Do stock markets value intangibles in the New Economy? Evidence from a panel of Indian software firms

Transcription

1 Do stock markets value intangibles in the New Economy? Evidence from a panel of Indian software firms PAPER FOR PRESENTATION AT THE 2007 AUSTRALASIAN MEETING OF THE ECONOMETRIC SOCIETY Supriyo De Faculty of Economics and Business, The University of Sydney, Sydney, NSW 2006, Australia ( sude7043@mail.usyd.edu.au, s.de@econ.usyd.edu.au) Abstract The success of the Indian software industry has been driven by intangible assets like human capital and corporate organizational capabilities. Stock market valuations should reflect these intangible factors. This paper analyses the impact of intangible assets on the market value of Indian software companies. Innovative methods are used to measure intangible assets. Measures of physical capital, human capital and organization capital are constructed using firm-level panel data. The primary hypothesis of the research is that, intangible assets have a significant effect on the market value of New Economy firms. The effect of intangible assets and sales growth on market value is analysed in a dynamic panel data econometric model. The dynamic specification captures the evolution of market value. The estimation technique uses a combination of system Generalised Method of Moments (GMM) and Minimum Distance Estimation (MDE). This methodology accounts for unobserved firm heterogeneity, endogenous explanatory variables and persistent variables. The econometric estimation also allows for an autoregressive valuation shock and serially uncorrelated measurement errors. The results conclusively show that intangible assets have a significant impact on market values of Indian software firms. Keywords: Dynamic panel data, Indian software industry, intangible capital, New Economy, market valuation. JEL classification: C23, G12 1

2 1. Introduction A major part of India s recent economic growth is attributed to the dramatic rise of the information technology (IT) software industry. The IT industry accounted for over 28 % of Gross Domestic Product (GDP) growth between 2000 and 2002 (Athreye, 2005). Software services export was estimated at $12.2 billion for financial year and increased at an annual rate of 28% (Ministry of Information Technology, 2004). 1 India s share in global software services exports was around 17% in Indian software exports grew at a rate of about 51% over the period (World Bank, 2004). The buoyancy of the sector has also been reflected by the spectacular increase in market values of Indian IT firms. This research examines the market valuation of Indian software firms with regard to the role played by intangible assets in the rise of the New Economy. The advent of the New Economy is characterised by the substantial use of information technology and intangible assets. There is broad agreement in the literature that intangible assets constitute the core of the New Economy (Bond and Cummins, 2000; Baddeley, 2003). Intangible assets do not have physical existence but play a crucial role in generating future revenues. Such assets include organizational architectures, innovation capabilities, training methods and management information systems. The inherently elusive nature of intangible assets makes them difficult to measure. This conundrum has engendered great debate in the literature. The market valuation puzzle lies at the centre of the controversy. According to one argument, if intangible assets contribute significantly to profit flows, they should be reflected in the valuation of the firm s shares. This approach recommends that the stock market value of firms in excess of physical capital replacement cost be used as a proxy for intangible assets (Hall, 2001; Brynjolfsson, Hitt and Yang, 2002). The counter-argument is that market valuations are subject to irrational impulses. If stock markets are inefficient, using market value to infer the stock of intangibles may lead to erroneous results (Bond and Cummins, 2000). The problem ultimately comes down to developing reliable measures of intangible assets. Recently, several concerted efforts have been made to measure intangible assets in 1 The Indian financial year commences on the 1 st of April of a calendar year and ends on the 31 st of March of the following calendar year. 2

3 the New Economy (Corrado, Haltiwanger and Sichel, 2005). One approach adopted for measuring intangible assets is based on the use of expenditure data. In this framework, intangible capital is estimated capitalizing expenditures that create long-lasting revenue flows from those whose returns dissipate quickly (Corrado, Hulten and Sichel, 2005). The measurement method used in this paper adheres to the aforesaid principle. Using the same idea, Lev and Radhakrishnan (2005) develop a measure of organization capital from sales, general and administrative (SGA) expenses. Adjustment cost measures derived from investment equations have also been used to estimate intangible assets (Cummins, 2005). Besides intangible organization capital, studies show that human capital also has an influence on market value. Abowd et al. (2005) develop detailed firmspecific measures of human capital. These measures have a positive influence on firm valuations. This study seeks to address the market valuation problem by using expenditurebased measures of intangible organization capital and human capital. The empirical relevance of these measures in a market value regression is examined. The primary hypothesis is that, expenditure-based measures of intangible assets have a significant effect on the market value of New Economy firms. Demonstrating the impact of these measures on the stock market value of firms validates the appropriateness of this measurement method. It also aims to address the market value puzzle by showing the effect of explicit measures of intangible assets on the stock market value of a firm. The major contributions of this paper are: (i) The development of comprehensive measures of intangible organization capital and human capital. (ii) The use of an innovative dynamic panel data framework to test the relevance of these measures. The empirical strategy employs a rigorous two stage estimation procedure that utilises system Generalised Method of Moments (GMM) and Minimum Distance Estimation (MDE). This empirical strategy accounts for the autoregressive evolution of the dependent variable, the possibility of measurement errors and the endogenous nature of the explanatory variables. 3

4 (iii) It seeks to address and put to rest the theoretical controversy regarding the impact of intangible assets on firm value. (iv) Finally, in demonstrating the validity of this approach in the context of the Indian software industry, the importance of intangible assets in emerging market economies is highlighted. The rest of the paper proceeds in the following manner. Section 2 provides the theoretical basis of the valuation equation. Section 3 deals with the empirical valuation equation and estimation strategy. Section 4 describes the data and measurement methods. The empirical results are discussed in section 5. Section 6 is conclusive in nature. 2. Theoretical model The substantial theoretical literature on valuation of capital demonstrates that the market value of a firm is related to the capital it owns (Hayashi, 1982; Hall, 1993; Bond and Cummins, 2000; Brynjolfsson, Hitt and Yang, 2002). The basic model used in this research draws upon the dynamic optimisation problem of the firm s investment decision as described by Brynjolfsson, Hitt and Yang (2002). The managers of the firm maximize the market value of equities (V) while investing in the different types of assets (I j, j = 1,, J) and incurring variable costs (N). The market value of equity (V) equals the present values of future cash flows π(t) based on a discount function u(t). Accumulation of various assets as reduced by the depreciation rates (δ j ) produces the stock of capital assets (K j, j = 1,, J). These assets include tangible as well as intangible elements. The optimisation problem is as follows: Max π ( t ) u ( t ) dt, (1) 0 I j, N where π ( t) = F( K, N, t) N I, (2) subject to j j dk dt j J = I δ K, for all j = 1,, J. (3) j j j j= 1 4

5 It is assumed that F(K j, N) is a homogenous function of degree 1 over K j, N, and I j (constant returns to scale) and twice differentiable. If all assets are accounted for and there are no adjustment costs involved in installing new capital, it can be shown that the market value of a firm has a direct relation with the current stock of capital assets: V J = K. (4) j= 1 j Even if the restrictive assumptions of a linearly homogenous production function implying constant returns to scale do not hold, there would be some correspondence between the market value of a firm and its stock of tangible and intangible assets. If the firm has market power this would influence the equity value. The value of a firm can be expressed as: V = V ( K, H, X, P ) (5) it it it it it where V it is the ith firm s value in time t, K it is plant and machinery, H it is human capital, X it denotes intangible organization capital, and P it is some measure of the market power of the firm, the movement of future prices of the output or future growth prospects. The empirical estimation equation can be derived from this formulation. 3. Empirical equation and estimation strategy The empirical estimation equation specifies the relationship between the value of a firm s equity and the stock of tangible and intangible assets in a log-linear framework (Hall, 1993; Abowd et al., 2005). Using a two-way error component panel data specification the valuation equation is as follows: v = β k + β h + β x + β y + γ + ( η + u + m ) it k it h it x it y i, t 1 t i it it u = ρu + e ρ < 1 (6) it i, t 1 it e, m ~ MA(0) it it 5

6 where v it is log of market value of firm i in year t, k it is log of plant and machinery, h it is log of human capital, and x it is log of organization capital. The log of lagged sales y i,t-1 is used to proxy the dynamics of output price, market power of the firm and growth prospects. The year specific intercept is γ t and the unobserved firm specific effect is η i. The valuation shock u it is considered autoregressive to capture the evolution of market value. Uncorrelated measurement errors are represented by m it. MA indicates a moving average process. The model has a dynamic common factor form: v = ρv + β k ρβ k + β h ρβ h + β x ρβ x + β y it i, t 1 k it k i, t 1 h it h i, t 1 x it x i, t 1 y i, t 1 ρβ y + ( γ ργ ) + ( η (1 ρ ) + e + m ρm ) y i, t 2 t t 1 i it it i, t 1 This can be represented as: v = π v + π k + π k + π h + π h + π x + π x + π y it 1 i, t 1 2 it 3 i, t 1 4 it 5 i, t 1 6 it 7 i, t 1 8 i, t 1 + π 9 yi, t 2 + γ t + ( ηi + wit ) (7) (8) where the non-linear (common factor) restrictions are π 3 = -π 1 π 2, π 5 = -π 1 π 4, π 7 = -π 1 π 6 and π 9 = -π 1 π 8. On obtaining consistent estimates of the unrestricted parameter vector π=(π 1, π 2, π 3, π 4, π 5, π 6, π 7, π 8, π 9 ) and the associated variance-covariance matrix, the restricted parameters (β k, β h, β x, β y, and ρ) can be estimated using minimum distance. Alternatively, the above equation may be considered an explicit dynamic approximation to an adjustment process and the restrictions need not be imposed. Given the log-linear specification, coefficients of the various capital variables represent the elasticity of market value with respect to that variable. Also * γ = γ ργ and η = η (1 ρ). The * t t t 1 serial correlation structure of the error term w it depends on the presence of measurement errors. This is discussed subsequently. The choice of this specification is appropriate in several respects. The substantial literature on production function estimation demonstrates that the presence of endogenous explanatory variables is a major problem. 2 Given the essential similarity between production functions and valuation equations, the problem of endogenous variables needs to be addressed (Cummins, 2005). The valuation equation estimated here i i 2 See Griliches and Mairesse (1995), Mairesse and Hall (1996), and Blundell and Bond (1998 and 2000). 6

7 is likely to have endogenous explanatory variables since all the capital inputs and lagged sales would probably be correlated with market value since the theory posits that investment depends on firm value. 3 Blundell and Bond (2000) show that this type of specification coupled with the system Generalised Method of Moments (GMM) estimator can control for endogenous explanatory variables and finite sample biases in persistent panel data series. 4 The non-linear common factor restrictions can be imposed and tested using Minimum Distance Estimation (MDE). This rigorous two-stage procedure forms the basis of the estimation methods used in this paper. 5 The autoregressive error term helps obtain valid lagged internal instruments for the GMM estimators. The autocorrelation coefficient has the economic interpretation of evolving change in value due to economic growth or technological advancements. The use of constructed measures for the inputs may lead to measurement errors. The inclusion of an explicit measurement error facilitates the detection of this problem. The estimation technique can be modified to account for measurement errors. The econometric method is also appropriate given the nature of the data available. This research uses firm-level panel data with a large number of cross-sectional units (N) but comparatively limited time span (T). The system GMM estimator is particularly suited to this type of data structure (Blundell and Bond, 2000; Baltagi, 2005). Alternative estimators namely, Ordinary Least Square (OLS), Within Groups and the Arellano and 3 Standard investment theory posits that capital investment depends on firm value (Hayashi, 1982). 4 In view of the results and recommendations described in Blundell and Bond (2000) and Bond (2002), a careful analysis of the time series properties of all the variables is carried out. The variables appear to be persistent. 5 The choice of estimation technique is also dictated by the peculiarities of the IT software sector and assumptions regarding the underlying data. Alternative methods that control for endogenous variables and unobserved shocks, like Olley and Pakes (1996) and Levinsohn and Petrin (2003) are inappropriate in the present scenario. The former method uses investment as a proxy to control for errors correlated with the inputs. However, since investment is subject to adjustment costs, the estimation may be problematic. The latter use intermediate inputs to address the problem. This technique is of doubtful value in case of services like software, where the use of intermediate inputs is negligible. The Blundell and Bond (2000) estimator places great demands on the data. The data series are constructed keeping these initial condition and stationarity requirements in perspective. This is discussed in the subsequent section. 7

8 Bond (1991) differenced GMM, are also used to determine the appropriateness of the results obtained from the preferred system GMM estimator. All the alternative estimators are known to display certain problems when used in dynamic models. Extensive studies using simulated and real data show that in dynamic panels, for the coefficients on endogenous explanatory variables, OLS estimates are likely to be biased upwards and Within Group estimates are likely to be biased downwards. Differenced GMM suffers from a weak instruments problem when the variable series are persistent. This leads to the coefficients on the endogenous variables to be biased towards zero (Blundell and Bond, 2000; Baltagi, 2005). The known biases and inconsistencies of the alternative estimators provide useful bounds that verify the accuracy of the preferred estimator. These alternative estimators also help establish the robustness of the basic estimation equation. The system GMM estimator is used to estimate the unrestricted models and MDE is employed to arrive at the restricted parameters. The system GMM estimator utilises a stacked system of equations in first differences and levels (Blundell and Bond, 2000). For the levels equations, lagged first differences are used as instruments. In case of the firstdifferenced equations, lagged levels are used as instruments. Instrument selection depends on the serial correlation structure of the model. The empirical specification of this paper results in the reflection of measurement errors in the serial correlation structure. In the absence of measurement errors (var(m it )=0) the error term is not serially correlated (w it = e it ~ MA(0)). If measurement errors are present, the error term is serially correlated (w it ~ MA(1)). In the former case, levels of the endogenous explanatory variables dated t-2 or earlier are valid instruments for the first-differenced equations. For the levels part of the system, first differences dated t-1 are appropriate instruments. In the presence of serial correlation, levels dated t-3 or earlier are appropriate instruments for the first-differenced equations. First differences dated t-2 are used as instruments for the levels equations. Arellano and Bond (1991) propose tests to check for the validity of instrument choice and model specification. These methods are employed to diagnose the unrestricted estimates. To avoid small sample over-fitting biases, a parsimonious instrument set is used. First the model is estimated by assuming that there are no measurement errors. For the first-differenced equations, the endogenous explanatory 8

9 variables in levels lagged by two to four periods (t-2 to t-4) are used as instruments. For equations in levels, the instruments are the first differences lagged by one period (t-1). Furthermore, to account for the possibility of measurement errors, the equation is also estimated with longer lagged instruments. For the first-differenced equations, the endogenous explanatory variables in levels lagged by three to five periods (t-3 to t-5) are used as instruments. For equations in levels, the instruments are the first differences lagged by two periods (t-2). The use of different instrument sets also helps test the robustness of the estimation results. The unrestricted parameters and variances obtained from the system GMM estimator are subjected to restrictions using the Minimum Distance Estimation method attributable to Chamberlain (1982). 6 This helps obtain the restricted parameters β k, β h, β x, β y, and ρ. The restrictions are subjected to a minimum distance test Data and measurement techniques The data used for this research are based on financial statements of major Indian IT companies drawn from a dataset maintained by the Centre for Monitoring Indian Economy (CMIE). The econometric estimates use an unbalanced panel of 164 crosssectional units. There are a minimum of four and a maximum of ten annual observations amounting to 979 observations. The estimation equation uses twice lagged variables. This reduces the maximum annual observations to eight years for the financial years to The total number of observations is then reduced to 651. The detailed structure and arrangement of the panel is given in Appendix A. 6 Chamberlain (1982), Arellano (2003) and Lee (2004) describe the principles underlying MDE. In essence, interest lies in estimating the f x 1 restricted parameter vector α given the k x f (f < k) restriction matrix R and the k x 1 unrestricted parameter vector β such that: β= Rα. The number of restrictions is k - f. Given b an estimator of β, the efficient estimator of α is a=(r V -1 R) -1 R V -1 b, where V is the variance-covariance matrix for the unrestricted estimates. The variance-covariance matrix for the restricted estimates is M=(R V -1 R) 1. 7 The appropriate test statistic is W= (Ra- b) V -1 (Ra- b) ~ χ 2 k-f. 9

10 Measures of market value (V it ), physical capital (K it ), human capital (H it ) organization capital (X it ) and sales (Y it ) are constructed from the aforementioned database. The principles and methods used to construct the variables are given below: Market value (V it ): The dependent variable is log of debt-adjusted market value (v it ). Market value is the closing share price at the end of the financial year multiplied by the number of ordinary shares. Debt-adjusted market value (V it ) is the market value of ordinary shares after adjusting for debt. 8 Organization capital (X it ): The main methodological innovation of this research is the measurement of intangible organization capital (X it ). Organization capital is a collection of business management systems and organizational practices that facilitate the extraction of the highest value of output from available human and physical capital. 9 Intangible organization capital is conceptualised as an asset arising out of the delivery of a stream of business services. For instance, a management consultancy provides certain services on a project basis and creates an efficient business model for the client. Similarly, software firms provide software solutions, which remain with clients as business assets. This implies that the value of the asset created can be imputed from the expenditure made to acquire the services. Even if the services were produced within the firm, some expenditure would be associated with it. Therefore, firm-level measures of expenditures can be used to infer the amount of intangible organization capital created. 10 This principle underlies the measurement strategy espoused by Corrado, Hulten and Sichel (2005). 8 The precise adjustment factor for debt depends on the required rate of return and the interest rate on debt. These are difficult to measure given the influence of other variables like risk. In this measure, the debt reported in financial accounts is added to market value. This debt would generally be valued according to conservative financial reporting principles. Measurement errors if any are accounted for by the econometric method. 9 For details see Evenson and Westphal (1995) and Lev and Radhakrishnan (2005). 10 There is substantial evidence indicating that Indian software firms are using external consultants and internal procedures to put in place quality control and other management structures (Arora and Asundi, 10

11 Lev and Radhakrishanan (2005), develop measures of intangible organization capital using sales, general and administrative expenses (SGA). In this research a more focussed variable, administrative expenses is used to measure intangible capital. Administrative expenses are capitalised to create measures of intangible organization capital. 11 Existing literature indicates that it takes time to build intangible assets (Corrado, Hulten and Sichel, 2005). But the resulting benefits erode slowly since rivals find it difficult to replicate these assets (Brynjolfsson, Hitt and Yang, 2002). Also, some administrative expenses would be routine in nature and should not be included in the intangible capital measures. Taking these aspects into account, 20% of administrative expenses are capitalised subject to a depreciation rate (δ X t ) of 10%, to create a measure of intangible organization capital. The expenses (E X it ) are deflated using the Indian Consumer Price Index for Urban Non-manual Employees [CPI (UNME)] ( =100), the most appropriate indicator of urban inflation in India. The deflator series (P t ) comes from the Reserve Bank of India (Reserve Bank of India, 2006). Investment in organization capital is as follows: 0.2 X X Eit Iit = Pt Thereafter, the perpetual inventory method is used to calculate the organization capital series in the following manner: X = (1 δ ) X + I X X it t i, t 1 it Human capital (H it ): The measure of human capital is derived from wages. The use of wages to proxy human capital accounts for unobservable factors like quality, experience, talent, education, etc. (Abowd, et al., 2005). These are not reflected in traditional measures of human capital based on educational attainment or labour measures based on time worked. It is well documented that experienced and talented individuals in the Indian software sector often earn well above the 1999; Athreye, 2005). Arora and Asundi (1999) indicate that large expenditure outlays are incurred to acquire these capabilities. 11 Instead of the direct instantaneous effect of SGA on organization capital implied by Lev and Radhakrishnan (2005) a slower build up of intangible capital is posited. 11

12 going rate (Heeks, 1996). Further, in the Indian IT sector, salaries are paid on a monthly basis and time based wage payments are rare. In the circumstances, wages provide a fair approximation of the human capital available to the firm. However, the rapid rise in IT sector salaries needs to be accounted for. In this regard, an index of average salaries of major IT companies is derived. This is used to deflate the wages. The usual indicator of urban inflation, the Indian Consumer Price Index for Urban Non-manual Employees [CPI (UNME)], is not used since IT sector salaries have grown much faster than urban inflation or even salaries of other industries. Using a slower rising deflator would overestimate the human capital used by the firm. Physical capital (K it ): It is assumed that physical capital is composed of computer equipment. In computing the series the differences in gross capital stock between G G two periods ( Kit Ki, t 1) is taken as investment. The investment amounts are deflated using a quality-adjusted series (P K t ) based on the United States (U.S.) Bureau of Economic Analysis (BEA) data as discussed below. Investment is then calculated as follows: G G Kit Ki, t 1 Iit = K Pt For computer equipment the perpetual inventory method is used to compute capital stock in the following manner: K = (1 δ ) K + I K it t i, t 1 it The annual geometric depreciation rate (δ K t ) is taken to be 30% broadly following estimates used by Jorgenson and Stiroh (2000). Indices developed by the U.S. Bureau of Economic Analysis (BEA) are used to arrive at quality-adjusted measures of capital equipment. This helps account for the rapid quality improvement and decline in prices of computer equipment Jorgenson and Stiroh (2000) and Grimm et al. (2005) discuss the issue of quality-adjusted prices of information technology hardware and software in detail. The deflators are appropriate for Indian data since much of the hardware used in Indian software firms are imported from the U.S. 12

13 Sales (Y it ): Sales data are also deflated using a measure of software price levels based on BEA deflators (Bureau of Economic Analysis, 2005). 13 The variables and measurement methods are summarised in Table 1: Table 1. Variables and Measurement Methods Variable Market value adjusted by debt (V it ) Physical Capital (K it ) Basic Source Closing price, number of ordinary shares and book debt reported by CMIE. Plant and machinery stocks reported by CMIE. Method Deflator Depreciation rate Closing share price at end Nil Nil of financial year multiplied by number of ordinary shares plus book debt. Differences in end of year Chain-type price 30% stock taken as investment. indices for Perpetual inventory computers (Bureau method used. Appropriate of Economic deflator used. Analysis, 2005). Human Capital (H it ) Organization Capital (X it ) Salaries and wages reported by CMIE. Administrative expenses reported by CMIE. Salaries and wages deflated by an appropriate measure. 20% of expenses taken as investment. Perpetual inventory method used. Appropriate deflator used. Index of average salaries of major IT companies derived from CMIE. Consumer Price Index (Reserve Bank of India, 2006). Nil 10% Sales (Y it ) Sales reported by CMIE. Appropriate deflator used. Chain-type price indices for custom software (Bureau of Economic Analysis, 2005). Nil The data series are constructed from periods well before the start of the sample. This ensures that the series are stable and meet the initial conditions requirements of the system GMM estimator. Since most of the companies in the data set were in existence well before the start of the sample, the stationary means requirement would be valid. Furthermore, the system GMM estimator is fairly robust to dilutions of this condition. Even with persistent series, if time dummies are included, the mean stationarity 13 These deflators are appropriate since most of the software produced is exported to the U.S. 13

14 requirement can be replaced with a weaker assumption of common evolution of the means (Blundell and Bond, 2000; Bond, 2002). Given the recent origin of the industry and its small size, the sample of 164 is fairly large and representative of the relevant population. NASSCOM, the premier industry group of Indian software and service companies reports a membership of 854 as on 31 December 2001 (NASSCOM, 2002). Considering that some of these members would be recent start-ups, business process outsourcing companies and foreign companies not registered in India, the sample covers a large part of the companies that have been in existence for a reasonable period of time. The variables in levels have the following descriptive statistics: Table 2. Descriptive statistics (Values in Rs. Crore, 1 crore=10 million) Variable Mean Standard Min. Max. Median deviation Market value (V it ) Physical Capital (K it ) Human Capital (H it ) Organization Capital (X it ) Sales (Y it ) The wide range of maximum and minimum values for all variables shows that the sample covers both large and small firms. The distribution is skewed with mean sales and physical capital (i.e. computers, etc.) being around nine to ten times larger than the respective medians. The mean market value is around five times larger than the median. The median sales are around Rs. 140 million while the mean is around Rs million. Considering the growth of the industry up to 2005, the descriptive statistics seem quite representative of the industry. For instance, median sales of Rs. 28 million for and Rs million for have been reported in a previous study using NASSCOM data (Athreye, 2005). 14

15 The measures of intangible assets, human capital and organization capital have a positive relation with market value. Figure 1 shows that the measure of human capital has a distinct positive correlation with market value. Market value has a strong positive correlation with organization capital as seen in Figure Market Value 4 0 Market Value Human Capital Organization Capital Figure 1. Market value and human capital Figure 2. Market value and organization capital Note: Each point represents a firm-year observation. The term market value represents log of debtadjusted market value, human capital is log of the measure of human capital, and organization capital denotes log of the measure of intangible organization capital. 5. Empirical results This section describes the results obtained from empirical estimation. The results of the alternative estimators are given in Table 3. The first and second columns report OLS and Within Groups estimates, respectively. The third and fourth columns give differenced GMM estimates using twice lagged (t-2 to t-4) and thrice lagged (t-3 to t-5) levels as instruments. The OLS and Within Groups estimates both display first order serial correlation. In the presence of fixed effects, the OLS estimates of the lagged dependent variable (v i,t-1 ) are known to have an upwards bias. On the other hand, the Within Groups estimate of the lagged dependent variable (v i,t-1 ) generally demonstrates a downward bias. 15

16 Table 3. OLS, Within Groups and differenced GMM estimation results (Dependant variable is v it ) Variable (1) OLS (2) Within Groups (3) Differenced GMM (t-2) v i,t (22.886)*** (2.317)** (-1.012) k it (2.209)** (1.255) (-0.181) k i,t (-1.119) (-1.632) (-0.511) h it (0.611) (1.134) (1.259) h i,t (-0.139) (1.519) (3.835)*** x it (2.852)*** (4.213)*** (6.338)*** x i,t (-2.553)** (-3.396)*** (-3.818)*** y i,t (0.286) (1.036) (-1.161) (-0.077) y i,t (0.714) (-0.467) (-3.592)*** (0.393) m m Sargan R Wald joint Wald (h,x) Instruments - - v i,t-2,, v i,t-4, k i,t-2,,k i,t-4, h i,t-2,,h i,t-4, x i,t-2,,x i,t-4, y i,t-3,,y i,t-5 (4) Differenced GMM (t-3) (-1.526) (-1.465) (1.148) (-1.614) (2.348)** (3.161)*** (-2.927)*** v i,t-3,, v i,t-5, k i,t-3,,k i,t-5, h i,t-3,,h i,t-5, x i,t-3,,x i,t-5, y i,t-4,,y i,t-6 DPD98 for Gauss (Arellano and Bond, 1998) is used for estimation and diagnostic tests. Asymptotic robust standard errors are used for t-statistics given in parenthesis ( ). *,** and *** denote statistical significance at the 10%, 5% and 1% levels, respectively. Two-step differenced GMM estimates are reported. m1 and m2 are tests for first order and second order serial correlation with a null of no serial correlation. P-values reported. Sargan denotes the Sargan test for validity of overidentifying restrictions with a null of instrument validity. P-values reported. 2 2 R indicates adjusted R Wald joint is a Wald test for joint significance of the explanatory variables. P-values reported. Wald (h,x) is a Wald test for significance of selected explanatory variables h it and x it. P-values reported. Year dummies included in all estimates. 16

17 As expected the OLS estimate of the coefficient on the lagged dependant variable is quite large (0.744) when compared to that obtained from the Within Groups estimate (0.104). Other variables that have an endogenous relationship with the explained variable are likely to show similar biases. The differenced GMM estimators when used in conjunction with persistent variables gives biased estimates for the endogenous variables. The bias term effectively scales the estimated coefficient towards zero (Baltagi, 2005). This estimator also yields low standard errors. The model demonstrates these typical characteristics when differenced GMM estimators are used. Remarkably though, the coefficient for organization capital is large ( ) and significant in all the alternative estimators. This clearly establishes the importance of this variable in determining the market value of New Economy firms. The OLS estimate also provides a broad guide to the goodness of fit of the basic model. The adjusted R- square of 0.85 shows that the model explains a significant amount of the behaviour of the dependant variable. Of course, as discussed before, the OLS estimator is not reliable for dynamic panels. Having determined the general performance of the model in terms of the known deficiencies of the alternative estimators, the estimation is carried out using the preferred system GMM estimator. The econometric estimates and diagnostic tests for the preferred system GMM estimators are reported in Table 4. The first column denoted by System GMM (One-step, t-2) gives one-step estimates using the system GMM estimator with instruments of lags t-2 to t-4 for the differenced equations. The second column, System GMM (One-step, t-3) gives one-step estimates using the system GMM estimator with instruments of lags t-3 to t-5 for the differenced equations. The third and fourth columns, System GMM (Twostep, t-2) and System GMM (Two-step, t-3), give two-step estimates corresponding to the first and second columns. The one-step estimators are known to be more reliable but are less efficient in finite samples. The two-step estimators are more efficient but less consistent in finite samples. Instead of choosing the difficult trade off between the desirable properties of efficiency and consistency, both results are reported. 17

18 Table 4. System GMM estimation results (Dependant variable is v it ) Variable (1) System GMM (One-step, t-2) v i,t (7.373)*** k it (0.219) k i,t (0.355) h it (0.505) h i,t (-0.424) x it (2.340)** x i,t (-1.522) y i,t (0.411) y i,t (2) System GMM (One-step, t-3) (5.165)*** (0.245) (1.116) (0.877) (-0.840) (1.171) (-0.818) (-0.794) (3) System GMM (Two-step, t-2) (26.569)*** (0.434) (1.778)* (3.856)*** (-2.750)*** (13.406)*** (-9.327)*** (2.876)*** (4) System GMM (Two-step, t-3) (15.085)*** (0.835) (6.011)*** (3.348)*** (-3.758)*** (4.769)*** (-3.428)*** (-3.477)*** (4.028)*** (0.888) (1.214) (3.247)*** m m Sargan Wald joint Wald (h,x) Instruments for differenced equations Instruments for levels equations v i,t-2,, v i,t-4, k i,t-2,,k i,t-4, h i,t-2,,h i,t-4, x i,t-2,,x i,t-4, y i,t-3,y i,t-5 ; v i,t-1, k i,t-1, h i,t-1, x i,t-1, y i,t-2. v i,t-3,, v i,t-5, k i,t-3,,k i,t-5, h i,t-3,,h i,t-5, x i,t-3,,x i,t-5, y i,t-4,,y i,t-6 ; v i,t-2, k i,t-2, h i,t-2, x i,t-2, y i,t-3. v i,t-2,, v i,t-4, k i,t-2,,k i,t-4, h i,t-2,,h i,t-4, x i,t-2,,x i,t-4, y i,t-3,y i,t-5 ; v i,t-1, k i,t-1, h i,t-1, x i,t-1, y i,t-2. v i,t-3,, v i,t-5, k i,t-3,,k i,t-5, h i,t-3,,h i,t-5, x i,t-3,,x i,t-5, y i,t-4,,y i,t-6 ; v i,t-2, k i,t-2, h i,t-2, x i,t-2, y i,t-3. DPD98 for Gauss (Arellano and Bond, 1998) is used for estimation and diagnostic tests. Asymptotic robust standard errors are used for t-statistics given in parenthesis ( ). *,** and *** denote statistical significance at the 10%, 5% and 1% levels, respectively. m1 and m2 are tests for first order and second order serial correlation with a null of no serial correlation. P-values reported. Sargan denotes the Sargan test for validity of overidentifying restrictions with a null of instrument validity. P-values reported. Wald joint is a Wald test for joint significance of the explanatory variables. P-values reported. Wald (h,x) is a Wald test for significance of selected explanatory variables h it and x it. P-values reported. Year dummies included in all estimates. 18

19 The model performs well with regard to the principal diagnostic tests. The consistency of the estimates relies heavily on the assumption of no second order serial correlation. If the disturbances w it are not serially correlated the first difference residuals would show first order serial correlation ( wˆ ˆ it wi, t 1). On the other hand, there would be no second order serial correlation. The m1 test checks for first order serial correlation. In all cases, the null of no first order serial correlation is rejected. The m2 test for second order serial correlation has a null of no second order serial correlation. In all cases, the null is not rejected at conventional levels of significance (α=0.05). The Sargan test of over identifying restrictions also remains valid in all cases. The null hypothesis of instrument validity is not rejected in all four cases. The basic estimates are therefore well specified and have valid instruments. It also indicates that substantial measurement errors are not present. Furthermore, even if second order serial correlation were present, the estimates with instruments of levels t-3 to t-5 for first difference equations and first difference t-2 for levels equations lead to consistent estimates. The coefficient on the lagged dependant variable is stable and significant in all the estimates. The estimates do not vary widely and are between and The validity of these estimates is further verified by the fact that they lie between the upwards-biased OLS estimate of and the downwards-biased Within Groups estimate of The coefficient of physical capital is small and insignificant in all the estimates. Human capital is significant in the two-step estimates. The coefficients range between and Intangible organization capital is significant in all estimates other than the one-step thrice lagged estimator. This variable appears to have a large effect with the coefficients being between and Lagged sales have significant effect only in the two-step estimates but the signs in the twice lagged and thrice lagged estimates are not in consonance. It appears that the effect of lagged sales on market value is small. The Wald test of joint significance for all the explanatory variables is not rejected. The Wald test for significance of human capital and organization capital is also not rejected. This substantiates the specification and validity of the model. Overall, the preferred system GMM estimators appear to give robust and reliable results. Having established the reliability of the system GMM estimates, Minimum Distance Estimation is used to impose non-linear restrictions on the system GMM 19

20 estimates. The restricted estimates recover the parameters of the basic econometric model described in equation 6 (ρ, β k, β h, β x and β y ). Table 5 reports these estimates. The four columns correspond to MDE results obtained by imposing restrictions on the four system GMM estimates reported in Table 4. Table 5. MDE results obtained by imposing restrictions on system GMM estimates. Parameter (1) MDE on System GMM (One-step,t-2) ρ (7.428)*** β k (1.009) β h (0.353) β x (3.542)*** β y (2) MDE on System GMM (One-step,t-3) (5.298)*** (2.631)*** (0.165) (2.478)** (3) MDE on System GMM (Two-step,t-2) (27.343)*** (2.601)*** 0.11 (4.445)*** (19.032)*** (4) MDE on System GMM (Two-step,t-3) (13.715)*** (4.664)*** (2.652)*** (10.597)*** (-0.823) (0.363) (-0.367) (5.653)*** COMFAC Gauss is used for estimation and COMFAC test. t-statistics given in parenthesis ( ). *,** and *** denote statistical significance at the 10%, 5% and 1% levels, respectively. COMFAC denotes a minimum distance test for the non-linear common factor restrictions imposed in the restricted models. P-values are reported. Year dummies included in all estimates. The COMFAC test indicates that the non-linear restrictions are valid for the twostep estimators (Columns 3 and 4). The one-step thrice lagged estimator (Column 2) performs marginally. The one-step twice lagged estimator (Column 1) does not show strong performance. The autocorrelation coefficient is statistically significant and stable. The estimates range between and The large and significant autocorrelation coefficient has interesting implications. It represents a dynamic process encompassing productivity growth, improvements in technology and increasingly optimistic expectations regarding the prospects of the industry. The log-linear specification of the model implies that the coefficient on a capital input represents the elasticity of market value with respect to that input. The physical capital coefficient is significant in three estimates (Columns 2, 3 and 4). The coefficients are between and The 20

21 human capital coefficient is significant in two estimates (Columns 3 and 4) with the coefficients ranging between and The coefficient of intangible organization capital is large and significant in all estimates. The estimates are stable ranging between and This conclusively establishes the importance of this variable in determining the market value of New Economy firms. The coefficient of lagged sales is not significant in all but one estimate. This parameter is therefore of marginal importance. Given the specifications of the model, the empirical estimates appear to be stable and reliable. The parameter estimates are reasonable. The major innovative feature of the model, the inclusion of an explicit measure of intangible organization capital is justified. The validity of the basic system GMM estimates also indicates that substantial measurement errors are not in evidence. This shows that the measures of market value, tangible and intangible capital accurately capture the underlying phenomena. It also indicates that the a priori assumptions of slow build up and erosion of organization capital posited by theoretical considerations are justified. Nevertheless, the results need to be viewed within the theoretical and empirical limitations of the model. The variable representing intangible organization capital captures only those aspects of the firm that are correlated with directly observed expenditures. Firm and time specific intangibles not related to expenditures, like unobservable managerial talent, are not reflected and would end up within the general error term. However, such intangible assets are likely to be negligible. Furthermore, the structure of the model, which wipes out firm specific effects and controls for time variant effects, is fairly robust to the presence of such unobserved variables. 6. Conclusions This study examined the influence of intangible assets, including human capital and organization capital, on market values of Indian software firms. Measures of physical capital, human capital and organization capital were constructed using firm-level panel data. The effects of these inputs and sales on market values were analysed using a 21

22 dynamic panel data econometric model. The results conclusively show that intangible assets have a large impact on market values of Indian software firms. The empirical and measurement methods demonstrated in this paper have several theoretical and practical applications. This paper validates the utility of system GMM estimation in dynamic panel data models. When used in conjunction with Minimum Distance Estimation, it provides a potent mechanism to deal with persistent variables, endogenous explanatory variables and autoregressive processes. While this methodology was initially developed to deal with problems in production function estimation, it can be extended to other persistent economic variables that are jointly determined by their own lagged values, and current and lagged values of other variables. Besides the example of market valuation demonstrated in this research, other important economic variables like investment, household consumption, industry wages and inventories may display these properties. The methods demonstrated herein can therefore be extended to other areas of research. Panel data methods are particularly useful in emerging economies where rapid structural changes invalidate the use of longer time series. For instance, in case of the Indian economy, due to liberalization policies pursued since 1991, data prior to that period are virtually useless for drawing conclusions regarding the present trajectory of the economy. The availability of a large cross-sectional element allows reasonable statistical inference despite the unavailability of long time span data. The study also demonstrates the usefulness of panel data methods in capturing the dynamic nature of economic phenomena. The significance of organization capital measures validates the use of expenditure-based measures of intangible capital. Such measures can be incorporated in valuation equations to forecast share values. This method can be used for valuing initial public offerings and deciding share prices for mergers, acquisitions, etc. The relevance of certain expenditures for long-term value creation has interesting implications. It sheds new light on corporate cost management issues. Myopic cost cutting measures, especially in key areas like human resources can be detrimental. On the other hand, organizational restructuring while incurring high initial costs may have long-term benefits. Furthermore, if intangible assets are incorporated in market values, significant deviations from values 22

23 that account for intangibles may represent irrational asset bubbles. This idea can be used to predict and deal with stock market bubbles. Therefore, this research has potential benefits for managers, shareholders, financial analysts, market regulators and policymakers alike. The methods and results arrived at in this research can also be extended to other industries and countries. The importance of intangible capital in determining the stock market valuation of Indian software companies yields interesting insights. It shows that the New Economy phenomenon that originated in advanced economies has started having an impact in certain emerging market economies. As a consequence, some firms in these countries are probably adapting to changing circumstances by paying greater attention to intangible assets like organization capital. 23

24 Appendix A: DATA APPENDIX The data collation process is initiated with an unbalanced panel of 359 firms for the financial years to The data are then used to construct variables for physical capital, human capital, organization capital and sales. The data are cleaned and arranged for the financial years to so that each cross-sectional unit has at least four continuous years of data. Only companies with complete records for all major variables are included. There is no interpolation of missing data. This results in an unbalanced panel sample of 164 firms with a maximum of ten and a minimum of four annual observations. Consequently a total of 979 data points are available. Two years are lost due to the lagged structure of the model. This leaves 651 data points for a maximum time span of eight years. The structure of the unbalanced sample is given in Table A1 below: Table A1 Number of records for each cross-section Number of cross-sectional units