโˆ’ ๐œ€๐‘–,๐‘กโˆ’1) but correlated with (๐‘ฆ๐‘–,๐‘กโˆ’1โˆ’ ๐‘ฆ๐‘–,๐‘กโˆ’2). Anderson and Hsiao (1982) suggest that as long the error terms are not serially correlated ๐‘ฆ๐‘–,๐‘กโˆ’2 is the obvious choice for an instrument for (๐‘ฆ๐‘–,๐‘กโˆ’1โˆ’ ๐‘ฆ๐‘–,๐‘กโˆ’2).

The Anderson and Hsiao (1981) estimator (henceforth termed the AH estimator) when the dimension of a panel is (๐‘ ร— ๐‘‡) can be written as

๐›ฟ๐ด๐ป = (๐‘1๐‘‹)โˆ’1๐‘1๐‘Œ ) โ€ฆ โ€ฆ โ€ฆ โ€ฆ ๐ธ๐‘ž๐‘ข๐‘Ž๐‘ก๐‘–๐‘œ๐‘› 10

where ๐‘ is a ๐พ ร— ๐‘ (๐‘‡ โ€“ 2) matrix of instruments, ๐‘‹ is a ๐พ ร— ๐‘ (๐‘‡ โ€“ 2) of regressors and ๐‘Œ is a ๐‘ (๐‘‡ โ€“ 2) ร— 1 vector of dependent variables.

๐‘ = [

๐‘Œ๐‘–,1 โˆ†๐‘ฅ๐‘–,3

. .

. .

๐‘Œ๐‘–,๐‘‡โˆ’2 โˆ†๐‘ฅ๐‘–,๐‘‡

] ๐‘‹ = [

โˆ†๐‘Œ๐‘–,2 โˆ†๐‘ฅ๐‘–,3

. .

. .

โˆ†๐‘Œ๐‘–,๐‘‡โˆ’1 โˆ†๐‘ฅ๐‘–,

] ๐‘Œ = [

โˆ†๐‘Œ๐‘–,3 . .

โˆ†๐‘Œ๐‘–,๐‘‡

] โ€ฆ โ€ฆ โ€ฆ โ€ฆ ๐ธ๐‘ž๐‘ข๐‘Ž๐‘ก๐‘–๐‘œ๐‘› 11

๐‘ = [ ๐‘1

. . ๐‘๐‘

] ๐‘‹ = [ ๐‘‹1

. . ๐‘‹๐‘

] ๐‘Œ = [ ๐‘Œ1

. . ๐‘Œ๐‘

] โ€ฆ โ€ฆ โ€ฆ โ€ฆ ๐ธ๐‘ž๐‘ข๐‘Ž๐‘ก๐‘–๐‘œ๐‘› 12

The IV estimation produces consistent estimates if the error term in levels is not serially correlated. However, its weakness is that it fails to use all the available moments, which means that it does not necessarily result in more efficient estimates.

GMM in first differences as advanced by Arellano and Bond (1991) produces more efficient and consistent estimates, hence its preference over the AH estimator. It deploys additional instruments obtained by applying the moment conditions that exist between the lagged dependent variable and the disturbances. The number of moment conditions depends on ๐‘‡, the time periods, which are derived from the first differenced equation. Generalised Method of Moments uses the lagged dependent variables plus the lagged values of exogeneous regressors

116

as instruments and a weighting matrix which takes into account the moving averages (MA) (1) process in the differenced error term and the general heteroscedasticity. As a result, the Generalised Method of Moments estimates result in smaller variances than those associated with the AH type instrumental variable estimators.

The Generalised Method of Moments estimator can be expressed as follows ๐œƒฬ‚๐บ๐‘€๐‘€ = (๐‘‹1๐‘โˆ—๐ด๐‘๐‘โˆ—1๐‘‹)โˆ’1๐‘‹1๐‘โˆ—๐ด๐‘๐‘โˆ—1๐‘Œ โ€ฆ โ€ฆ โ€ฆ โ€ฆ ๐ธ๐‘ž๐‘ข๐‘Ž๐‘ก๐‘–๐‘œ๐‘› 13

Where ๐œƒ ฬ‚ is vector of coefficient estimates of exogeneous and endogeneous regressors, ๐‘‹ฬ… and ๐‘ฆฬ… are the vectors of first differenced regressors and dependent variables, respectively, ๐‘ is a vector of instruments and ๐ด๐‘ is a vector used to weight the instruments.

๐‘‹ = [ ๐‘‹1

. . ๐‘‹๐‘

] ๐‘Œ = [ ๐‘Œ1

. . ๐‘Œ๐‘

] โ€ฆ โ€ฆ โ€ฆ โ€ฆ ๐ธ๐‘ž๐‘ข๐‘Ž๐‘ก๐‘–๐‘œ๐‘› 14

Generalised Method of Moments uses an instrument matrix of the form,

๐‘๐‘– = [

[๐‘ฆ๐‘–0, โˆ†๐‘ฅ2โ€ฒ 0 โ‹ฏ 0

0 [๐‘ฆ๐‘–0, ๐‘ฆ๐‘–2, โˆ†๐‘ฅ3โ€ฒ] โ‹ฑ 0

โ‹ฎ 0

0 0 0 [๐‘ฆ๐‘–0, โ€ฆ , ๐‘ฆ๐‘–๐‘‡โˆ’2โˆ†๐‘ฅ๐‘‡โ€ฒ]

โ€ฆ โ€ฆ โ€ฆ โ€ฆ ๐ธ๐‘ž๐‘ข๐‘Ž๐‘ก๐‘–๐‘œ๐‘› 15

where the rows correspond to the first differenced equation for the period ๐‘ก = 3, 4, โ€ฆ , ๐‘‡ for individual ๐‘– and exploit the moment conditions,

๐ธ[๐‘๐‘–โ€ฒโˆ†๐œ€๐‘–] = 0 โ€ฆ โ€ฆ โ€ฆ โ€ฆ ๐ธ๐‘ž๐‘ข๐‘Ž๐‘ก๐‘–๐‘œ๐‘› 16 for ๐‘– = 1, 2 , โ€ฆ , ๐‘ where โˆ†๐œ€๐‘– = (โˆ†๐œ€๐‘–3, โˆ†๐œ€๐‘–4, โ€ฆ , โˆ†๐œ€๐‘–๐‘‡) โ€ฒ.

Arellano and Bond (1991) proposed two estimators; the one-step estimator and the two-step estimator (henceforth termed GMM1 and GMM2, respectively). GMM2 is the optimal estimator. GMM1 turns out to be optimal when the residuals are homoscedastic. If there is heteroscedasticity, GMM1 of instrumental variables continues to be consistent; however,

117

carrying the estimation in two steps increases efficiency. The weight matrix of a GMM1 is given by;

๐ด1๐‘ = (1

๐‘โˆ‘ ๐‘๐‘–โˆ—โ€ฒ

๐‘

๐‘–=1

๐ป๐‘๐‘–โˆ—) โˆ’1โ€ฆ โ€ฆ โ€ฆ โ€ฆ ๐ธ๐‘ž๐‘ข๐‘Ž๐‘ก๐‘–๐‘œ๐‘› 17

where H is a T โ€“ 2 square matrix with twos in the main diagonals, minus ones in the first sub diagonals, and zeros otherwise.

The weight matrix of a GMM2 if given by, ๐ด๐‘ = (1

๐‘โˆ‘ ๐‘๐‘–โˆ—โ€ฒ

๐‘

๐‘–

โˆ†๐œ€ฬ‚โˆ†๐œ€๐‘– ฬ‚๐‘–โ€ฒ๐‘๐‘–โˆ—) โˆ’1 โ€ฆ โ€ฆ โ€ฆ โ€ฆ ๐ธ๐‘ž๐‘ข๐‘Ž๐‘ก๐‘–๐‘œ๐‘› 18

Where โˆ†๐œ€ฬ‚๐‘– = โˆ†๐œ€ฬ‚๐‘–โ€ฆ , โˆ†๐œ€ฬ‚๐‘–๐‘‡ are the residuals from a consistent GMM1 of โˆ†๐‘ฆ๐‘–.

5.3 WORKING CAPITAL AND FIRM VALUE RELATIONSHIP ESTIMATION MODEL