WPS7352 Policy Research Working Paper 7352 Do Poor Countries Really Need More IT? The Role of Relative Prices and Industrial Composition Maya Eden Paul Gaggl Development Research Group Macroeconomics and Growth Team June 2015 Policy Research Working Paper 7352 Abstract Conventional wisdom suggests too little information technology adoption lags and ICT-labor substitutabil- and communication technologies (ICT) in poor coun- ity, there is little empirical support for these hypotheses. tries. Indeed, within 70 countries at various levels of Instead, the paper establishes that this regularity can be development, there is a positive relationship between fully accounted for by (a) relatively higher ICT prices income per capita and the capital share of ICT. While in low-income countries and (b) industrial composition. this regularity is consistent with explanations based on This paper is a product of the Macroeconomics and Growth Team, Development Research Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at meden@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Do Poor Countries Really Need More IT? The Role of Rela- tive Prices and Industrial Composition JEL: O14, O33, O57, E22 Keywords: ICT adoption, industrial composition, ICT capital stocks Maya Eden1 Paul Gaggl1 The World Bank University of North Carolina at Charlotte Development Research Group Belk College of Business Macroeconomics and Growth Team Department of Economics 1818 H St. NW 9201 University City Blvd Washington, DC 20433 Charlotte, NC 28223-0001 Email: meden@worldbank.org Email: pgaggl@uncc.edu 1 This paper was prepared as background work for the World Bank’s World Development Report 2016 on the topic of the Internet. We thank Asad Zaman for helpful research assistance and Robert Inklaar for assistance with technical details in the Penn World Table 8.1. We are further grateful for helpful comments by Uwe Deichmann, Aart Kray, Deepak Mishra, Indhira Santos, Marc Schiffbauer, as well as Luis Serven and for generous financial s upport f rom a World Bank RSB grant. 1. Introduction In 2013, the per-capita number of broadband internet subscriptions in the US was 0.3. In India, the corresponding number was 0.01. In high income countries, over 75% of households owned a computer; in low and middle income countries, the corresponding number was only 27.6%.1 These disparities suggest that the proliferation of information and communication technologies (ICT) has been vastly different in rich and poor countries. At the same time, a substantial literature documents that ICT was one of the key drivers of growth in advanced economies (e.g., Colecchia and Schreyer, 2002; Basu, Fernald, Oulton and Srinivasan, 2003; Bloom, Sadun and Van Reenen, 2012). In light of this, the dramatic cross-country differences in ICT abundance across countries may seem alarming. Motivated by these observations we address two important questions: first, is ICT capital truly scarce in low income countries relative to other capital? And second, if so, what can explain this? Notice that our emphasis on the ICT stock relative to other capital is important, since poor countries not only have fewer internet subscriptions and PCs but also fewer refrigerators and cars. We begin our analysis by compiling a new dataset that allows us to measure ICT and non- ICT (henceforth NICT) capital stocks for a sample of 70 countries at various levels of economic development.2 Using these data we document that, in low income countries, the value of ICT capital indeed represents a smaller share of the aggregate capital stock. In fact, the differences are even larger in real terms, as we document that ICT capital goods are relatively more expensive in low income countries. This suggests that the scarcity of ICT in developing countries is likely not explained simply by lower capital-labor ratios. What else can explain the relationship between income per capita and the abundance of ICT capital? One potential mechanism relates to delayed technology adoption. Due to frictions in learning and adopting new technologies, low income countries may be slower to accumulate ICT capital (e.g., Comin and Hobijn, 2010; Gust and Marquez, 2004). A second explanation relates to 1 These numbers are based on the World Telecommunication/ICT Indicators database provided by the International Telecommunication Union (ITU). For more information on this datasource see Appendix A. 2 To the best of our knowledge, our dataset is the most comprehensive account of ICT stocks around the world to this date. See Section 2.1 for a comparison with other international datasets containing information on ICT. 2 the relative prices of capital and labor. If labor is to some degree substitutable with ICT capital (e.g., Autor and Dorn, 2013; Karabarbounis and Neiman, 2014), then low-income countries will likely opt for lower ICT capital stocks, as labor is relatively cheaper in these countries. While these mechanisms are plausible, we find evidence in support of a much simpler expla- nation based on cross-country differences in industrial composition. Heterogeneity in industrial composition may imply that production in developing countries is relatively less intensive in ICT capital. For example, if agriculture has a larger share in output in developing countries and agri- culture is inherently less ICT-intensive than other industries, one would expect there to be rela- tively less ICT in developing countries. While the first two explanations would imply that, in a given industry, production in developing countries relies less on ICT, our findings suggest that, after controlling for industrial composition, the strong association between income per capita and relative ICT abundance breaks down. To illustrate this point, we use a simple theoretical framework in which ICT and non-ICT capi- tal intensities differ by industry. Since we are interested in gauging the degree to which industrial composition alone can explain differences in ICT abundance, we assume that sector specific ICT intensities are the same across countries and calibrate these based on data from the US. We then use sector-specific value added for a wide range of countries to predict the aggregate ICT capital share. The exercise suggests that after accounting for the portion of relative ICT abundance that is predicted by cross-country heterogeneity in industrial composition, there is no longer a system- atic relationship between GDP per capita and the share of ICT out of total capital. This implies that differences in specialization can fully account for the strong correlation between income per capita and relative ICT abundance. As an additional piece of evidence for this result we further document that there is no sys- tematic relationship between a proxy for ICT abundance and income per capita within industries. More precisely, while it is well documented that capital labor ratios within the same industry vary widely across countries (e.g., Davis and Weinstein, 2001), we do not find a systematic relationship between income per capita and ICT spending as a fraction of the capital stock. In sum, our results suggest that the variation across countries in ICT abundance is predomi- 3 nantly between-industry variation rather than within-industry variation. This suggests that any frictions associated with the accumulation of ICT are reflected in changes in industrial compo- sition, rather than in changes in the production structure within industries. This reasoning is consistent with a Heckscher-Ohlin model of trade specialization. Despite this potential source of endogeneity, we find similar results when using predicted values based on industrial composition in 1980, which is at the beginning of the ICT revolution and therefore less likely endogenous to ICT adoption. However, the extent to which industrial composition is endogenously determined by ICT adoption remains an open question, which is beyond the scope of this paper. While in- dustrial composition may be, to some degree, endogenous to ICT accumulation, it is also likely that a number of other factors determine industrial composition, such as the level of development, comparative advantage, and path dependence. The remainder of this paper is organized as follows: we begin with empirical evidence on the distribution of ICT capital around the world in Section 2, together with the details of the construction of our dataset. Section 3 presents several possible explanations for the observed correlation between ICT abundance and income per capita. In Section 4 we present a simple model that highlights the relationship between ICT abundance and industrial composition and we provide empirical evidence consistent with its predictions. Section 5 illustrates a benchmarking procedure implied by our analysis, which can be used to assess whether ICT is over- or under- abundant in a given country. We offer some concluding remarks in Section 6. 2. Measuring ICT Capital Around the World We begin our analysis by estimating the stock of ICT and non-ICT (henceforth, NICT) capital for 70 countries at various levels of development. To accomplish this, we use data on ICT spend- ing from the World Information Technology and Services Alliance (WITSA) as well as the Inter- national Telecommunication Union (ITU).3 WITSA is currently the most widely used source for data on ICT spending on a global scale and is assembled using a combination of various surveys, 3 These two datasources are described in detail in Appendix A. 4 vendor supply analysis and other statistics.4 Specifically, WITSA reports ICT spending for four categories: (1) computer hardware, (2) computer software, (3) computer services, and (4) commu- nications. The sum of these four categories gives a fairly comprehensive picture of ICT expenditure around the world. However, as we are interested in constructing measures of the physical stock of ICT capital, it is important to notice that, conceptually, some of these WITSA spending mea- sures represent investment spending but others consist primarily of rental fees. For example, while spending on internet subscriptions or telecommunication fees may comprise a substantial amount of ICT spending, it does not constitute investment: from a macro perspective, these are transfers between users of ICT capital and owners of ICT capital, more appropriately viewed as rental fees. From an aggregate perspective, an internet subscription does not require the sacrifice of resources today for the purpose of increasing aggregate production capacity tomorrow, which is the defining characteristic of investment. More specifically, of the four WITSA spending categories, computer services is in fact the only category that consists primarily of true aggregate investment spending, taking the form of custom software development and equipment maintenance. This category also includes some services that may be more appropriately viewed as rental payments, such as web hosting, but these likely represent a small share of spending in this category. The categories of computer hardware and computer software include the total value of pur- chases and leases. Ideally, one would like to count hardware and software investment as the purchase of new machinery or software. However, the WITSA measure includes secondary mar- kets as well, as it takes into account the value of leases. Bluntly, if a computer is purchased and then leased, it is double counted. We therefore adopt an approach similar to Vu (2005) and assume that hardware investment is 0.57 times computer hardware spending, which is roughly the coeffi- cient of proportionality in US data.5 The coefficient of proportionality for software is greater than one, suggesting that software spending is lower than software investment in the United States. This is probably due to the omission of computer services spending, which includes some forms 4 For instance, both the Penn World Table and the Conference Board’s Total Economy Database use WITSA as the main source for information on ICT spending. 5 See the Appendix of Vu (2005) for year-by-year estimates of this factor of proportionality. 5 of software investment. Since we include computer services in our ICT investment measure, we assume that the remaining software investment is equal to WITSA software spending. It is perhaps worth noting that the distinction between software leases and software invest- ment is somewhat blurred. The software spending category consists of the total value of pur- chased or leased packaged software. While purchasing software is investment from the firm’s perspective, from a macro perspective this is perhaps more appropriately viewed as a rental fee. The creation of new software is similar to investment in research and development (R&D). The returns to writing new software are the dividends from selling or leasing the rights to use that software. From a timing perspective, the value of the initial investment is the costs of program- mers and associated capital costs for producing new software. The returns to the investment are the sales of software licenses, either permanent (purchases) or temporary (leases). From a macroe- conomic perspective, software investment should be counted as the costs associated with devel- oping new software (similar to R&D investment). However, given that this data is not available, we stick with the commonly adopted micro perspective and assume that software investment is software purchases and leases.6 The fourth WITSA category, communication technology (CT), is defined as the total value of voice and data communication services and equipment. Conceptually, communication services (such as internet subscriptions or payments for phone usage) represent rental fees for communica- tion infrastructure, rather than investment. Since we are interested in a pure investment measure, we substitute this category with a direct measure of CT investment from ITU. Taken together, our final measure of nominal (current USD) ICT investment is the sum of TC investment (ITU), computer services spending (WITSA), adjusted computer hardware spending (WITSA), and computer software spending (WITSA). To construct ICT capital stocks we deflate each of these nominal investment series with the ICT price deflator estimated by Eden and Gaggl (2014) based on the BEA’s fixed asset accounts and employ the perpetual inventory method (PIM) 6 Note that the high depreciation rate of software implies that there is no big difference between permanent purchases and temporary leases. Generally, most attempts to construct capital stocks take this perspective. The BEA’s computations for the NIPA tables are one example. See the official NIPA documentation for details: http://www.bea.gov/national/pdf/chapter6.pdf. 6 to construct ICT stocks.7 Following the methodology in the construction of the Penn World Table (PWT) starting with version 8.0 (Feenstra, Inklaar and Timmer, 2013; Inklaar and Timmer, 2013) we use constant depreciation rates that differ for telecommunication (11.5% p.a.) and other infor- mation technology (31.5% p.a.).8 To estimate NICT investment, we subtract nominal ICT investment from aggregate investment measured in the PWT. Similar to the procedure for ICT, we deflate the resulting investment series using the BEA-based NICT price index from Eden and Gaggl (2014) and compute NICT capital stocks using the PIM, with a depreciation rate of 6% p.a., as estimated by Eden and Gaggl (2014) for the US.9 2.1. Comparison With Existing Datasets At this point, it is perhaps useful to compare our measurement strategy with existing datasets that include measures of ICT capital. There are two main datasets containing ICT capital measures: (a) the Groningen Growth and Development Center’s KLEMS datasets, and (b) the Conference Board’s Total Economy Database (TED). The key difference with the KLEMS datasets is country coverage. The EU KLEMS covers 27 high income countries between 1970-2013 (O’Mahony and Timmer, 2009) and the WORLD KLEMS database provides additional ICT/NICT data for Canada (Gu, 2012) and Russia (Voskoboynikov, 2012). In contrast, our dataset on ICT/NICT capital stocks covers 70 countries at various levels of development (see Panel D of Table 1) and is—to the best of our knowledge—the most compre- 7 Note that we explicitly choose to use a single price index to deflate all four investment series, even though we could in principle estimate the disaggregated price indexes from the BEA’s fixed asset accounts (e.g., Figure 2). We do so for two reasons: first, it allows us to aggregate the four categories simply by adding the four resulting constant dollar stocks within each category. With time varying investment-type specific price indexes aggregation would be less straight forward. Moreover, while Figure 2 shows that there were some differences in the trends of price indexes for different ICT categories, the general trend is relatively uniform. Second, and more importantly, it is not clear how representative the differential trends between detailed investment categories in the US are for other (especially poor) countries. 8 While Eden and Gaggl (2014) estimate time varying depreciation rates for the US, and we could have used their procedure to estmate these here as well, we choose to stick with the same methodology as in the PWT for two reasons: first, at several points in our data construction we use PWT aggregates to imupte residual quantities, and second, we would like to keep our dataset conceptually comparable to the PWT. 9 Note that we kept our description of the data construction brief here and we provide additional details in Appendix A. 7 Table 1: Summary Statistics Year Obs. Mean. Std. Dev. Min. Max. A. The value of ICT as a Fraction of Total Capital Value (%) 1995 70 3.95 2.65 0.17 12.40 2011 70 3.88 1.56 1.60 8.92 B. The value of ICT relative to NICT (%) 1995 70 4.19 2.95 0.17 14.16 2011 70 4.07 1.73 1.63 9.80 C. PPP Units of ICT per Unit of NICT (%, only countries with ICP data in 2011) 1995 68 0.89 0.70 0.10 3.13 2011 68 5.43 2.86 1.46 14.96 D. Log Real GDP Per Person 1995 70 -4.76 1.04 -8.41 -3.38 2011 70 -4.28 0.96 -6.42 -2.61 Notes: The table reports raw sample summary statistics for 70 countries (and 68 for the real ratios) in 1995 and 2011. hensive account of ICT and NICT stocks at this point. That said, there are currently numerous WORLD KLEMS projects under construction to expand converge. The TED, on the other hand, has very comprehensive country coverage, yet it only contains measures of the growth in ICT capital services for the period 1990-2014 and does not specifically attempt to measure ICT capital stocks. While our measurement efforts are clearly complementary to Jorgenson and Vu (2005), who also use WITSA and ITU data to estimate ICT, there are some differences. Specifically, they as- sume that ICT investment is proportional to ICT spending, while we try to construct an invest- ment measure directly. Furthermore, most previous work does not count the category of “capital services” as an investment category, and rather focuses on projected values of hardware, software and telecommunications spending on hardware, software and telecommunications investment. Since the services category consists of some software investment (such as custom software or website design), the ratio of software spending and software investment in the US is above two (Vu, 2005). Our view is that the category “ICT services” represents pure investment spending and should be counted as such. Another important difference is that the data provided in his paper is data on ICT capital growth rather than on the stock of ICT. 8 Further, most existing studies focus on investment in constant USD but do not take into ac- count the differences across countries in the relative prices of ICT and NICT goods (see Jorgenson and Vu, 2005; Vu, 2005, and references therein). We devote Section 2.3 to explore the role of cross country variation in this relative price by constructing a measure of ICT units per NICT units that is comparable across countries. Finally, we keep our methodology conceptually close to that of the PWT (Feenstra et al., 2013; Inklaar and Timmer, 2013). Note that, starting with version 8.0, the PWT constructs aggregate capital stocks by adding estimated capital stocks of six different asset types, among them comput- ers, communication equipment and software. These are also based on the PIM, run separately for these categories, with depreciation rates that are similar for computers and software (31.5%) but substantially lower for communications (11.5%). As mentioned in the previous section, we also adopt these assumptions. 2.2. ICT Abundance Around the World Using the procedure described above, we estimate the stocks of ICT and NICT for a balanced sample of 70 countries over the period 1993-2011.10 Table 1 summarizes our estimates for 1995 and 2011, which suggest a number of interesting conclusions. First, in most countries, ICT capital constitutes a small fraction of the aggregate value of the capital stock. In our sample of countries, the value of ICT capital out of total reproducible capital is on average around 3.88% in 2011 (see panel A of Table 1). Second, the summary statistics reported in panel A of Table 1 show that this range is relatively stable throughout our sample, with a mean of 3.95% in 1995, only marginally higher than in 2011. Panel B of Table 1 illustrates a similar picture for the value of ICT as a fraction of NICT, a measure we will use to proxy the relative abundance of ICT for the remainder of this paper. Third, in relatively poorer countries, ICT constitutes a smaller fraction of the total value of cap- ital. Panel A of Figure 1 summarizes this finding, which relates our measure of “ICT abundance” to log real GDP per capita in 1995 and 2011. To assess this observation more formally we run 10 See Appendix A for details on the construction of this dataset. 9 Figure 1: ICT Abundance and Income Per Person A. ICT Value out of Non-ICT Value (B.1) 1995 (B.2) 2011 20 20 Value of ICT/Value of NICT (PIM, 1995,%) Value of ICT/Value of NICT (PIM, 2011,%) 15 15 10 10 5 5 0 0 -5 -5 -10 -10 -8 -7 -6 -5 -4 -3 -6 -5 -4 -3 -2 PWT: Log Real GDP Per Capita PWT: Log Real GDP Per Capita C. Units of ICT per Unit of Non-ICT (C.1) 1995 (C.2) 2011 Units of ICT per Unit of Non-ICT (PIM, 1995,%) 20 Units of ICT per Unit of Non-ICT (PIM, 2011,%) 20 16 16 12 12 8 8 4 4 0 0 -8 -7 -6 -5 -4 -3 -6 -5 -4 -3 -2 PWT: Log Real GDP Per Capita PWT: Log Real GDP Per Capita Notes: The figures plot measures of ICT abundance across countries in relation to log output per capita. Panel A illustrates ICT values as a fraciton of NICT capital values; panel B shows units of ICT per unit of Non-ICT (ratio of quantity indexes). Column 1 shows this relationship in 1995 and column 2 in 2011. regressions of our ICT abundance measure on log real GDP per capita and a complete set of time effects. Panel A of Table 2 shows the point estimates for the full sample of 1994-2011 as well as broken down into three sub periods: 1994-2000, 2001-2005, and 2006-2011. These point estimates convey the patterns observed in Figure 1 and suggest three main insights: (a) there is a signifi- cantly positive relationship between income per capita and ICT abundance; (b) this relationship is becoming weaker over time, suggesting that developing countries are catching up in terms of ICT accumulation; and (c) the level of ICT/NICT values is relatively stable throughout 1993-2011. We draw this final observation from the estimated time effects. We normalized income per capita 10 Table 2: ICT Abundance: Values & Quantities A. % $ICT/$NICT B. % Units of ICT per Unit of NICT 1993-2011 1993-2000 2001-05 2006-11 1993-2011 1993-2000 2001-05 2006-11 Log Real GDP/L 0.870*** 1.331*** 0.759*** 0.279 0.680*** 0.499*** 0.764*** 0.876** (0.239) (0.286) (0.259) (0.281) (0.174) (0.0832) (0.181) (0.365) t=1993 5.105*** 5.261*** 1.034*** 0.972*** (0.508) (0.498) (0.113) (0.0938) t=1994 4.723*** 4.866*** 1.048*** 0.992*** (0.403) (0.394) (0.104) (0.0850) t=1995 4.471*** 4.598*** 1.082*** 1.032*** (0.339) (0.331) (0.0978) (0.0810) t=1996 4.380*** 4.495*** 1.163*** 1.117*** (0.303) (0.296) (0.0966) (0.0822) t=1997 4.334*** 4.440*** 1.297*** 1.255*** (0.288) (0.283) (0.102) (0.0896) t=1998 4.372*** 4.479*** 1.526*** 1.484*** (0.282) (0.276) (0.113) (0.104) t=1999 4.536*** 4.633*** 1.750*** 1.712*** (0.286) (0.280) (0.125) (0.119) t=2000 4.707*** 4.780*** 1.977*** 1.948*** (0.294) (0.287) (0.139) (0.138) t=2001 4.508*** 4.492*** 2.157*** 2.169*** (0.266) (0.268) (0.144) (0.145) t=2002 4.393*** 4.380*** 2.374*** 2.384*** (0.260) (0.260) (0.158) (0.158) t=2003 4.170*** 4.160*** 2.573*** 2.581*** (0.245) (0.243) (0.169) (0.169) t=2004 3.936*** 3.931*** 2.883*** 2.886*** (0.225) (0.223) (0.184) (0.184) t=2005 3.761*** 3.764*** 3.316*** 3.314*** (0.219) (0.217) (0.212) (0.211) t=2006 3.752*** 3.792*** 3.757*** 3.744*** (0.223) (0.219) (0.239) (0.241) t=2007 3.719*** 3.784*** 4.222*** 4.201*** (0.217) (0.212) (0.257) (0.262) t=2008 3.694*** 3.778*** 4.626*** 4.598*** (0.223) (0.215) (0.281) (0.289) t=2009 3.967*** 4.035*** 4.732*** 4.710*** (0.227) (0.218) (0.281) (0.286) t=2010 3.966*** 4.060*** 4.952*** 4.921*** (0.232) (0.226) (0.298) (0.306) t=2011 3.943*** 4.053*** 5.303*** 5.267*** (0.231) (0.226) (0.327) (0.334) Obs. 1292 544 340 408 1292 544 340 408 F-Stat. 32.3 37.7 74.1 67.8 22.7 28.3 51.0 49.9 Notes: The table reports coefficient estimates from linear regressions of our ICT abundance measures on log GDP per captia and time effects. Panel A reports regressions for the ratio of ICT values relative to NICT values and panel B reports units of ICT per unit of NICT. In these regressions we re-center log real GDP per capita so that it is zero for the median income country in our sample, which ensures that the time effects reflect the time specific average for the median income country. We note that the point estimates for the slopes are virtually unaffected by this normalization but give the time effects a clear-cut interpretation. Standard errors are clustered on country and reported in parentheses below each coefficient. Significance levels are indicated by * p < 0.1, ** p < 0.05, and *** p < 0.01. 11 such that it is zero for the median income country, which implies that the estimated time effects trace out the level of relative ICT abundance for the median income country. 2.3. The Relative Price of ICT & Physical ICT Abundance While we have so far focused on the market value of ICT relative to NICT, it is also instructive to compare the physical abundance of ICT, that abstracts from cross country differences in the relative price of ICT. This exercise is especially important as a number of studies have shown that the relative price of investment goods varies dramatically across countries and is significantly higher in poor countries (e.g., Hsieh and Klenow, 2007). Since ICT is likely imported to a large degree in many poor countries, it is natural to expect that ICT is more expensive in these countries relative to other, domestically produced capital goods. Intuitively, we would like to capture the number of “computers” (units ICT capital goods) relative to “buildings and other equipment” (units of non-ICT capital goods) to measure true abundance of ICT. To accomplish this, we use two sources of information on ICT prices: (1) US price data from the BEA as in Eden and Gaggl (2014) to estimate the path of relative prices over time, and (2) item-level price data from the International Comparison Program (ICP) to measure the cross-country distribution of the relative price of ICT. A wide literature has documented a precipitous decline in the relative price of ICT which is reflective of technological progress that reduces the costs of computations and communication (e.g., Autor and Dorn, 2013, and references therein). Figure 2 illustrates these patterns for the US based on estimates by Eden and Gaggl (2014) using the BEA’s detailed fixed asset accounts. The Figure clearly illustrates the dramatic fall in ICT prices relative to NICT prices over the past 30 years. Our construction of the capital stock series assumes that this trend is shared by all countries in our sample. This assumption is commonly applied in this context, and follows the procedure in the PWT, among others. While plausible, data limitations prevent direct assessment of this hypothesis. In this section, we present some evidence suggesting that there is no strong relationship between the relative price of ICT and income per-capita. We utilize two waves of item-level price data from the International Comparison Program 12 Figure 2: The Price of ICT in the US (A) US Price of ICT and NICT (B) US Price of Harw., Softw., and R&D 2.5 2.5 Non-ICT Capital Price (Chain Index, 1983=1) Non-ICT Capital Price (Chain Index, 1983=1) ICT Capital Price (Chain Index, 1983=1) ICT: Hardware Capital Price (Chain Index, 1983=1) ICT: Software Capital Price (Chain Index, 1983=1) 2 2 ICT: R&D Capital Price (Chain Index, 1983=1) Price Index (1983=1) Price Index (1983=1) 1.5 1.5 1 1 .5 .5 0 0 1940 1960 1980 2000 2020 1940 1960 1980 2000 2020 Year Year Notes: Panel A illustrates the price if ICT and NICT based on Eden and Gaggl (2014). Panel B disaggregates ICT into software, hardware, and R&D based on the same data and procedures as in Eden and Gaggl (2014). (ICP)—2005 and 2011.11 Unfortunately, these data are only available for a small number of coun- tries in 2005 and even in 2011 there are many missing observations for prices at the item level. Nevertheless, we are able to construct a fairly comprehensive country specific measure for the price of ICT relative to that of NICT, defined as i,j p Ei∈ICT pi,us rpj = i,j p (1) Ei∈N ICT pi,us where pi,j denotes the price of item i in country j and j = us indicates the US. This measure uses US prices as a benchmark, and compares prices of ICT and non-ICT items relative to the US. We use the US as a benchmark because, due to limited data availability, we cannot construct ICT and non-ICT investment bundles that are comparable across countries. The role of the comparison with the US is to remove item fixed effects. To see the importance of this, consider a hypothetical scenario in which there are two countries. In country 1, we have data on the price of computers (an ICT item) and a sewing machine (a non-ICT item). In country 2, we have data on the price 11 We manually classify items as ICT and non-ICT investment goods. Our classification is in line with the usual definitions of ICT and details are available upon request. 13 Figure 3: The Price of ICT relative to NICT and Income Per Person (A) Relative Price (2011) (B) Change in Rel. Price. (2005-2011) 2.5 1 Change in Price of ICT Relative to NICT: 2005-2011 Price of ICT Relative to NICT 2 .5 1.5 0 1 -.5 .5 -8 -6 -4 -2 -7 -6 -5 -4 -3 Real GDP Per Capita Real GDP Per Capita Notes: Panel A illustrates the relationship between rpj as defined in equation (1) for 160 countries based on ICP price data from 2011. Panel B illustrates the change in the relative price if ICT. Panel A uses the USA as the benchmark country. In panel panel B we use the UK as the benchmark country because we do not have ICP price data for the US in 2005. of computers but we do not have data on the price of a sewing machine; instead, we have data on the price of a vehicle (a larger non-ICT item). It would be meaningless to compare the ratio of the computer price to the non-ICT item, because the vehicle represents a more expensive item. However, if we compare the price of computers relative to a benchmark country to the price of the non-ICT item relative to a benchmark country (where we use the US as a benchmark), we are capturing some notion of whether there is a premium associated with ICT items, on average.12 Panel A of Figure 3 documents a negative correlation between rpj and log real GDP per capita. This suggests that, relative to non-ICT capital goods, ICT capital goods are relatively more expen- sive in low income countries. We further compute the growth rate in rpj between 2005 and 2011 for a small number of countries for which item level data is available in both ICP waves.13 Using this measure, Panel B of Figure 3 suggests that, while there is a weak negative correlation between the change in rpj and log income per-capita, the relationship is not statistically significant. Note 12 We have also experimented with alternative reference countries and find similar results. We report the US here because most of the extant research on ICT prices has thus far focused on the US and it therefore serves as a natural benchmark. 13 Since data is not available for the US in 2005, we use the UK as a benchmark for this exercise. 14 that the confidence intervals do not reflect the uncertainty regarding the estimation of rpj , which is usually constructed based on a small set of items. Thus, we cannot reject the common working assumption that the dynamics of the price of ICT relative to NICT are uniform across countries at different levels of development. Moreover, it is perhaps worth noting that we show price dynam- ics between 2005 and 2010 here—a period in which relative ICT prices weren’t falling dramatically even in the U.S. (see Figure 2). Thus, our finding of no systematic worldwide decline in ICT prices between 2005 and 2010 is perfectly consistent with the assertion that the US price trend is relevant more globally. In sum, these findings suggests a simple strategy to translate the nominal shares reported in panel A of Figure 1 and Table 2 into a measure of ICT units per NICT unit that is comparable across countries. Specifically, we model country and time specific ICT and NICT prices as p(ICT, i, t) = p(ICT, us, t) · p(i, k, t) · p(ICT, i) (2) p(N ICT, i, t) = p(N ICT, us, t) · p(i, k, t) · p(N ICT, i), (3) where p(i, k, t) is some relative price of capital in country i at time t, which may vary over time. p(ICT, i) and p(N ICT, i) are country-specific ICT and non-ICT components that are time invari- ant. Given these definitions, the physical units of ICT per unit of NICT capital can be written as K (ICT ) p(ICT, i, t) · K (ICT, i, t) p(N ICT, i, t) = · (4) K (N ICT ) p(N ICT, i, t) · K (N ICT, i, t) p(ICT, i, t) p(ICT,i,t)·K (ICT,i,t) Notice that p(N ICT,i,t)·K (N ICT,i,t) is the nominal ratio reported in panel A of Figure 1 and Table 2, K (ICT ) and we can therefore easily translate it into the real ratio K (N ICT ) by multiplying it with p(N ICT, i, t) p(N ICT, us, t) · p(i, k, t) · p(N ICT, i) = p(ICT, i, t) p(ICT, us, t) · p(i, k, t) · p(ICT, i) p(N ICT, us, t) p(N ICT, i) = · , (5) p(ICT, us, t) p(ICT, i) where we proxy the first fraction with our US relative price measures based on the BEA (Eden and Gaggl, 2014) and the second fraction using our country specific relative price based on equation 15 Figure 4: The Rise in the Quantity of ICT A. 25-th Percentile B. Median Income Country C. 75-th Percentile 10 10 10 Avg. ICT Value/ NICT Value Avg. ICT Value/ NICT Value Avg. Units of ICT per Unit of NICT Avg. Units of ICT per Unit of NICT Avg. Ratio at 25-th Income Percentile (%) Avg. Ratio at 50-th Income Percentile (%) Avg. Ratio at 75-th Income Percentile (%) +/- 2 Std. Err. +/- 2 Std. Err. 8 8 8 6 6 6 4 4 4 2 2 2 Avg. ICT Value/ NICT Value Avg. Units of ICT per Unit of NICT +/- 2 Std. Err. 0 0 0 1995 2000 2005 2010 1995 2000 2005 2010 1995 2000 2005 2010 Year Year Year Notes: The figure shows a comparison of the nominal and real share of ICT over time. The lines are plots of the time effects from regressions analogous to the ones in the first columns in panels A and B of Table 2. These time effects are normalized so that they reflect countries at the 25-th percentile (panel A), the median (panel B), and the 75-th percentile of log real GDP per capita in our sample. That is, for each panel we run a separate log linear regresssion in which we re-center log real GDP per capita such that it is zero at the respective percentile cutoff. (1). Our resulting estimates of “real” ICT abundance are summarized in panel B of Figure 1 and Table 2, suggesting the following insights: first, the ratio of ICT capital goods to non-ICT capital goods is significantly higher in richer countries. Second, the difference in this measure between the poorest and the richest countries has increased significantly between 1993 - 2011. This is reflected both in the significantly steeper slope of our fitted regression line during the later periods in our sample (see panel B of Table 2) as well as the comparison of the maximum and minimum values in panel C of Table 1. These observations are also visually apparent in panel B of Figure 1 as well as Figure 4, which plots the time effects for three alternative versions of the full-sample regressions reported in Panels A and B of Table 2: ones in which income per capita is respectively normalized to be zero for the country at the 25-th, 50-th, and 75-th percentile of the real income distribution. This final figure particularly highlights that, while the nominal value of ICT as a fraction of NICT has remained roughly constant (or if anything declined mildly), the physical abundance of ICT has increased significantly around the world. Moreover, this pattern is relatively uniform across different levels of development. In sum, these findings suggest that lower physical abundance of ICT capital in developing countries may partly be accounted for by its higher relative price. Similar to the argument laid 16 out by Hsieh and Klenow (2007), lower labor costs will tend to imply higher costs of imported goods. Since ICT capital goods are likely to have a larger imported component than NICT capital goods, the higher relative price may contribute to the lower physical abundance of ICT. However, since the relative price of ICT capital is higher, the implications for the relative value of ICT capital are, in principle, ambiguous, and depend on the elasticity of substitution between ICT and NICT capital. 3. Why is ICT Capital Relatively Less Abundant in Poor Coun- tries? There are several potential explanations for the relatively lower ICT abundance in low-income countries. In Section 4, we will present evidence relating this correlation to systematic differences in industrial composition between rich and poor countries. However, before proceeding, it is use- ful to illustrate that there are a variety of plausible mechanisms consistent with this correlation, all of which would also imply a systematic correlation between income per-capita and ICT abun- dance within industry—a prediction we reject in Section 4. To illustrate, it is useful to consider a production framework that utilizes ICT and NICT capital, denoted kn and ki respectively, as well as labor (l). We will interpret l as consisting either of total labor or more specifically of labor performing automizable tasks which are substitutable with ICT (commonly referred to as “routine labor”). In the spirit of Autor and Dorn (2013), Krusell, ıos-Rull and Violante (2000), or Eden and Gaggl (2014) we assume that the production Ohanian, R´ function is given by: 1−α α σ y = kn (Aki + lσ ) σ (6) where α, σ ∈ (0, 1). The above production function is a Cobb-Douglas in NICT capital and a constant elasticity of substitution (CES) aggregate of ICT capital and labor. The assumption that σ ∈ (0, 1) implies that ICT and labor are more substitutable than Cobb-Douglas. The parameter A captures the relative productivity of ICT capital. If the efficient utilization of ICT capital requires some know-how, a higher A captures better technology adoption. 17 Assuming that the interest rate is close to 0 (for expositional purposes), the producer’s problem can compactly be written as follows:14 1−α α σ max kn (Aki + lσ ) σ − δn pn kn − δi pi ki − wl (7) ki ,kn ,l,x ∗ where δn , δi are depreciation rates and pn and pi are the prices of capital relative to output. Let ki ∗ be the ICT and NICT capital levels that solve the above optimization problems. and kn The following lemma establishes that, under realistic assumptions regarding the relationship between ICT productivity, relative input prices and income per-capita, this production framework implies a positive relationship between ICT abundance and income per-capita: Lemma 1. Let h = 1, .., n be a set of countries, ordered by income per-capita (where h = 1 is the country with the lowest income per-capita). Let Ah , pi,h , pn,h and wh denote the ICT productivity, ICT capital ∗ and k ∗ denote the solution to the producer’s price, NICT capital price and wage in country h, and let ki,h n,h optimization problem in country h. Assume that: pi,h (a) The ratio Ah pn,h is decreasing in h pi,h (b) The ratio Ah wh is decreasing in h Then, pi,h ki,h 1. The ratio pn,h kn,h is increasing in h. 2. The labor share ( wly ) is decreasing in h. The proof is provided in Appendix B. Note that the conditions of this lemma are satisfied if (a) ICT is more expensive in low income countries relative to both labor and NICT capital, and if (b) ICT productivity is relatively lower in low-income countries. The first condition is consistent with the empirical evidence that we document regarding the relative prices of ICT and NICT capital goods, as well as with well-known evidence regarding the higher relative price of capital in low income countries (e.g., Hsieh and Klenow, 2007, and references therein). The second condition is consistent with the view of technology adoption lags: if low-income countries are slower in 14 More elaborately, assuming that capital is purchased a period in advance and must be financed by borrowing at an interest rate r, capital enters the producer’s maximization problem as (1 − δ )pk − pk ∗ (1 + r) (if changes in the price of capital are ignored). The above is equal −δpk if we assume that r = 0. These assumptions are not crucial and are made merely for expositional purposes. 18 adopting new technologies, we would expect A to be relatively lower in low-income countries (e.g., Comin and Hobijn, 2010). The second part of the lemma establishes an additional testable prediction regarding the rela- tionship between income per-capita and the labor income share. We directly test this prediction, and establish no strong relationship between the labor income share and income per capita. Even a more generous interpretation of the model, that restricts attention to labor in occupations that are relatively more substitutable with ICT, is not supported by the data. To evaluate this prediction, we use data from the World Bank’s International Income Distri- bution Database (I2D2). This database covers a standardized set of demographic, education, la- bor market, household socioeconomic and income/consumption variables for over 150 countries drawn from more than 1000 nationally representative household surveys, with the earliest data starting in 1960. We use this data source to measure the total annual wage bill by occupation and then classify occupations as performing “routine” and “non-routine” tasks following Acemoglu and Autor (2011). That is, we classify (1) managerial, professional and technical occupations as “non-routine cognitive”; (2) sales, clerical and administrative support occupations as “routine cognitive”; (3) production, craft, repair, and operative occupations as “routine manual”; and (4) service occupations as “non-routine manual”. This then allows us to calculate the routine (groups 2 and 3) and non-routine (groups 1 and 4) wage bill as a percent of the total wage bill within each country. Finally, we obtain routine and non-routine income shares by multiplying the resulting wage bill shares with the aggregate labor income share taken from the PWT. Figure 5 illustrates that, if anything, both routine and non-routine labor income shares (and hence the aggregate labor income share) are slightly positively correlated with income per-capita, rejecting part two of Lemma 1. Finally, note that the above model assumes a single final good and an aggregate production technology. It can also be interpreted as a production technology for a particular industry. Thus, this model implies a positive correlation between ICT abundance and income per-capita at the industry level—a prediction we reject based on empirical evidence provided Section 4. 19 Figure 5: Income Shares: Routine/Non-Routine Labor (A.1) Non-Routine Share 2011 (A.2) Routine Share 2011 70 25 60 20 Income Share (%) Income Share (%) 50 15 40 30 10 20 5 -7 -6 -5 -4 -3 -2 -7 -6 -5 -4 -3 -2 Log Real GDP Per Capita Log Real GDP Per Capita Notes: The figures illustrate measures of the routine and non-routine income shares. Earnings shares are taken from the I2D2 database and the aggregate labor share is taken from the PWT 8.1. Self-employed workers are included and the patterns are robust to their exclusion. 4. ICT Abundance and Industrial Composition Inspired by the discussion in the previous section we present a simple quantifiable model to il- lustrate the relationship between industrial composition and the relative values of ICT and NICT capital stocks. Based on the predictions of this model we then evaluate the degree to which the cross-country differences in ICT abundance are accounted for by variation in industrial composi- tion. To motivate the analysis, Figure 6 documents that low-income countries produce dispropor- tionately in sectors that utilize ICT less intensively. Panel A is based on country and industry specific value added as reported in the Groningen Growth and Development Center’s 10 sector database (GGDC10, Vries, Vries and Timmer, 2014). While this measure provides a relatively high level of disaggregation, it is available only for a relatively small number of countries. As an alter- native, we consider a more coarse measure of industrial composition by using data on the value added in agriculture, industry excluding manufacturing, manufacturing and services, based on the World Bank’s World Development Indicators (WDI) for a wider set of countries. Panel B of Figure 6 illustrates a similar distribution of industrial composition based on this alternative data source. To highlight differences between high and low income countries we illustrate the value 20 Figure 6: Industrial Composition & Income (A.1) GGDC10: Industrial Composition & Income (A.2) WDI: Industrial Composition and Income Construction Agriculture Agriculture Wholesale/Retail/Hosp. Manufacturing Industry excl. Manuf. US ICT Share (2000) Finance/Ins./RE US ICT Share (2000) Y(i)/Y: 1-st quartile Y(i)/Y: 1-st quartile Y(i)/Y: 4-th quartile Minidng/Quarrying Y(i)/Y: 4-th quartile Manufacturing Utilities Government/Services Services Transp./Storage/Comm. 0 10 20 30 0 20 40 60 80 Fraction of GDP (%) Fraction of GDP (%) Notes: Panel A shows the industrial composition for the top and bottom income quartile in 2000 based on he GGDC10 Database. Industry shares are averaged over 2000 and 2010 and industries are orderd by their ICT Share in the US in 2000. Panel B shows an analogous graph based on WDI sector data. added shares of countries in the lowest (first) and highest (fourth) quartile of the per-capita income distribution within our sample. Most notably, agriculture, wholesale, retail and hospitality constitute the largest sectors in low income countries. Even in the US, these sectors do not use ICT intensely. In contrast, services and manufacturing constitute the largest sectors in high income countries, and the ICT intensity of these sectors is substantially higher. This suggests that perhaps one reason for lower ICT stocks in low income countries is lower demand for ICT in production. To set ideas, consider the following simple model. Economies are indexed h = 1, ..., H . There are n industries indexed j = 1, ..., n. Output in industry j in country h is given by j,i j,n α α 1−αj,i −αj,n Yh,j = Ah,j kh,j,i kh,j,n lh,j , (8) where Ah,j is normalized such that the output price of industry j in country h is 1. This specifica- tion assumes that ICT and NICT capital intensities are common across countries within the same industry. The rationale behind this assumption is that we wish to fix the intensity with which a given industry uses ICT and ask to what extent the industrial composition alone affects ICT abun- dance in the aggregate. The first order conditions with respect to ICT and NICT capital in sector j 21 yield M P Kh,j,i = ph,i (r + δi ) and M P Kh,j,n = ph,n (r + δn ), (9) where MPK here denotes the physical marginal product of each type of capital. Using the assump- tion of a Cobb-Douglas production technology (equation (8)) and multiplying the above relations by the appropriate capital stocks yields αj,i Yh,j = ph,i kh,j,i (r + δi ) and αj,n Yh,j = ph,n kh,j,n (r + δn ). (10) Using kh,i and kh,n to denote aggregate ICT and NICT capital stocks and summing the above relations across industries delivers n n αj,i Yh,j = ph,i kh,i (r + δi ) and αj,n Yh,j = ph,n kh,n (r + δn ). (11) j =1 j =1 ph,i kh,i r+δn Dividing the two expressions yields predicted values for ph,n kh,n as a function of r + δi and the industrial composition as reflected by the country and industry specific levels of output, Yh,j : n ph,i kh,i r + δn j =1 αj,i Yh,j = n (12) ph,n kh,n r + δi j =1 αj,n Yh,j Equation (12) allows us to conduct a simple test for the hypothesis that cross-country hetero- geneity in the industrial composition may play an important role in explaining the cross-country differences in ICT abundance. Specifically, we use equation (12) to compute predicted values for ICT abundance that vary across countries exclusively because of differences in industrial com- position. To accomplish this we first estimate industry specific capital income shares for the US based on the BEA’s fixed asset accounts to proxy αj,i and αj,n . These estimates are constructed based on the aggregate estimates by Eden and Gaggl (2014). Specifically, we use estimates of the aggregate marginal products of ICT and NICT and multiply them by the industry specific ICT and NICT capital output ratios.15 We then assume that δn = 0.06 and δi = 0.2, the average values 15 Notice that this approximation results in an implied aggregate capital share that is more than 100% for a few countries and industries, since the aggregate marginal product is only an approximation of the true industry specific 22 Figure 7: ICT Abundance and Industrial Composition (A.1) Model Fit: GGDC10 (A.2) Model Fit: WDI .15 .25 45-Degree Line 45-Degree Line 2000 until 2000 2010 after 2000 .2 ICT Value /NICT Value ICT Value /NICT Value .1 .1 .15 .05 .05 0 0 .02 .04 .06 .08 .1 .02 .04 .06 .08 .1 Industry Predicted ICT Value/NICT Value Industry Predicted ICT Value/NICT Value (B.1) Ind. Pred. and Real Income: GGDC10 (B.2) Ind. Pred. and Real Income: WDI .1 .1 2000 until 2000 2010 .08 after 2000 Industry Predicted ICT/NICT Industry Predicted ICT/NICT .08 .06 .06 .04 .04 .02 .02 -8 -7 -6 -5 -4 -3 -8 -6 -4 -2 Log Real GDP Per Capita Log Real GDP Per Capita Notes: Panel A displays the observed ratios against the ones predicted by equation (12). Panel B plots the industry predictions against log real GDP per capita. Column 1 is based on GGDC10 data while column 2 is based on WDI data, respectively. observed in the US over the period 1993-2011, again based on Eden and Gaggl (2014). Finally we assume r = 0.03, which resembles the average real return consistent with equation (9). As de- scribed above, we consider two alternative measures for Yh,j , one based on the GGDC10 database and one based on the WDI. Again, it is worth emphasizing that, by construction, the only source of cross-country variation in these predicted values is due to heterogeneity in industrial composition measured by value added. Panel A of Figure 7 illustrates the “fit” of these predictions based on industrial composition. marginal product. Specifically, this sometimes happens in very NICT intensive industries in which we likely overesti- mate the marginal product of NICT based on our aggregate approximation. 23 Table 3: GGDC10: ICT Abundance and Industrial Composition Ind. Obs. A. 2000 B. 2010 C. 1994 & 2011 Pred. ln( Pred. ) 1980 (A.1) (A.2) (A.3) (B.1) (B.2) (B.3) (C.1) (C.2) (C.3) (D) (E) ln(GDP/L) 0.32*** 0.034 0.029 -0.088 0.19** -0.0081 -0.048 -0.037 (0.075) (0.14) (0.11) (0.12) (0.074) (0.099) (0.11) (0.056) Ind. Pred. 1.16*** 1.06** 0.48** 0.64** 0.86*** 0.88*** 0.79*** (0.20) (0.42) (0.18) (0.27) (0.17) (0.30) (0.27) t=2000 -2.23*** -0.80* -0.79 -1.33*** -0.58** (0.35) (0.48) (0.53) (0.41) (0.29) t=2010 -2.41*** -0.81 -0.78 -1.41*** -0.55** (0.31) (0.49) (0.58) (0.44) (0.25) Constant -1.66*** 0.025 -0.093 -3.13*** -1.88*** -1.80*** (0.35) (0.57) (0.66) (0.45) (0.54) (0.56) Obs. 32 32 32 32 32 32 64 64 64 62 64 F-Stat. 17.9 33.3 17.2 0.1 7.1 3.9 984.6 1116.7 827.3 764.7 17.2 R2 0.4 0.4 0.4 -0.0 0.1 0.1 1.0 1.0 1.0 1.0 0.4 ph,i kh,i Notes: The table reports coefficient estimates from linear regressions of ln ph,n kh,n on log real GDP per capita and predicted values based on country specific industrial composition, both separately and jointly. Predictions based on industrial composition are n r +δn j =1 αj,i Yh,j based on equation (12) and computed as ln r +δi · n , with r = 0.03, δn = 0.06, and δi = 0.2. Robust standard errors j =1 αj,n Yh,j are reported in parentheses below each coefficient. Significance levels are indicated by * p < 0.1, ** p < 0.05, and *** p < 0.01. Since our estimates for ICT start in 1993 and the GGDC10 database reports values every 10 years, we can only use the years 2000 and 2010 to conduct our cross country analysis for this data source. Panel A.1 suggests a remarkably close fit of our predicted values for both years. Likewise, Panel A.2 reveals a similar fit based on the much coarser WDI sectors, yet for a much larger number of countries and reported at an annual frequency for the period 1993-2011. Given this fairly tight fit it is perhaps not surprising that panel B of Figure 7 suggests a strong positive correlation between our predicted values and log income per-capita for both data sources. Inspired by this graphical illustration we use our predicted values to evaluate the degree to which cross country variation in ICT abundance is accounted for by industrial composition. Table 3 reports a regression of log nominal ICT/NICT on log real GDP per capita and the log of our predicted values from equation (12) based on the GGDC10 database. While this database has fairly detailed information on value added by sector, we only have data for 32 countries in 2000 and 2010 in overlap with our ICT measures. Panel A displays the results for 2000, panel B for 2010, and panel C pools both years but con- trols for year effects. Columns (A.1) and (A.2) in panel A indicate that, in 2000, both income and 24 our predictions based on industrial composition were positively correlated with ICT abundance. However, once we include both regressors jointly, only the industrial prediction regressor has any explanatory power. In fact, the point estimate does not significantly differ from the estimate in column (A.2) and is in the neighborhood of one, as our model would predict. The results in panel B mirror those of panel A. The pooled regressions in panel C also confirm the results from panels A and B. Notice further that the time effects in panes C.2 and C.3 are insignificant (i.e. statistically speaking zero on average), directly lending support to our expression derived in equation (12). While these results provide some evidence for the predictions of our model, there are many potentially omitted regressors that might cause both ICT accumulation and industrial composition (e.g., low skill labor) and therefore bias our results. To address this concern, panel D of Table 3 shows an additional regression in which we proxy the the predicted values based solely on industrial composition in 1980. That is, we “fix” the industrial composition within each country prior to the main thrust of the so-called “ICT revolution”. This alternative specification confirms the results from our main estimates in panel C. Finally, panel E reports one additional specification in which we regress the log ratio of ob- served and predicted relative ICT/NICT values on log real GDP per capita. Again, this regression also suggests that there is no systematic relation between relative ICT abundance and real income after industrial composition is accounted for. Specifically, notice the substantially lower R2 (and adjusted R2 ) for this final specification. Given the limited number of countries and years in the GGDC10 database we repeat the above exercise with data from the WDI, which provides many more years and countries but offers a much more coarse classification for the sectors of production. Table 4 shows that our estimates based on WDI data are very close to those based on the GGDC10 database. Specifically, notice that all the time effects in specification (C.1) are highly significant while the bulk of the variation is absorbed by our predictions based on industrial composition in specifications (C.2) - (E). Again, like in our analysis based on GGDC10 data, the coefficient on our industry predictions is close to one in all specifications. Finally, once we divide the observed values of relative ICT abundance by our industry predictions (column E), real income and time effects have virtually no explanatory 25 Table 4: WDI: ICT Abundance and Industrial Composition Ind. Obs. A. 2000 B. 2010 C. 1994 – 2011 Pred. ln( Pred. ) 1980 (A.1) (A.2) (A.3) (B.1) (B.2) (B.3) (C.1) (C.2) (C.3) (D) (E) ln(GDP/L) 0.37*** 0.15 0.069 0.066 0.27*** 0.14* 0.037 0.12** (0.058) (0.12) (0.061) (0.063) (0.061) (0.074) (0.10) (0.058) Ind. Pred. 1.67*** 1.11* 0.13 0.047 1.31*** 0.88*** 0.93*** (0.30) (0.60) (0.23) (0.21) (0.26) (0.30) (0.34) t=1993 -2.19*** 0.56 -0.12 -0.51 0.17 (0.30) (0.79) (0.76) (0.70) (0.28) t=1994 -2.21*** 0.47 -0.19 -0.51 0.10 (0.29) (0.78) (0.75) (0.71) (0.28) t=1995 -2.18*** 0.50 -0.16 -0.52 0.12 (0.29) (0.78) (0.75) (0.71) (0.27) t=1996 -2.18*** 0.44 -0.21 -0.49 0.075 (0.28) (0.76) (0.73) (0.71) (0.27) t=1997 -2.16*** 0.41 -0.22 -0.47 0.057 (0.28) (0.75) (0.72) (0.71) (0.27) t=1998 -2.14*** 0.35 -0.25 -0.45 0.017 (0.28) (0.74) (0.70) (0.71) (0.27) t=1999 -2.09*** 0.46 -0.16 -0.40 0.11 (0.28) (0.75) (0.72) (0.71) (0.26) t=2000 -2.05*** 0.45 -0.16 -0.34 0.11 (0.28) (0.73) (0.71) (0.71) (0.26) t=2001 -2.06*** 0.35 -0.23 -0.35 0.031 (0.27) (0.72) (0.69) (0.71) (0.26) t=2002 -2.07*** 0.43 -0.19 -0.35 0.082 (0.27) (0.73) (0.71) (0.72) (0.26) t=2003 -2.11*** 0.37 -0.24 -0.39 0.026 (0.27) (0.73) (0.71) (0.72) (0.25) t=2004 -2.16*** 0.29 -0.31 -0.43 -0.050 (0.26) (0.72) (0.70) (0.72) (0.25) t=2005 -2.19*** 0.31 -0.31 -0.45 -0.047 (0.26) (0.73) (0.71) (0.72) (0.25) t=2006 -2.19*** 0.43 -0.23 -0.42 0.049 (0.26) (0.75) (0.74) (0.73) (0.25) t=2007 -2.18*** 0.43 -0.23 -0.40 0.047 (0.26) (0.75) (0.73) (0.73) (0.25) t=2008 -2.18*** 0.46 -0.21 -0.39 0.069 (0.25) (0.75) (0.74) (0.74) (0.25) t=2009 -2.11*** 0.75 0.0072 -0.31 0.31 (0.26) (0.80) (0.79) (0.74) (0.25) t=2010 -2.11*** 0.63 -0.077 -0.31 0.21 (0.25) (0.77) (0.76) (0.74) (0.25) t=2011 -2.11*** 0.58 -0.11 -0.32 0.17 (0.26) (0.76) (0.75) (0.75) (0.25) Constant -1.60*** 1.48* 0.58 -2.98*** -2.89*** -2.85*** (0.25) (0.83) (1.24) (0.25) (0.68) (0.62) Obs. 66 66 66 67 67 67 1226 1226 1226 702 1226 F-Stat. 40.1 31.8 19.3 1.3 0.3 0.7 338.2 416.6 415.4 310.6 151.0 R2 0.3 0.3 0.4 0.0 0.00 0.0 1.0 1.0 1.0 1.0 0.4 Adj. R2 0.3 0.3 0.3 0.0 -0.0 -0.001 1.0 1.0 1.0 1.0 0.4 ph,i kh,i Notes: The table reports coefficient estimates from linear regressions of ln ph,n kh,n on log real GDP per capita and predicted values based on country specific industrial composition, both separately and jointly. Data are taken from the WDI. Predictions based n r +δn j =1 αj,i Yh,j on industrial composition are based on equation (12) and computed as ln r +δi · n , with r = 0.03, δn = 0.06, and j =1 αj,n Yh,j δi = 0.2. Standard errors are clustered on country and reported in parentheses below each coefficient. Significance levels are indicated by * p < 0.1, ** p < 0.05, and *** p < 0.01. 26 power, with R2 substantially lower than in columns (C.1) through (D). Notice, however, that (de- spite the overall low explanatory power) this is the only specification in which we get a marginally significant coefficient on income per capita. This result is consistent with column (C.2), which sug- gests that the coefficient on our industry predictions is slightly above one. 4.1. Reduced Form Within Industry Evidence As a final test of our model’s predictions we investigate whether there is a notable within-industry relation between ICT abundance and real income. Note that, if cross-country differences in ICT abundance were truly driven primarily by variation in industrial composition, then we should not expect a strong relationship between ICT abundance and real income per capita within a given industry. Due to data limitations we cannot perform a direct test of this prediction but we provide some suggestive evidence using reduced form analysis based on an approximation. Specifically, we assemble a sector by country by year panel with a proxy for ICT abundance based on WITSA’s detailed tables on sector specific total ICT spending and total capital stocks based on the World Input Output Database (WIOD).16 Note that we would ideally like to construct relative ICT stocks as in Section 2. However, at the sector level, we only have aggregate WITSA ICT spending data. We do not have any data on telecommunication spending from the ITU at the sector level. Thus, as a crude proxy for ICT abundance we simply divide WITSA’s measure of sector-specific nominal ICT spending by the corresponding nominal sector-specific capital stock reported in the WIOD (in current USD).17 The resulting panel spans the period 2003-2009 for a total of 14 WITSA sectors in 32 countries. Table 5 summarizes our results. The first column reports the within industry correlation of our relative ICT abundance proxy and log real income per capita. While there is some overall correlation (measured by the correlation coefficient pooled over all years), we only find a statisti- cally significant relationship for two out of fourteen sectors when controlling for time fixed effects. Considering that the dependent variable in this exercise is only a very crude proxy for true ICT 16 This datasource is publicly available at http://www.wiod.org. 17 Note that we manually build a concordance between the SIC sectors in the WIOD and the sectors reported in WITSA. 27 Table 5: Within Sector Correlation of ICT Abundance and Real Income Regression with time effects ρ β s.e. p-value N Manufacturing 0.562 0.008 0.002 0.000 195 Hospitality and Leisure 0.405 0.021 0.006 0.002 195 Retail Trade 0.261 0.010 0.006 0.111 195 Financial Services 0.268 0.040 0.028 0.173 195 Educational Services 0.241 0.003 0.002 0.187 195 Energy & Utilities 0.223 0.001 0.001 0.223 195 Healthcare 0.187 0.007 0.006 0.247 195 Telecommunications -0.222 -0.061 0.066 0.363 195 Construction 0.156 0.005 0.006 0.416 195 Government 0.133 0.004 0.006 0.505 195 Transportation -0.096 -0.002 0.005 0.628 195 Wholesale Trade 0.073 0.002 0.006 0.791 195 Professional Services 0.042 0.000 0.001 0.806 195 Natural Resources 0.011 0.000 0.001 0.903 195 Notes: For each WITSA sector, the first column reports the within-sector cor- relation of Xict /K and ln(GDP/L) over the years 2003-2009 based on 32 countries, where Xict represents sector specific WITSA ICT spending and K is the sector specific capital stock reported in the WIOD. The remaining columns report the results from a regression of Xict /K on ln(GDP/L) and a complete set of time effects. Standard errors are clustered on country. abundance, we take this as an additional piece of suggestive evidence in favor of our results from the previous section. 5. Does India Have Too Little ICT? An Illustrative Example We opened this paper with a comparison of the percent of households that subscribe to broadband internet in the United States and in India, suggesting that internet subscriptions are 30 times more abundant in the United States. Is this 30-fold difference too large or too small? In other words, is ICT under-utilized or over-utilized in India? Assuming that internet subscriptions are proportional to the physical abundance of ICT, we can construct a benchmark level of ICT capital based on the capital labor ratio, the ICT price and industrial composition using the following decomposition: pi ki p ki = ·k· , (13) pk pi pi ki where ki is the physical stock of ICT capital per-capita; pk is the nominal share of ICT capital out 28 Table 6: Industrial Composition and ICT Intensity: United States & India WDI Value Added Share (%) United States India ICT share (%) NICT share (%) Agriculture 1 18 9 91 Industry Excluding Manufacturing 8 12 2 98 Manufacturing 13 15 5 25 Services 78 55 6 30 Notes: The table reports value added shares and ICT/NICT income shares in the four major WDI sectors, both for India and the United States in 2010. Value added shares are based on WDI data and income shares are approximations based on aggregate estimates of the capital specific marginal product based on Eden and Gaggl (2014) and industry specific ICT and NICT output ratios. p of total capital; k is the aggregate capital stock per-capita, and pi is the ratio of the capital price and the price of ICT capital. Our model from Section 4 suggests that we can further decompose pi ki pk to arrive at the following expression: n p j =1 αi,j Yj r + δn ki = k · · n · (14) pi j =1 (αi,j + αn,j )Yj r + δi r + δn where j = 1, .., n are industries. Note that the ratio r+δi is an approximation, using the fact that ICT represents a relatively small share of capital (otherwise, the numerator is some weighted average of δn and δi ). Applying these expressions to the specific case of India and the United States then gives rise to the following benchmark ratio: p kus ( pi )us Ius BR = · p · (15) kind ( p i )ind Iind n j =1 αi,j Yj where us and ind are subscripts for the US and India, and I = n . ( j =1 αi,j +αn,j )Yj Using estimates from the Penn World Table 8.1, the dollar value of the capital stock per capita is about 28.5 times higher in the United States than in India, explaining a large part of this disparity kus (k ind = 28.5). The relative price of aggregate capital relative to the price of ICT capital—roughly equal to the relative price of NICT and ICT capital, given that ICT constitutes a small share of the p (p )us capital stock—is about 1.6 times higher in the US than in India, i.e. p (p i )ind = 1.6. i To compute I , we use the values displayed in Table 6, which presents the value added shares 29 by industry in the United States and in India, together with the breakdown of the ICT and NICT shares by industry in the US. Notice that, while agriculture has the highest ICT share in absolute terms, it also has a substantial NICT share. The same is true for industry excluding manufacturing. Put differently, the relative expenditure share of ICT and NICT is much higher in manufacturing and services than it is in agriculture and industry excluding manufacturing. Since India is produc- ing much more intensively in the latter two sectors (compared to the US) this suggests that India will use less ICT in the aggregate, simply based on its industrial composition. More precisely, to calculate Iind , we compute 18 ∗ 9 + 12 ∗ 2 + 15 ∗ 5 + 55 ∗ 6 Iind = (16) 18 ∗ (9 + 91) + 12 ∗ (2 + 98) + 15 ∗ (5 + 25) + 55 ∗ (6 + 30) Ius We derive Ius analogously and arrive at approximately Iind = 1.4. The benchmark ratio of ICT in the US relative to ICT in India is then given by BR = 28.5 · 1.6 · 1.4 = 63. Thus, for the specific example of India, an observed 30-fold difference in broadband subscrip- tions relative to the US is actually suggestive of ICT utilization rates that are high compared to our benchmark. Given the capital labor ratios, the price differentials and the industrial composition, India appears to be utilizing ICT beyond what would be expected. It is worth emphasizing that the above benchmarking procedure does not utilize our estimates of ICT and NICT capital stocks. The procedure provides a benchmark for ICT prevalence, that requires as inputs only aggregate capital-labor ratios, value added by industry, and estimates of the relative prices of ICT and NICT capital goods. Since capital-labor ratios are available widely from the PWT and industrial composition is available widely from the WDI, the only input that is not readily available is the relative ICT price. Our ICP-based estimates for relative prices are available for a wide variety of countries and can be utilized in this benchmarking procedure. 6. Concluding Remarks This paper documents a correlation between income per capita and the ICT capital stock. This correlation is not surprising: technology adoption lags would imply that capital that embeds rel- 30 atively new technologies would be less abundant in low income countries. This is especially the case if the new technology is substitutable with labor (e.g., Autor, Levy and Murnane, 2003), a factor that is relatively abundant in low income countries. Despite these plausible economic mechanisms, our findings suggest that the correlation be- tween ICT abundance and income per capita is entirely accounted for by differences in industrial composition. Predicted values based on domestic industrial composition and US industry level ICT intensity exhibit the same relationship with income per capita. Moreover, we present evidence suggesting that, within industries, ICT spending is similar across countries of different levels of development. The suggested implication is that technology adoption lags are not very important for the proliferation of ICT. On average, after controlling for industrial composition, the value of ICT capital per worker is the same as the general capital-labor ratio. A higher relative price of ICT capital in low income countries, which is perhaps reflective of higher costs of tradables, implies that the corresponding quantity of ICT capital goods is lower in low income countries. Of course, our analysis here “explains” cross country differences in ICT abundance only in an accounting sense. This accounting exercise illustrates that in order to understand the source of the differences, one must look deeper into the fundamental questions of development economics, and understand why the capital-labor ratios are lower in developing countries; why non-tradables are relatively cheaper; and why production is concentrated disproportionately in agriculture and other industries that do not require much ICT capital. References Acemoglu D, Autor D. 2011. Skills, Tasks and Technologies: Implications for Employment and Earnings, volume 4 of Handbook of Labor Economics, chapter 12. Elsevier, 1043–1171. URL http://ideas.repec.org/h/eee/labchp/5-12.html Autor DH, Dorn D. 2013. The growth of low-skill service jobs and the polarization of the us labor market. American Economic Review 103: 1553–97. URL http://ideas.repec.org/a/aea/aecrev/v103y2013i5p1553-97.html Autor DH, Levy F, Murnane RJ. 2003. The skill content of recent technological change: An empirical exploration. The Quarterly Journal 31 of Economics 118: 1279–1333. URL http://ideas.repec.org/a/tpr/qjecon/v118y2003i4p1279-1333.html Basu S, Fernald JG, Oulton N, Srinivasan S. 2003. The Case of the Missing Productivity Growth: Or, Does Information Technology Explain why Productivity Accelerated in the US but not the UK? NBER Working Papers 10010, National Bureau of Economic Research, Inc. URL http://ideas.repec.org/p/nbr/nberwo/10010.html Bloom N, Sadun R, Van Reenen J. 2012. Americans Do IT Better: US Multinationals and the Productivity Miracle. American Economic Review 102: 167–201. URL http://ideas.repec.org/a/aea/aecrev/v102y2012i1p167-201.html Colecchia A, Schreyer P. 2002. ICT Investment and Economic Growth in the 1990s: Is the United States a Unique Case? A Comparative Study of Nine OECD Countries. Review of Economic Dynamics 5: 408–442. URL http://ideas.repec.org/a/red/issued/v5y2002i2p408-442.html Comin D, Hobijn B. 2010. An Exploration of Technology Diffusion. American Economic Review 100: 2031–59. URL http://ideas.repec.org/a/aea/aecrev/v100y2010i5p2031-59.html Davis DR, Weinstein DE. 2001. An Account of Global Factor Trade. American Economic Review 91: 1423–1453. URL http://ideas.repec.org/a/aea/aecrev/v91y2001i5p1423-1453.html Eden M, Gaggl P. 2014. The substitution of ICT capital for routine labor: Transitional dynamics and long-run implications. SSRN Scholarly Paper ID 2432313, Social Science Research Network, Rochester, NY. URL http://papers.ssrn.com/abstract=2432313 Feenstra RC, Inklaar R, Timmer M. 2013. The Next Generation of the Penn World Table. NBER Working Papers 19255, National Bureau of Economic Research, Inc. URL http://ideas.repec.org/p/nbr/nberwo/19255.html Gu W. 2012. A comparison of official and euklems estimates of mfp growth for canada. Technical report, Statistics Canada, Economic Analysis Division. Gust C, Marquez J. 2004. International comparisons of productivity growth: the role of information technology and regulatory prac- tices. Labour Economics 11: 33–58. URL http://ideas.repec.org/a/eee/labeco/v11y2004i1p33-58.html Hsieh CT, Klenow PJ. 2007. Relative Prices and Relative Prosperity. American Economic Review 97: 562–585. URL http://ideas.repec.org/a/aea/aecrev/v97y2007i3p562-585.html Inklaar R, Timmer M. 2013. Capital, labor and tfp in pwt8. 0. University of Groningen (unpublished) . Jorgenson DW, Vu K. 2005. Information technology and the world economy*. Scandinavian Journal of Economics 107: 631–650. ISSN 1467-9442. URL http://dx.doi.org/10.1111/j.1467-9442.2005.00430.x 32 Karabarbounis L, Neiman B. 2014. The global decline of the labor share. The Quarterly Journal of Economics 129: 61–103. URL http://qje.oxfordjournals.org/content/129/1/61.abstract ıos-Rull JV, Violante GL. 2000. Capital-Skill Complementarity and Inequality: A Macroeconomic Analysis. Krusell P, Ohanian LE, R´ Econometrica 68: 1029–1054. URL http://ideas.repec.org/a/ecm/emetrp/v68y2000i5p1029-1054.html O’Mahony M, Timmer MP. 2009. Output, Input and Productivity Measures at the Industry Level: The EU KLEMS Database. Economic Journal 119: F374–F403. URL http://ideas.repec.org/a/ecj/econjl/v119y2009i538pf374-f403.html Voskoboynikov IB. 2012. New measures of output, labour and capital in industries of the Russian economy. GGDC Research Memo- randum GD-123, Groningen Growth and Development Centre, University of Groningen. URL http://ideas.repec.org/p/dgr/rugggd/gd-123.html Vries Kd, Vries Gd, Timmer M. 2014. Patterns of Structural Change in Developing Countries. GGDC Research Memorandum GD-149, Groningen Growth and Development Centre, University of Groningen. URL http://ideas.repec.org/p/dgr/rugggd/gd-149.html Vu K. 2005. Measuring the impact of ict investments on economic growth. Manuscript, Harvard University. Appendix A. Capital Stocks Based on WITSA & ITU As described in Section 2, we construct nominal ICT investment series based on data from the the World Information Technology and Services Alliance (WITSA) and the International Telecom- munication Union (ITU) databases. We then treat NICT investment as the residual between total capital investment in the PWT and our measure of ICT investment. WITSA produced reports (Digital Planet 1998, 2000, 2002, 2004, 2006, 2008, 2010) that provides data on Information Communication Technology investment from 1993 to 2011. The number of countries varies across the reports (55-75 countries). WITSA uses data provided by International Data Corporation (for reports 1998, 2000, 2004) & Global Insights, Inc. (for reports 2004, 2006, 2008, 2010). For more details please see http://www.witsa.org. ITU publishes the World Telecommunication/ICT Indicators database covering data on 150 telecommunication/ICT statistics from 1975 to 2013 for over 200 countries. ITU collects it data from annual questionnaire that are sent to official country contacts. For more details please see http://www.itu.int/en/ITU-D/Statistics/Pages/publications/wtid.aspx 33 Taken together WITSA and ITU provide US with data on IT and TC investment for an unbal- anced sample of 70 countries that also appears in the PWT for the period 1993-2011. About 40 countries have have complete data series for all years from all sources but for many countries (es- pecially for poor countries) the data is much more sparse. Therefore, like in the construction of the PWT capital investment series (Inklaar and Timmer, 2013), we employ two layers of interpolation and two layers of extrapolation. First, we interpolate ”in sample” (1993-2011) missing values in proportion to total WITSA ICT spending where available, and aggregate PWT capital investment otherwise. To facilitate the construction of starting values for our ICT stocks in 1993 we further extrapolate the ”in-sample” investment series backward to 1950 in proportion to aggregate PWT investment and a time trend where PWT investment is available (many countries in PWT only have investment data starting in 1970) and any remaining missing values are extrapolated based purely on a log linear time trend.18 To compute the stock of ICT and NICT we deflate the resulting investment expenditures (from 1950-2011) using capital specific price indexes as estimated by Eden and Gaggl (2014). We then use Ic,0 a version of the standard (Solow) steady state condition, Kc,0 = ¯c +δc,0 , g to estimate an initial value ¯c represents country specific long run in 1950, where Ic,0 is real ICT investment in the first period, g growth, and δc,0 is the depreciation rate in the initial period. We use our extrapolated value in 1950 as the initial value for capital specific investment and use the associated implied constant growth ¯c . Based on this initial capital stocks we then use the perpetual inventory rate as a proxy for g method separately for IT, TC, and NICT capital and iterate on the following equation: Kc,t+1 = Ic,t + (1 − δc,t )Kc,t (A.1) where we assume the following depreciation rates: 31.5% for IT, 11.5% for TC (see Inklaar and Timmer, 2013), and 6% for NICT (as in Eden and Gaggl (2014)).19 We have experimented with a 18 There are three countries for which we use a time trend only to extrapolate backward: Nigeria, Sri Lanka, and Japan. We make this exception because the PWT investment series are poor predictors of ICT investment trends for these countries. 19 Our choice of mapping the NICT depreciation rate to the estimate in Eden and Gaggl (2014) rather than to the PWT is due to aggregation issues, as the PWT distinguishes between several types of NICT assets with different depreciation rates. 34 number of version of this procedure and have found little sensitivity to our results. Appendix B. Proof of Lemma 1 ∂y ∂y ∂y The standard first order conditions imply that ∂kn = δn pn , ∂ki = δi pi and ∂l = w. Given the Cobb-Douglas structure of the production function, NICT capital is paid a fraction α of output: αy = δn pn kn (B.1) The first order conditions with respect to ICT capital and labor can be written as: 1−α α σ −1 σ −1 (1 − α)kn (Aki + lσ ) σ Aki = δi pi (B.2) 1− α −1 α (1 − α)kn σ (Aki + lσ ) σ lσ−1 = w (B.3) Dividing the two yields: ki σ−1 δi pi ki δi pi ki ( ) = ⇒ A( )σ = (B.4) l Aw l wl pi ki The first part of the above equation illustrates that if Aw is higher, the optimal choice of l is lower ki (note that σ < 1). The second part of the above equation illustrates that, if the ratio l is lower, the pi ki relative expenditure on ICT capital ( δiwl ) is lower. Given the Cobb-Douglas assumption, it is easy to establish that: δi pi ki + wl = (1 − α)y (B.5) pi,h Let si denote the ICT share. Given that Ah wh is decreasing in h, it follows that sl,h is increasing sl,h in h. Thus, the ratio α is increasing in h, and therefore (under the assumption that depreciation pi,h ki,h rates for ICT and NICT are the same across countries) pn,h kn,h is increasing in h. 35