A journey into uncharted territory: How to determine who spends what on infrastructure in a context of scarce data and resources Summary We provide the first consistently estimated dataset on infrastructure investments in low and middle- income countries (LMICs). To do so, we identify three possible proxies for infrastructure investments— two are variants on gross fixed capital formation from National Account systems following ADB 2017, another is based on fiscal data from the World Bank’s BOOST database—all but one, relying on the World Bank’s PPI database to capture the private share of infrastructure investments. Given the limitations of each of these proxies, we rely on a number of transformations to derive a lower bound estimate for infrastructure investments in LMICs of 3.40 percent of LMICs’ GDP, a central estimate of around 4 percent, and an upper bound estimate of 5 percent all for 2011. Corresponding absolute amounts are $0.82 trillion, $1.00 trillion, and $1.21 trillion, respectively. The public sector largely dominates, accounting for 87-91 percent of infrastructure investments, but with wide variation across regions from a low of 53-64 percent in South Asia, to a high of 98 percent in East Asia. Given the absence of fiscal or national account data capturing investments in infrastructure, estimates such as ours are likely to be the best available in the near future. Nevertheless, we propose some possible avenues for future improvements (including an update when 2015 data is made available by the International Comparison Project) building on the excellent existing MDB collaboration around this issue. 2 Contents 1. The state of play on estimating infrastructure spending...................................................................... 6 2. Three possible proxies based on four datasets .................................................................................... 8 2.1 Option one: GFCF of general government (GFCF_GG) combined with private sector infrastructure investments from the PPI database......................................................................................................... 10 2.2 Option two: GFCF of the whole public sector but including only civil engineering (GFCG_CE) ....... 11 2.3 Option three: Budgetary data from the BOOST data set (BOOST) combined with the PPI data set 12 2.4 Pros and cons of our three proxies—including how much work they entail .................................... 13 3. Too hot, too cold, just right… .............................................................................................................. 16 3.1 Comparing across our estimates ....................................................................................................... 16 3.2 A robustness check ............................................................................................................................ 20 4. Why settle with what we have? Refining our estimates ................................................................... 21 4.1 Refinement 1: Fitted values using all our proxies ............................................................................. 21 4.2 Refinement 2: BOOST+PPI supplemented by the minimum of the two GFCF estimates.................. 23 4.3 Refinement 3. Attempting to correct for the inclusion of non-infrastructure sectors..................... 24 4.4 Refinement 4. Attempting to improve BOOST data by using SPI data on SOE shares ..................... 25 4.5 Pulling it all together: triangulating to define a range ...................................................................... 26 5. Making the most of rich data sets: what else can we learn about infrastructure spending by exploring the BOOST database ................................................................................................................... 29 5.1 BOOST insights: low rates of execution and frequent mis-classification of recurrent expenditures as capital spending .................................................................................................................................. 29 6. Conclusion ........................................................................................................................................... 33 References .................................................................................................................................................. 35 Annex 1: Country level results .................................................................................................................... 37 Annex 2: Addressing data quality through BOOST Sample description ..................................................... 41 Annex 3. Improving Infrastructure Investment Estimates Using Disaggregated Data .............................. 45 3 Acknowledgments This report was produced by a team composed of Hyoung Il Lee and Massimo Mastruzzi (co-TTLs), Sungmin Han and Moonkyoung Cho, under the leadership of Marianne Fay. We are grateful to Deblina Saha for many exchanges as well as for sharing the results of her work on publicly reported data on infrastructure projects sponsored by public entities (a key input for one of our “refinements�). We are grateful for the support extended to us by our Asian Development Bank colleagues, Rana Hasan and Kaushal Joshi who kindly shared the methodology, knowledge, and experience accumulated while developing their Asia infrastructure estimates. We also want to acknowledge the extensive collaboration of our InterAmerican Development Bank colleagues, Tomas Serebrisky, Ancor Suarez Aleman, and Cinthya Pastor Vargas. Finally, we wish to thank the many additional reviewers who have helped sharpen the focus of this study: Luis Andres (Lead economist, GWAGP), Jens Kristensen (Lead public sector specialist, GGOSC), Vivien Foster (Senior Economic Advisor, GGIVP), Punam Chuhan-Pole (Lead economist, DCFII), and Jordan Z. Schwartz (Director, GTPDR). 4 A journey into uncharted territory: Who spends what on infrastructure in a context of scarce data and resources “The infrastructure gap is large: 1.2 billion individuals are without electricity, 663 million lack improved drinking water sources, 2.4 billion lack improved sanitation facilities, 1 billion live more than 2 kilometers from an all-weather road, and uncounted numbers are unable to access work and educational opportunities due to the absence or high cost of transport services. Infrastructure in low- and middle-income countries (LMICs) falls short of what is needed for public health and individual welfare, environmental considerations, and climate change risks —let alone economic prosperity or middle-class aspirations.� Beyond the Gap—How countries can afford the infrastructure they need while protecting the planet, J. Rozenberg and M. Fay (Eds), The World Bank, forthcoming 2019. As the above quote makes abundantly clear, economic infrastructure continues to represent a significant challenge for low- and middle-income countries (LMICs). In fact, the challenge is worse than the pure access gap described above as quality is often also an issue. Being connected to the grid does not ensure that electrons will flow reliably and continuously; having a water connection in the house does not guarantee that the water is safe or will flow 24/7. The result is that infrastructure services are a major impediment to both growth and improved welfare in many LMICs (Rozenberg and Fay 2019 forthcoming). The infrastructure gap has in fact been the subject of much attention in recent years, with a plea for countries to invest more and efforts to attract the private sector in the hope that it will bring fresh capital and greater efficiency to infrastructure services. Yet, despite all this attention, remarkably little is known about how much countries should be spending on infrastructure to get the services they want or about how much they actually are spending on infrastructure—the two sides of the “spending gap� that is often highlighted in the public debate on infrastructure. The World Bank report cited above addresses the first question by providing careful and well documented estimates of what countries need to spend on infrastructure, making the point that this depends on countries’ goals and spending efficiency (Rozenberg and Fay 2019 forthcoming). The present report tackles the question of how much countries actually spend. Infrastructure economists are regularly asked how much countries spend on infrastructure, and it is the subject of endless surprise that this data is not generally available. The issue is that infrastructure spending spans different functional and economic classifications in fiscal accounts; it can be undertaken by different actors—different ministries, central or local governments, or state-owned enterprises (SOEs) 5 that are not usually included in readily available government statistics, or by the private sector. Infrastructure spending includes both capital and recurrent expenditure, which can be difficult to classify or track in fiscal accounts, even where detailed budgetary data is available. And finally, infrastructure is both everybody and nobody’s business—no single agency is concerned with tracking infrastructure as with WHO for health, UNESCO for education, or FAO for agriculture. In other words, figuring out how much countries spend on infrastructure is challenging. As such, the approach we use here is an old-fashioned “triangulation� one, where we look at several different possible methods and data sources, each with some advantages and disadvantages, and different errors of inclusion and errors of exclusion in what they cover. The approach builds on three initiatives—one, known as the BOOST database, managed by the World Bank with financing from the Gates foundation, involved collecting the full fiscal accounts of 55 countries and painstakingly analyzing that data; the other, follows the methodology developed by the Asian Development Bank (ADB) for its report on Meeting Asia’s Infrastructure Needs (ADB 2017) relies on national accounts data. A third is a World Bank effort involving collecting publicly reported data on infrastructure projects sponsored by public entities (Saha 2019) The result is a reasonable set of estimates—a high, a low, and a preferred one—for 120 countries; a detailed analysis of infrastructure spending and budgeting challenges for the 55 countries for which BOOST fiscal data is available; and a better understanding of how, going forward, it might be possible to develop a set of reliable estimates on how much countries spend on infrastructure. The report is structured as follows. Section 1 discusses previous efforts at estimating infrastructure spending. Section 2 explains the four different datasets available as well as their relative advantages. Section 3 compares the results obtained using these data sets, while Section 4 propose methods to combine them to “triangulate� and improve accuracy. Section 5 uses the BOOST fiscal data to provide some trend analysis and some discussion of budgeting challenges. The last section concludes and discusses potential future directions for further strengthening our understanding of what countries spend on infrastructure and the use that can be made of such data. 1. The state of play on estimating infrastructure spending The most common and easiest—if inexact—way to estimate public spending infrastructure involves relying on the gross fixed capital formation of general government (henceforth GFCF_GG). This is readily available from the IMF’s Investment and Capital Stock Dataset and covers all capital investments (dwellings, civil engineering, machinery and equipment) of central, state or local governments. It is available for most countries in a time series format but does not allow for any sectoral disaggregation. Unfortunately, and as discussed in more details below, GFCF _GG is unlikely to be a good measure of public investment in infrastructure—although whether it is an under or over-estimate will vary across countries and over time. It excludes SOEs (which tend to be significant investors in infrastructure) but it includes non-infrastructure sectors (health, education, mining, etc.) 6 On the private side of infrastructure investments, the situation is rosier given the existence of the Public Private Investment in Infrastructure (PPI) database, a World Bank initiative that harks back to the 1990s when private investment in infrastructure took off. The PPI database collects information on both the public and private share of public-private partnerships in infrastructure. Its main limitation is that it records commitments rather than actual investments. In addition, several efforts were made to look into more details at regional spending on infrastructure, typically through painstaking efforts to work with public authorities to collect fiscal data. Perhaps the earliest one was done for Africa in the context of the Africa Infrastructure Country Diagnostic (Foster and Briceño-Garmendia 2010). It found that in 2007 Africa spent approximately $25 billion or about 1.5 percent of its regional GDP on infrastructure capital investments (public and private) in addition to about $20 billion on operations and maintenance. The report’s conclusion was that after potential efficiency gains, Africa’s infrastructure funding gap was approximately $31 billion, mostly in the power sector. A similar effort, also at the World Bank, was conducted for South Asia (Andres, Biller, and Herrera 2013). It found that infrastructure spending in South Asia had increased from 4.7 percent of GDP in 1973 to 6.9 percent in 2009, but fluctuating considerably during that time period. However, service gaps remain significant, leading the report to conclude that a mix of investments and supportive reforms were needed. More recently, the InterAmerican Development Bank (IDB), collaborating with the UN’s economic Commission for Latin America (ECLAC) and CAF (Development Bank of Latin America), worked with budgetary authorities of the countries of the region to carefully exploit fiscal accounts and develop estimates of public investment in infrastructure. These are combined with the PPI database and reported in the INFRALATAM database, providing infrastructure spending data for 15 Latin America countries from 2008 to 2015, disaggregated by sectors and public or private source.1 The most recent effort was undertaken by the ADB (ADB 2017), which experimented with three different approaches by combining country specific estimates where available, two different national accounts indicators based on gross fixed capital formation data, and the PPI database. ADB (2017) offers a detailed discussion of the methodology and the strengths and weaknesses of the different estimates. This approach is the one we build on, as explained in more details in the next section. Based on these studies and compilations of other (much rougher) estimates, Fay and others (2017) provided a very tentative estimate of global spending on infrastructure (table 1). Table 1. Summarizing previous estimates of infrastructure investments among developing regions Region East Asia Central Latin America Middle East South Sub- and Asia and the and North Asia Saharan Pacific Caribbean Africa Africa Infrastructure spending 7.7 4.0 2.8 6.9 5.0 1.9 (share of GDP, 2014) Source: Fay and others 2017 1 http://www.infralatam.info/ 7 Finally, the World Bank recently developed a regional baseline of public spending by leveraging the wealth of micro fiscal data collected by the BOOST initiative in over 55 countries (with another 15 in progress). This allows to examine annual trends, execution rates, funding sources and levels of capital expenditure by the general government across infrastructure sectors. The BOOST database covers 25 countries in Africa—which has enabled the World Bank to develop a regional baseline of annual public spending across infrastructure sectors in Sub-Saharan Africa (World bank 2017). It also includes 11 Latin American and Caribbean countries, which the World Bank and IDB teams are using to derive investment estimates ground-truthed in the IDB’s country specific fiscal analysis. The hope is that, in the future, estimates could be derived from BOOST data instead of requiring costly country visits. 2. Three possible proxies based on four datasets Infrastructure typically includes Transport, Energy, and Water and Sewerage, with some debate about whether information and communication technology (ICT), flood defense, and irrigation should be included. In what follows, we include the following four sectors: • Transport includes civil engineering works on highways, bridges, streets, roads, railways, tunnels, airfield runways, ports/harbors, waterways, and related harbor and waterway facilities, among others, as well as non-residential buildings. Related machinery and equipment, including ICT, are also included. • Energy encompasses non-residential buildings and civil engineering works for power plants, power stations, hydroelectric dams, electricity grids, long-transmission lines, power lines, transformer stations, and gas and oil pipelines, among others. Related machinery and equipment, including ICT, are also included.2 • Water and Sewerage includes non-residential buildings, civil engineering works and machinery and equipment for dams, irrigation and flood control waterworks, local water and sewer mains, local hot-water and steam pipelines, sewage, and water treatment plants. Related machinery and equipment, including ICT, are also included. However, BOOST does not include irrigation. • Information and Communication Technology comprises non-residential buildings and civil engineering works for telephone and internet systems, land- and sea-based cables, communication towers, and telecommunication transmission lines, among others. Related machinery and equipment, including ICT, are also included. PPI data excludes fully privatized investment that is not part of public infrastructure project. We use five different data sets to construct four different estimates of capital spending in infrastructure (Figure 1): 2 The World Bank standard definition of economic infrastructure typically would exclude gas and oil pipelines, but fiscal data and national accounts data does not allow us to separate our power (electricity). 8 • Two estimates based on systems of national accounts.3 Here we follow ADB (2017) and rely on GFCF as the key macroeconomic aggregate in national accounts relevant to assessing infrastructure investment. GFCF measures the total value of fixed assets that are used in production for more than one year plus certain specified expenditures on services that add to the value of non-produced assets (European Commission et al. 2009). As depicted in Table 2, in standard national accounts, an economy’s GFCF can be further decomposed by the type of institution making the investment—general government, non-financial corporate (which in national accounts includes SOEs), financial corporate, non-profit institutions serving the household sector, and the household sector. GFCF can also be classified by type of asset (buildings, civil engineering, and machinery and equipment). Infrastructure investments are only a subset of a country’s GFCF or overall investments (other sectors include health, education, mining, defense, etc.) as depicted in Table 2. The available data limits us to two (imperfect) options, which are discussed below: (i) GFCF of the general government, complemented with the PPI database to capture private investment; and (ii) the civil engineering portion of total GFCF (which includes the GFCF of the general government, SOEs, and the private sector). • An estimate based on national Treasury systems using the BOOST database, which has produced over 70 national and subnational datasets containing well-classified and highly disaggregated budget data. • An estimate of private investment in infrastructure using the PPI database. Figure 1. Four data sets to derive three proxies Table 2. Infrastructure (hatched) is a subset of an economy’s capital investments (shaded) 3 This paragraph is based on ADB (2017). 9 Note: The shaded area represents examples Gross Fixed Capital Formation and the hatched part is the infrastructure subset. These different options are discussed in the next subsections. 2.1 Option one: GFCF of general government (GFCF_GG) combined with private sector infrastructure investments from the PPI database GFCF_GG captures gross fixed capital formation by central, state, and local governments as reported in the IMF’s Investment and Capital Stock Dataset. It is expressed here as a share of GDP (both in constant 2011 international dollars). While it has the advantage of ready accessibility, wide coverage (119 emerging economies), and long time-series, it unfortunately suffers from both: • Errors of exclusion: o SOEs—a major issue given that SOEs often dominate the water and sanitation and electricity sectors, as well as much of transport (ICT however tends to be largely private). o Private sector investments (which we offset by combining GFCG_GG with the private share of PPI investment from the PPI database as described in Box 1.) • Errors of inclusion: residential dwellings and non-infrastructure sectors (including health, education, defense, mining, etc.). In other words, GFCF_GG could either underestimate or overestimate public investment in infrastructure depending on the relative importance of infrastructure SOEs vs non-infrastructure sectors’ share in public investment. An additional drawback is that the data available to us has no sectoral breakdown Private investment in infrastructure is obtained from the World Bank’s PPI database. It is disaggregated by sectors and collected annually. However, the PPI data records commitment rather than actual spending, and does not track fully privatized investment (e.g. telecom network facility 10 investment by a fully private company, or captive infrastructure). Importantly, since it typically reports total commitments associated with a public-private partnership, we subtract the public portion to avoid double counting (Box 1). To transform commitments into actual spending, we annualize commitments for the following five years—an admittedly crude adjustment.4 Box1: Combining the PPI and GFCF_GG data sets while avoiding double counting The PPI database collects information on infrastructure investment through public private partnerships. As such they include both private and public investments. We therefore remove the public (national or donor) share of the investments for each project as per the information available from the database, a non-trivial task, but one that is possible given the project information available in the PPI database. Box Table 1.1 Private spending in infrastructure through PPPs in 2011 (annualized), by regions and sectors (US Million $ and percent of GDP in parenthesis) Transport Water & Energy ICT Total Sewerage 761 599 12 5,237 6,608 Africa (0.07) (0.06) (0.00) (0.50) (0.64) East Asia and 6,038 1,833 378 2,834 11,083 the Pacific (0.06) (0.02) (0.00) (0.03) (0.12) Europe and 6,354.0 1,241 20 9,901 17,517 Central Asia (0.17) (0.03) (0.00) 0.26) (0.46) Latin America 9,753 634 13,149 6,997 30,533 and the (0.19) (0.01) (0.25) (0.13) (0.59) Caribbean Middle East 777 461 667 2,211 4,115 and North (0.05) (0.03) (0.05) (0.16) (0.29) Africa 20,355 9,205 38 8,493 38,091 South Asia (0.90) (0.41) (0.00) (0.37) (1.68) 47,323 23,409 1,543 35,673 107,947 Total (0.20) (0.10) (0.01) (0.15) (0.45) Source: World Bank, Private Participation in Infrastructure Database, as of November 2017. The number in parenthesis shows the weighted regional average share of regional GDP. 2.2 Option two: GFCF of the whole public sector but including only civil engineering (GFCG_CE) The International Comparison Program (ICP) database of the World Bank Group provides data on GFCF on construction excluding buildings, which mainly includes expenditures on civil engineering works. This 4 The team understands that five years spread for all the sectors may not be suitable in some sectors. This is another line of future research. 11 constitutes our second GFCF-based approach. We use GFCF_CE from the ICP 2011 data set expressed as a share of GDP (both in current local currency unit). It has the advantage of including SOEs. As with GFCF_GG, GFCF_CE suffers from errors of inclusion and exclusion, albeit different ones: • Errors of exclusion: non-residential buildings (e.g. airport terminals, railway stations) and machinery and equipment (e.g. turbines, locomotives) • Errors of inclusion: civil engineering works of non-infrastructure sectors (including mining, irrigation, recreational facilities, etc.)5 Here again, we could either overestimate or underestimate infrastructure spending depending on the relative importance of investments in non-infrastructure sectors vis-à-vis machinery and equipment. The data does not allow for sectoral breakdowns, and is currently only available for 2005 and 2011, although 2017 data will be available in late 2019 or early 2020.6 As with GFCF_GG, the advantage of this measure is its wide coverage (114 emerging economies) and the fact that it is measured consistently across countries. 2.3 Option three: Budgetary data from the BOOST data set (BOOST) combined with the PPI data set Our third approach relies on expenditure flows from Treasury systems as captured in the BOOST dataset. After extracting executed budget on infrastructure from treasury data, we tagged annual capital spending for 2009-2016 to infrastructure sub-sectors, using economic, administrative, and functional classification. The data was then smoothed over the entire period to produce an annual average. Data quality issues and challenges are discussed in Annex 3 along with ongoing efforts to harmonize estimates with those of partner organizations. While a much richer dataset, this, again, is an imperfect one: • Errors of exclusion: o Private sector investments (which we offset by combining with the private share of PPI investment from the PPI database) o Under-identification of sectoral spending when national classification does not clearly identify sectoral spending o Spending by SOEs except for national capital transfers (see annex 2 for a discussion) o Other important off-budget spending (due to missing execution data of foreign funded spending in some high aid countries) 5 Irrigation is often considered infrastructure—however, since it is not included in our other estimates we refer to its inclusion as an error of inclusion. 6 The 2017 round is ongoing, with the final outputs are scheduled for release in 2019. 12 • Errors of inclusion: potential mis-tagging of non-infrastructure spending, although this risk is minimal Typically, this approach would tend to under-estimate overall spending especially when the prevalence of off-budget spending (SOEs or donor funding) is high. On the other hand, the data allows for in depth sectoral analysis (as presented in Section 5) and is available on an annual basis since 2009. One of the key advantages of this approach is that once the initial investment of tagging the raw data has been completed, subsequent annual updates are easy to compute provided the BOOST data sets are updated on a regular basis. 2.4 Pros and cons of our three proxies—including how much work they entail Our three proxy options enable us to achieve a wide geographic and economic coverage: the 114 countries covered by the GFCF_CE option and the 118 covered by GFCF_GG+PPI represent more than 95 percent of low and middle-income countries GDPs (Tables 3 and 4). The BOOST data only covers 55 countries (50 of which overlap with the other data sets) and 24 percent of LMICs GDP. However, it offers good coverage for Africa and Latin America.7 Finally SPI + PPI data includes 110 countries. Table 3. Country coverage by estimation methodologies (number of countries) GFCF_GG GFCF_CE BOOST Africa 44 44 25 East Asia and the Pacific 11 11 5 Europe and Central Asia 21 19 9 Latin America and the 25 24 11 Caribbean Middle East and North 10 9 1 Africa South Asia 7 7 4 Total 118 114 55 Table 4. Country coverage by estimation methodologies (% share of regional GDP covered) GFCF_CE BOOST GFCF_GG Africa 98.6 98.6 25.7 East Asia and the Pacific 99.6 99.6 0.8 Europe and Central Asia 99.2 97.9 7.2 Latin America and the Caribbean 98.7 89.3 83.8 Middle East and North Africa 92.8 90.3 2.9 South Asia 99.2 99.2 15.9 Total 99.1 96.5 24.0 7 Kiribati, Solomon Islands, Kosovo, Argentina, and Afghanistan do not have GFCF estimates due to a lack of national account. 13 Table 5 summarizes the different characteristics and sources of the five data sets. Table 5. Data structure used by three options Data Year Characteristic sources BOOST Annual average based on -Executed investment World Bank 2009-2017 data -Sectoral breakdown http://boost.worldbank.org/ GFCF_GG 2011 -Executed investment IMF -Headline numbers only (Investment and Capital Stock Dataset) GFCF_CE 2011 -Executed investment World Bank -Headline numbers only (International Comparison Program) PPI 2011 -Planned investment World Bank (annualized over five -Sectoral breakdown https://ppi.worldbank.org/ years) -Public portion subtracted In conclusion, all options present some advantages and some drawbacks. GFCF_CE and GFCF_GG+PPI may either underestimate or overestimate actual investments depending on: • The relative importance of infrastructure SOEs. Where these are big investors then GFCF_GG+PPI and BOOST would likely be under-estimates, and GFCF_CE would be preferable. Where infrastructure SOEs are small, BOOST (where available) would be the preferred option, followed by GFCF_GG+PPI. • The relative importance of civil engineering investments relative to non-residential buildings and machinery and equipment. Where civil engineering is a small share of public investment, GFCF_CE may be an underestimate. Where civil engineering dominates, either GFCF_CE or BOOST are likely to be the preferred estimate depending on the importance of SOEs. • The share of non-infrastructure sectors in civil engineering. Where these dominate national investment, both GFCF-based estimates could be over estimates, and BOOST may be preferable (depending on the importance of SOEs). More generally, we can conjecture that BOOST will typically be a lower bound estimate, reflecting the fact that SOE investments may not be captured. GFCF_GG+PPI is likely to offer an upper bound estimate except in those countries where SOEs control an important share of public investment. Figure 2 offers a visual comparison among the four different options while table 6 summarizes the pros and cons of each estimate. 14 Figure 2. A visual comparison among four methodologies capturing infrastructure spending Table 6. The pros and cons of different estimates Errors of: Other advantages Other drawbacks Exclusion Inclusion GFCF_GG + -SOEs - Widely available - No sectoral breakdown PPI -Fully privatized (119 countries) - Requires time consuming PPI investment (e.g. - Time series data cleaning to isolate Non- telecom) private share infrastructure GFCF_CE - Non-residential Widely available - No sectoral breakdown sectors buildings (114 countries) - Very limited time series - Machinery and equipment BOOST + PPI -SOEs* n.a. Sectoral - Time consuming -Fully privatized breakdown (BOOST & PPI data investment (e.g. available cleaning) telecom) - Limited sample (55 countries with 15 more in progress) Note: * BOOST captures public capital transfers to SOEs but not self or donor financed capital investments by SOEs. Finally, the different approaches vary tremendously as to the amount of effort they require to produce. The easiest method is GFCF_CE which is available from the International Comparison Project. GFCF_GG is readily available from the IMF, but needs to be combined with the annualized, private element of the PPI dataset. The latter requires time-consuming data transformation of the relevant five-years’ worth of projects to remove the public financing share of the project and avoid double counting. The initial production of BOOST estimates involving combing through fiscal data to ensure the right expenditures 15 are tagged might be time consuming in settings with insufficient functional classification. However, once this is done, producing future annual estimates can be largely automated8. 3. Too hot, too cold, just right… We look at our three proxies in turn. In the case of our two GFCF results, we provide global results both with and without China given its weight—due to both its size (it accounts for about 30 percent of LMICs’ GDP in 2011) and the fact that China’s GFCF numbers clearly include much more than infrastructure. 3.1 Comparing across our estimates The GFCF_GG+PPI option estimates LMICs infrastructure investment to be a somewhat unbelievable 8.61 percent of LMIC GDP in weighted global average, equivalent to some $2.08 trillion in 2011 current $(Figure 3). This drops to 5.24 percent (and $0.87 trillion) without China--whose GFCF_GG+PPI is 16 percent. Apart from China, there are seven other outliers with observations above 15 percentage—all of which are either very small countries (e.g., Dominica, Kosovo, Sao Tome and Principe), or countries where a large mining or hydroelectricity project is likely to account for the high investment (e.g., Angola, Democratic Republic of Congo, Lao PDR). Among the regions, East Asia and the Pacific exhibits the highest level of spending both with China (13.7 percent) and without (5.1 percent). Europe and Central Asia exhibits the lowest investment (3.3 percent). Figure 3. Country and regional results using GFCF_GG+PPI for 118 countries, in 2011 Note: regional averages are weighted using GDP shares. See Annex1 for country data. Source: own estimates Surprisingly, the LMIC average for GFCF_CE is a remarkably similar 8.62 percent of GDP equivalent to $2.03 trillion (figure 4). This drops to 5 percent ($0.81 trillion) without China. Apart from China, there were only 8 The only exception resides in the few countries with no functional classification and little administrative disaggregation. In these cases, estimates were computed by identifying individual capital projects which will need to be identified each year. 16 three other observations above 15 percent, suggesting this proxy generates fewer outliers than GFCF_GG+PPI. East Asia and the Pacific was again found to exhibit the highest spending levels with China (14.64 percent) or without China (9.16 percent), driven particularly by Indonesia’s high estimate of 15.34 percent. Sub-Saharan Africa displayed the lowest estimate (3.86 percent of GDP), a troubling finding given the region’s high infrastructure needs and low GDP. Figure 4. Country and regional results using GFCF_CE for 114 countries, in 2011 o The BOOST+PPI option yielded a much more reasonable, but likely underestimated 2 percent of GDP across its 55-country sample (Figure 5). Applying this share to LMIC GDP yields an estimate of $0.49 for investment in infrastructure. This much lower result is likely driven by country coverage (notably the absence of China) and the incomplete treatment of SOE spending.9 Regional weighted averages are only shown for Latin America and the Caribbean and Africa, as the country coverage is too low to be regionally representative for the other regions. Comparing the three proxies for the 50 countries for which the 2011 data is available, we find, as expected that the BOOST+PPI estimates offers a lower bound compared to National Account estimations with GFCF_GG+PPI an upper bound and GFCF_CE typically in between the other two (Figure 6): 80 percent of BOOST estimates are the lowest; 60 percent of GFCF_CE estimates lie between GFCF_GG+PPI and BOOST+PPI while 80 percent of GFCF_GG+PPI estimates are higher than the other two estimates. 10 This pattern likely reflects the fact that the inclusion of non-infrastructure capital spending (such as dwellings, mines, industrial plants, etc.) by government more than offsets the exclusion of SOE capital 9 If China spending was computed according to the shares identified through National Accounts—as opposed to applying the weighted average computed for the BOOST sample - then the global baseline would exceed $1 trillion. 10 Among the 55 BOOST countries, 5 countries do not have either GFCF_GG+PPI or GFCF_CE estimaties. (Kiribati, Solomon Islands, Kosovo, Argentina, and Afghanistan) 17 spending. Analyzing underlying drivers for these patterns across regions would be an interesting future research project to be able to better capture over and under-estimation biases in these measurements.11 Figure 5. Country and regional results using BOOST+PPI for 55 countries, in 2011 9.0 8.0 7.0 6.0 AFR (3.03) 5.0 LAC 4.0 (1.98%) 3.0 2.0 1.0 0.0 This pattern generally holds even when comparing across the full sample as in Figures 3, 4, and 5, with the exception of East Asia (Table 7). Table 7. Summarizing results using full sample Africa East Europe Latin Middle South All Sample Asia and and America East and Asia available size (# the Central and the North countries countries) Pacific Asia Caribbean Africa GFCF_GG+PPI 6.19 13.71 3.30 4.39 8.238.18 8.61 118 (2011) (5.1) (5.24) GFCF_CE 3.86 14.64 4.91 3.96 5.41 4.77 8.62 (5.07) 114 (2011) (9.16) BOOST+PPI 3.02 n.a. n.a. 2.16 n.a. n.a. 2.00 55 (avg 2009-15) Note: Numbers in parenthesis are average that exclude China. The SPI numbers are not strictly comparable with the others as they cover different years. Note also that we would expect SPI to be larger than the BOOST estimate given that BOOST does not capture SOE investments. 11The regional weighted average for the 50 countries with all three 2011 estimates are shown below for Africa and Latin America (the only regions for which the sample is representative). Africa Latin America and the All Caribbean countries GFCF_GG+PPI 11.93 3.70 4.43 GFCF_CE 5.01 3.72 3.86 BOOST+PPI 3.02 2.16 2.14 Sample size 25 10 50 18 Figure 6. Comparing infrastructure spending proxies across the 50 countries for which fiscal data (BOOST) was available, 2011 19 3.2 A robustness check We test the robustness or accuracy of our options by comparing them with more detailed data from the European Investment Bank (EIB) (Revoltella and Brutscher 2017). A recent data update from Eurostat allowed EIB to distinguish between GFCF in the infrastructure and other sectors and do so by asset types, and capture (1) investment in civil engineering works by infrastructure sectors and (2) investment in non- residential buildings by infrastructure sectors for the 22 countries covered by Eurostat (Annex 1-2). While the data is still not perfect (it does not include machinery and equipment), it is still a better approximation than what we have, in as much as it covers only infrastructure and includes both civil engineering and non-residential structures. Unfortunately, however, this data is only available for 22 high income European countries. The comparison suggests that, at least for high income countries, GFCF_CE is generally a more accurate measure than GFCF_GG, with 20 out of 22 observations of GFCF_CE closer to the EIB estimates than the GFCF_GG+PPI ones (Figure 7). However, assuming the EIB estimate is the better one (even if it is an underestimate given the omission of machinery and equipment), this comparison suggests both GFCF_GG and GFCF_CE generally overestimate infrastructure spending. Even the lower one (GFCF_CE) is about 40- 50 percent higher than the EIB one suggesting that the inclusion of non-infrastructure sectors is an issue. Unfortunately, the EIB database overlapped with BOOST for only one country (Bulgaria) which precluded a comparison. Figure 7. Comparing EIB and three methodologies, for 22 European countries, 2011 10.0 EIB data 9.0 GFCF on CE 8.0 GFCF of GG + PPI 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 We also compared our estimates with the results of two country studies. In the case of Kenya, data was collected for fiscal year 2015-16 on public, SOE, and private investments in transport, water and sanitation, and energy. The resulting estimate was 6.1 percent of GDP, close to the 4.8 percent GFCF_GG+ PPI but considerably higher than the 3.77 percent GFCF_CE one, and 3.2 percent for BOOST+PPI—which admittedly were estimated for a different year (2011). 20 A second comparison was carried out with data collected in Indonesia by the World Bank from budgets reports of central and local governments, SOE financial reports, and PPI data, covering transport, water and sanitation, energy, and irrigation (World Bank 2015). The 2011 result was 2.6 percentage of GDP, reasonably close to the 3.1 percent of GFCF_GG+PPI but substantially less the 15.3 percent of GFCF_CE. It is hard to conclude from these results whether anyone of our three estimates is better than the other. In the case of European countries, GFCF_CE appears to be the better approximation. But the opposite holds with our two country study comparisons. This variation in performance suggest that a hybrid approach that finds a common ground between the three methodologies might be the best approach to enhance the accuracy and robustness of a global baseline. 4. Why settle with what we have? Refining our estimates The above suggests there is scope for further refinement of our measurements. Above all, China’s importance in the overall results calls for more accurate data—we therefore use ADB’s estimate of public infrastructure investment for China (6.31 percent of GDP) in lieu of a BOOST estimate. We then experiment with three more refinements. The first entails using regression analysis over the sample of 50 countries with common coverage between BOOST and the two GFCF estimates to infer missing values for countries not covered by BOOST (refinement 1). The second is a simpler way to combine the data sets and eliminate outliers by using BOOST where available, and the minimum value of GFCF_GG+PPI and GFCF_CE (refinement 2), where it isn’t. The third involves adjusting GFCF_CE’s tendency towards overestimation (likely due to the inclusion of non-infrastructure investment), by adjusting it downward using a crude estimate of the share of non-infrastructure investments in total public investments derived from BOOST data (refinement 3). The fourth attempts to address the omission of SOE investment in BOOST by augmenting it using the regional average share of SOEs in public infrastructure from the SPI database on infrastructure projects (Saha 2019) which combines newly collected data on public infrastructure projects with the PPI database and identify the share of commitments that is financed by SOEs (refinement 4). 4.1 Refinement 1: Fitted values using all our proxies We estimate a regression model for BOOST methodology to get a fitted value for infrastructure spending for each country. The model specification is as follows: Y = α + β1 ∗ Xgg + β2 ∗ Xce + β3 ∗ logGDP + β4 ∗ dummy + ε Here Y is the BOOST estimates (excluding PPI), Xgg and Xce are estimates from GFCF_GG (excluding PPI) and GFCF_CE respectively, log GDP is the logarithm of current GDP in 2011, and the dummy variable identifies whether a country has a federal government system or not.12 12 The team tried a different model that included PPI (e.g. BOOST+PPI and GFCF_GG+PPI) but that resulted in a worse fit. 21 When this regression model is estimated using the sample of 50 countries with all three estimates, all coefficients are significant (Table 7) with an R-square of 0.56. When some of the independent variables in the model are omitted, then some of coefficients are not significant, leading the research team to conclude that the model specification is reasonable. But because GFCF_CE, GFCF_GG, and GDP are correlated, implying multicollinearity, we cannot interpret each coefficient directly. Table 7. Regression result Coefficients Standard Error t Stat P-value Intercept 4.586 2.715 1.689 0.098 GG 0.164 0.034 4.837 0.000 CE 0.153 0.047 3.243 0.002 LogY -0.461 0.261 -1.770 0.084 Federal 1.397 0.682 2.047 0.047 With this regression model combined with the PPI data, we estimate infrastructure spending for countries without BOOST data. This enables us to derive estimates for the large emerging economies not included in BOOST such as India and Russia which we estimate to 4.2 and 1.9 percent of GDP respectively. This approach yields an LMICs average of 3.40 percent of GDP equivalent to $0.82 trillion (Figure 8). It is higher than the original BOOST+PPI estimates but still smaller than either of the GFCF proxies. Refinement 1 suggests that East Asia at 5.36 percent is the region that spends the most and Europe and Central Asia (1.53 percent) and Middle East and North Africa (1.70 percent) the regions that spend the least. Figure 8. Refinement 1 (BOOST or Fitted values) for 120 countries, in 2011 Note: Laos remains an outlier because of a large PPI value of 11.9 percent even though fitted value for BOOST is 2.3 percent. The number of estimated countries increases to 120 thanks to the combination of datasets. 22 4.2 Refinement 2: BOOST+PPI supplemented by the minimum of the two GFCF estimates Refinement 2 uses BOOST+PPI where it is available supplemented by the minimum value of the two GFCF estimate where it is not. Since BOOST+PPI is normally a lower bound, using BOOST as the primary source and complementing it with GFCF estimates for missing values after eliminating outliers makes the new dataset quite conservative. For example, spending is estimated at 6.4 percent for Lao PDR instead of 17.7 percent (GFCF_GG+PPI), 3.1 percent for Indonesia instead of 15.3 percent (GFCF_CE), and 6.3 percent for Equatorial Guinea instead of 20.6 percent (GFCF_GG+PPI). However, if a country has two GFCF estimates with similar bias, then this approach does not necessarily remove outliers. This was the case for the Democratic Republic of the Congo for instance where both GFCF_GG+PPI and GFCF_CE were extremely high (16.5 percent and 25.9 percent of GDP, respectively). Refinement 2 results in a weighted LMIC average of 4.12 percent of GDP, amounting to $1 trillion invested in infrastructure in 2011 (Figure 9). This is very close to the GFCF_GG+PPI estimates—which is unsurprising as it is the dominant data source. When China is omitted, global weighted average is significantly reduced, dropping to 3.13 percent. With China, East Asia and the Pacific is the region who spends the most (5.72 percent) while Africa is the region who spends the least (2.56 percent). If China estimate is omitted, then the East Asia and the Pacific regional weighted average drops to 3.52 percent and Middle East and North Africa becomes the region who spends the most. These results once again emphasize the importance of a deeper investigation to accurately account for China’s high impact on global estimates. Figure 9. Refinement 2 (BOOST or Min[GFCF_GG+PPI, GFCF_CE]) for 121 countries, in 2011 Note: The number of estimated countries increased to 121 because Refinement 2 includes Lebanon that has only GFCF+GG+PPI. 23 4.3 Refinement 3. Attempting to correct for the inclusion of non-infrastructure sectors Refinement 3 entails revising GFCF_CE estimates by attempting to isolate non-infrastructure shares. In the absence of disaggregated data for GFCF_CE, we use the full original BOOST data (which includes complete general government budgets) to identify non-infrastructure spending in the GFCF_CE—namle health and agriculture facilities. This amounted to an average of about 10 percent of total capital expenditure. We then applied this to our original GFCF_CE estimates, reducing it by 10 percent.13 The research team understands that this approach needs to be more rigorously addressed with micro data to account for country idiosyncrasies - but at this juncture it offers the best result given data availability.14 With this refinement, we estimate infrastructure investment at 4.99 percent of GDP, equivalent to $1.21 trillion (Figure 10). Without China, the weighted average drops to 4.39 percent. East Asia and the Pacific (6.71 percent) is the region who spends the most and Latin America and the Caribbean (3.22 percent) is the region who spends the least.15 Figure 10. Refinement 3 (Revised GFCF_CE) for 114 countries, in 2011 13 The regional averages of non-infrastructure portion are Africa 0.89, Latin America 0.94. The team does not think the averages of the remaining regions have representability due to lack of enough data. Please see the table 3 and 4. 14 Each Refinement implies some risks or limitation depending on the nature of refinement. For example, refinement 1 carries the risk that the regression still cannot capture SOE investment because its dependent variable is BOOST which does not have SOE’s investment. And whenever new BOOST data arrives, the regression coefficients will change, and the extrapolations should be recalculated. In addition, other econometric techniques, such as principal component analysis, may be a better approach. Refinement 2 is a simple merged series of three different sources that does not guarantee a significant improvement for a country with similar value for the two GFCF estimates. Finally, refinement 3 has limitation given that the reduced weight is computed on the basis of shares of health and agriculture sectors. Excluding others like mining, industrial plants. etc. which might be significant in some countries. 15 When China estimate is omitted, East Asia and the Pacific regional weighted average increase to 8.23 percent due to Indonesia’s extremely high GFCF_CE (13.79 percent even after refinement). 24 4.4 Refinement 4. Attempting to improve BOOST data by using SPI data on SOE share Refinement 4 addresses the omission of SOE investments in BOOST by using the regional share of SOE investments from the SPI dataset of Saha (2019). The reason for using a regional share is because the SPI dataset yields some rather extreme and simply not believable estimates for a number of countries—with either all investments done by SOEs or all by the general government and none at all by SOEs.16 We then use these regional shares to calculate an estimated total public (SOE and general government) investment from BOOST or the fitted values. 17 We then add the PPI data to get total investment for each country. Regional SOE shares of infrastructure investment from the SPI data set are 0.26 for Africa, 0.46 for East Asia and the Pacific, 0.38 for Europe and Central Asia, 0.25 for Latin America and the Caribbean, 0.35 for Middle East and North Africa, and 0.35 for South Asia. The result is an estimated infrastructure investment of 3.89 percent of GDP in weighted global average, equivalent to $0.94 trillion (Figure 11). If China is omitted, the global weighted average drops to 2.98 percent. East Asia and the Pacific (5.61 percent) is the region that spends the most and Europe and Central Asia (2.18 percent) is the region that spends the least.18 Figure 11. Refinement 4 (Augmented BOOST) for 120 countries, in 2011 16 6 countries record of zero general government investments which is unlikely to be true (Botswana, Malaysia, Thailand, Moldova, Paraguay, and Algeria). And some countries such as China have records close to 1 (China 0.95, Maldives 0.93, Azerbaijan 0.93). We expect this reflects greater projectization of SOE investments. 17 The formula we use is I= I +I = gI+sI = I /(1-s) with ,I , and I denoting total public, general government, and SOE G S G G S infrastructure investments respectively, and g and s the share of total infrastructure investments attributable to the general government and SOEs respectively (so that g+s=1). As such, for any country with a very high “s� (e.g. China, Azerbaijan), w e would have to multiply IG by an absurdly large number (e.g., by twenty for China since its s=0.95) tending to infinity as “ s� approaches 1. 18 When China estimate is omitted, East Asia and the Pacific regional weighted average increase to 2.97 25 4.5 Pulling it all together: triangulating to define a range Our results, shown in Figure 12 and table 8, are a lower bound estimate of 3.40 percent of LMICs’ GDP, a central estimate of 3.89-4.12, and an upper bound estimate of 4.99 percent. Corresponding total infrastructure spending in 2011 amounts to $0.82 trillion, $0.94-$1.00 trillion, and $1.21 trillion, respectively. Rounding it off, this suggest high and low estimates of 3.4 and 5 percent (0.8 and 1.2 trillion) respectively and a central estimate of 4 percent ($1 trillion). Regional weighted averages from the four new resulting estimates are shown in Table 8. Among the regions, East Asia and the Pacific is the region who spends the most regardless of the method. Latin America and the Caribbean and Africa are below average spenders in the “refinement 2� and “refinement 3� scenarios. Comparing our results to those of previous studies (Fay and others 2017), and accounting for the fact that the studies covered different years, we find generally comparable results. In term of regional variations, this paper and Fay and others (2017) show significant consistency, especially for East Asia and the Pacific (top spender), and Africa (bottom spenders), but less so for Middle East and North Africa region which was in fact the least well documented of the regional estimates reported in Fay and others. Figure 12. Weighted regional averages on infrastructure spending by methodologies in 2011 9.0 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 AFR EAP ECA LAC MENA SAR Refinement 1 Refinement 2 Refinement 3 Refinement 4 Fay and others(2017) Table 8. Summary of estimation on infrastructure spending by methodologies Lower bound Central Upper bound Refinement 4 Fay and Refinement 1 estimate Refinement 3 (SOE- others (2017) (fitted values) Refinement 2 (0.9 GFCF_CE) augmented (BOOST or Min BOOST & of two GFCFs) fitted values) (Percent GDP) Africa 1.87 2.54 3.47 2.35 1.9 26 East Asia and the 5.36 5.72 6.71 5.61 7.7 Pacific Europe and 1.51 2.73 4.36 2.16 4.0 Central Asia Latin America 2.02 2.39 3.22 2.52 2.8 and the Caribbean Middle East and 1.67 4.79 4.73 2.45 6.9 North Africa South Asia 3.59 4.42 4.25 4.71 5.0 Global weighted 3.38 4.11 4.99 3.88 average (%) Total spending 0.82 1.00 1.21 0.94 1.5 (trillion $) Target Year 2011 2011 2011 2011 Various Note: All refinements use ADB data for China. The results reported in Fay and others were drawn from a variety of sources, covered different years and derived with varying methodologies and degrees of care. Another way of comparing these four methods is to compare their basic statistics using a box-whisker plot that show median, minimum, maximum, quantile 1 and 3, and outliers (Figure 13). Figure 13. Characteristics of estimation by methodologies Note: x mark is the mean, middle line in the box is the median, boundaries of the box are the lower and upper quartile, upper or lower whisker is the largest or the smallest value of inner points, dots are outlier points (defined as values outside 1.5 times the interquartile range above the upper quartile and below the lower quartile) We derive infrastructure spending amounts for each region using our central estimate applied to regional 2011 GDPs (Figure 14). This suggests that East Asia and the Pacific spent about $549 billion in 2011, 27 accounting for some 54 percent of total infrastructure spending by LMICs with Sub-Saharan Africa accounting for less than 5 percent. Figure 14. Regional distribution of infrastructure spending using central estimate (refinement 2), 2011 SAR, 102 AFR, 39 MENA, 70 LAC, 134 EAP, 549 ECA, 105 Finally, we derive an estimation of the public & private share of infrastructure investments (Table 9). The public sector clearly dominates, accounting for 87-91 percent of total infrastructure investments—but with considerable variation across regions from a low of 53-64 percent in South Asia to a high of 98 percent in East Asia and the Pacific. Table 9. Public share of infrastructure investments (share of total, percent) Lower bound Central Upper bound Refinement4 (refinement 1 estimate (refinement 3) (refinement 3) Africa 66 75 82 73 East Asia and the Pacific 98 98 98 98 Europe and Central Asia 70 83 89 79 Latin America and the 71 75 82 77 Caribbean Middle East and North 83 94 94 88 Africa South Asia 53 62 60 64 All countries 87 89 91 88 Note: these were derived using estimated private share of infrastructure financing from the PPI database as explained in Box table 1.1 and total investments from Table 8. 28 5. Making the most of rich data sets: what else can we learn about infrastructure spending by exploring the BOOST database While BOOST does have the drawback of not being comprehensive both in terms of countries and sources of spending, it does have the distinct advantage of providing perfectly disaggregated data, over time, and for most countries both as budgeted and actual spending. This makes it a great database to learn more about the business of public spending on infrastructure. 5.1 BOOST insights: low rates of execution and frequent mis-classification of recurrent expenditures as capital spending The main output of the BOOST analysis was a global baseline of capital spending in infrastructure for 55 countries, disaggregated by sub-sectors (road/water/air transport, electricity, telecoms, and water and sanitation), spending ministries and economic classification for the period 2009-2016. This allows us to explore issues of financing, credibility, drivers and efficiency. Most importantly it developed a methodological approach that can facilitate annual updates. Public infrastructure investments, at 1.6 percent of GDP on average between 2009 and 2017, is rather small among the 55 countries covered by BOOST (figure 15), partly reflecting its greater coverage of two regions known for relatively low spending in infrastructure (Latin America and Eastern Europe). Importantly, this also reflects difficulties in executing budgeted spending, particularly in Sub-Saharan Africa: overall spending in infrastructure was considerably lower than capital allocations during the same period (around 2.4 percent of GDP), reflecting substantial under-execution of such investments. Figure 15: Infrastructure spending by regions, 2009-2017 % GDP Approved Executed 4.0% 3.5% 3.0% 2.5% 2.0% 1.5% 1.0% 0.5% 0.0% Other Regions Europe & Central Asia Latin America & Sub-Saharan Africa Caribbean 29 BOOST also allows to look at the evolution over time of public spending on infrastructure, which has stagnated over the past 7 years, particularly in the main three sub-sectors. Total capital budget allocations in roads has been decreasing after reaching a peak of 1.8 percent of GDP in 2011. Similarly, capital allocations for water supply and sanitation have reverted to 2009 levels after peaking around 0.6 percent of GDP. Energy allocations have also gradually decreases following a peak of 0.5 percent of GDP in 2012, with increasing reliance on externally funded investment projects and limited domestic revenue mobilization towards this sector. Figure 16: Annual capital flows in main sub-sectors, 2009-16 % GDP 1.80% 1.60% 1.40% 1.20% 1.00% 0.80% 0.60% 0.40% 0.20% 0.00% 2009 2010 2011 2012 2013 2014 2015 2016 energy roads water and sanitation There are however significant differences among the countries in the sample (figure 17). In terms of capital budget allocation, countries like Afghanistan, Fiji, Paraguay, Dominican Republic and Togo experienced increases of 3 percentage points (in terms of GDP) or more between 2009 and 2017s, driven mostly by increases in roads allocations. On the other end of the spectrum, Sao Tome, Angola and Benin recorded the largest percentage decreases in public capital spending in infrastructure from 2009 levels. While Sao Tome and Angola continue to experience large spending levels overall (5.6 and 4.5 percent of GDP respectively), the slowing pace of capital spending in Benin – driven mostly by roads and water and sanitation – is more problematic in light of lower spending levels overall and acute infrastructure gaps. Under-execution affects efficiency of public investment. Analyzing under-execution of public investment (deviation from original approved budget) can help assess the degree to which residual approaches to capital spending might be at play—and more generally to assess the level of credibility built into approved capital budgets. Given the uneven availability of spending of foreign funded aid, the analysis relies on examining deviations between capital allocations and execution of domestically funded projects. This provides a better assessment of the ability of national systems to implement capital projects. 30 Figure 17: Change in allocations (% GDP), 2009-2016 Uruguay Pakistan (Punjab) Trinidad & Tobago Bhutan St. Lucia Bangladesh Afghanistan Sao Paulo Ukraine Rio Grande Do Sul Tajikistan Peru Poland Paraguay Moldova Minas Gerais Macedonia Mexico Kyrgyz republic Haiti Kosovo Bulgaria Guatemala Armenia El Salvador Albania Dominican Republic Solomon Islands Costa Rica Myanmar Chile Mongolia Brazil (federal) Kiribati Argentina Fiji -1.5% -1.0% -0.5% 0.0% 0.5% 1.0% 1.5% -6.0% -4.0% -2.0% 0.0% 2.0% 4.0% 6.0% 8.0% Uganda Togo Tanzania Sierra Leone Senegal Sao Tome and Principe NIger Namibia Mozambique Mauritius Mauritania Mali Liberia Lesotho Kenya Guinea Bissau Guinea Gabon Ethiopia Cape Verde Cameroon Burundi Burkina Faso Benin Angola -12.0% -10.0% -8.0% -6.0% -4.0% -2.0% 0.0% 2.0% 4.0% 6.0% 31 The analysis revealed very large levels of under execution, particularly in Sub-Saharan Africa (Figure 18). On average, over 30 percent of total domestic allocations were not executed every year from 2009 to 2016. This amounted to almost 1 percent of GPD remaining unspent on capital infrastructure every year with roads accounting for two third of overall underspending. This is particularly worrisome since most of domestic resources are mobilized towards roads with capital spending in electricity and water and sanitation still dominated by foreign assistance. Several reasons could account for these high levels of under-execution. Historically, most African countries have taken a residual approach to capital spending which often results in mid-year decrease in disbursement in typical scenarios of over-optimistic revenue projections leading to in year prioritization towards recurrent expenses. Further explanations include manifestations of late releases typical of countries without a proper medium-term capital commitment framework, implementation gaps due to absorptive capacity and other institutional bottlenecks (i.e. inefficient procurement systems), and pervasiveness of corruption and lack of accountability leading to sub-optimal delivery of infrastructure. Figure 18: Under-execution of domestic-funded capital investments 0% Burundi Guinea Under-execution rates (% of GDP) -1% Sierra Leone Guinea Bissau Namibia Cameroon Mali Tanzania -2% Gabon Senegal Burkina Faso -3% Uganda Mozambique NIger Kenya Benin -4% -5% Angola Togo -6% -7% Lesotho -8% GDP per capita A second key finding of the BOOST analysis is that a very large share of reported capital expenditures is in fact recurrent expenditures. A preliminary analysis of four francophone African countries concluded that to be close to 50 percent of capital expenditures were in fact recurrent expenditures. Combining these findings with the high under-execution rates, this analysis provides compelling evidence of the need for looking beyond budget allocation since these would likely overestimate actual capital formation. 32 Figure 19: Composition of Public Investment in selected Countries, 2014 70% 60% 50% 40% 30% 20% 10% 0% Cameroon Senegal Mali Niger 2014 Effective capital expenditure 2014 Recurrent expenditure 6. Conclusion The primary purpose of this study was to develop a consistent set of estimates of infrastructure spending and financing. We explored the robustness and consistency of four alternative methodological approaches to estimate actual capital spending in infrastructure, highlighting the merits and flaws of each. We estimated and analyzed infrastructure spending in about 120 countries according to each methodology. Finally, we established a global baseline on infrastructure investments along with an analysis of stylized facts and implications, recognizing persistent deficiencies in these estimate as well as their catalytic value in motivating further advances in the field. Based on the four methodologies, a reasonable range of infrastructure spending was developed. We combined the informative content of each dataset into representative series and estimate total infrastructure spending to range between $0.8 trillion and $1.2 trillion across the developing world with significant regional variations. In absolute terms, East Asia and the Pacific was found to account for more than half (54 percent) of total infrastructure investment by LMICs followed by Latin America and the Caribbean as a distant second (about 15 percent of total). More work is needed to refine these estimations. Possible ways forward include: - Expanding on the BOOST database to include more countries and complement it with investment data from (at least) large SOEs. Such data has been collected for the electricity sector, and a pilot effort is under way to gauge the feasibility of expanding the data collection to cover water and transport. In addition, collaboration will continue with the other MDBs to ground truth the analysis. - Expand time series for GFCF_CE estimate, which will be possible in 2019 when 2017 data from the International Comparison Program will become available. 33 - As recommended by ADB (2017) and discussed in Annex 3, the ideal would be to access the rich set of disaggregated data that underpins GFCF estimates. This would follow the EUROSTAT approach and collect infrastructure data from national accounts disaggregated by sector and asset type—a task that would require an effort on the part of national statistics bureaus. While ideal, this approach would require a new global initiative to support the national statistics bureaus. 34 References Andres, L., D. Biller, and M. Herrera Dappe (2013). Infrastructure Gap in South Asia: Infrastructure Needs, Prioritization, and Financing. the World Bank. Washington, DC. Asian Development Bank. 2009. Special-Report: Infrastructure for a Seamless Asia. Asian Development Bank. Manila, Philippines. Asian Development Bank. 2017. Meeting Asia’s Infrastructure Needs. Asian Development Bank. Manila, Philippines. Aschauer, D. A. 1989. “Is Public Expenditure Productive?� Journal of Monetary Economics 23 (2): 177–200. Bonaglia, F., E. La Ferrara, and M. Marcellino. 2000. “Public Capital and Economic Performance: Evidence from Italy.� Giornale degli Economisti e Annali di Economia: 221–244. European Commission, International Monetary Fund, Organization for Economic Cooperation and Development, and United Nations. 2009. The System of National Account 2008. New York. Fay, M., L. Andres, C. Fox, U. Narloch, S. Straub, and M. Slawson. 2017. Rethinking Infrastructure in Latin America and the Caribbean – Spending better to achieve more. the World Bank. Washington, DC. Foster, V. and C. Briceño-Garmendia. 2010. Africa’s Infrastructure – A time for transformation. the Agence Française de Development and the World Bank. Washington, DC. Hulten, C. R. and R. M. Schwab. 1991. “Public Capital Formation and the Growth of Regional Manufacturing Industries.� National Tax Journal: 121–134. Holtz-Eakin, D. 1994. “Public-sector Capital and the Productivity Puzzle.� The Review of Economics and Statistics 76 (1). International Monetary Fund. Investment and Capital Stock Dataset, 1960-2015. https://www.imf.org/external/np/fad/publicinvestment/ (accessed on November 2nd, 2017). Kim, Il Tae and Bong Jin Kim. 2008. “Public Infrastructure and Production Efficiency in Regional Economy.� The Korea Spatial Planning Review 59: 103–117 (in Korean). Schwab, K. 2015. “The Global Competitiveness Report 2015-2016.� World Economic Forum. Revoltella, D. and P. Brutscher. 2017. Infrastructure Investment in Europe - New Data, market dynamics, policy actions and the role of the European Investment Bank. the European Investment Bank, Kirchberg, Luxembourg. Rozenberg, J. and M. Fay. 2019. Beyond the gap – How countries can afford the infrastructure they need while protecting the planet. forthcoming. Warner, A. 2014. “Public Investment as an Engine of Growth.� IMF Working Paper 14/148. 35 World Bank. 2015. “Technical note: Estimating Infra Investment and Capital Stock in Indonesia .� Processed., World Bank, Private Participation in Infrastructure Database, http://ppi.worldbank.org/ (accessed on November 2nd, 2017) World Bank Group. 2017. Africa's Pulse, No. 15, April 2017. World Bank, Washington, DC. © World Bank. https://openknowledge.worldbank.org/handle/10986/26485 License: CC BY 3.0 IGO.� World Bank, International Comparison Program 2011 World Bank, BOOST initiative, boost.worldbank.org. World Bank, World Development Indicators, data.worldbank.org. 36 Annex 1: Country level results 1. Consolidated Global Infrastructure Spending Estimation in 2011 with new series Refinement Refine- Refinement2 GG+ BOOST 1 (BOOST ment3 Country PPI CE (Boost or PPI +PPI or_Regress (Revised Min(GG,CE)) ion) CE) Angola 0.22 22.63 6.55 3.33 3.33 3.33 5.89 Benin 2.32 6.62 1.31 3.17 3.17 3.17 1.18 Botswana 0.49 11.33 4.89 2.91 4.89 4.40 Burkina Faso 0.76 10.11 5.16 1.78 1.78 1.78 4.64 Burundi 0.14 6.45 3.17 0.68 0.68 0.68 2.85 Cabo Verde 9.98 15.00 2.71 2.71 2.71 13.48 Cameroon 0.55 4.88 4.60 1.57 1.57 1.57 4.14 Central African 0.33 4.03 1.94 1.52 1.94 1.74 Republic Chad 0.42 4.54 7.26 2.15 4.54 6.53 Comoros 2.33 2.02 1.24 2.02 1.82 Congo, Republic of 16.50 25.90 6.49 16.50 23.29 Cote d'Ivoire 1.21 2.55 0.69 1.44 0.69 0.62 Dem. R. of the Congo 0.43 4.39 3.16 1.36 3.16 2.84 Equatorial Guinea 20.59 6.33 4.18 6.33 5.69 Ethiopia 0.00 14.55 5.39 4.56 4.56 4.56 4.85 Gabon 0.63 10.24 1.13 2.25 2.25 2.25 1.02 Gambia, The 6.58 2.35 1.90 2.35 2.12 Ghana 0.75 6.65 1.94 1.72 1.94 1.75 Guinea 1.89 5.38 1.14 2.63 2.63 2.63 1.03 Guinea-Bissau 2.05 4.03 0.63 3.23 3.23 3.23 0.57 Kenya 0.68 4.84 3.77 3.18 3.18 3.18 3.39 Lesotho 0.35 12.63 5.09 4.73 4.73 4.73 4.58 Liberia 6.70 13.48 0.46 8.40 8.40 8.40 0.41 Madagascar 0.73 3.46 0.84 1.29 0.84 0.76 Malawi 1.39 5.59 0.71 2.21 0.71 0.64 Mali 0.86 6.39 3.10 2.02 2.02 2.02 2.79 Mauritania 6.72 9.05 1.93 1.93 1.93 8.13 Mauritius 0.67 6.02 3.89 1.68 1.68 1.68 3.50 Mozambique 1.08 13.06 7.32 3.81 3.81 3.81 6.58 Namibia 9.94 5.83 1.35 1.35 1.35 5.24 Niger 1.21 5.87 5.50 2.57 2.57 2.57 4.94 Nigeria 0.55 2.79 0.78 1.67 0.78 0.70 37 Rwanda 1.04 9.61 3.50 3.04 3.50 3.14 Sao Tome & Principe 25.46 1.17 6.95 6.95 6.95 1.06 Senegal 1.00 5.93 6.02 2.94 2.94 2.94 5.41 Sierra Leone 1.33 7.07 3.35 1.45 1.45 1.45 3.01 South Africa 0.29 3.34 5.55 0.87 3.34 4.99 Sudan 0.58 3.64 0.00 0.68 0.00 0.00 Swaziland 0.09 8.78 2.09 1.96 2.09 1.88 Tanzania 1.22 6.63 6.36 2.47 2.47 2.47 5.72 Togo 2.03 6.72 2.19 4.07 4.07 4.07 1.97 Uganda 2.58 8.27 1.73 3.99 3.99 3.99 1.55 Zambia 0.63 4.26 0.95 1.18 0.95 0.86 Zimbabwe 1.46 4.40 4.58 2.59 4.40 4.12 Cambodia 5.97 1.89 1.20 1.89 1.70 China 0.03 16.00 16.11 6.31 6.31 6.31 Fiji 0.40 6.16 4.34 3.29 3.29 3.29 3.91 Indonesia 0.27 3.12 15.34 2.17 3.12 13.79 Kiribati 0.35 0.35 0.35 Lao PDR 9.52 17.69 6.43 11.86 6.43 5.78 Malaysia 0.25 9.78 4.24 1.77 4.24 3.81 Mongolia 0.00 9.92 6.24 2.57 2.57 2.57 5.61 Myanmar 0.00 11.21 5.31 1.63 1.63 1.63 4.78 Philippines 0.85 2.69 2.19 0.84 2.19 1.97 Solomon Islands 1.72 1.72 1.72 Thailand 0.22 5.22 3.54 0.84 3.54 3.19 Vietnam 0.67 7.43 9.00 2.62 7.43 8.09 Albania 0.64 5.82 14.79 3.49 3.49 3.49 13.29 Armenia 2.05 3.56 8.07 3.39 3.39 3.39 7.26 Azerbaijan 0.24 5.27 2.54 1.06 2.54 2.29 Belarus 0.23 1.47 4.36 0.72 1.47 3.92 Bosnia and 0.04 5.84 3.25 1.34 3.25 2.92 Herzegovina Bulgaria 0.80 3.80 4.77 2.73 2.73 2.73 4.29 Croatia 3.61 3.38 0.72 3.38 3.04 Georgia 0.70 3.89 2.06 1.44 2.06 1.85 Kazakhstan 0.17 2.98 6.53 1.01 2.98 5.87 Kosovo 0.53 11.72 6.26 6.26 6.26 Kyrgyz Republic 0.43 2.02 1.94 0.92 0.92 0.92 1.74 Macedonia 1.40 6.98 3.04 2.41 2.41 2.41 2.74 Moldova 0.44 2.31 5.54 2.51 2.51 2.51 4.98 Montenegro, Rep. of 0.38 3.19 6.94 2.04 3.19 6.24 38 Romania 0.50 7.13 6.35 1.96 6.35 5.70 Russia 0.34 2.69 5.74 1.92 2.69 5.16 Serbia 0.21 2.92 5.28 1.13 2.92 4.74 Tajikistan 1.36 6.79 5.99 1.38 1.38 1.38 5.38 Turkey 0.66 4.00 2.44 0.68 2.44 2.20 Ukraine 0.65 1.96 4.93 1.02 1.02 1.02 4.43 Uzbekistan 0.53 0.84 0.25 0.84 Argentina 0.29 3.22 0.53 0.53 0.53 Belize 3.10 1.38 1.08 1.38 1.24 Bolivia 8.83 3.44 1.78 3.44 3.09 Brazil 0.68 3.06 3.15 1.62 1.62 1.62 2.84 Colombia 0.51 6.93 7.61 2.00 6.93 6.84 Costa Rica 0.17 2.96 3.07 0.39 0.39 0.39 2.76 Dominica 20.96 8.77 5.37 8.77 7.89 Dominican Republic 0.12 2.71 1.74 0.82 0.82 0.82 1.56 Ecuador 0.37 10.15 2.43 1.92 2.43 2.19 El Salvador 0.28 2.20 2.36 1.01 1.01 1.01 2.12 Grenada 7.48 1.18 1.90 1.18 1.06 Guatemala 0.42 2.71 2.70 1.97 1.97 1.97 2.43 Haiti 0.50 12.12 0.00 0.87 0.87 0.87 0.00 Honduras 1.27 4.19 1.78 1.88 1.78 1.60 Jamaica 0.24 0.24 2.44 0.51 0.24 2.19 Mexico 0.34 4.99 5.24 3.35 3.35 3.35 4.71 Nicaragua 1.19 4.52 3.13 2.20 3.13 2.81 Panama 1.05 6.99 6.61 2.77 6.61 5.94 Paraguay 0.19 3.76 0.33 1.85 1.85 1.85 0.30 Peru 0.78 4.54 4.04 2.93 2.93 2.93 3.63 St. Kitts and Nevis 5.88 1.71 1.72 1.71 1.54 St. Lucia 9.11 3.23 0.98 0.98 0.98 2.90 St. Vincent and the 8.70 1.45 2.17 1.45 1.31 Grenadines Suriname 6.39 1.88 1.48 1.88 1.69 Venezuela 10.64 3.64 2.99 3.64 3.27 Algeria 0.27 9.85 7.96 2.44 7.96 7.16 Djibouti 17.72 1.64 3.56 1.64 1.48 Egypt 0.46 2.00 3.83 0.65 2.00 3.45 Iran 0.04 11.91 5.13 1.94 5.13 4.61 Iraq 0.38 6.99 6.12 1.80 6.12 5.50 Jordan 1.29 4.83 0.81 1.76 0.81 0.73 Lebanon 1.65 0.00 1.65 39 Morocco 0.35 5.30 6.96 1.74 5.30 6.26 Tunisia 0.47 7.19 4.64 1.34 1.34 1.34 4.17 Yemen 0.33 4.03 2.53 1.06 2.53 2.27 Afghanistan 0.18 2.54 2.54 2.54 Bangladesh 0.59 5.65 2.24 1.14 1.14 1.14 2.01 Bhutan 12.18 13.65 7.94 7.94 7.94 12.27 India 1.90 9.10 5.08 4.19 5.08 4.57 Maldives 0.00 9.64 18.50 4.68 9.64 16.63 Nepal 0.20 3.64 3.77 1.19 3.64 3.39 Pakistan 0.89 3.04 2.61 1.38 1.38 1.38 2.34 Sri Lanka 0.46 4.89 7.41 1.93 4.89 6.66 2. Comparison of infrastructure spending in 2011 between EIB data and three methodologies EIB data This paper Gov’t Private GFCF_CE + GFCF_CE + GFCF_CE + Other GFCF_GG BOOST country GFCF_CE Other Building Other Building Building + PPI + PPI (A) (B) (C=A+B) Austria 0.646 0.781 1.426 1.016 3.137 Bulgaria 1.207 1.188 2.395 4.775 3.650 2.588 Cyprus 0.546 1.534 2.080 3.637 5.108 Czech Republic 1.396 1.121 2.517 3.854 4.071 Denmark 0.401 0.849 1.250 1.876 4.267 Estonia 1.098 3.092 4.190 4.653 5.200 Finland 0.624 0.797 1.420 1.932 4.191 France 0.816 0.096 0.912 2.307 4.121 Germany 0.414 0.501 0.915 1.508 2.164 Greece 1.013 0.024 1.037 0.992 2.330 Hungary 0.940 0.782 1.723 2.860 2.858 Ireland 0.705 0.518 1.223 1.581 3.302 Italy 0.533 1.091 1.624 1.924 3.539 Latvia 1.266 1.497 2.763 3.587 4.340 Luxembourg 1.080 0.896 1.976 2.074 8.898 Malta 0.566 0.206 0.772 1.615 0.896 Netherlands 1.247 0.275 1.522 2.504 4.150 Portugal 0.952 1.422 2.375 4.440 3.972 Slovakia 1.426 0.899 2.325 2.395 3.316 Slovenia 0.770 0.515 1.285 2.168 4.130 Spain 1.198 0.920 2.118 3.162 4.190 Sweden 0.790 1.116 1.905 1.904 4.127 40 Annex 2: Addressing data quality through BOOST Sample description The BOOST program is a World Bank led-effort launched in 2010 to provide quality access to budget data. The initiative strives to make well-classified and highly disaggregated budget data available for policy makers and practitioners within government, researchers, and civil society and promote their effective use for improved budgetary decision making, analysis, transparency and accountability. The program has designed and delivered over 70 national and subnational BOOST datasets in standardized formats, whose contents are country specific. Using the Government’s own data from public expenditure accounts held in Governments’ Financial Management Information Systems (FMIS), and benefiting from a consistent methodology, the program transforms highly granular fiscal data into accessible and readily available formats to facilitate expenditure analysis. Each dataset typically allows for approved, revised and executed budgets to be cross-referenced across years with categories such as: • Government levels (e.g., central or local) • Administrative units (e.g., ministries, departments, agencies, schools, hospitals, etc.) • Subnational authorities (e.g., districts, municipalities, other local government units) • Economic classification categories (e.g., staff salaries, procurement of goods, etc.) • Sources of funding (e.g., budget funds, off-budget funds, external financial, etc.) • Budget programs (if the country uses a program-based budgeting system) Efforts were made to ensure consistency between amounts computed through BOOST and those computed separately by partner organizations (e.g., the IDB for its INFRALATAM database). Consistent with emerging international standards, the team employed a two-pronged approach to quantify capital infrastructure from BOOST databases. First, using the economic classification, only pure gross capital formation was used to identify true capital spending. In other words, recurrent expenditures classified under capital budgets were removed from calculations when possible. Secondly, the team used functional classification to identify each individual infrastructure subsector: rail, road, water and air transport, energy, water and sanitation and telecoms. This was not always a straightforward process, as discussed below. Inevitably these figures were not always consistent with the ones computed by the other multilaterals. Therefore, several measures were taken to ensure consistency of these efforts while maximizing long term sustainability. These included working sessions with staff members of IADB to identify source of discrepancy between country figures and achieve consensus over optimal framework moving forward using BOOST as the main instrument. For instance, in Paraguay discrepancies were connected to different treatment of inter-governmental transfers as well as under-estimation of road investments which were 41 not properly identified by the functional classification “Road transport� since most projects were recorded as “Public works�. After these adjustments, estimates between the two organizations converged. 19 Further interactions with both IADB and ADB staff would be welcomed to further refine estimates and achieve greater consistency/convergence across partner’s estimates. Several data challenges undermining the potential for better capturing public capital spending in infrastructure were encountered during the BOOST exercise. On one hand, a large number of countries had minimal identification of functional purpose, making it particularly hard to accurately identify spending across the various sectors, often relying on administrative classification or availability of project information. This was the case in countries such as Sierra Leone, Burundi, Togo and Niger among others. In fact, in Sub-Saharan Africa region, only 7 countries presented accurate and sufficiently granular project level information in their Treasury systems. On the other hand, a significant number of countries – particularly low income – did not present sufficient disaggregation of economic categories that would allow to neatly identify pure gross capital formation from operation expenditures. This problem was particularly acute in cases of foreign funded projects presenting no economic disaggregation, therefore likely over-estimating the amount of true capital spending built into such projects. Coverage of source of funding was also inconsistent, preventing a thorough analysis of foreign vs. domestic funding dynamics across sub-sectors. Box 2 presents an overview of data challenges encountered in Sub-Saharan Africa. Quality issues – such as proliferation of special and manual procedures in francophone countries - also presented challenges in terms of collecting executed amounts through Treasury systems, further complicating efforts to compile reliable fiscal statistics. The work carried out in Togo, Burundi and Niger pointed to the proliferation of special procedures as an important area to focus on during data quality vetting processes. If they are not regularized, these special procedures are not captured in the FMIS, therefore underestimating the true amount of expenditures in a given year. While these issues were resolved through technical assistance activities workshops which supported ministry of Finance officials in regularization of such procedures so that payments would regularly be recorded on Treasury systems (as in the case of Niger) they still present challenges in countries like Togo and Burundi among others. Foreign funded investments in infrastructure play a big role, particularly in water and sanitation and electricity. However, in most countries, a substantial amount of foreign aid remains off-budget particularly as it relates to disbursements and execution phases. In countries such as Togo, Guinea, Guinea Bissau, Sierra Leone and Liberia, budget data is recorded in the system but there is no trace of execution amounts. In others such as Uganda, information on execution of certain foreign funded projects is either absent or incomplete. This affects not only measurements of total spending in a sector (under-estimated) but also affect potential deviation analysis (difference between originally planned and executed) leading to spurious upward bias. While innovative approaches are being pioneered to integrate execution data into national systems (box 2), the magnitude of the amount of foreign aid currently not being captured on budget calls for caution in interpreting overall amounts of spending. 19In other countries typical difference aroused from using different expenditure parameters (Mexico), institutional coverage (Guatemala) and functional tagging (i.e. missing irrigation in water and sanitation in Mexico). 42 Similarly, capital spending in infrastructure by State owned enterprises (SOEs) is complex to capture. This spending might be significant particularly in energy and water/sanitation sectors. In particular, two challenges need to be addressed. On one hand, little information is available about spending by SOEs given the opacity of financial reporting of these entities. Secondly, without proper identification of funding source, it might be difficult to disentangle spending by SOE financed by central government (which is typically captured in BOOST) from spending financed through own revenues and/or commercial financing. Finally, only a subset of countries20 presented significant geographical disaggregation of capital spending, undermining potential equity analysis of public investments. While this did not affect the primary objective of this study to derive a global baseline of infrastructure capital spending, it does however limit the informational and analytical potential. Given the well-known existence of wide within-country variations in terms of economic need and performance, it is critical for public spending data to provide geographical disaggregation to allow for systematic assessment of equitable targeting of public investments in infrastructure and analyze more systematically execution patterns across national and subnational spheres. More generally, fragmentation of financial management systems and duality of roles between Planning and Finance Ministries often prevents accessing more integrated views of project data across its full cycle. In best case scenarios, Treasury systems capture annual budget, allocation and expenditure amounts dissected by basic economic categories across all government levels. Detailed project information such as appraisal information (when available), expected cost, timeframe and outputs among others however is scattered amongst various departments and ministries and often in formats inconsistent with treasury data. This inhibits systematic analysis of project performance and stronger evidence-based assessment of efficiency in Public Investment management. Significant efforts will be needed to be devoted to ensure greater integration and inter-operability of information systems across developing countries. 20Albania, Ukraine, Moldova, Poland, Romania, Kyrgyzstan, Bulgaria, Peru, Mexico, Guatemala, El Salvador, Brazil, Kenya, Tanzania, Angola, Mali, Mozambique, Uganda and Ethiopia 43 Annex Box 2.1 Data challenges in tagging infrastructure spending in Sub-Saharan Africa Data quality issues made the task of properly tagging capital expenditure in infrastructure particularly hard, while at the same time constraining the range and quality of analytics. These include the following: • With few exceptions (Myanmar, Costa Rica) only general government expenditures were included along with capital transfers to SOEs – potentially underestimating capital spending in key infrastructure sub-sectors. • Only information captured in Treasury systems was included in the analysis – foreign funded spending in infrastructure which is kept off budget is therefore not included in this analysis and might affect totals particularly for high aid countries. To enhance integration of these funds on budget, the BOOST program is rolling out a new approach that would enable to complement existing national aid management systems with other sources of donor data – particularly IATI - to be able to better align off budget donor spending to the national. A current pilot is currently under way in Haiti. • In some francophone countries, budget classifications are not fully integrated with the Chart of account making it difficult to ensure a seamless tracking of transactions covering a typical ELOP expenditure chain. Countries like Mali use bridge tables to link accounting data with budgetary operations effectively retrieving execution data from their systems fully consistent with budget nomenclature. Others however like Togo, Mauritania and Niger do not have such tables and as a result only present data at committed and payment order stage but not actual payments. • Existing data for Equatorial Guinea and Zimbabwe provided only minimal information at administrative and economic level which did not allow to properly identify infrastructure spending and were therefore excluded from the analysis. These will be featured once the full BOOST is completed.Guinea Bissau treasury systems do not capture execution amount of capital expenditures – as such only approved amounts were used for the analysis. • Several francophone African countries merge accumulated spending by multiple ministries into one administrative unit (“depense communes�) which do not allow for proper functional identification. This wa s the case in Guinea and Mali for instance leading to potential under-estimation of spending. • In several countries such as Togo, Benin and Guinea, foreign funded capital expenditure was only available at budget level – for these countries, deviation analysis was only conducted for domestic funded levels. In several francophone countries (Haiti, Tunisia), development expenditures were presented on aggregate form by project – this prevented the team from isolating recurrent shares of reported capital amounts 44 Annex 3. Improving Infrastructure Investment Estimates Using Disaggregated Data (Kindly contributed by Kaushal Joshi, Asian Development Bank) When gross fixed capital formation (GFCF) is broken down by asset type and industrial sector of the investor, infrastructure investment can be estimated more accurately (Box table 3.2.1). Specifically, asset classes for each infrastructure sector are identified as follows: • Transport—Civil Engineering Works, Buildings other than Dwellings, and Information and communication technology (ICT) and rail-related machinery and equipment; • Energy—Civil Engineering Works, Buildings other than Dwellings, and machinery and equipment not transport related; • Water supply and irrigation—Civil Engineering Works, Buildings other than Dwellings, and Machinery and equipment not transport related; • Telecommunications—Civil Engineering Works, Buildings other than Dwellings, and Machinery and equipment not transport related. Adding up asset-sector-specific infrastructure investment gives the total infrastructure investment. The example of Fiji has two implications: i. The majority of infrastructure investment—approximately 80%—went to civil engineering works. This lends support to using GFCF on construction excluding buildings (GFCF-CE) to approximate infrastructure investment when there is no better alternative. ii. A non-trivial amount of infrastructure investment is not captured by GFCF-CE. This is mainly on machinery and equipment (used mainly in telecommunications, energy, and water infrastructure)—accounting for about 20% of Fiji’s infrastructure investment (with about a half in ICT equipment). Non-residential buildings are also missing from GFCF-CE, but the amount is small. 45 One problem with this approach is that not all road investment is classified as transport. For example, Public Works (or Public Administration) may also contain information on road investment. The practice seems to vary by country. Nevertheless, the measure described above offers a conservative, or lower- bound, estimate. One way to address this issue is to include civil engineering works of all sectors, while keeping the investment in machinery/equipment and non-residential buildings in energy, water, and telecommunications unchanged. This creates an upper bound (higher estimate) as some non- infrastructure components in civil engineering would also be included (mines and industrial plants, mining construction, other construction for manufacturing, outdoor sport and recreation facilities, and other civil engineering works such as satellite launching sites and defense). Comparing alternative estimates for infrastructure investment in India, Pakistan, and Fiji show interesting results (Box table 3.2.2). First, combining information from the alternative estimates may provide a more refined measure of infrastructure investment. For example, in Pakistan the BUDGET-PPI, a conservative estimate, and the upper-bound of the GFCF Breakdown approach are very close, suggesting that the actual infrastructure investment is near 2.1%. Second, even using detailed GFCF data, the constructed lower and upper bounds may still show a fairly large gap, especially in India. For this to narrow, statistics on road investment in all relevant sectors, such as Public Works, need to be available. Box table 3.2.2: Comparison of Alternative Estimates of Infrastructure Investment (% of GDP) Fiji, 2011 India, 2013 Pakistan, 2011 Total Infrastructure Investment Ideal Measure: GFCF Breakdown [5.58, 6.46] [4.03, 8.39] [1.23, 2.15] Measure 1: Budget + PPI 3.78 5.50 2.14 Measure 2: GFCF-GG + PPI 5.96 7.78 3.29 Measure 3: GFCF-CE 5.48 5.79 2.21 GDP = gross domestic product; GFCF = gross fixed capital formation; GFCF-CE = gross fixed capital formation in construction excluding buildings; GFCF-GG = general government GFCF; PPI = Private Participation in Infrastructure Database. Source: ADB staff estimate; Country sources. 46