Indian Software Exports and Engineering College Capacity

Transcription

1 The Indian Software Industry: the Human Capital Story Ashish Arora & Surendrakumar Bagde Heinz School of Public Policy and Management Carnegie Mellon University Pittsburgh PA {ashish, Abstract: The Indian software industry has been widely studied. Previous studies have recognized the role of skilled labor in the growth of the industry, but have not empirically investigated the role of skilled labor. In this study we focus on the role of human capital in the regional location of the software industry. Specifically, we study the effect of the engineering college capacity for undergraduate engineering studies in different Indian states on the regional growth of the software exports industry between 1990 and We find significant effect of human capital on the growth of software exports even after controlling for other relevant factors. In addition the initial as well as current size of electronics hardware industry also plays important role in the regional growth of the software exports industry.

2 Introduction The Indian software industry, which was almost non-existent till late 1980s, grew at tremendous pace after early 1990s. The Indian software industry's export was about US$128 millions in and grew to US$485 millions in By the software exports had increased to US$ 12.2 billions. The bulk of Indian software industry is concentrated in a few clusters; indeed Bangalore has often been branded as the Silicon Valley of India in press accounts. Some scholars have highlighted the role of policies adopted by federal and state governments. The infrastructure initiatives of the federal government, especially Software Technology Parks (STP) scheme , provided reliable internet connectivity and single window clearance for various government permissions to software export firms. In 1990s many state governments have provided suitable infrastructure by setting up information technology (IT) parks (see Nasscom, various years). The importance of availability of skilled manpower in the success of Indian software exports has been recognized by several research studies (e.g., Lakha, 1994). A noteworthy feature of Indian software industry is predominant share of engineers amongst software professionals in India. Arora et al. (2001) have argued that large share of south and west region in engineering engineering college capacity, spurred by the growth of private (not publicly funded) engineering colleges, was one of main reasons for growth of software industry in those regions. However, these claims have not been systematically tested. In this paper we empirically investigate how software exports by the fourteen major states of India are influenced by local levels of human capital. The simple point of the paper is that some states were favored locations for software because they had higher stocks of human 2 Software Technology Parks of India (STPI) is a society set up by the Department of Communication & Information Technology, Government of India in 1991, with the objective of encouraging, promoting and boosting the Software Exports from India. The software companies are given tax incentives and have duty free imports for some of their equipment and they register themselves with STPI for availing such benefits. There were other schemes like export processing zones which offered similar incentives to firms locating in such zones. However, STP scheme offers much higher level of flexibility to firms in their location choices and was targeted to software export firms. Firms could locate anywhere and were required to register with designated STP office to avail various incentives. 5 Uttar Pradesh and Madhya Pradesh reflect the geographic boundaries as in 1990 and not the current boundaries.

3 capital, and they had higher stocks of human capital because they allowed private engineering colleges to operate earlier than other states. There are empirical challenges in such research. First, there may be unobservable state characteristics that are correlated with software industry and engineering education capacity. We address this potential source of bias by using a panel dataset and controlling for state fixed effects. Second, there is issue of endogeneity of engineering college capacity. We address the issue of endogeneity in two ways. We estimate effect of engineering college capacity before the birth of the industry on software exports. We also instrument for engineering college capacity. The instrument, discussed in greater detail below, is based on the policy on private, non-granted engineering colleges, of neighboring states. Our research covers fourteen states of India, for the period According to official statistics, these fourteen states population share in country s population was percent in 2003, have 78 percent geographic area of India and share in net state domestic product in was 79.2 percent 7. As well, the available data require that our measure of software exports is a broad one, including not only the export of software, but also affiliated IT services, including the so-called IT enabled service. IT enabled services are relatively unimportant in terms of revenues for much of our sample and only become significant after the turn of the century. As well, such services, particularly low-end services such as call centers rarely employ engineers. Their exclusion would therefore only strengthen our results. These, and other issues related to our data are discussed in more detail in the appendix. This research is contributes to the literature on the globalization of the software industry (Arora and Gambardella, 2005). It is also related to a recent literature that examines inter-state variation in institutions and policies. Bhaumik et al. (2006) analyze inter-state variation in firm entry into 3-digit industry between 1984 and They conclude that during the 1980s industry level factors largely explained the variation in entry rates, but post-1991 reforms, variation in entry rates during 1990s were explained largely by state level institutions and legacy factors. Aghion et al. (2006) study the impact of delicensing and state labor market regulations on the performance of manufacturing industry. Intriguingly, they find that only states with relatively pro-employer policies benefited from de-licensing. Thus there was reallocation of industrial production from states with pro-worker labor institutions to states with pro-employer labor institutions. Kochar et al. (2006) conclude that state level institutions and policies have 7 The source for population and area share is the Census of India, and for the GDP share is Government of India, Ministry of Ministry of Statistics and Programme Implementation (

4 become more important determinants of industrial production and growth than was earlier the case. As in these studies, we find that state level policies, specifically regarding the entry of private engineering colleges, appear to play a very important role in the location of the Indian software industry. Unlike them, we focus on a single industry, albeit one that has become very visible in recent years, and one whose impact upon the Indian economy arguably far exceeds its share in GDP or employment (cf. Arora and Gambardella, 2005). Finally, this research is also related to the literature on human capital and regional growth. Many studies have found that the city s later growth is influenced by the initial level of human capital. Glaeser et al. (1995) find that human capital level in 1960 influences growth of the cities between 1960 and Similarly, Simon et al. (2002) found that cities that have higher level of human capital initially grow faster in the long run. The higher level of human capital may enhance the regional growth through direct as well as indirect channels (Mathur, 1999). For example, the existing stock of knowledge influencing creation of new knowledge (Simon, 1998). Regional differences in level of human capital also explain geographic differences in firm formation rates with regions endowed with higher level of human capital having higher firm formation rates (Acs, 2003). Our focus is not on regional growth per se, but on the growth of a particular industry, albeit an industry which is especially human capital intensive. As well, we do not explore mechanisms such as knowledge spillovers or new firm formation. Finally, this research is also related to the literature on industry agglomeration. Agglomerations can be resource-based or interaction-based (Wolter, 2004). Firms may gravitate towards a region or city if they have access to scarce resource (e.g. trained engineers, a key input for Indian software companies) and some firms may cluster to take advantage of technology transfers from the academic and research institutions. Individual firms may benefit from information spillover due to proximity to other firms within an agglomeration. However, Arora et al. (2001) find inconclusive evidence of agglomeration economies in Indian software industry though they recognize the possibility of it arising due to a thick labor market or good infrastructure. Our results support this possibility. The remainder of this paper is organized as follows. Section 2 describes the character, size and regional spread of software exports industry. Section 3 describes the size, growth of technical education in India, its regional dimension and importance of private sector. Section 4 describes the role of human capital in regional development and model of labor supply is developed. Section 5 develops empirical model. Section 6 presents summary statistics. We present results in section 7, and discuss important findings of paper in section 8.

5 II. The Indian Software Industry: Indian software exports were a mere US $4 millions in 1980 and rose to US $27.7 millions in 1985 (Heeks, 1996). The industry grew very rapidly in the decade of 90s and exports reached US $128 millions in 1990 and US $12200 millions in A key factor in the success of the Indian software industry is its absolute advantage in supply of skilled software professionals. Table 1 shows comparative salaries for software professionals in The estimated wage costs in India were about 1/3 rd to 1/5 th of the corresponding US levels for comparable work (Arora et al., 2001). The cost advantages are smaller for onsite work (done on the client s site overseas) but there is nonetheless potential for lower labor costs. The wages of an Indian employee working on client s site would be percent lower than wages of similar US employee 9. The projects executed in the offshore development center offer even greater cost savings to clients. The Indian firms have also succeeded in executing large projects, which has helped transition from mostly onsite work to a favorable mix of onsite and offshore work (Athreye, 2005b). Table 1:Software professionals: comparative salaries, 1997, US $ per year a Designation United States India b Programmer System analyst Programmer analyst Network administrator Database administrator Help-desk support technician Software developer Source: INFAC (1998), Mumbai, cited in Arora et al., a Figures are starting salaries for large establishments employing more than 50 software professionals. Salaries for a particular designation would vary due to factors such as educational and experience profile of the professional, platform of operation, nature of assignment (contract/full-time), location of the employer and additional technical/professional certification. b Converted at exchange rate of Rs /US$. The growth in software exports revenue has been due to increase in revenue per employee as well as growth in the number of professionals (see Table 2). The number of 9 Interview of Phiroz Vandrevala, Executive Vice-President of Tata Consultancy Services (TCS) in Business World ( 15 Source: accessed on 11/06/2005 at 6.51PM.

6 professionals (including IT enabled services and call centres) was merely 6800 in 1985 and increased more than ten fold times in the next five years to 80,000. The growth was at smaller pace in next decade and number of professionals rose to 841,500 in The number of professionals in the software exports sector has increased more slowly in recent years in comparison to those in the IT enabled services sector (ITES-BPO), from 110,000 in 1990 to 270,000 by Table 2: Growth of Manpower Employed in Indian Software Industry Software-export sector 110, , , ,000 Software-domestic sector 17,000 20,000 25,000 28,000 Software- in-house captive staff 115, , , ,000 ITES-BPO 42,000 70, , ,500 Total , ,500 Source: NASSCOM's Strategic Review of 2003, 2004, 2005 An important feature of the Indian software industry is preference for engineers. Arora et al., 2001, for instance, find that in 1998, over 80% of those employed by Indian software firms had engineering degrees. This is not just electrical or computer engineers but engineers in any discipline with a 4 years of undergraduate degrees, though often followed by specialized, non degree training in software tools. There are some important reasons why firms prefer undergraduate engineers. First, the high competition for admission to engineering colleges implies selectivity and the higher quality (Arrow, 1973). The undergraduate engineering degree thus acts a screening device. These engineering graduates also get good exposure to fundamentals of computers and also learn basic programming and some even master advanced programming language during the course of their studies thus reducing need for longer duration training. As well four years of engineering education imparts a set of problem solving skills, methods of thinking logically and learning tools that help quick adaptation to changes in technology, domains and tasks (Arora et al., 2001). Perhaps equally important, initially the bulk of software exports consisted of software professionals working on client s site in the US on temporary work permits, or H-1 B visas. US visa requirements meant that it was (and is) easier for engineers to qualify for H-1 B visas. Uneven regional growth of the Software Industry: In the very early period the software industry was mainly concentrated in Mumbai, the capital of the state of Maharashtra (Heeks, 1996) and as size of the exports grew industry started spreading to other cities and states. Bangalore attracted many multinational companies

7 after Texas Instruments set up its development center in By 1990 the states of Maharashtra (Bombay), Karnataka (Bangalore), Tamil Nadu (Chennai) and Delhi were the ones with large share of exports and states of Uttar Pradesh (NOIDA), Andhra Pradesh (Hyderabad) and West Bengal (Kolkata) also had software exports, albeit at lower levels, and were later jointed by the other states included in this study (see figures 1-3, and table A1). Initially, software exporting companies grew by expanding their operations in current locations. Many multinational companies (MNCs) set up their subsidiaries after foreign investment norms were liberalized by the federal government in 1991 (Athreye, 2005a). These MNCs locations were typically in the leading software centers such as Bangalore, Hyderabad, Chennai, Bombay and Delhi-NOIDA. Therefore the states which had head-start continued to grow rapidly in 1990s. This resulted in very heavy regional concentration of industry. Seven states contributed 95% of total software exports in , but only 48% of the country s population, 47% of the net state domestic product (NSDP) and 57% of the industrial production in the country. In the other seven states software exports are growing very rapidly as new firms and established large firms set up their development centers in these states. However, despite high growth rates the absolute size is still small. III. Undergraduate Engineering Education in India: In this section we discuss how India s undergraduate engineering education sector has evolved in past couple of decades. There are three main points. First, there is substantial regional variation in engineering college capacity, especially at the birth of the software export industry in the late 1980s. Second, this regional variation is due to differences in private engineering colleges. Finally, the differences in private engineering college capacity are due mainly to when the private colleges were allowed in the state. In India, higher education, particularly technical education, had been provided mostly by the government run institutions, except in last two decades. The majority of the institutions were set up and funded by various state governments. The number of institutions offering undergraduate degree in engineering has increased over the years as also the total intake capacity of these institutions 16. Table 3 shows the growth in intake capacity in undergraduate engineering colleges between 1951 and In 2004 the engineering college capacity was The engineering college capacity of a college/institution is the number of students it can admit in a given academic year in all the disciplines. There are discipline-wise upper limits fixed for each academic year by AICTE. Any increase in the engineering college capacity requires approval of AICTE. The All India Council for Technical Education (AICTE), set up in 1945 as an advisory body, was given statutory status in 1987 through an Act of Parliament. The AICTE grants approval for starting new technical institutions and for introducing new courses or programs.

8 times that of Even accounting for population growth, Table 3 shows the engineering college capacity per million of population grown from 13 in 1951 to 405 in Table 3 : Undergraduate capacity in Indian engineering colleges, Year Population in millions Engineering college capacity Engineering college capacity per million of population Source: Ministry of Human Resources Development, Government of India, AICTE, NTMIS. Regional Variation in engineering college capacity Table 4 shows the sanctioned intake capacity (for undergraduate engineering degree programs) by state. There are two important points to be noted about this table. We see a large inter-state variation in capacity. In fact share of four states of Andhra Pradesh, Karnataka, Maharashtra and Tamil Nadu was almost 75% in as compared to 29% of the population. As other states added capacity, the share of these states has declined, but continues to be around 63% in As table 4 shows, the growth in capacity has varied over time and across states. Consider the period from 1990 to Only three states, Karnataka, Maharashtra and Tamil Nadu were adding capacity. In other states the capacity did not increase perceptibly during these years. Also some states have experienced a sudden jump in the capacity, albeit in different years. Role of private self-financed colleges: The inter-state disparities in engineering college capacity are mostly due to differences in the timing and growth of the private sector colleges. Engineering college capacity in a state can increase by two ways: either by expanding capacity in existing institutions or by opening new institutions. The new institutions can be in the public sector or the private sector. Much of the actual increase has been through new private colleges. 19 Their contribution to engineering college capacity was similar even in

9 Table 4 : Sanctioned intake capacity in undergraduate technical institutions Year AP Delhi GJ HR KA KL MP MH OA PN RJ TN UP WB AP: Andhra Pradesh, GJ: Gujarat, HR: Haryana, KA: Karnataka, KL: Kerala, MP: Madhya Pradesh, OA: Orissa, PN: Punjab, RJ: Rajasthan, TN: Tamil Nadu, UP: Uttar Pradesh, WB: West Bengal In 1981, the vast majority of institutions were in the public sector, as was also true of engineering college capacity. Tuition fees were very low and the vast bulk of the expenses were met from the budgets of the respective state governments, with the exception of the few institutes and colleges directly supported by the central (federal) government. Budget constrained state governments faced severe limits on increasing capacity. Therefore capacity expansion in the public sector has been infrequent, and mostly limited to accommodating new disciplines such as computer science in 1990s and information technology in early 2000s. Furthermore, capacity expansion in existing institutions requires approval of All India Council for Technical Education (AICTE). AICTE limits intake capacity of a college right from its inception. For example the maximum capacity per discipline is 60 and college can have maximum of 4 disciplines in first year of its operation. The total capacity can increase by 60 to 300 in the second year and finally to a maximum of 420 in the fourth year. Any increase beyond 420 requires that an institution meets stringent quality standards. 20. The effect of regulation on capacity expansion combined with states (public sector) constraints in adding capacity means new private colleges have been the main source of growth. Analysis of college-level data between 1981 and 2004 support this conclusion. The number of colleges in the entire country increased from 246 in 1987 to 353 in 1995 and over 1100 in Eighty percent of new colleges added between 1987 and 1995 were in the private sector and the share of private colleges was even higher at 94 percent for colleges added between 1995 and Source: AICTE Handbook for Approval Process, and

10 Karnataka was among the first state to permit the private sector in undergraduate engineering education. The first such college opened in Karnataka in Thereafter one in 1962 and two in 1963 started their operation in the state. Then a large number of private colleges entered, beginning 1979, with nine colleges opening in 1979 and eleven in The first private college started in 1977 in Andhra Pradesh and in 1983 in Maharashtra after the government introduced policy permitting such institutions to operate. By 1986, only six states had such institutions. As a result, Andhra Pradesh, Karnataka, Maharashtra and Tamil Nadu accounted for almost 75 percent of total engineering college capacity in the entire country in Beginning in 1992, other states began to allow private self-financed institution and by 1999 all fourteen states had allowed private engineering colleges. As a result the share of private colleges has steadily increased over the years, from 62% in 1995 to more than 82% in IV. Regional Development and Human Capital: Many studies have found that a region s growth is influenced by the initial level of human capital. Glaeser et al. (1995) find that human capital level in 1960 influences growth of the cities between 1960 and Similarly, Simon et al. (2002) found that cities that have higher level of human capital initially grow faster in the long run. Thus initial level of human capital seems to be advantage cities and regions, perhaps by attracting knowledge-intensive industries. The regional differences in level of human capital also explain geographic differences in firm formation rates with regions endowed with higher level of human capital having higher firm formation rates (Acs, 2003). This is one mechanism through which initial differences in human capital can affect industry growth. We now develop a simple model of a two region economy that illustrates how local supply of human capital conditions the location of the software industry. Model of supply of engineers and size of industry: Let N1 and N 2 be the engineering stock in region 1 and 2. We assume that everybody joins the labor market and is willing to work at prevailing wages, w1 in region 1 and w2 in region 2. We further assume that the utility for engineer i from region 1 if they work in region 1 is w 1. The utility in region 2 for i would be w2 Ci, where Ci is the migration cost (the migration cost includes whatever utility loss there is from moving). Similarly, for engineer j educated in region 2, the utility is w2 when working in region 2 and w1 C j when working in 21 Manipal Institute of Technology (

11 region1. We assume that Ci and C j are all drawn from a distribution F. We do not specify lower bound for C and it can take negative values as well. Then, the fraction of people moving from region1 to 2 is F( w 2 w1 ), and the fraction moving from 2 to 1 is F( w 1 w2 ). Let x = w 2 w1. So, total labor supply in region 1 is N 1 F( x)) + N F( ), and N 1 F( x)) + N F( ) for 1( 2 x 2 ( 1 x region 2. Labor demand: There are M firms, which are price takers. We assume that firms can locate anywhere they want. We also assume free transport of output. They have some production function Q (L). It is immediate that with price taking firms, we must have w 2 = w1 in equilibrium. The way labor demand and supply will be equilibrated will be through the percent of firms in each region. So, total labor supply in 1 is N 1 F(0)) + N (0). If y percent of firms 1( 2F locate in region1, and if we normalize the price of the output p to 1, then each firm employs m (w) workers, given by Q '( m) = w / p = w (Recall w is same in each region). The labor demand is Mym in region 1 and M ( 1 y) m in region 2. This yields two equilibrium conditions for the labor market to clear in both regions: Mym w) = N (1 F(0)) + N (0) (1) ( 1 2F M 1 y) m( w) = N (1 F(0)) + N (0) (2) ( 2 1F 1 2 By adding (1) and (2), we get Mm ( w) = N + N (3) Equation (3) gives total demand. Substituting for Mm(w) in (1), we get N1 (1 F(0)) + N 2F(0) y = (4) N + N 1 2 If we let N1 = θ * N 2, then (4) becomes ( 1 F(0)) * θ + F(0) y = (5) 1+ θ In order to understand how share of firms, y respond to changes in capacity imbalance in two regions we differentiate (5) w.r.t. toθ, which equals dy 1 2F(0) = 2 dθ (1 + θ ) This means dy > 0 dθ (6) if F ( 0) < 1/ 2 (7)

12 The regional variations in employment opportunities, career growth prospects and cost of living differences, etc. may result in distribution of Ci being different for each region. Let Ci is distributed F(.) in region 1 and G (.) in region 2. In the equilibrium, the wages in two regions are same. Thus, dy > 0 dθ dy 1 ( F(0) + G(0)) =, so that 2 dθ (1 + θ ) if F ( 0) + G(0) < 1 (7 ) It is obvious from (7 ) that holding F( 0), G(0), N 2 constant, an increase in N1 would increase the share of firms y in region 1 provided G ( 0) + F(0) < 1 respectively. V. Empirical Specification and Results This simple model suggests that as long as there is some stickiness in the labor market, local endowments of human capital attract software firms. Engineers being the key input for software, the size of a state s engineering college capacity would be an important determinant in the location decisions of firms. In other words, it suggests that we specify a model with software exports as a function of the stock of engineers. We lack a measure of the stock of engineers in a state over time. However, our measure of engineering college capacity is arguably closely related to changes in the stock. Specifically, if there were limited mobility of engineers across states, then the growth in the stock of engineers in state i would be equal to the lagged engineering college capacity in the state. Since available anecdotal evidence suggests that such mobility is in fact small, we use this as a proxy for a change in the stock of engineers. To the extent that there is mobility, we have measurement error. As discussed in greater detail below, we also instrument for engineering college capacity to address biases due to measurement error, as well as problems posted by potential endogeneity. In other words, if S it are software exports in year t for state i, and K it is the corresponding stock of human capital, we have S it = a it + g t + βk it + ε it. (8) By taking first differences over time (represented by ) we have S it = a it + g t + β K it + ε it (9) Note that (8) allows for each state to have a different time trajectory for exports, so that the state effect varies by time. For feasible estimation, we assume that a it = α i i.e., states differ in both

13 the level and change in their exports. However, the change in exports per year (for a given state) does not systematically vary over time. Letting g t = γ t yields S it = α i + γ t + β K it + ε it (10) In other words, the annual increase in software exports are equal to a state fixed effect, a year effect and β times the engineering college capacity. This is our benchmark specification. However, we also estimate a related specification where we use as the dependent variable the level of software exports rather than the change. Although this specification is not theoretically as well grounded, it is plausible that with rapidly growing demand, local of firms should depend not merely on the level of human capital stock but also its growth. Both specifications find support in the data, as discussed below. Also, in both specifications, we exploit the variations in state policy allowing private engineering colleges to develop an instrument for engineering college capacity. Data We obtain data on engineering college capacity from the Annual Technical Manpower Review (ATMR) reports published by National Technical Manpower Information System NTMIS. These reports are prepared by a state-level nodal center of NTMIS and give details of sanctioned engineering college capacity and outturn for all undergraduate technical institutions in the state. The Handbook of Engineering Education, a publication of the Association of Indian Universities has also been used as a supplementary source. 23 Data on software exports are obtained from the Electronics and Comptuers Software Export Promotion Council (ESC), which is the apex government trade promotion organization for this sector, for the years For earlier years, ESC does not provide state level export data. Accordingly, we used export revenues by NASSCOM member firms, allocating the export revenues of each firm to the state where its headquarter is located. Till 1995, virtually all firms were located in a single state. Thus, this approximation is a reasonable one. As further described in the data appendix, we verified NASSCOM figures where possible from Dataquest, a trade magazine that covers Indian IT firms and provides data on sales, exports and employment for the leading firms, since For a couple of leading firms which operated in multiple states, we were able to obtain data on employment by state and allocated export revenues in proportion. Exploratory analysis does not indicate that that these two data sources yield different results, with the possible exception 23 In a few cases, the data from these reports are inconsistent, typically involving decreases or large increases in capacity or where capacity is markedly inconsistent with the number of graduating engineers. In such cases the other sources have been used to arrive at the acceptable figures of sanctioned intake.

14 of the switchover period where the change in software exports between 1998 and 1997 yields odd results for some states. Although the STPI is another potential source of state level export data, for the earlier years only a small fraction of software exports appear to by companies registered through the STPI, although towards the end, exports through the STPI are about over 90% of the software exports. Carrying out the analysis at the level of the state raises some additional issues. In particular, Delhi is bordered by the states of Haryana and Uttar Pradesh. Software exports from the latter two are concentrated very near their border with Delhi, in Gurgaon and Noida respectively. Since firms can move across the three locations, this results in large jumps and dips in software exports. We chose not to smooth the jumps and dips. Data on control variables like population, per capita power consumption, industrial output, teledensity, per capita income and number of students passing 12 th grade is obtained from various publications and websites of concerned departments of Government of India as detailed in the data appendix. Table 5 shows the descriptive statistics for the variables used in the regressions. Software exports, industrial output and per capita net state domestic product (NSDP) are in constant prices. Table 5: Summary Statistics Mean Std Dev Min Max Software export (Rupees Million, constant prices) Change in Software export (Rupees Million, prices) Intake Capacity (number) Outturn (number of graduating engineers ) Population ( 000s) Teledensity (no. of telephone lines/100 persons) Per Capita Power Consumption (Kilo Watt hours/year) Per Capita Income a (Rupees, prices) Industrial Output b,c (Rupees Million, prices) Electronics Production (Rupees Million, prices) No. of Students Graduating 12 th Grade N = 196. (14 states x 13 years. Some states have missing observations.) a b Lagged by four years. c Net value addition by all manufacturing units in a given state for each year. Results We begin with some simple descriptive relationships. Figure 1, which shows the log of software exports by state for three years, points the persistence of export leadership: states which were the early export leaders retain their leadership even after nearly a decade and a half. Figure 2 shows the relationship between software exports and engineering college capacity. For all each of the three years, we see a positive correlation between a state s engineering college capacity

15 and its software exports (in logs). Finally, figure 3 shows the change in software exports by state in various years against the year in which the state first allowed private engineering colleges. As can be seen, states which allowed private colleges to enter earlier are also those where the software exports have increased the most. These figures suggest that states which allowed private engineering colleges to enter early, were also favored early destination of software exporters, and their advantage appears to have persisted even as engineering college capacity in other states has rapidly expanded. Figure 1: Software exports by state, 1990, 1995, Log of Software Export (Rs BN of ) Karnataka Tamil Nadu Maharashtra Andhra Pradesh Delhi Haryana Uttar Pradesh State West Bengal Orissa Kerala Madhya Pradesh Gujarat Punjab Rajasthan Figure 2A: Software Exports and Engineering college capacity,

16 Log of Software Exports (BN of Rupees) dl wb upap kl hr pn oa rj mp gj tn mh ka Lagged Engineering Intake Capacity (in thousands) Graphs by year Figure 2b: Software exports and engineering college capacity, Log of Software Exports (BN of Rupees) dl rj hr wb up oa pn mp kl gj 2002 ap ka dl hr oa up wb kl gj pn rjmp Lagged Engineering Intake Capacity (in thousands) Graphs by year Legend: ap: Andhra Pradesh, dl: Delhi, gj: Gujarat, hr: Haryana, mh: Maharashtra, mp: Madhya Pradesh, ka: Karnataka, kl: Kerala, oa: Orissa, pn: Punjab, rj: Rajasthan, tn: Tamil Nadu, up: Uttar Pradesh, wb: West Bengal tn ap tn mh ka mh

17 Fig.3: Change in software exports , by year of policy change allowing private engineering colleges Log (Software Export Software Exports 1990) (Rs BN of KA TN MH AP DL HA UP Year of Policy Change Fig.4: Log of software exports for 1990, 1995, 2003, by year of policy change allowing private engineering colleges WB OA KL MP GJ PN RJ Log of Software Export (Rs BN of ) (KA) 1984 (TN) 1983 (MH) 1977 (AP) 1999 (DL) 1995 (HA) 1995 (UP) 1996 (WB) 1986 (OA) Year of Policy Change 1992 (KL) 1986 (MP) 1995 (GJ) 1993 (PN) 1998 (RJ)

18 We further explore these patterns through regression analysis. Consider first the long term impact of initial engineering college capacity. In table 6 we show the impact of engineering college capacity in 1987 on the level of software exports and also on the increase in software exports between 1990 and It is worth pointing out that the total software exports in 1987 were $54 million dollars (Athreye, 2005b) and therefore engineering college capacity in a state in 1987 was unlikely to be influenced by software exports industry. We control for several other factors that might have facilitated growth of the software exports in the state. Electronics production in 1990 reflects the size of hardware electronics industry before software industry achieved significant size. We include it our regression as it has been argued that in the initial years software industry also relied on experienced professionals working in the electronics industry to meet its manpower requirement (Lateef, 1997). This also controls for a variety of other influences. For instance, Klepper has argued that related industries are more likely to spawn successful firms. However, firms in many sectors, such as Banking, Finance and Manufacturing are also significant producers of software (primarily for their own use), this factor may not be as salient. Accordingly, we control for industrial production as well. Table 6: Initial differences in engineering college capacity and software exports Effect of Engineering Capacity in 1987 on Difference in Software Export in 2003 and 1990 Log (Software Exports Software exports ) Software exports 1990 Eng. Coll. Capacity (1.00) Electronics Production (0.50) Lagged Industrial Output (0.15) Constant 6096 (4956) R No. of obs. 14. As table 6 shows, initial levels of college capacity in the state have a significant and sizable effect upon software exports in 2003, nearly a decade and a half later. The limited number of observations makes it difficult to use more controls, and it is possible that this long lasting influence is merely a reflection of unobserved state characteristics. Accordingly, we exploit the within state-variation in capacity over time in table 7, which use both year and state fixed effects. In addition, we control for per capital income, and teledensity, and use the state s population to control for size effects.

19 Table 7 reports on the specification implied by equation (10). Note that we lag engineering college capacity by four years as it takes four years to complete undergraduate engineering degree. This makes it unlikely that our effects reflect the feedback effect of software growth, except possibly towards the end of our sample period. Other controls such as electronic production, industrial output, per capita income and teledensity are also lagged, albeit by one year. Further, the standard errors are cluster corrected to account for the non-independence of errors within a state. Table 7 shows two specifications, with and without time varying state characteristics such as per-capita income, population and electronics and industrial production. Although controlling for time varying characteristics reduces the impact of engineering college capacity on software exports. In specification 1, a unit increase in capacity increases exports by Rs 340,000 or about $8,000. To put this in perspective, average revenue per employee in the software industry in India in the mid 1990s was of the order of $15,000. If one takes into account the less than full capacity utilization, attrition, movement into non-software activities and migration to other states, the quantitative impact is even more plausible. Table 7: Dependent Variable: Annual change in software exports, in Rs millions, constant prices (1) (2) Lagged Eng. Coll. Capacity 0.34 (0.1) 0.20 (0.07) Lagged Electronics Production 0.40 (0.24) Lagged Industrial Output (0.023) Lagged Per Capita Income (0.61) Population (0.16) Constant -371 (1308) (11914) State-fixed effects Yes Yes Year-fixed effects Yes Yes R Note: Cluster corrected std. errors in parenthesis. No. of obs Despite the state fixed effects and despite the four year lags, it is possible that capacity is endogenous. A growing software industry may create the expectation of growth in future demand for engineers. This will bias our estimate upwards. On the other hand, it is likely that capacity is an imperfect measure for the change in the state level stock of human capital, which

20 would imply an attenuation bias to zero. To address both issues we therefore use instrument for engineering college capacity. Our instrument is based when neighbouring states first allow private engineering colleges to operate. Specifically, we create a dummy variable, policy, for each state, which is one if private engineering colleges are operating in that state in that year, and zero otherwise. Our instrument is the average of policy for neighbouring states. This identification strategy assumes that when neighbouring states allow private colleges to operate may depend upon anticipated demand in that state, it is independent of the anticipated demand in neighbouring states. Put differently, we assume that although state governments may respond to local software firms, they are not responsive to software firms in neighbouring states. However, they do respond to policy changes in neighbouring states. It was not until the 1990s that Indian states actively began to compete to attract businesses to locate. Before that, states frequently viewed private business with some suspicion. More business friendly states might, prior to the 1990s, have offered tax concessions or regulatory relief, they were extremely unlikely to offer to make significant policy changes to address business concerns. Thus, this is a plausible instrument. Table 8a shows the results of the first-stage regression of college capacity on average neighbour policy, with state and year fixed effects, with and without time varying controls. Though the neighbour policy measure is significant, the F statistic is only around 4.5 with time varying controls, and around 6 without them, although these are after cluster correction. Table 8a: First-stage Results Dependent Variable: Lagged Eng. College capacity (1) (2) Average neighbour policy (2618) (2429) Lagged electronics production 0.37 (0.18) Lagged industrial output 0.06 (0.03) Lagged per capita income 0.1 (0.82) Population (0.11) Constant (11098) (4918) State-fixed effects Yes Yes Year-fixed effects Yes Yes F-statistic R Note: Cluster corrected std. errors in parenthesis. No. of obs. 182.

21 Table 8b: Two Stage Least Square Results Change in Software Exports (2SLS) (1) Lagged Eng. College 0.62 Capacity (0.36) Lagged Electronics Production Lagged Industrial Output Lagged Per Capita Income Change in Software Exports (2SLS) (2) 0.74 (0.50) 0.21 (0.23) (0.05) (0.67) Software Exports (OLS) (3) 1.31 (0.27) 1.60 (0.80) (0.08) 0.98 (1.68) (0.56) Software Exports (2SLS) (4) 2.02 (1.95) 1.35 (0.56) (0.20) 0.83 (1.54) Population (0.14) (0.35) Constant (4489) (11527) (37591) (34762) State-fixed effects Yes Yes Yes Yes Year-fixed effects Yes Yes Yes Yes R Note: Cluster corrected std. errors in parenthesis. No. of obs. 182 Table 8b presents the corresponding instrumented estimates. Columns 1 and 2 present results where the dependent variable is change in software exports over the previous year. Note that the estimated coefficient increases three fold as compared to the OLS estimate, suggesting that upward bias was unlikely and that measurement error is more likely. However, the coefficient is imprecisely estimated, possibly because the instrument is weak. We also present analogous results where we use the log of software exports, rather than the annual change in software exports, in columns 3 (OLS results) and 4 (2SLS). The estimated coefficient of capacity in this case also increases upon instrumenting for it, although the increase is not as large. Once again, the estimated coefficient has a large standard error. The other noteworthy point is that lagged electronics production also has a positive and significant impact on the level software exports, but not on the annual change in software exports. Other time varying controls are statistically insignificant. Specifically, per capita income and industrial output do not play any role in explaining software exports growth. Consistent with this, some of the richer states like Punjab and Gujarat lag in software exports compared with Karnataka and Andhra Pradesh. VIII. Discussion and conclusions: The importance of human capital -- skilled and creative workers -- to a high-tech industry is routinely acknowledged but often public policy discussions tend to focus on more trendy prescriptions such technology parks, venture capital, incubators and university industry

22 centers. Software, perhaps more than any other high-tech industry, relies more intensively upon human capital. Software services, the engine of the Indian software sector, is arguably even more human capital intensive than software products. Thus, few would question the role of human capital stocks in the rise of the Indian software industry. What is less clearly appreciated is that there are significant variations across Indian states in stocks of human capital, specifically engineers, and that these have played an important part in conditioning where the software industry has flourished. Even less well understood are the reasons for this regional disparity in human capital stocks. We find that these variations exist even after controlling for factors such as how rich or large the state is, and measures of industrial production, electronics production or telecommunication investment. Since engineering education has been controlled, and in the main, provided by state funded colleges, differences in the willingness of states to invest in engineering colleges could, but do not, explain the bulk of the inter-state variation. Instead, it is the role of private engineering colleges which is the key the puzzle. Simply put, states which allowed private engineering colleges to enter early were able to get a head start and, this early advantage has persisted for nearly a decade and a half. This persistence is noteworthy, and may be the result of a number of factors. Klepper s studies of automobiles, tyres and televisions indicates that early entrants are likely to be the source of eventual market leaders, accounting for large shares of output and revenue (see Arora, Gambardella and Klepper, 2005). In the Indian software industry, there is evidence that larger firms are able to attract larger (and more lucrative) contracts (Athreye, 2005). Thus, our results are consistent with the notion that states relatively abundant in engineering talent, by attracting software production, were able to ride the coat-tails of these firms, as (some) grew into the industry leaders. These effects could be amplified if employees of existing software firms are also a significant source of new firm formation and tend to locate in the same region. The persistence also points to the potential importance of positive feedback forces. Gain in reputation is an obvious one, as exemplified by the tremendous positive publicity Bangalore has received. It is possible that software firms located in Bangalore had an easier time attracting foreign contracts as a result. Similarly, a firm entering the software industry in Bangalore may have found it easier to attract talented and experienced software developers and project managers. Finally, as noted earlier, the enhanced visibility of the industry in the leading states may have attracted eager students and equally eager educational entrepreneurs wanting to educate them, to those leading states. Indeed, a topic for further research is indeed to model the process by which educational entrepreneurs decide to start an engineering college.

23 Our results assume limits on geographical mobility of software workers. Potential software workers can change location at two stages, once when they decide upon where to go to college, and then after graduating from college. Local levels of human capital matter to the extent workers are imperfectly mobile after graduating. Alternatively, it must be assumed that software firms incur lower search costs (or matching costs) with local graduates. The empirical and theoretical foundations for these assumption remains for future research. Our unit of analysis is state even though the industry we are analyzing is mostly located in urban centers in India. The possible extension of this work would also include developing theoretical model for modeling interregional migration for engineering education and for work and its effect on regional labor supply and consequently on growth of the industry. Our results corroborate the findings in the literature about the increasingly important role of institutional and policy differences across states. But they also illuminate how successful policies are rapidly imitated. Every state in our sample had adopted policies allowing the entry of private engineering colleges by 1998!. This underscores the importance of economic decentralization in a large country such as India, and of allowing states to act as laboratories for policy experiments. Reference: Acs, Z.J. and C. Armington (2004), The impact of geographic differences in human capital on service formation rates, Journal of Urban Economics, Vol. 56, pp Aghion, Phillippe, Robin Burgess, Stephen Redding, and Fabrizio Zilibotti (2006), The Unequal Effects of Liberalization: Evidence from Dismantling the License Raj in India, National Bureau of Economic Research, Working Paper All Indian India Council for Technical Education (AICTE) (2006), Approved Institute, accessed on 02/14/2006. Arora, A. et al (2001), The Indian Software Industry, Research Policy, Vol. 30 pp Arora, A., et al. (2005a), Organizational Capabilities and the Rise of Software Industry in the Emerging Economies: Lessons from the History of some US Industries, in Ashish Arora and Alfonso Gambardella (Ed.), From Underdogs to Tigers, Oxford University Press. Arora, A. and Gambardella, A., (2005b), Bridging the Gap: Conclusions, in A. Arora and Alfonso Gambardella (Eds.), From Underdogs to Tigers, Oxford University Press. Arrow, K. (1973), Higher Education as a Filter, Journal of Public Economics Vol. 2, pp Athreye, Suma S. (2005a), The Indian Software Industry, in A. Arora and Alfonso Gambardella (Eds.), From Underdogs to Tigers, Oxford University Press. Athreye, Suma S. (2005b), The Indian Software Industry and its Evolving Service Capability, Industrial and Corporate Change, Vol. 14, pp Banerjee, A.V. and Esther Duflo (2004), Growth through the Lens of Development Economics, MIT, Department of Economics. Bhaumik, Sumon K., Shubhashis Gangppadhyay, and Shagun Krishnan (2006), Reforms, Entry and Productivity: Some Evidence from the Indian Manufacturing Sector", Institute for the Study of Labor (IZA), Discussion Paper No Carrington, W.J. et al. (1996), Migration with Endogenous Moving Costs, The American Economic Review, Vol. 86, pp