2 Telecommunications Infrastructure and Economic Growth: A Cross-Country Analysis M A X I M O TORERO, S H Y A M A L K. CHOWDHURY, A N D A R J U N S . B E D I In recent years, the potential of information and communications technologies (ICT) to facilitate economic development, especially in low-income countries, has attracted considerable attention. Several commentators (Pohjola 2001, for example) have argued that development of these new technologies, in terms of proliferation and accessibility, should be integral to country-level development strategies and that ICT investments are essential in the process of enhancing liv­ ing standards. Nevertheless, detractors also exist. They counter with the primary argument that developing countries have far more pressing investment priorities, and that investing scarce resources in ICT does not fulfill the needs of the poor (Roche and Blaine 1996; Saith 2002). Whether additional investments in these technologies are justified and whether they have the potential to increase incomes and alleviate poverty can be determined only by empirical investigation. I f these new technologies are to command the continued interest of the developing world and justify additional investments, a convincing demonstration of their effects on economic perform­ ance is required. On this basis, and given limited existing empirical evidence, this chapter examines the impact of terrestrial telecommunications infrastructure— by far the most prevalent communication technology in developing countries— on aggregate economic output. The conceptual framework that forms the basis of the book is once again presented in Figure 2.1, this time highlighting the area of analysis dealt with in this chapter. For the purposes of the empirical investigation, a substantial data set cov­ ering 113 countries was assembled for the period 1980-2000. The empirical framework explicitly accounts for the two-way relationship between tele­ communications infrastructure and economic output (meaning that the frame­ work factors in issues of endogeneity). The four-equation framework used by Roller and Waverman (2001) was adopted and, in the first instance, replicated. However, the estimation methodology and empirical work in this study goes beyond Roller and Waverman's specification. In particular, this study examines the time-series properties of the data set used and corrects for the presence of unit-roots (an attribute of the time series within the statistical model). 21 22 Maximo Torero et al. F I G U R E 2.1 Conceptual framework: Area of analysis dealt with in Chapter 2 Impact Driving Supply (teledensity) forces and design •a 9 _ Infrastructure and /a te , io na ! service provision Pu bl ic , p ri ' in te rn al t Pu bl ic , p ri ' in te rn al Content Demand (utilization) Area under analysis in Chapter 2 Impact at the global level Impact at the microeconomic level NOTE: Teledensity indicates the number of telephone lines per 100 inhabitants. There are several ways that a country's ICT infrastructure can affect eco­ nomic growth. Apart from making a direct contribution to gross domestic prod­ uct (GDP), investments in these technologies are assumed to have pervasive impacts throughout the economy—for example, by reducing transaction costs, improving organizational functioning, and enhancing the spread and develop­ ment of factor and product markets (along with many other potential effects that have already been outlined by several authors, such as Saunders, Warford, and Wellenius 1983; Leff 1984; and Norton 1992), and therefore are not elaborated on further here. It is worth noting, however, that ICT is unlike other forms of infrastructure in that its expansion generates benefits for both new and existing users. This externality suggests that the effects of ICT on economic growth may be subject to the attainment of a critical mass, meaning that, unless the said infrastructure reaches a certain minimum level within a given country or region, the growth effects may not be discernible. Put in simple economic terms, a non­ linear relationship may exist between telecommunications infrastructure and economic growth. Thus, the empirical work in this chapter addresses two is­ sues. First, does telecommunications infrastructure have an impact on economic growth? Second, i f a growth effect does exist, how does it vary with infra­ structure and income levels across countries and regions? This chapter presents a summary of the literature on infrastructure and growth, a description of the data used (presenting correlations between GDP Telecommunications Infrastructure and Economic Growth 23 and the availability of ICT), and conclusions. For those interested, the detailed econometric model and empirical results are included in appendix form at the end of the chapter. A Review of Infrastructure Development and Growth Overview Early work on economic growth and development highlighted the necessity of adequate infrastructure. Defining the scope of social overhead capital (SOC), Hirschman (1958, 89) writes "SOC is usually defined as those services without which primary, secondary, and tertiary production activities cannot function." He goes on to write that in its wider sense SOC "includes all public services from law and order through education and public health to transportation, communi­ cations, power, and water supply." Although infrastructure implies a wide variety of services, such services have several traits in common. First, while these ser­ vices yield direct benefits, it is often their indirect contribution—as intermediate inputs enhancing the productivity of all other inputs—that is often considered more important. Second, the development of these services is usually subject to increasing returns to scale. Third, while private-sector participation in the provision of infrastructure has increased, it is still largely funded and provided by the public sector. The rationale for public provision of these services is well known and is usually justified on the basis of a combination of externalities, nonrival consumption, and nonexcludable characteristics of such services. The appropriate level and composition of public expenditure on different types of infrastructure services is an active area of debate, particularly because of their potential to spark growth and influence economic outcomes. For the purposes of assisting and informing public expenditure decisions, it is impor­ tant to know the overall impact of infrastructure on economic output, as well as the relative effects of different types of infrastructure. The importance of this question for public policy has motivated a number of authors to examine the macroeconomic link between public capital and output. While there are micro- oriented studies of the infrastructure-growth output link, the economywide effects and potential externalities ascribed to such investments suggest the need for a macroeconomic approach.1 The infrastructure-output literature consists of several branches; this study, however, reviews only the relevant portions of the existing evidence. The discussion initially focuses on single countries, either over time or across regions, examining the effect of overall infrastructure capital on private 1. Micro-oriented studies lead to a deeper understanding of how infrastructure enhances out­ put, but when the main goal is to establish the overall effect of infrastructure capital on output and productivity, a macroeconomic analysis is the more logical choice. 24 Maximo Torero et al. output. This is followed by a look at cross-country evidence of the impact of infrastructure capital on economic performance (meaning GDP growth over time). This review of the more general infrastructure literature is followed by a review of studies focusing on the links between telecommunications and output. General Infrastructure Literature Despite the existence of estimates of the link between infrastructure and output that predate Aschauer (1989), interest in the effects of public capital on output can be traced to this work. Aschauer used annual U.S. time-series data for the period 1949-85 and estimated that a 1 percent increase in the ratio of public to private capital stock was associated with a 0.39 percent increase in private- sector total factor productivity (TFP). A similar effect was reported by Munnell (1992). Such large effects—at a time when growth in productivity and in the stock of public infrastructure was declining—prompted Aschauer to propose that the declining growth of public infrastructure was an important determinant in the productivity slowdown. To examine whether these results could be ex­ trapolated more generally, Ford and Poret (1991) used time-series data from several countries of the Organisation for Economic Co-operation and Develop­ ment (OECD) to estimate production function relationships as similar as pos­ sible to those studied by Aschauer. Ford and Poret's estimates of the elasticity of output from infrastructure stock spanned a wide and implausible range (from 1.00 for Canada to -0.55 for Norway) and did not support the idea that a de­ cline in infrastructure growth was responsible for the decline in TFP. Despite this lack of support for the infrastructure-productivity link from studies based on other countries, support for Aschauer (1989) was provided by studies that relied on panel data to examine the effect of publicly provided in­ puts on the economic performance of U.S. states. Although his results were smaller in magnitude, Munnell (1992) reported that public capital was a statis­ tically significant link in the determination of differences in productivity across U.S. states. Rather than focusing on public capital in its entirety, Garcia-Mila and McGuire (1992) used observations from the 48 contiguous states during 1969-83 to examine the effect of the stock of highway capital and educational expenditure on gross state product. While their results were smaller than the elasticities based on time-series data, they did show that these two variables played a substantial role in explaining statewide productivity differences. The preceding studies focused exclusively on developed countries; the paucity of data makes it difficult to carry out single-country studies for devel­ oping countries. Clues to the link between infrastructure stock and economic de­ velopment and growth in developing countries stem mainly from cross-country studies. Antle (1983) uses data from a sample of developing and developed countries to examine the extent to which intercountry differences in agricultural productivity can be explained by country-level investments in transportation and communications. For both developing and developed countries, Antle's Telecommunications Infrastructure and Economic Growth 25 analysis supports the conclusion that additional investments in infrastructure play a larger role than agricultural research and education in explaining inter- country differences in agricultural productivity. The broad inference from the above studies is that the stock of public infra­ structure plays a causal role in determining growth and productivity in both de­ veloping and developed countries. Policy implications drawn from such studies are fairly clear and support the argument for additional investments in physical infrastructure. Nevertheless, policy prescriptions based on the results of these studies have attracted strong criticism, primarily on the basis that they do not adequately account for the possibility of reverse causality and simultaneous determination of output and public capital—meaning that, while it is tempting to infer a causal relationship from public capital to output, it is equally likely that the direction of causality goes from output to public capital. In the context of time-series analysis, the estimated coefficient may reflect a spurious correlation between output and public capital stock that is driven by a common time-trend and not by any underlying relationship between the two variables. In short, the data may be what is economically termed "nonstationary" and inferences based on these data may be misleading. Another problem primarily afflicting panel- data studies stems from omitted variables. The first-generation panel-data studies usually ignore the possibility of unobserved state- or country-specific variables that may influence both output and the stock of public capital. Ignor­ ing such fixed effects is quite likely to lead to an exaggeration of the effect of infrastructure on output. The more recent literature in this area accounts for both of these problems —simultaneity bias and fixed effects—and presents a rather different picture. Holtz-Eakin (1994) uses panel data from the United States to estimate a variety of production functions in levels and in first differences. These estimates allow for fixed and random effects. Regardless of variations in the specifications, Holtz- Eakin finds no evidence that public capital is involved in productivity differ­ ences across states and concludes that the previous large positive findings "ap­ pear to be the artifact of an inappropriately restrictive framework" (Holtz-Eakin 1994, 20). In a reconsideration of some of their earlier studies, Garcia-Mila, McGuire, and Porter (1996) use a panel data set drawn from the United States to estimate the effect of public capital investments in highways, water and sewage systems, and all other infrastructure investments on private output. Their results confirm the conclusions reached by Holtz-Eakin (1994). From a cross-country perspective, Craig, Pardey, and Roseboom (1997) control for country-level fixed effects and conclude that, for developing countries, differences in road density are not responsible for differences in agricultural productivity. Thus, in marked contrast to the first-generation studies, these second- generation studies find no evidence of high returns to investments in infra­ structure. Despite these results, it would not be correct to argue that investments in infrastructure are not necessary. These studies show that within the context 26 Maximo Torero et al. of a narrow production-function framework there are no indirect productivity effects associated with public capital. These results do not, however, detract from the large direct effects that investments in infrastructure services generate. Furthermore, any investment in public infrastructure calls for a project-specific cost-benefit analysis and should not be based on aggregate analysis. Telecommunications Infrastructure Literature Turning to studies more closely related to the topic of this chapter, Jipp (1963) and Hardy (1980) are some of the earliest that focus on the telecommunications- economic growth link.2 Hardy (1980), for instance, uses data from 15 developed and 45 developing countries for the period 1960-73, regressing per capita GDP on lagged per capita GDP, lagged telephones per capita, and the number of ra­ dios. Hardy's results support the idea that the greater availability of telephones has a positive effect on GDP. By demonstrating the strong correlation between teledensity (the number of telephone lines per 100 inhabitants) and GDP, these early studies drew attention to the potential role of telecommunications in in­ fluencing growth. However, similar to the first-generation infrastructure litera­ ture, these early studies ignored the econometric issues outlined above. A more sophisticated example of this genre is Norton (1992). Using data from a sample of 47 countries from the post-World War I I period until 1977, Norton investigates the effects of telephone infrastructure on growth rates and also attempts to identify the channels through which the availability of this infrastructure leads to growth (that is, the effect of telephone infrastructure on the mean investment ratio, and consequently income growth). The empirical framework replicates Kormendi and Meguire (1985) but includes additional variables to capture telecommunications infrastructure. The inclusion of a more comprehensive set of macroeconomic regressors is designed to reduce the possibility of overestimating the effect of telecommunications infrastructure on growth. The two telecommunications infrastructure measures used are tele­ density in 1957 and mean teledensity over the sample timeframe. The use of these two infrastructure variables is an attempt to address the endogenous nature of teledensity and growth. Norton argues that a measure of teledensity prevalent during the early years of the sample is less susceptible to endogeneity bias than a variable that captures the mean teledensity over the entire time period. 2. An emerging body of literature that examines the effect of the stock of computer hard­ ware, software, and labor and other information technology-related measures in influencing eco­ nomic growth and output. Since the main focus of this study is developing countries, where the stock of such capital is quite small, this body of evidence is not reviewed in detail here. Recent macroeconomic evidence on the United States is provided by Gordon (2000), Oliner and Sichel (2000) , and Stiroh (2002). Results based on data from other countries and a cross-country analysis of the effect of information technology expenditure on economic growth are available in Pohjola (2001) . Telecommunications Infrastructure and Economic Growth 27 Norton's results show that the two measures of telecommunications infra­ structure are statistically significant and exert positive effects on mean growth rates. For instance, increasing the 1957 teledensity by one standard deviation (9.909) leads to an increase in mean GDP growth of around 0.73 percent. The effect of the average density variable is greater but potentially more suscep­ tible to reverse causality. The second set of estimates examines the effect of the telecommunications infrastructure variables on the mean investment-output ratio, and, similar to the earlier results, the impact is positive. Increasing tele­ density by one standard deviation leads to an increase in the investment ratio of around 3.5^4.5 percent. While Norton controls for the endogeneity of telecommunications infrastructure and growth, the large effects reported in the study are reminiscent of the effects reported in early infrastructure literature. Norton does not control for country-level fixed effects, and it is likely that this omission is responsible for the large estimated effects of telecommunications infrastructure. Madden and Savage (2000) follow the framework used by Mankiw, Romer, and Weil (1992) to estimate the effect of telecommunications on the level and growth of GDP for a cross-section of 43 countries (including 16 developing countries) for the period 1975-90. The authors present estimates based on or­ dinary least squares (OLS) and mention that they use instrumental variables estimation to control for the possible endogeneity between telecommunications capital and GDP. Their results are not sensitive to alterations in the estimation methodology, and they report large effects of telecommunications capital on the level of GDP. Once again their estimates may be upwardly biased because they do not control for country-level fixed effects. As mentioned in Chapter 1, Roller and Waverman (2001) present cross­ country estimates of the effect of telecommunications on output based on a cross-section of 21 OECD countries over a period of 20 years. They tackle the endogeneity problem by estimating a four-equation model that endogenizes telecommunications infrastructure, and they control for country-level fixed ef­ fects. Their results indicate that estimates allowing for fixed effects lead to a reduction in the teledensity elasticity from 0.15 to 0.045. Despite this drop, the growth effect attributed to telecommunications infrastructure is still quite large. The elasticity implies that about one-third of the economic growth in OECD countries between 1971 and 1990 may be attributed to growth in telecommu­ nications infrastructure. An interesting element of Roller and Waverman's work is the investigation of whether a nonlinear relationship exists between tele­ density and economic output. They found that teledensity begins to exert an influence on output only when it is universally available—more specifically, when a teledensity threshold or critical mass of about 40 percent is reached. While the robustness and generality of the threshold effect may be ques­ tioned, these results do suggest that enhancements in telecommunications in­ frastructure may generate higher growth effects in developed countries than in 28 Maximo Torero et al. developing countries. In addition, given low teledensity in developing countries (the 1995 average was around 4.0), it appears that marginal improvements in telecommunications infrastructure may not generate the desired growth effects. Thus, developing countries may require substantial investments in telecommu­ nications infrastructure before they can benefit from the growth-generating ef­ fects of these technologies. In this chapter we use a framework and a specification that is as similar as possible to Roller and Waverman (2001). We endogenize telecommunications by estimating a demand and supply model for telecommunication investments and simultaneously estimate a macro-production function. Our primary aim is to examine whether the idea of a critical mass is valid for a larger sample of nations and for a longer timeframe. Our analysis emphasizes an examination of this relationship for developing countries. Data and Summary Statistics To construct a comprehensive data set, the inclusion of a country was determined by the availability of time series data of long duration covering the variables re­ quired for the analysis. The countries included are presented in Appendix 2A, which also provides details of the income categories and the regional groupings used.3 The data set for this chapter is tailored to the needs of the empirical frame­ work and contains information on economic variables such as output, labor force, capital stock, and budget deficit (surplus). The telecommunications-related variables are teledensity, revenue per fixed telephone line, and annual invest­ ment in telecommunications. In addition to these telecommunications-related variables, the data set also contains information on the availability of other ICT. Table 2.1 provides a list of the variables and their descriptive statistics, also indicating the sources of the variables. With respect to variable sources, while most of the variables are readily available from publicly accessible databases, such as the World Development Indicators (WDI) and the International Telecommunications Union (ITU) data­ base (World Bank 2002 and ITU 2002), some of the data—in particular tele­ communications capital stock and total physical capital stock—required estima­ tion. Construction of the telecommunications capital stock series is based on annual investment in telecommunications data available in WDI (World Bank 3. In terms of income categories, our data set consists of 36 low-income countries, 27 lower middle-income countries, 21 higher middle-income countries, 8 high-income non-OECD coun­ tries, and 21 OECD countries. We divided the countries across six geographic regions based on WDI classifications (World Bank 2002). The data set comprises 40 countries from Africa, 24 coun­ tries from the Asian/Middle Eastern region, 4 countries from the Australian region, 23 countries from the European region, 13 from the Central and North American region (including the Carib­ bean), and 9 countries from South America. Q t/3 o t~- o O ON r- 00 O N co in n OO 00 >n NO in CN 00 o © © o © o O © © o © o ON oo CN CN O N oo CN ON * — • r- CN o CN O N '—< CN n >n CN CO CO o © 00 OO CN < 00 © CO ^ o CN d CN d CN d ON CN NO r- o CO CN ON o CO 00 CN CN in O ON r- ON NO CO "X •—^ N O CN CN —T >n" r-" O N oo" CN" r- CO oo" CN CN 00 NO >n CO O N O N N O in CN r- oo ON NO N O CO O t- CN 'r. o CO CO O N O N C N CO C N i> I - i co" co" ON" NO" oo" in" co" in" •o 3 6 T 3 o X S J3 (Hi 32 Maximo Torero et al. F I G U R E 2.3 Availability of I C T in various regions, 2000 (Millions) 20 • ISPs 0 Africa Asia Australia Europe North America South America SOURCE: Calculated by authors from study data set. NOTE: For details of the countries within the six regional groupings, see Appendix 2A. ISP indi­ cates Internet service provider; PCs, personal computers. in 2000 was 25 times the low-income country average, and the number of cel­ lular telephones in Indonesia in 2000 was more than 12 times the low-income country average. In addition to limited variety of ICT available in low-income countries, an imbalance also exists between the availability of fixed telephone lines and the availability of more recent technologies such as the Internet. While the ratio be­ tween fixed telephone lines and the Internet was 1.85 to 1 on average in OECD countries in 2000, it was 5 to 1 in low-income countries. Figure 2.3 shows the availability of ICT across the six different regional categories for the year 2000, based on the country average per region. Like the income groups, there are regional differences in ICT levels. While countries in Africa have the lowest penetration rates on average, countries in the Central and North American region have the highest. For example, while the number of Internet hosts in Africa is 0.59 percent of the world average, the number in the Central and North American region is seven times the world average. In Europe, in contrast to other geographic regions, cellular telephony already outnumbers fixed telephony. These averages mask individual differences within regions. Not surpris­ ingly, the larger the country (or countries) within a region in terms of area or population, the larger the stock of ICT. In the case of Africa, for example, the number of Internet hosts in South Africa is 35 times the African average, and the number of fixed telephone lines in Egypt is 12 times the African average; in Telecommunications Infrastructure and Economic Growth 33 the case of Asia, the outliers are Japan and Korea for all types of ICT and India and China for fixed telephony. Differences in the other regions also follow income and country size (either by area or population). Table 2.2 shows compound an­ nual growth rates (CAGR) of per capita GDP and fixed line teledensity for the period 1980-2000. Over this time, the 113 countries (representing "the world") reported positive growth in both per capita income and fixed telephony. While per capita income grew at 1.1 percent on average, the growth rate for fixed telephony per 100 inhabitants was 6.2 percent. The last column of the table in­ dicates the correlation between the CAGRs for per capita GDP and fixed line teledensity (Table 2.2). From 1980 to 2000, per capita income in low-income countries virtually stagnated, while all other income groups experienced substantial growth. A l l income groups achieved growth in fixed telephony per capita, but the growth rate was higher in low-income than in high-income countries. Despite this, the per capita number of fixed telephone lines in low-income countries was only 1.8 in 2000 compared with 58.2 in high-income countries. The growth rate of fixed telephone lines was higher in middle-income countries than in either low- or high-income countries, and per capita GDP growth was also higher for middle-income countries compared with low-income countries. Despite the dif­ ferences in growth rates among the various income groups, for all groups—with the exception of high-income non-OECD countries—a positive and statistically significant correlation was demonstrated between telecommunications infra­ structure and GDP growth (Table 2.2). Some regional differences were evident. During 1980-2000, per capita GDP in the Asian and European regions grew at a higher rate than in the coun­ tries of other regions. African countries lagged behind, followed by the countries of South America. Asian countries displayed high per capita GDP growth, as well as attaining the highest per capita growth in the number of fixed telephone lines. African countries also achieved a high rate of growth in fixed telephony over the same period; nevertheless, in 2000, fixed line teledensity was only 3.6 in Africa compared with 19.2 in Asia. In South American countries the tele­ density increased substantially over the 20-year period, reaching 14.1 in 2000. Although there are differences across regions, and with the exception of the Australian region, a high positive correlation exists between per capita GDP growth and per capita fixed telephony growth for all regions (Table 2.2).4 Figures 2.4a and 2.4b show the relationship between fixed line teledensity and per capita GDP for the years 1980 and 2000, respectively. In both figures, 4. When the four countries comprising the Australian region—two OECD countries (Aus­ tralia and New Zealand) and two lower middle-income countries (Papua New Guinea and Fiji)— were separated into their income groups, the result produced a correlation coefficient that was positive, significant, and close to one. F I G U R E 2.4a Fixed telephone lines and per capita gross domestic product, 1980 GDP per capita (1995 US$) 50,000 | F I G U R E 2.4b Fixed telephone lines and per capita gross domestic product, 2000 GDP per capita (1995 US$) 60,000 80 Fixed telephone lines per 100 inhabitants SOURCE: Calculated by authors from study data set. Telecommunications Infrastructure and Economic Growth 35 per capita GDP expressed in constant 1995 U.S. dollars is the dependent vari­ able, and fixed line teledensity is the independent variable. In 1980, there was a positive relationship between growth of fixed telephony and per capita GDP; a similar positive relationship was found to exist in 2000 as well. Although both per capita GDP and fixed line teledensity grew at different rates (1.1 percent per year and 6.2 percent per year, respectively), the relationship remained strong and positive over the period (Figures 2.4a and 2.4b).5 Because there are regional variations in the relative and absolute avail­ ability of ICT, as well as among income groups, additional figures detail regional subdivisions based on geographic and income classifications (see Appendix 2C). These supplementary figures show that the relationship between fixed line teledensity and per capita GDP has both regional and income characteristics. While the relationship is relatively weak for low-income countries, it is partic­ ularly strong for the lower middle-income and high-income non-OECD coun­ tries; in terms of regional groupings, the relationship is relatively high.6 As a means of assessing the relationship between the availability of mod­ ern forms of ICT and GDP, the relationship between per capita GDP and both PCs per 1,000 inhabitants and Internet users per 1,000 inhabitants was con­ sidered (Figures 2.5a and 2.5b). Given limited data availability, results are presented for the year 2000 only. The relationship between PCs per 1,000 in­ habitants and per capita GDP is very strong and positive. A linear regression of per capita GDP on PCs per 1,000 inhabitants explains more than 85 percent variation in per capita GDP. Although weaker compared with the relationship to PCs, a positive relationship also exists between per capita GDP and the number of Internet users per 1,000 inhabitants. This regression explains about 73 percent of the variation in GDP. Table 2.3 provides further evidence on the positive relationship between ICT availability and per capita GDP. In this instance, both traditional and modern forms of ICT are included (fixed and cellular telephones, the Internet, and PCs) for both income and regional groups. In addition, the bottom row of the table presents the combined relationships for the 113 sample countries. With the exception of fixed telephone lines, all the correlation coefficients are for the year 2000. There is a positive relationship between teledensity and per capita GDP, and, with the exception of high-income non-OECD countries,7 the correlation 5. A simple linear regression of per capita GDP on fixed line teledensity explains 75 percent of the variation in per capita GDP in 1980 and 78 percent of the variation in per capita GDP in 2000. 6. A linear regression of per capita GDP on fixed line teledensity explains at least 70 per­ cent of the variation in per capita GDP. 7. Among the eight countries in this group, Kuwait and United Arab Emirates are oil-rich countries that suffered from the decline of oil prices. The exclusion of these two countries would result in a correlation coefficient of 0.9. F I G U R E 2.5a Personal computers and per capita gross domestic product, 2000 GDP per capita (1995 US$) 80,000 • 50,000 Japan • * 40,000 - USA 30,000 20,000 10,000 Ghana 4 Peru • A - UK A • i i i • i 0 100 200 300 400 500 600 700 Number of PCs per 100 inhabitants F I G U R E 2.5b Internet users and per capita gross domestic product, 2000 GDP per capita (1995 US$) Bang! 0 100 200 300 400 500 600 Number of Internet users per 100 inhabitants SOURCE: Calculated by authors from study data set. o g M © 3 ^ 1 © o o G r-< 3 S.^ § O g 5 O CJ o O o © ^ o •a ,-T o is C3 fl 6 0-3 tS g CQ £ £ 2 2; PH >- 60 CS OO c 'I 3 •5 c id -J 3 OH O ffl 43 00 K Hi § 1 e3 • a l l "o o 5 mow I S 2 (U PH O £ H « « -5 I -a « w is " «s .S 'S § on ja •3 | m pi Z g : S . | ° Pi PH .9 « I 3 J | '3 PH t/5 t/3 H T3 o « s - » » ™ o OA Bl «2 t i B •s g _ M OH 0 |^ *H 8 .a O jO P 'JS I S J 2 oo ^ 3 6 ho CD o 'cd 'U 0 0 H O cs 1 m 0 0 3 I O 3 S H £ 3 u S n CO OO >—\ CO CN ON NO in 00 r- CO in in < in CN T—< NO CN CN NO od in ON CO r-- t - HKn> TELECOM,, t). (1) This equation is empirically estimated as follows: \og{GDPit) = aw + a, \og(Ku) + a2 log(7X^) + a3 \og(PENu) + a4t + u\, (1') where GDP is the real gross domestic product, K is a measure of the real capi­ tal stock net of telecommunication capital as mentioned in the data section. TLF is the total labor force, which is a proxy for human capital, and t is a linear time trend. The variable PEN, that is, the penetration rate, is defined by the number of fixed lines per hundred inhabitants. This variable is a proxy for the stock of telecommunications infrastructure (TELECOM). The demand for telecommunications infrastructure is treated as a function of per capita GDP and the price of telephone service: TELECOMit = h(GDPtIPOPt, TELPt). (2) Given that the objective is to measure the demand for telecommunications, in this equation TELECOM is approximated by the sum of the penetration rate and the waiting list per hundred habitants (WL). The price for telephone service is approximated by the total service revenue per fixed line (TELP) and per capita GDP measures income. The empirical counterpart of (2) is given by lag(PENu + WLU) = b0 + bl log(GDPu/POPu) + b2 \og(TELPt) + ufr (2') The third equation corresponds to the supply of telecommunications in­ vestment. It is treated as a function of the price of telephone service (TELP) and other variables specific to the country: TTI=g(TELPt,Zit). (3) The empirical counterpart of this equation is given as: log(7T4) = c0 + c, \og(GAu) + c2GDu + \og(TELP)u + u\, (3') where (as in the demand equation) service revenue per fixed line is used as a proxy for price. The scale of the country and the economic well-being of the country is measured by the geographic area in thousands of square kilometers (GA) and the government surplus (deficit) in billions of 1985 U.S. dollars (GD), respectively. 60 Maximo Torero et al. Finally the telecommunications infrastructure production function meas­ ures the relationship between investment in telecommunications infrastructure and the change in the stock of telecommunications infrastructure: TELECOMit - TELECOM.^ = (TTLt, Rit). (4) To empirically estimate this equation, the change in the stock of telecommuni­ cations is approximated by the change in penetration as a function of invest­ ment in telecommunications infrastructure and the geographic area: logiPENJPEN^ ) = d0 + dl log(7T4) + d2 log(04;,) + u*. (4') Since Equations (2)-(4) involve the demand for and supply of telecommunica­ tions infrastructure, they endogenize telecommunications infrastructure. A l l four equations may be estimated as a system or using a two-step estimation procedure. Roller and Waverman (2001) estimate the empirical model outlined above using variables in levels, and with and without country-level fixed effects. An issue that Roller and Waverman do not take into account is that variables such as GDP and capital may follow a random walk. I f these variables do follow a random walk, then a regression of one on the other may lead to spurious re­ sults. De-trending the variables before running the regression may not help because the de-trended series may still be nonstationary. It is likely that only first-differencing wi l l yield stationary series. We attempt to solve this problem by estimating equations in first differ­ ences, and in our empirical work we assume that the error terms in the four equations follow an error component model, uu = Vt + vu> (5) where [i ~ IID(0, o2|J.) and vit ~ IID(0, o 2 v ) are independent of each other and represent unmeasured time-invariant country and country/year effects, respec­ tively. Given this error structure, lagged values of the dependent variable (let's call the generic dependent variable Y9) wi l l be correlated with the error terms in equations ( l ) - (4) . Even though the first difference transformation mitigates this correlation problem by eliminating the individual effect, OLS estima­ tion of the differenced model would also be inconsistent because now A 7 t l and Av; are correlated (given that Yit x and u i t l are correlated). Anderson and Hsiao (1981) observed that as long as u. is not serially correlated, AYU_2 (which de­ pends on the second and further lags of uit) is clearly correlated with AYitt but not with Av. (which only depends on vit and v f t l ) . Therefore, AY.f_2 is a valid instrument for AYU1 and may be used to estimate the model consistently. 9. Y represents the four dependent variables in the four equations: log(GZ)P), log(PSV + WL), log(777) and logiPENJPEN,^). Telecommunications Infrastructure and Economic Growth 61 Arellano and Bond (1991) observed that, when the number of periods is small and the number of groups in the panel is large, in order to gain efficiency, the number of valid instruments grows with the number of available periods. For example, for t = 3 the only valid instrument is Yt v but for t = 4 both Yitl and Yi2 are valid instruments. Consequently, for any given period T the set of valid instruments becomes (Ya, YUt_1,..., YjT2). Because the exogenous vari­ ables (called xu for simplicity, though they actually include the model's other explanatory variables such as capital and labor force) may be predetermined and correlated with \ip the valid set of instruments is [xjV, xa,, . . . , x^s xy], given that E(xuvjs) ^ 0 for s < ? and otherwise equals zero. With the instruments obtained from the lagged 7 and the lagged explana­ tory variables, a matrix of instruments W10 can be obtained, so that EiW/Av^) = 0. Using generalized method of moments (GMM), the one-step estimators of a and 8 would be = ([AY_lAX]'WVN 1W'[AY_lAX]y\[AY_1AX]'WV-1WW), (6) where ^ v = & ' ( A v . ) ( A v . ) ' ^ . (7) i=i This GMM estimator does not require any knowledge of the initial condi­ tions or distributions of v. and However, to correct for the presence of un­ observed firm heteroskedasticity, we operationalize this procedure by replacing Av with differenced residuals obtained from the preliminary consistent estima­ tor 8j . This yields the more efficient, two-step GMM estimator (Arellano and Bond 1991, 1998) used in our study. This second-step estimation provides ro­ bust standard errors. In the absence of such standard errors, test statistics could be highly misleading, especially in long panel data sets such as the ones with which we are working.1 1 References Anderson, T. W., and C. Hsiao. 1981. Estimation of dynamic models with error compo­ nents. Journal of the American Statistical Association 76 (375): 598-606. Antle, J. M . 1983. Infrastructure and aggregate agricultural productivity: International evidence. Economic Development and Cultural Change 31 (4): 609-619. 10. The transformed Wt are deviations from individual means. 11. For the hypothesis that there is no second-order serial correlation for the disturbances of the first-differenced equation, we follow the test by Arellano and Bond (1991,282), which assumes the consistency of the GMM estimation because it relies on the following: £TAv.(Av. (_2] = 0. F i ­ nally, we performed Sargan's (1958) test of overidentifying restrictions, suggested by Arellano and Bond (1991). 62 Maximo Torero et al. Arellano, M . , and S. Bond. 1991. Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. Review of Economic Studies 58 (2): 277-297. . 1998. Dynamic panel data estimation using DPD98 for Gauss. Working Paper No. 88/15, Institute for Fiscal Studies, London. Aschauer, D. A. 1989. Is public expenditure productive? Journal of Monetary Economics 23 (2): 177-200. Craig, B. J., P. G. Pardey, and J. Roseboom. 1997. International productivity patterns: Accounting for input quality, infrastructure, and research. American Journal of Agricultural Economics 79 (4): 1064-1076. Doherty, A. 1984. Empirical estimates of demand and cost elasticities of local telephone service. In Changing patterns in regulated markets and technology: The effect of public utility pricing. East Lansing, Mich., U.S.A.: Michigan State University In­ stitute of Public Utility Pricing, Institute of Public Utilities. Duncan, G., and D. Perry. 1994. IntalLATA toll demand modelling a dynamic analysis of revenue and usage data. Information Economics and Policy 6: 163—178. Ford, R., and P. Poret. 1991. Infrastructure and private-sector productivity. OECD Eco­ nomic Studies 17 (Autumn): 63-89. Garcia-Mila, T., and T. J. McGuire. 1992. The contribution of publicly provided inputs to states' economies. Regional Science and Urban Economics 22 (2): 229-241. Garcia-Mila, T., T. J. McGuire, and R. H. Porter. 1996. The effect of public capital in state-level production functions reconsidered. Review of Economics and Statistics 78 (1): 177-180. Gatto, J. J., H. Kelejian, and S. Stephan. 1988. Stochastic generalizations of demand sys­ tems with an application to telecommunications. Information Economics and Pol­ icy 3 (4): 283-309. Gatto, J. J., L. Hooper, P. Robinson, and H. Tyan. 1988. Interstate switched access de­ mand analysis. Information Economics and Policy 3 (4): 333-358. Gordon, R. J. 2000. Does the new economy measure up to the great inventions of the past? Journal of Economic Perspective 14 (4): 49-74. Harberger, A. 1978. Perspectives on capital and technology in less developed countries. In Contemporary Economic Analysis, M . J. Artis and A. R. Nobay, eds. London: Croom Helm. Hardy, A. 1980. The role of the telephone in economic development. Telecommunications Policy 4 (4): 278-286. Hirschmann, A. O. 1958. The strategy of economic development. New Haven, Conn., USA: Yale University Press. Holtz-Eakin, D. 1994. Public-sector capital and the productivity puzzle. Review of Eco­ nomics and Statistics 76 (1): 12-21. ITU (International Telecommunications Union). 2002. Online database, (accessed February 2002). Jipp, A. 1963. Wealth of nations and telephone density. Telecommunications Journal 20: 199-201. Kormendi, R., and P. Meguire. 1985. Macro-economic determinants of growth: Cross­ country evidence. Journal of Monetary Economics 16: 141-163. Leff, N . H. 1984. Externalities, information costs, and social benefit-cost analysis for economic development: An example from telecommunications. Economic Devel­ opment and Cultural Change 32 (2): 255-276. Telecommunications Infrastructure and Economic Growth 63 Levy, A. 1996. Semi parametric estimation of telecommunications demand. Ph.D. dis­ sertation, University of California-Berkeley, Berkeley, Calif, USA. Madden, G., and S. J. Savage. 2000. Telecommunications and economic growth. Inter­ national Journal of Social Economics 27 (7-10): 893-906. Mankiw, N. G., D. Romer, and D. N . Weil. 1992. A contribution to the empirics of growth. Quarterly Journal of Economics 107: 407-437. Munnell, A. H. 1992. Policy watch: Infrastructure investment and economic growth. Journal of Economic Perspectives 6 (4): 189-198. Nehru, V., and A. Dhareshwar. 1993. A new database on physical capital stock: Sources, methodology and results. Revista de Andlisis Economico 8(1): 37-59. Norton, S. W. 1992. Transaction costs, telecommunications and the microeconomics of macroeconomic growth. Economic Development and Cultural Change 41 (1): 175-196. Oliner, S. D., and D. E. Sichel. 2000. The resurgence of growth in the late 1990s: Is in­ formation technology the story? Journal of Economic Perspectives 14 (4): 3-22. Pasco-Font, A., J. Gallardo, and V. Fry. 1999. La demanda residencial de telefonia basica en el Peru. In Estudio en Telecomunicaciones No. 4. Lima: Organismo Supervisor de la Inversion Privada en Telecomunicaciones (OSIPTEL). Panzar, J. C. 2000. A methodology for measuring the costs of universal service obliga­ tions. Information Economics and Policy 12 (3): 211-220. Pohjola, M . , ed. 2001. Information technology, productivity and economic growth: In­ ternational evidence and implications for economic development. World Institute for Development Economics Research, United Nations University, Studies in De­ velopment Economics. Oxford and New York: Oxford University Press. Roche, E. M. , andM. J. Blaine. 1996. Information technology, development and policy. Avebury, UK: Aldershot. Roller, L.-H., and L. Waverman. 2001. Telecommunications infrastructure and economic growth: A simultaneous approach. American Economic Review 91 (4): 909-923. Saith, A. 2002. ICT: Hope or hype. Paper presented at ICTs and Indian Development, held in Bangalore, India, December. Saunders, R. J., J. J. Warford, and B. Wellenius. 1983. Telecommunications and economic development. Baltimore: Johns Hopkins University Press. Stiroh, K. J. 2002. Are ICT spillovers driving the new economy. Review of Income and Wealth 48 (1): 33-57. Torero, M . , S. Chowdhury, and V. Galdo. 2003. Willingness to pay for the rural tele­ phone service in Bangladesh and Pern. Information Economics and Policy 15 (3): 327-361. World Bank. 2002. World development indicators. Online database (accessed February 2002). Zona, J. D., and R. Jacob. 1990. The total bil l concept: Defining and testing alternative views. Presented at the Bellcore/Bell Canada Industry Forum, "Telecommunications Demand Analysis with Dynamic Regulation," held in Hilton Head, S.C., in April 1990. Cambridge, Mass., USA: National Economic Research Associates.