WPS6186 Policy Research Working Paper 6186 MTEFs and Fiscal Performance Panel Data Evidence Francesco Grigoli Zachary Mills Marijn Verhoeven Razvan Vlaicu The World Bank Poverty Reduction and Economic Management Network Public Sector & Governance Unit September 2012 Policy Research Working Paper 6186 Abstract In the last two decades more than 120 countries have country variation in MTEF adoption in a dynamic panel adopted a version of a Medium-Term Expenditure framework to estimate their impacts. The analysis finds Framework (MTEF). These are budget institutions that MTEFs strongly improve fiscal discipline, with more whose rationale it is to enable the central government advanced MTEF phases having a larger impact. Higher- to make credible multi-year fiscal commitments. This phase MTEFs also improve allocative efficiency. Only paper analyzes a newly-collected dataset of worldwide top-phase MTEFs have a significantly positive effect on MTEF adoptions during 1990–2008. It exploits within- technical efficiency. This paper is a product of the Public Sector & Governance Unit, Poverty Reduction and Economic Management Network. 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 zmills@worldbank.org and mverhoeven@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 MTEFs and Fiscal Performance: Panel Data Evidence* Francesco Grigoli† Zachary Mills‡ Marijn Verhoeven§ Razvan Vlaicu¶ Abstract In the last two decades more than 120 countries have adopted a version of a Medium- Term Expenditure Framework (MTEF). These are budget institutions whose rationale it is to enable the central government to make credible multi-year fiscal commitments. This paper analyzes a newly-collected dataset of worldwide MTEF adoptions during 1990- 2008. It exploits within-country variation in MTEF adoption in a dynamic panel framework to estimate their impacts. The analysis finds that MTEFs strongly improve fiscal discipline, with more advanced MTEF phases having a larger impact. Higher-phase MTEFs also improve allocative efficiency. Only top-phase MTEFs have a significantly positive effect on technical efficiency. JEL Classification: E62, H51, H62. Keywords: budget institutions, MTEF, fiscal discipline, allocative efficiency, technical efficiency, dynamic panel data analysis. * We would like to thank Abdul Abiad, Nataliya Biletska, Jim Brumby, Stefania Fabrizio, Richard Hemming, Malcolm Holmes, Sam Ouliaris, Philip Keefer, Steve Knack, Ajay Tandon, and participants at the PREM Knowledge & Learning Forum and the World Bank MTEF Workshop. Haishan Yuan provided excellent research assistance. The views expressed in this paper are those of the authors and should not be attributed to the World Bank or its affiliated institutions, the IMF, its Executive Board, or its management. † IMF, Fiscal Affairs Department, 1900 Pennsylvania Ave. NW, Washington, DC 20431, USA. E-mail: fgrigoli@imf.org. ‡ The World Bank, Europe and Central Asia Department, 1818 H Street NW, Washington, DC 20043, USA. E-mail: zmills@worldbank.org. § The World Bank, Poverty Reduction and Economic Management Vice Presidency, 1818 H Street NW, Washington, DC 20043, USA. E-mail: mverhoeven@worldbank.org. ¶ University of Maryland, Department of Economics, and Department of Government and Politics, 3105 Tydings Hall, College Park, MD 20742, USA. E-mail: vlaicu@econ.umd.edu. 1 Introduction s long-run growth It is generally accepted that …scal performance is a key factor in a country’ prospects.1 Understanding the determinants of …scal performance has thus become a central topic of research. A recent strand of literature has emphasized the role of budget institutions in a¤ecting …scal performance. Budget institutions are the formal rules and procedures according to which budgets are drafted, approved, and implemented. They can take the s conduct of …scal form of either (i) laws establishing ex ante constraints on the government’ policy, such as balanced budget and debt ceiling provisions, or (ii) bargaining rules between the executive and the legislature, such as line-item executive veto or legislative amendment rules. In the last two decades more than 120 countries have adopted laws instituting multiyear …scal targets, known as Medium-Term Expenditure Frameworks (MTEFs ). First introduced in a small set of developed countries to contain expenditure overruns in the 1980s, MTEFs spread rapidly during the 1990s and 2000s, being in operation in 132 countries at the end of 2008 (see Figure 1 and Table 2). MTEFs translate macro-…scal objectives and constraints into broad budget aggregates as well as detailed expenditure plans by sector. The rationale of this budget institution is to enable the central government to more adequately incorporate future …scal challenges into the annual budgets, thereby reducing an undue emphasis on short-term goals. The key public …nance problem that MTEFs are intended to overcome is dynamic …scal ine¢ ciency.2 Whether it takes the form of strategic obstruction of future political opponents (Alesina and Tabellini 1990), or electoral manipulation through budget cycles (Drazen 2000, Brender and Drazen 2005), dynamic common pool (Velasco 1999), or time-inconsistent voters (Bisin, Lizerri, and Yariv 2011), government spending and borrowing deviates from the social planner level, resulting in suboptimally high de…cits and debt. Dynamic ine¢ ciency seems particularly inherent in a yearly cycle of budget planning and implementation. Wildavsky (1986, p. 317) makes this point as follows: “Multiyear budgeting has long been proposed as a reform to enhance rational choice by viewing resource allocation in a long-term perspective. One year, it has s expendi- been argued, leads to short-sightedness, because only the next year’ 1 See, e.g., Fischer (1993), Easterly and Rebelo (1993), Easterly, Irwin, and Serven (2008). 2 Other public …nance distortions include static common pool problems (Weingast, Shepsle and Johnsen 1982), rent seeking (Persson, Roland and Tabellini 2000, Besley and Smart 2007), and clientelism (Keefer and Vlaicu 2008). 2 tures are reviewed; overspending, because huge disbursements in future years are hidden; conservatism, because incremental changes do not open up large future vistas; and parochialism, because programs tend to be viewed in isolation rather than in comparison with their future costs in relation to expected revenue.� At the basis of an MTEF is a commitment by the budget actors to a medium-term, typically two to four years, …scal trajectory. Thus, it can be seen either as an ex ante constraint on the government, similar to a balanced budget requirement (Alesina and Perotti 1996), or as a "contract approach" to centralizing the budget process through a broad-based political agreement, as opposed to a "delegation approach" where the executive receives enhanced powers (von Hagen and Harden 1995). While the theoretical underpinnings of this institution are well understood, the empirical evidence on its impact is scarce. The main obstacle has been the shortage of data on MTEF adoption. An additional impediment has been the lack of sources of exogenous variation in national budget institutions (Acemoglu 2005).3 This paper reports on the …rst large-sample empirical study of the MTEFs’impacts on …scal performance. As part of a larger World Bank (2012) study, we collect and analyze MTEF adoption data for a panel of 181 countries over the period 1990-2008, the most comprehensive dataset to date on worldwide MTEF adoption. Following World Bank (2012) we classify MTEFs into three phases, based on their level of sectoral disaggregation, and generate testable hypotheses about the e¤ects of each phase on multiple dimensions of …scal performance. The rich time variation in the data allows us to model the dynamics of the …scal adjust- ment process, as well as address the endogeneity of MTEF to …scal performance. We use a Di¤erence Generalized Method of Moments (D-GMM) approach to estimate dynamic panel data models of …scal performance. These models are designed for "small T large N " panels and, when correctly applied, generate valid internal instruments that hold the promise of overcoming identi…cation issues due to the absence of strong external instruments, a typical hurdle with country-level data.4 When an MTEF is implemented well we should observe (i) spending that is limited by resource availability (…scal discipline), (ii) budget allocations that re‡ect spending pri- 3 One way to circumvent econometric identi…cation issues has been to study similar institutions operating at sub-national levels of government. See Besley and Case (2003) for a review of the literature that has employed U.S. state-level data. 4 The D-GMM approach was …rst proposed by Holtz-Eakin, Newey, and Rosen (1988) and later developed by Arellano and Bond (1991). Recent re…nements include Windmeijer (2005) and Roodman (2009). 3 orities (allocative e¢ ciency), and (iii) public goods delivery that is cost e¤ective (techni- cal e¢ ciency). We classify MTEFs into three phases : Medium-Term Fiscal Framework (MTFF, which establishes the aggregate resource envelope), Medium-Term Budget Frame- work (MTBF, which focuses on the allocation of spending across sectors, programs and agencies) and Medium-Term Performance Framework (MTPF, which sets sectoral perfor- mance targets). These three MTEF phases are “nested�: an MTPF contains an MTBF, which in turn contains an MTFF.5 The new data reveal patterns in the timing of MTEF adoption across regions and levels of development. OECD countries were the …rst to adopt MTEFs, and by the early 1990s most countries in this group had an MTPF in place. The bulk of MTEF reforms in Sub- Saharan African countries took place in the 1990s, Latin American countries adopted MTEFs in the late 1990s, and Eastern Europe and the former soviet republics joined the trend in the 2000s. We exploit di¤erential patterns of MTEF adoption across regions to construct external instruments that complement the standard GMM-style internal instruments. The empirical results show that MTEFs strongly improve …scal discipline and that the e¤ect is larger the more advanced the MTEF phase. The e¤ect varies between 1 and 3 percentage points of central government balance as a percentage of GDP. We also …nd that MTBFs and MTPFs improve allocative e¢ ciency, measured by the volatility in per capita health spending in purchasing power parity dollars (PPP$), and that MTPFs contribute to technical e¢ ciency, measured as technical e¢ ciency scores from a stochastic frontier model of public health delivery. These results are robust to excluding highly autocratic and highly developed countries. Our results are more favorable about MTEF e¤ectiveness than the conclusions of prior work. Bevan and Palomba (2000), La Houerou and Talierco (2002), Holmes and Evans (2003), and Oyugi (2008), based on case studies of about a dozen African countries, conclude that the budget process has generally not improved after the adoption of an MTEF, while McNab, Martinez-Vasquez, and Boex (2000) and Oxford Policy Management (2000) raise questions of adequate implementation. However, Gleich (2003) and Ylaoutinen (2004) …nd that MTEFs in Central and Eastern Europe alleviated the de…cits and debts that emerged in the second half of the 1990s. Wescott (2008) and Filc and Scartascini (2010), using data from Central and Latin America, found mixed results and emphasized the importance of piloting MTEFs in areas where they are likely to deliver the largest payo¤.6 5 This taxonomy follows World Bank (2012) and borrows from Castro and Dorotinsky (2008) with the nesting concept as an added innovation. 6 Drawing on extensive operational experience with MTEF implementation in developing countries, 4 Alesina, Hausmann, Hommes, and Stein (1999) and Fabrizio and Mody (2006) include MTEFs in aggregate indexes of budget institutions using data from Latin America and Eastern Europe, respectively.7 Our results complement this empirical literature by provid- ing evidence of MTEF impacts from a worldwide sample. While previous studies rely on small samples and either cross-sectional or static panel models our empirical methodology takes advantage of the time variation to estimate dynamic panel models. We also propose a new classi…cation of MTEF phases, based on the level of disaggregation of the central s …scal objectives. government’ Our paper is also related to the broader empirical literature on budget institutions.8 The most studied institutions have been balanced budget amendments, debt ceilings, tax and expenditure limitations, and supermajority requirements for tax increases. Examples include: Bayoumi and Eichengreen (1995), von Hagen and Eichengreen (1996), Stein, Talvi, and Grisanti (1999), Hallerberg and von Hagen (1999), Perotti and Kontopoulos (2002), Fatas and Mihov (2003), von Hagen and Wol¤ (2006), and Hallerberg and Ylaoutinen (2010), all using country-level data; and Poterba (1994), Bohn and Inman (1996), Kiewert and Szakaly (1996), Poterba and Rueben (1999), Knight (2000), Knight and Levinson (2000), and Fatas and Mihov (2006) using state-level data. Important lessons from this literature are that numerical constraints have limited e¤ectiveness because they can be circumvented, that the e¤ect of reduced …scal discretion on macroeconomic volatility remains an open question, and that the political environment matters for the e¤ectiveness of budget institutions. We contribute to this literature by proposing a dynamic panel approach (di¤erence GMM) that models the …scal adjustment process while at the same time addressing the issue of insti- tutional endogeneity. Apart from employing standard internal instruments, we also propose external instruments that improve estimation e¢ ciency, based on the time-varying degree s geographic region. Also, in addition to macroeconomic of MTEF penetration in a country’ e¤ects, which have been the focus of this literature, we provide evidence of sectoral e¤ects. The structure of the paper is as follows. Section 2 provides a background discussion of the MTEFs, their adoption trends, and their expected e¤ects on …scal performance. Section 3 discusses the data and the empirical strategy. Section 4 presents the empirical results. Section 5 summarizes the paper and suggests directions for future research. Schiavo-Campo (2009) puts forward conceptual arguments supporting a gradual introduction of these insti- tutions and emphasizes the potential downsides of instant reform. 7 In these two papers the MTEF component of the index is weighted by 1/10 and 1/12, respectively. 8 See the NBER volume edited by Poterba and von Hagen (1999), as well as von Hagen (2006), for reviews of the budget institutions literature. 5 2 Background This section takes a …rst look at the worldwide MTEF adoption data collected for this paper by presenting stylized facts of MTEF global growth and regional penetration during 1990-2008. It also discusses the rationale behind MTEFs as budget institutions designed to improve …scal performance. This discussion helps generate theoretical expectations. We subject these conjectures to empirical scrutiny in the next section. 2.1 MTEF Phases We classify MTEFs into three phases, based on the following criteria.9 Medium-Term Fiscal Framework (MTFF): the government has rolling aggregate, ex- penditure, revenue, and other …scal forecasts. Features include the availability of a macro-…scal strategy, macroeconomic and …scal forecasts, and debt sustainability analysis. Medium-Term Budgetary Framework (MTBF): budget, spending agency or other re- ports explain aggregate and sectoral expenditure objectives and strategies, budget circulars detail medium-term expenditure ceilings and revenue forecasts, and budget documents contain some detail about medium-term estimates.10 Medium-Term Performance Framework (MTPF): budget, spending agency or other reports explain program objectives and strategies, and list speci…c agency and/or pro- gram output or outcome targets, as well as results. These three phases are "nested" in the sense that a higher-phase MTEF contains the lower- phase MTEF just below it. 2.2 Stylized Facts Although some forms of medium-term expenditure projections existed in OECD countries as early as the 1960s, the …rst application of a coherent system of multiyear budgeting occurred in Australia, where an MTEF was introduced in the 1980s (see Folscher 2007). MTEFs 9 We use these de…nitions to code each country-year in our sample as falling into each of these mutually exclusive categories; see Section 3 and the Data Appendix for a description of variable construction. 10 We coded countries that introduced a "pilot" MTBF in a few sectors as MTFF since the health sector, our focus in the analysis below, might not be one of the piloted sectors. 6 have since been adopted by a large number of low and middle-income countries as a central element of public …nancial management reform. While MTEFs began to spread across industrial countries and Africa in the early 1990s, it was not until the late 1990s and 2000s that they took o¤ in emerging market economies; see Figure 1. An average of 10 countries per year introduced an MTEF between 1996 and 2008. By the end of 2008, 132 countries, or about two-thirds of the globe, had an MTEF. Figure 2 shows temporal patterns of MTEF growth by continent. The regional trends are evident, with Europe leading the pack, followed by Africa and the Americas. Initially, most MTEFs were of the …rst phase, or MTFF, and until recently about two- thirds of the increase in MTEFs has been in the form of new MTFFs. However, there has been a recent uptick in the number of MTBFs and MTPFs. In 2008 there were 71 MTFFs, 42 MTBFs, and 19 MTPFs. Table 2 shows that the shift to MTBFs and MTPFs has been mainly through transitions from a lower MTEF phase to a higher one.11 MTEF coverage varies signi…cantly. Advanced economies had achieved almost complete coverage (96%) by the end of our sample period. MTEF adoption in advanced countries occurred in two waves. In the late 1980s and early 1990s only a few advanced economies s lead in MTEF adoption. Then, in the late 1990s MTEFs were introduced followed Australia’ in the European Union to support budgetary targets set as pre-conditions for monetary union. By the end of 2008, 46% of the MTEFs in advanced economies were MTPFs. The relatively low fraction of the second-phase MTEF (the MTBFs) in these countries suggests that when advanced economies decide to move beyond an MTFF introducing a performance focus is a natural development, re‡ecting their more sophisticated budgeting systems. MTEFs have also achieved broad coverage in Europe and Central Asia. The spread was more rapid and consistent in Central and Eastern Europe than in the Former Soviet Union. This may re‡ect e¤orts made in Central and Eastern Europe for quick integration with Western Europe.12 Building on an early start in Botswana and Uganda, MTEFs spread rapidly across Sub-Saharan Africa in the 2000s. MTEFs are more numerous in Francophone Africa than Anglophone Africa. MTEFs have also been adopted by most countries in South Asia, with Nepal and Sri Lanka having implemented MTBFs. MTEFs are less widespread in other regions, despite a recent spurt of adoptions in East Asia and the Paci…c, including MTBFs in Cambodia and Thailand. The picture is similar 11 Three countries (Bulgaria, Canada and Norway) performed a full transistion - from an MTFF to an MTBF to an MTPF - during this period. 12 The countries remaining without an MTEF in this region are: Azerbaijan, Belarus, Montenegro, and Turkmenistan. 7 in Latin America and the Caribbean, where a number of countries have introduced MTFFs following years of managing …scal policy under IMF programs. Only four countries have moved beyond an MTFF and introduced an MTBF: Argentina, Colombia, Nicaragua and s budgeting system has recognizable MTBF characteristics. In St. Lucia, although Brazil’ the Middle East and North Africa MTEFs are a very recent innovation. Only Algeria and Jordan have an MTBF, while major oil exporting countries such as Saudi Arabia and United Arab Emirates, as well as Egypt, had not yet adopted MTEFs. Despite pronounced di¤erences between regions patterns of MTEF adoption have been relatively uniform across income and development levels. Apart from the widespread adop- tion of MTEFs in high-income countries, there is little di¤erence in penetration across upper middle, lower middle, and lower-income countries. MTEF adoption does not appear to fol- low a monotonic relationship with respect to income per capita or the human development index; see Figures 3 and 4. 2.3 MTEF Objectives MTEFs represent a multiyear approach to budgeting that addresses the shortcomings of annual budgeting noted above in the Introduction. Most public programs require funding and yield bene…ts over a number of years, but annual budgeting largely ignores future costs and s budget and modify bene…ts. Annual budgets take as their starting point the previous year’ it in an incremental manner, making it di¢ cult to re-prioritize policies and spending.13 MTEFs take a strategic forward-looking approach to establishing spending priorities and resource allocation. They also look across sectors, programs and projects to see how spending can be restructured to best serve national objectives, which contrasts with the narrow self- interest of spending agencies and bene…ciaries that dominates resource allocation under annual budgeting (World Bank 1998). Insofar as an MTEF constrains spending to resource availability, makes budget alloca- tions re‡ect spending priorities, and generates cost e¤ectiveness in the delivery of public goods and services, it should contribute directly to …scal discipline, allocative e¢ ciency, and technical e¢ ciency.14 Moreover, there are synergies among these three dimensions of …scal 13 While incremental budgeting can work well in times of revenue growth, it comes under particular pressure when revenue falls, becomes more volatile, or reaches its natural limit. In these instances expenditure prioritization takes on increased importance. 14 There is also a link to broader economic development. With improved …scal outcomes, growth should be higher, in‡ ation lower, and macroeconomic volatility reduced. Moreover, as the quality of spending improves, higher incomes should be accompanied by lower poverty rates, while better infrastructure should contribute to even higher growth and further poverty reduction. 8 performance. With …scal discipline secured, governments should be free to focus on the microeconomic challenges of improving spending e¢ ciency and not preoccupied with hav- ing to address the adverse macroeconomic consequences of persistent …scal imbalances.15 It should also be easier to maintain …scal discipline when improvements to both allocative and technical e¢ ciency reduce abuse and waste. Moreover, against a background of …scal disci- pline, new expenditure needs are more likely to prompt spending reallocations as opposed to requests for additional funding. Finally, both …scal discipline and expenditure e¢ ciency create …scal space for productive spending on economic and social infrastructure, and for responding to …scal risks. MTFFs can promote …scal discipline by addressing the root causes of de…cit bias. By specifying an overall "top-down" resource constraint, an MTFF reins in the political ten- dency to over-commit resources (the common pool problem). By imparting a medium-term perspective to budgeting and taking into account the future …scal costs of government policies and programs, an MTFF can …ll information gaps that allow politicians to renege on com- mitments to implement a¤ordable policies (the time consistency problem). A medium-term perspective also encourages governments to conduct discretionary stabilization in a symmet- ric, counter-cyclical manner, rather than asymmetrically which leads to rising de…cits and debt (Kumar and Ter-Minassian 2007).16 Since MTBFs and MTPFs incorporate an MTFF, they should have a stronger e¤ect on …scal discipline compared to an MTFF alone. This is in part because countries that have the administrative capacity to implement these higher phases will likely also have greater …scal discipline. But it is also a consequence of better prioritization and more emphasis on performance, which can bring the payo¤ to …scal discipline into sharper focus. Prioritization guided by longer-term sector strategies should improve allocative e¢ ciency. Insofar as spending agencies prepare sector strategies, identify their resource needs, and allocate their budgets according to strategic priorities, this "bottom-up" prioritization should produce a shift to spending with higher economic and social returns. However, the full payo¤ to prioritization requires that choices are also made as to how resources should be 15 It can be argued that, in fact, large …scal imbalances prompt better expenditure prioritization; however, the lessons from …scal adjustments around the world is that spending cuts are borne disproportionately by high-priority spending, and especially public investment in infrastructure, with adverse consequences for future growth (Easterly, Irwin, and Serven 2008). Lewis and Verhoeven (2010) report that the growth of social spending has dipped as the global …nancial crisis has put …scal positions under pressure, which risks setting back achievement of human development goals, because these depend on the rapid spending increases achieved in the 1990s and the earlier part of the 2000s. 16 On the downside, if spending agencies view MTEFs as minimum entitlements, rather than constraints, ceilings, or forward estimates, MTFFs could actually be a source of …scal indiscipline and de…cit bias (Schick 2010). 9 allocated across sectors, which is done as part of the reconciliation between the "top-down" and "bottom-up" approaches involving a lead agency, normally the Ministry of Finance, and spending agencies, and in connection with which less strategic guidance may be available, especially in the absence of national medium-term planning.17 The outcome of e¤ective prioritization should be a change in the allocation of spending. In the short term, spending volatility by sector may increase following MTEF implementation as spending is reallocated to more productive sectors and programs. Thereafter, insofar as spending decisions are guided by strategic priorities with a longer-term focus, sectoral spending should become less volatile, especially in the high-priority areas of health and education. The payo¤ coming from an MTBF should be even higher with an MTPF since this last phase goes further by setting within-sector and within-program performance targets.18 A third dimension of …scal performance is technical e¢ ciency. The better the economic and social outcomes achieved by spending programs from a given amount of budget resources, or the fewer resources used to achieve given outcomes, the more technically e¢ cient is gov- ernment spending. Improved technical e¢ ciency may follow from an MTFF, but is more likely a consequence of an MTBF and MTPF, with the latter possibly having the largest e¤ect as budgets are linked to results in the form of outcomes or outputs. Based on these considerations we state the expected MTEF e¤ects on …scal performance in the following hypotheses (see also World Bank 2012): (H1) MTFF, MTBF, and MTPF improve …scal discipline, with higher-phase MTEFs having larger e¤ects. (H2) MTBF and MTPF improve allocative e¢ ciency, with MTPF having a larger e¤ect. (H3) MTPF improves technical e¢ ciency. The rest of the paper examines the evidence for these conjectures. 3 Data and Empirical Strategy This section discusses the choice of variables for the empirical analysis and takes a …rst look at the statistical properties of our data. It then outlines our empirical strategy for identifying and estimating the MTEFs’e¤ects on …scal performance. 17 Moreover, di¢ cult decisions have to be made to cut low-priority, but often politically sensitive, spending. 18 A shift away from unproductive spending should also be observed. Poor-quality investment, distor- tionary and untargeted subsidies, bloated civil services, and the like should not survive scrutiny under the MTEF, while productive spending on economic and social infrastructure, health and education services, and other growth- and development-promoting activities should be favored. 10 3.1 Data The dataset contains both cross-sectional and time-series variation in MTEF presence. The sample consists of 181 countries over the period 1990-2008. The country sample re‡ects data availablility on MTEF status. The period sample re‡ects data availability on public …nances. Here we brie‡y discuss the key variables. Section A.3 of the Appendix contains the complete list of variables together with their data sources. The construction of the MTEF indicators relied upon an extensive data collection e¤ort as no single type of document su¢ ciently describes the existing institutional arrangements for all countries or even individual countries. Thus, the data were compiled from a large number of sources, including IMF Article IV country reports, IMF Reports on the Obser- vance of Standards and Codes (ROSC), …scal transparency modules, World Bank Public Expenditure Reviews (PERs), World Bank Country Financial Accountability Assessments (CFAAs), OECD documents, donor case studies, and country websites. Additionally, World Bank and IMF public sector specialists supplemented the above information with technical details.19 We measure …scal discipline, an indicator of macro …scal performance, by the central s overall balance. Although the literature suggests alternative indicators, e.g., government’ primary balance and debt, data availability limited the choice to the overall balance. More- over, by including government borrowing, the overall balance is a good indicator of the state of public …nances.20 Allocative e¢ ciency does not have a universally accepted de…nition. Potential proxies for allocative e¢ ciency are budget composition volatility and volatility of core spending (health and education). Since volatility in these sectors jeopardizes long-term objectives, health care and public education spending should be largely una¤ected by short-term ‡uctuations in GDP. In other words, allocative e¢ ciency implies that spending in core sectors where needs are fairly constant does not behave in a volatile manner. Given data constraints and the requirement that the public good category should be reasonably comparable across countries, we choose to work with the volatility of per-capita health spending, in PPP$. We de…ne volatility of a time series yi;t for country i as the absolute yearly growth rate of the 19 For the purposes of this paper, we refrain from making judgments to distinguish between an MTEF present in the law (de jure) and a well-functioning MTEF (de facto). Such a distinction would introduce a signi…cant amount of subjectivity into the analysis. 20 It could be argued that the overall balance does not account for the e¤ect of in‡ ation on interest payments and that interest payments are a function of the accumulated debt and not the present …scal stance. 11 detrended series: y~i;t V olatilityi;t = log 100 (1) ~i;t 1 y 1 PT where y ~i;t = yi;t t T k=1 k is the detrended series for yi;t : Technical e¢ ciency is typically measured using technical e¢ ciency scores from a Stochas- tic Frontier Analysis (SFA). This is the approach we adopt. The SFA approach relies on a reduced-form relationship between inputs and outputs. The country with the highest health output after controlling for inputs is the most e¢ cient, and the e¢ ciency level of the other countries is measured with respect to the most e¢ cient country.21 In particular, we compute technical e¢ ciency scores in the health sector using a parsi- monious version of the model estimated in Greene (2005). The outcome of interest is life expectancy, and the input is health spending per capita. The model is: log(Lif e_Expi;t ) = 0 + 1 log(Health_Spendi;t ) + (2) + 2 Densityi;t + 3 OECDi;t + t + vi;t ui;t where s t’ are year …xed e¤ects, vi;t N (0; v ), and ui;t = jUi;t j Exp(0; u ): The controls are population density and OECD membership.22 The parameters are estimated by maximum likelihood. The estimates of vi;t ui;t are translated into an estimate of ui;t using the standard Jondrow, Materov, Lovell, and Schmidt (1982) formula. Technical e¢ ciency then is simply: u ^i;t T ech_Ef f iciencyi;t = e (3) ^i;t is the maximum likelihood estimate of ui;t : Table 4 presents the estimation results. where u The coe¢ cients follow the same pattern noticed in prior work using di¤erent data. The asymmetry parameter = u= v is also within the range of variation reported previously. Estimated mean e¢ ciency is 86.48. Following Baltagi, Demetriades, and Law (2009), who study the e¤ect of …nancial open- ness on banking sector development, we introduce MTEF regional penetration as external 21 The SFA was inspired by Farrell (1957), who de…ned technical e¢ ciency as the ability to produce the maximum possible output from a given set of inputs, and measured it as the di¤erence between maximum attainable output and observed output. Ine¢ ciencies might arise from waste or because the most cost- e¤ective set of programs is not implemented. 22 Greene (2005) also includes education spending per capita, as an input, and controls for government voice and accountability, government e¤ectiveness, share of government …nancing, the Gini coe¢ cient, and GDP per capita. 12 instruments to supplement the internal GMM-style instruments. In a region consisting of n countries, penetration for country i is de…ned as the fraction of countries in the surrounding region that already have an MTEF: Pn M T EFj;t j =1;j 6=i M T EF _Re gional_P enetrationi;t = (4) n 1 for each of the MTEF phases. We use the classi…cation of the world into the twenty-two geographic regions de…ned by the United Nations Statistics Division. Table 1 reports summary statistics. All variables display considerable variation both between and within countries, justifying the use of panel estimation techniques. An exception is MTPF, which has small within variation due to the few adoptions of this top phase, and OECD membership. Table 3 reports pairwise correlations between the main variables. The correlation coe¢ cients are within plausible ranges and support our choices of regressors in the next subsection. 3.2 Empirical Strategy Estimating the impact of a budget institution on …scal performance requires that we ad- dress several identi…cation challenges: reverse causality, omitted variable bias, and errors-in- variables. First, reverse causality arises because …scal stress, e.g., a …nancial crisis, may have prompted a country to restrain spending, adopt an MTEF, or strengthen an existing one. Von Hagen (2006, p. 474) notes that "Historical experience suggests that governments make e¤orts to centralize the budget process to overcome sharp …scal crises." If MTEFs have pos- itive e¤ects on …scal performance, and poor …scal performance increases the probability of adopting an MTEF, then the reverse causality bias is probably negative. In this case, the estimates are still useful as a lower bound for the actual e¤ect.23 Second, omitted variable bias arises due to the failure to account for a factor that a¤ects both the adoption of an MTEF and …scal performance. For instance, strong economic growth may reduce the pressure on a government to reform budget institutions, and, at the same s …scal outcomes, thus leading to negative omitted variable time, improve the government’ bias. As suggested by Fabrizio and Mody (2006), a partial solution to this problem is to 23 The endogeneity of budget institutions with respect to …scal performance is extensively discussed in the cross-country literature (see Alesina and Perotti 1999, Stein, Talvi, and Grisanti 1999, Perotti and Kontopoulos 2002, and Fabrizio and Mody 2006) yet none of these papers proposes an instrument that in‡uences the probability of …scal reform while being exogenous to …scal performance. 13 use within-country variation in …scal institutions. This approach, in e¤ect, eliminates the unobservable country speci…c …xed e¤ects that may in‡uence budget de…cits. The problem of omitted variables is thus alleviated; however it is not eliminated.24 ; 25 Finally, if some of the variables in the analysis are not measured accurately, there is the potential for errors-in-variables bias, which usually dampens the e¤ect of interest. Although in our empirical model the primary explanatory variable, MTEF status, can be observed with a reasonably high degree of precision, there is still scope for measurement error in the other explanatory variables. Our empirical strategy exploits within-country variation in …scal institutions. Annual data allow us to account for the possibility that observed …scal performance in a given year may not represent long-run equilibrium values, because of incomplete adjustment in other variables. For example, as revenues cannot be perfectly anticipated budget balance in a given year ‡uctuates around the equilibrium balance level. To allow for the possibility of partial adjustment, we use a dynamic speci…cation with a lagged dependent variable: X L yi;t = + l yi;t l + 1 M T F Fi;t + 2 M T BFi;t + 3 M T P Fi;t + xi;t + "i;t (5) l=1 where yi;t is a measure of …scal performance, L is the number of lags of the dependent variable, M T F Fi;t ; M T BFi;t ; M T P Fi;t are indicators of the three MTEF phases, xi;t is a vector of covariates, and "i;t is an error term that contains country and year …xed e¤ects: "i;t = 'i + t + i;t (6) with the idiosyncratic error i;t assumed to be mean zero. While this dynamic panel formulation allows for a richer model of …scal adjustment, the presence of a lagged dependent variable introduces new sources of endogeneity. First, without controlling for the …xed e¤ects, the model in equation (5) has a built-in positive bias in the …rst lag of the dependent variable.26 Second, the di¤erenced version of the equation eliminates this bias, but has a built-in negative bias due to the fact that yi;t = yi;t is negatively correlated with yi;t 1 = yi;t 1 : Thus an unbiased estimate of the lagged 24 Most studies have not been able to use this method because either budget institutions do not change much over time or because changes are di¢ cult to measure. When it has been implemented with U.S. state- level data (e.g., Knight and Levinson 2000) the results are typically di¤erent, indicating that the problem of omitted variables is relevant. 25 Additional omitted variables could include political institutions. Evidence from Europe shows that institutional design responds to political factors and events (Hallerberg, Rolf, and von Hagen 2009). 26 Nickell (1981) has shown that this "dynamic panel bias" disappears only when T approaches in…nity. 14 dependent variable coe¢ cient should lie in the range between the FE estimate and the OLS estimate. This bracketing range thus provides a natural speci…cation check (Bond 2002). The D-GMM solution to this endogeneity problem is to instrument for yi;t 1 using yi;t 2 and possibly earlier lags. Notice that the lagged values of the dependent variable are useful instruments unless yi;t is close to a random walk, in which case past levels convey little information about future changes. Table 5 presents unit root test results for the three measures of …scal performance. The IPS test statistic safely rejects the null of unit root in each case.27 In the same fashion, one can instrument M T F Fi;t , which is also endogenous to yi;t , with M T F Fi;t 2 and possibly earlier lags (and similarly for the higher MTEF phases). In this way D-GMM generates internal instruments for the budget institutions. To improve estimation e¢ ciency we supplement the internal instruments with external instruments based s geographic region, computed according to equation (4). on MTEF penetration in a country’ The external instruments are inspired by the regional penetration patterns noted in the previous section; see Subsection 2.2.28 q deviations transform Because our panel is unbalanced, we use the orthogonal (Arellano Ti;t 1 P and Bover 1995) in the baseline speci…cations: yi;t = Ti;t+1 yi;t Ti;t s>t yi;s , where Ti;t is the number of available future observations; we also report estimates based on the di¤erence transform yi;t = yi;t yi;t 1 in the alternative speci…cations. Orthogonal deviations, instead of subtracting the previous observation from the current observation, subtracts the average of all future available observations from the current observation. This maximizes sample size in panels with gaps. The moment conditions are based on orthogonality between the transformed errors and the lagged values of the dependent variable. To test for this condition one can run two diagnostics, namely …rst-order and second-order serial correlation in the idiosyncratic error i;t : The test should reject the null of no …rst-order serial correlation, and not reject the absence of second-order serial correlation. We include year …xed e¤ects in all speci…cations to strengthen the case for the assumption of no correlation in the idiosyncratic errors across countries. 27 An alternative approach would be to use the System GMM estimator proposed by Arellano and Bover (1995) and Blundell and Bond (1998); this estimator requires an additional identifying assumption, namely that …rst di¤erences of the instrumenting variables are exogenous to the …xed e¤ects. This is the strategy we employ in World Bank (2012). There we also include a larger set of covariates which should mitigate omitted variable bias, but may also increase the possibility of additional endogeneity bias. The consistency in …ndings between the two approaches adds to the robustenss to our results. 28 By themselves the external instruments are not strong enough to justify a di¤-in-di¤s IV strategy. The lack of strength is driven in particular by insu¢ cient variation in MTPF adoption. 15 The number of moment conditions increases with T: To test for over-identifying restric- tions we use a Hansen J test. Too many moment conditions introduce bias while increasing e¢ ciency. Thus, it is important to keep the number of internal instruments in check. In the baseline speci…cations we use one lag for each lagged dependent variable, and two lags for each MTEF indicator. In the alternative speci…cations we also report results with collapsed instrument matrix, as recommended by Roodman (2009).29 To restrain the number of instruments we only include covariates that satisfy two condi- tions: (i) have the potential to act as conduits in transmitting regional trends, and (ii) are not endogenous to the dependent variable, at least in the short run. The …rst condition is needed because our external instruments are regional variables. The second condition sim- ply insures that we do not have to introduce internal instruments for the covariates as well. Openness and con‡ict are one economic and one political variable that satisfy these criteria and are available for our full sample, thus xi;t = (Opennessi;t ; Conf licti;t ) in equation (5). Our baseline speci…cation computes two-step D-GMM estimates with standard errors corrected with the Windmeijer (2005) procedure.30 We also report alternative speci…cations with one-step D-GMM estimates, in which case we report cluster-robust standard errors, i.e., robust to heteroskedasticity and arbitrary patterns of correlation within countries. 4 Estimation Results The main results of the paper are contained in Tables 6 through 11. The tables contain estimates of …scal discipline (Tables 6 and 7), allocative e¢ ciency (Tables 8 and 9), and technical e¢ ciency (Tables 10 and 11) regressions using the dynamic panel model discussed in the previous section. In each case the …rst table presents the baseline results, namely two- step D-GMM with Windmeijer corrected standard errors. In these speci…cations we treat the MTEF indicators as endogenous. The second table presents alternative speci…cations, namely one-step robust standard errors, collapsed instruments, di¤erence transform, and treatment of the MTEF variables as lagged and predetermined instead of contemporaneously endogenous. We preserve the structure of both the baseline and the alternative speci…cations across all three measures of …scal performance. 29 In large samples collapsing the instruments reduces statistical e¢ ciency, however in small samples it may alleviate the bias created when the number of instruments times the number of periods approaches the number of panel units. 30 The two-step standard error correction is needed because the original formula for the variance produces standard errors that are severely downward biased when the number of instruments is large. 16 4.1 Fiscal Discipline Table 6 presents the baseline regressions for …scal discipline, measured as the central gov- s budget balance as a percent of GDP. Columns (1) and (2) start with OLS and ernment’ …xed e¤ects estimates, which determine the bracketing range for the lagged dependent vari- able coe¢ cient, namely 0.379 0.481. As expected, government balance follows a distinct adjustment process as evidenced by the highly signi…cant coe¢ cients on the two lags of the dependent variable. Columns (3) and (4) present the D-GMM estimates of the MTEF e¤ects. The internal instruments are the second and third lags of the budget balance and of the MTEF indicators. Column (4) augments column (3) with the external regional penetration instruments. In both speci…cations the e¤ect of each MTEF phase is positive and statistically signi…cant at conventional levels. Moreover, the e¤ects are economically meaningful, ranging between 1.305 and 4.577 percentage points. In our sample the average …scal balance among countries without an MTEF is 2:87%. Taking the estimates in column (4) at face value implies that only introducing the top-phase MTPF will put this average country in the black. The increase in the coe¢ cients with more advanced MTEF phases lends support to hypothesis (H1). These baseline speci…cations have lagged dependent variable coe¢ cients in the FE-OLS bracketing range, not raising any speci…cation issues. The model with external instruments is somewhat more precise than the one without. The D-GMM estimates of the MTEF e¤ects are larger than the FE and OLS estimates, suggesting that the latter are depressed by a potential negative reverse causality bias, as when countries tend to adopt an MTEF in response to a …scal crisis.31 For both models the diagnostic tests are satisfactory. The absence of …rst-order serial correlation in errors is rejected, while the absence of second-order serial correlation is not. The Hansen test does not reject the over-identi…cation restrictions. We conclude that D- GMM is an internally consistent estimator in this model, and can be relied upon to carry out statistical inference for the hypotheses of interest. Table 7 reports alternative speci…cations. Overall the results uphold the conclusions drawn from the baseline speci…cations of Table 6 columns (3) and (4). One exception is the model in column (3) that uses the di¤erence transform instead of the orthogonal deviations 31 This is consistent with the fact that international assistance organizations such as the World Bank, the UK’ s Department for International Development (DFID), the Asian Development Bank (ADB), and to a lesser extent the IMF, have recommended these reforms as part of a sound public …nancial management strategy. 17 transform. The lagged dependent variable coe¢ cient is below the bracketing range, sug- gesting that just di¤erencing the data produces a model with speci…cation problems. The estimates are also very stable across speci…cations, with the exception of column (4) where the MTEFs enter lagged, increasing the size of their coe¢ cients by a factor of between 1.5 and 2. As expected, collapsing the lagged dependent variable instruments in column (2) reduces statistical e¢ ciency, yet the coe¢ cients remain close to the baseline. We also note that in some speci…cations the con‡ict variable has a signi…cant and large negative coe¢ - cient, consistent with the notion that armed con‡ict is costly on the budget. The openness variable has a very small e¤ect and does not attain statistical signi…cance. 4.2 Allocative E¢ ciency Table 8 presents the baseline regressions for allocative e¢ ciency, measured as the volatility in per-capita health spending in PPP$; see equation (1) above. The table structure mimics the previous …scal discipline regressions. The sample period shortens due to the lack of health spending data in the …rst half of the 1990s, as well as the loss of one year to compute the volatility measure. The FE-OLS bracketing range for spending volatility, based on columns (1) and (2), is 0.009 0.308: None of the MTEF coe¢ cient estimates in these …rst two columns are signi…cant, moreover they seem small relative to the sample average spending volatility of 6.69%. Based on hypothesis (H2) above, we expect the advanced micro MTEF phases (MTBF and MTPF) to reduce spending volatility, with the top-phase reducing it more.32 Columns (3) and (4) present the D-GMM estimates of these e¤ects. The internal in- struments are the second lag of budget balance and the second and third lags of the MTEF indicators. Compared to the previous two columns, the estimated MTEF e¤ects become more negative and increase in magnitude. The MTBF e¤ect is around 6 percentage points and statistically signi…cant at conventional levels. The MTPF e¤ect is about twice larger, al- though only weakly signi…cant. Overall, these results provide support for hypothesis (H2).33 The statistical properties of these baseline models do not raise model speci…cation issues. The lagged dependent variable coe¢ cient estimate is in the bracketing range, although not reaching the signi…cance threshold. The …rst-order and second-order serial correlation tests s assumptions about the idiosyncratic error term. The Hansen test does support the model’ 32 Some countries chose to pilot an MTBF in the health sector before extending it to other sectors. We have been unable to systematically identify the countries that follow this particular sequencing of reform. 33 This result can be seen as a sectoral counterpart of the …nding in Fatas and Mihov (2003) that "dis- cretionary" …scal policy, i.e., variation in spending unrelated to economic fundamentals, increases aggregate output volatility. 18 not reject the over-identi…cation restrictions. Table 9 reports alternative speci…cations of the allocative e¢ ciency regressions. Overall the results uphold the conclusions drawn from the baseline speci…cations of Table 8 columns (3) and (4). The MTBF and MTPF coe¢ cient estimates maintain their prior patterns, although again the di¤erence transform model in column (3) performs somewhat poorly. The covariates openness and con‡ict do not display any signi…cant e¤ects on health spending volatility. 4.3 Technical E¢ ciency Table 10 presents the baseline regressions for technical e¢ ciency, measured as e¢ ciency scores from a stochastic frontier model of health delivery; see equations (2), (3), and Table 4. As expected, this indicator of …scal performance is much more persistent than the previous ones. The FE-OLS bracketing range, based on columns (1) and (2), climbs to 0.858 0.999, indicating strong persistence. National health delivery is a complex system that may take decades to fully internalize the bene…ts of a given reform. The MTEF coe¢ cient estimates are all negative, small and far from statistically signi…cant. Hypothesis (3) above predicts that the top-phase MTEF (the MTPF) increases technical e¢ ciency. Columns (3) and (4) present the D-GMM estimates of the MTEF e¤ects. The internal instruments are the second lag of technical e¢ ciency and the second and third lags of the MTEF indicators. Compared to the previous two columns, the estimated MTEF e¤ects turn positive and increase in magnitude. The speci…cation with external instruments is more precise. The MTPF e¤ect is 1.015 and statistically signi…cant at the 10 percent level. The lack of precision should be expected since the within-country variation in technical e¢ ciency is much smaller relative to the …rst two measures of …scal performance (1.67 vs. 9.59, 6.35). As before, the diagnostic tests perform well, increasing con…dence in the model speci…cation. Overall, these baseline results provide moderate support for hypothesis (H3). Table 11 reports alternative speci…cations of the technical e¢ ciency regressions. The Hansen test suggests that the di¤erence transform and predetermined MTEFs in columns (3) and (5) respectively are not adequate speci…cations. The remaining three models have satisfactory diagnostics and display a pattern of coe¢ cients similar to the baseline. Among the alternative speci…cations only the speci…cation with collapsed instruments attains sta- tistical signi…cance in the MTPF coe¢ cient (p-value 0.059). As in the baseline results, the covariates openness and con‡ict do not display measurable e¤ects on technical e¢ ciency. 19 4.4 Democracy and Development Finally, to rule out the possibility that the results are driven by subgroups of countries with extreme characteristics, we restrict the sample in two ways. First, we exclude highly autocratic countries, de…ned as those whose Polity IV score in 1990 takes the extreme value –10 ("strongly autocratic" in the language of the score producers). Second, we exclude highly developed countries, de…ned as those that are classi…ed by UNDP in 1990 as having “very s Human Development Index (HDI). The high human development� based on the country’ two subgroups of countries are listed in Table 14. Note that the most developed countries are also the most democratic. Tables 12 and 13 present the results with the restricted samples. The tables report the two early D-GMM baseline speci…cations for each of the three measures of …scal performance. Overall, the prior patterns are preserved, strengthening the support for hypotheses (H1)- (H3). The government balance regression in Table 12 column (2) shows stronger MTEF e¤ects both in magnitude and in statistical signi…cance. This result seems consistent with the notion that MTEFs as commitment devices are particularly well suited for democratic settings which Mueller and Stratmann (2003) argue are more prone to …scal indiscipline.34 The results in Table 13 excluding the most developed countries are comparable to the baseline results in Tables 6, 8, and 10, both in magnitude and statistical signi…cance, providing evidence that the MTEFs’ positive e¤ects on …scal performance are not a phenomenon speci…c only to the most developed countries.35 5 Conclusion In the last two decades more than 120 countries have moved toward a multiyear budget process. Although there has been much debate in the literature as to whether MTEFs are a worthwhile budget institution, a systematic empirical analysis of their impacts has been lacking due to insu¢ cient data on MTEF adoption around the world. This paper is the …rst to empirically investigate the MTEFs’ impacts on …scal performance in a large sample of 34 Ideally, to study how democracy and development condition MTEF impacts we would have interacted each MTEF indicator with measures of democracy and development. However, that would introduce three new endogenous variables, increasing the number of internal instruments above the number of available countries, and biasing the estimates toward their OLS counterparts. 35 We also explored how the results change when we exclude more countries. Generally, the results are more sensitive when excluding more autocratic countries than when excluding more developed countries. The estimation strategy limits the degree to which we can restrict the sample through the requirement that the number of instruments times the number of periods remain below the number of panel units. See Roodman (2009). 20 countries. In order to disentangle the e¤ects of the di¤erent MTEF phases (MTFF, MTBF, and MTPF) from other factors and to correct for reverse causality we apply a dynamic panel data approach to a newly-collected panel dataset of 181 countries over the period 1990-2008, the most comprehensive dataset to date on worldwide MTEF adoption. The econometric …ndings suggest that, unlike in previous small-sample and case-study analyses, MTEF adoption is associated with strong improvement in …scal discipline, the ef- fects increasing with each successive MTEF phase. The adoption of an MTBF and an MTPF decrease the volatility of health spending per capita, which we interpret as an improvement in allocative e¢ ciency. Finally, the MTPF seems to be the only MTEF phase that exerts a signi…cant e¤ect on technical e¢ ciency in the health sector, although due to insu¢ cient within-country variation in technical e¢ ciency this e¤ect is less precisely estimated. Over- all these results are more favorable than the conclusions of prior work, and suggest that budget institutions that restrain short-term incentives to manipulate the budget can have measurable bene…ts for …scal performance. Our analysis may be limited by the fact that an MTEF could be in place only in law (de jure) and not in practice (de facto). However, if this phenomenon were widespread it would induce an attenuation bias and our estimates could still be regarded as a lower bound on the actual e¤ect. Being in e¤ect commitment mechanisms, transparency and enforcement are critical components of MTEFs. Studying which features of the broader civic, juridical, and political environment enhance MTEF e¤ectiveness may lead to a better understanding of these institutions. Also, our analysis of MTEF impacts on allocative e¢ ciency and technical e¢ ciency is limited to the health sector. Whether the results carry over to other categories of productive spending is an interesting question for future research. To create more pre- cise measures of these dimensions of …scal performance one needs to identify disaggregated categories of spending that are comparable across a broad range of countries. 21 References Acemoglu, D. (2005) "Constitutions, Politics, and Economics," Journal of Economic Literature 43(4): 1025-1048. Alesina, A., R. Hausmann, R. Hommes, and E. Stein (1999) "Budget Institutions and Fiscal Performance in Latin America," Journal of Development Economics 59: 253-273. Alesina, A. and R. Perotti (1996) "Fiscal Discipline and the Budget Process," American Economic Review Papers and Proceedings : 86(2), 401-407. Alesina, A. and R. Perotti (1999) "Budget De…cits and Budget Institutions," in J. Poterba and J. von Hagen (eds.) Fiscal Institutions and Fiscal Performance, Chicago: U. of Chicago Press. Alesina, A., Tabellini G. (1990), “Voting on the Budget De…cit,� American Economic Review : 80(1): 37-49. Arellano, M. and S. Bond (1991) "Some Tests of Speci…cation for Panel Data: Monte Carlo Evidence and an Application to Employment Equations," Review of Economic Studies 58: 277-297. Arellano, M. and O. Bover (1995) "Another Look at the Instrumental Variables Estima- tion of Error-Components Models," Journal of Econometrics 68: 29–51. Baltagi, B.H., P.O. Demetriades, and S.H. Law (2009) "Financial Development and Open- ness: Evidence from Panel Data," Journal of Development Economics 89: 285-296. Bayoumi, T. and B. Eichengreen (1995) "Restraining Yourself: The Implications of Fiscal Rules for Economic Stabilizations," IMF Sta¤ Papers 42: 32-67. Besley, T. and A. Case (2003) "Political Institutions and Policy Choices: Evidence from the United States," Journal of Economic Literature 41(1): 7-73. Besley, T. and M. Smart (2007) "Fiscal Restraints and Voter Welfare," Journal of Public Economics 91(3-4): 755-773. Bevan, D. and G. Palomba (2000) "Uganda: The Budget and Medium-Term Expenditure Framework Set in a Wider Context," background paper for Poverty Reduction Support Credit with DFID …nance. Bisin, A., A. Lizzeri, and L. Yariv (2011) "Government Policy with Time Inconsistent Voters," working paper, NYU. Blundell, R. and S. Bond (1998) "Initial Conditions and Moment Restrictions in Dynamic Panel Data Models," Journal of Econometrics 87: 115–143. Bohn, H. and R.P. Inman (1996) "Balanced-Budget Rules and Public De…cits: Evidence from the U.S. states," Carnegie-Rochester Conference Series on Public Policy 45(1): 13-76. 22 Bond, S. (2002) "Dynamic Panel Data Models: A Guide to Micro Data Methods and Practice," working paper, Institute for Fiscal Studies. Brender, A. and A. Drazen (2005) "Political Budget Cycles In New Versus Established Democracies," Journal of Monetary Economics 52: 1271-1295. Castro, I. and W. Dorotinsky (2008) "Medium-Term Expenditure Frameworks: Demys- tifying and Unbundling The Concept," Journal of Public Budgeting, forthcoming. Drazen, A. (2000) "The Political Business Cycle after 25 Years," NBER Macroeconomic Annual 15: 75-117. Easterly, W. and S. Rebelo (1993) "Fiscal Policy and Economic Growth: An Empirical Investigation," Journal of Monetary Economics 32(3): 417-458. Easterly, W., T. Irwin and L. Servén (2008) "Walking up the Down Escalator: Public Investment and Fiscal Stability," World Bank Research Observer 23(1): 37-56. Fabrizio, S. and A. Mody (2006) "Can Budget Institutions Counteract Political Indisci- pline?" Economic Policy 21: 689-739 Farrell, M.J. (1957) "The Measurement of Productive E¢ ciency," Journal of the Royal Statistical Society 120(3): 253-282. Fatas, A. and I. Mihov (2003) "The Case for Restricting Fiscal Policy Discretion" Quar- terly Journal of Economics 118(4): 1419–1447. Fatas, A. and I. Mihov (2006) "The Macroeconomic E¤ects of Fiscal Rules in the U.S. States," Journal of Public Economics 90(1-2): 101-117. Filc, G. and C. Scartascini (2010) "Is Latin America on the Right Track? An Analy- sis of Medium-Term Frameworks and the Budget Process," working paper, Inter-American Development Bank, Research Department. Fischer, S. (1993) "The Role of Macroeconomic Factors in Growth," Journal of Monetary Economics 32(3): 485-512. Folscher, A. (2007) "Budget Methods and Practices," in A. Shaw (ed.) Budgeting and Budgetary Institutions, Washington, DC: World Bank. Gleich, H. (2003) "Budget Institutions and Fiscal Performance in Central and Eastern European Countries," European Central Bank research paper 215. Greene, W. (2005) "Distinguishing between Heterogeneity and Ine¢ ciency: Stochastic s Panel Data on National Health Care Systems," Health Frontier Analysis of the WHO’ Economics 13: 959-980. Hallerberg, M., S. Rolf, and J. von Hagen (2009) Fiscal Governance: Evidence from Europe, Cambridge: cambridge University Press. 23 Hallerberg, M. and J. von Hagen (1999) "Electoral Institutions, Cabinet Negotiations, and Budget De…cits in the European Union," in J. Poterba and J. von Hagen (eds.) Fiscal Institutions and Fiscal Performance, Chicago: U. of Chicago Press. Hallerberg, M. and S. Ylaoutinen (2000) "Political Power, Fiscal Institutions and Bud- getary Outcomes in Central and Eastern Europe," Journal of Public Policy 30: 45-62. Holmes, M. and A. Evans (2003) "A Review of Experience in Implementing Medium- Term Expenditure Frameworks in a PRSP Context: A Synthesis of Eight Country Studies," working paper, ODI. Holtz-Eakin, D., W. Newey, and H.S. Rosen (1988) "Estimating Vector Autoregressions with Panel Data," Econometrica 56: 1371-1395. Jondrow, J., I. Materov, K. Lovell, and P. Schmidt (1982) "On the Estimation of Tech- nical Ine¢ ciency in the Stochastic Frontier Production Function Model," Journal of Econo- metrics 19(2/3): 233-238. Keefer, P. and R. Vlaicu (2008) "Democracy, Credibility and Clientelism," Journal of Law, Economics and Organization 24(2): 41-67. Knight, B. (2000) "Supermajority Voting Requirements for Tax Increases: Evidence from the States," Journal of Public Economics 76: 191-222. Knight, B. and A. Levinson (2000) "Fiscal Institutions in U.S. States," in R. Strauch and J. von Hagen (eds.) Institutions, Politics, and Fiscal Policy, Boston: Kluwer. Kumar, M. S. and T. Ter-Minassian (2007) "Fiscal discipline: key issues and overview," in M. S. Kumar and T. Ter-Minassian (eds.) Promoting Fiscal Discipline, Washington, DC: IMF. Le Houerou, P. and R. Taliercio (2002) "Medium-Term Expenditure Frameworks: From Concept to Practice (Preliminary Lessons from Africa)," working paper, World Bank. Lewis, M. and M. Verhoeven (2010) "Financial Crisis and Social Spending: The Impact of the 2008-2009 Crisis," World Economics 11(4): 79-110. Marshall, M.G., K. Jaggers, and T.R. Gurr (2010) Polity IV Project: Political Regime Characteristics and Transitions 1800-201 0, Center for Systemic Peace. McNab, R.B., J. Martinez-Vazquez, and L.F.J. Boex (2000) "Multi-Year Budgeting: A Review of International Practices and Lessons for Developing and Transitional Economies," Public Budgeting and Finance 20(2): 91-112. Mueller, D. and T. Stratmann (2003) "The Economic E¤ects of Democratic Participa- tion," Journal of Public Economics 87(9-10): 2129-2155. Nickell, S. (1981) "Biases in Dynamic Models with Fixed E¤ects," Econometrica 49(6): 24 1417-26. Oyugi, L. N. (2008) "Experiences with Medium-Term Expenditure Framework in Selected Southern and Eastern African Countries," SEAPREN Working Paper No. 7. Oxford Policy Management (2000) "Medium-Term Expenditure Frameworks— Panacea or Dangerous Distraction?" Oxford Policy Management Review 2: 1-4. Perotti, R. and Y. Kontopoulos (2002) "Fragmented Fiscal Policy," Journal of Public Economics 86(2): 191-222. Persson, T., G. Roland, and G. Tabellini (2000) "Comparative Politics and Public Fi- nance," Journal of Political Economy 108(6): 1121-1161. Poterba, J. (2004) "State Responses to Fiscal Crises: The E¤ects of Budgetary Institu- tions and Politics," Journal of Political Economy 102(4): 799-821. Poterba, J. and J. von Hagen (1999) (eds.) Fiscal Institutions and Fiscal Performance, Chicago: U. of Chicago Press. Roodman, D. (2009) "A Note on the Theme of Too Many Instruments," Oxford Bulletin of Economics and Statistics 71: 135-158. The’Medium-Term Expenditure Frame- Schiavo-Campo, S. (2009) "Potemkin Villages: ’ work in Developing Countries," Public Budgeting and Finance 29(2): 1-26. Stein, E., E. Talvi, and A. Grisanti (1999) "Institutional Arrangements and Fiscal Per- formance: The Latin American Experience," in J. Poterba and J. von Hagen (eds.) Fiscal Institutions and Fiscal Performance, Chicago: U. of Chicago Press. Schick, A. (2010) "Fiscal and Budget Institutions: Consolidators or Accommodators?" Washington, DC: IMF Fiscal Forum. Velasco, A. (1999) "A Model of Endogenous Fiscal De…cits and Delayed Fiscal Reforms," in J. Poterba and J. von Hagen (eds.) Fiscal Institutions and Fiscal Performance, Chicago: U. of Chicago Press. von Hagen, J. (2006) "Political Economy of Fiscal Institutions," in B.R. Weingast and D.A. Wittman (eds.) The Oxford Handbook of Political Economy, Oxford, UK: Oxford University Press. von Hagen, J. and I. Harden (1995) "Budget Processes and Commitment to Fiscal Dis- cipline," European Economic Review 39: 771-779. von Hagen, J. and G. Wol¤ (2006) What Do De…cits Tell Us about Debt? Empirical Evidence on Creative Accounting with Fiscal Rules in the EU," Journal of Banking and Finance 30(12): 3259-3279. Wescott, C. G. (2008) "World Bank Support for Public Financial Management: Concep- 25 tual Roots and Evidence of Impact," IEG working waper, World Bank. Weingast, B.R., K. Shepsle, and C. Johnsen (1982) "The Political Economy of Bene…ts and Costs: A Neoclassical Approach to Distributive Politics," Journal of Political Economy 89(4): 642-664. Wildavsky, A. (1986) Budgeting: A Comparative Theory of Budgetary Processes, New Brunswick, NJ: Transaction Books. Windmeijer, F. (2005) "A Finite Sample Correction for the Variance of Linear E¢ cient Two-Step GMM Estimators," Journal of Econometrics 126: 25-51. World Bank (1998) Public Expenditure Management Handbook, Washington, DC: World Bank Publications. World Bank (2012) Beyond the Annual Budget: Review of Global Experience with Medium Term Expenditure Frameworks, "Directions in Development" Series, Washington, DC: World Bank Publications, forthcoming. Ylaoutinen, S. (2004) "Fiscal Frameworks in Central and Eastern European Countries," Finnish Ministry of Finance discussion paper 72. 26 Appendix A1. Figures Figure 1: MTEF Growth Worldwide, 1990-2008 80 MTFF MTBF MTPF 60 Number of Countries 40 20 0 1990 1995 2000 2005 2010 Year Notes: Authors’ calculations based on the sample of 181 countries during 1990-2008 described in the Data Appendix. Figure 2: MTEF Growth by Region, 1990-2008 1 Africa Americas Fraction of Countries with MTEF in Region Asia .8 Europe Oceania .2 .4 0 .6 1990 1995 2000 2005 2010 Year Notes: Authors’ calculations based on the sample of 181 countries during 1990-2008 described in the Data Appendix. MTEF status indicates the presence of one of the three MTEF phases. Geographical regions as defined by the UN Statistics Division. 27 Figure 3: MTEF and Income 1 BWA KOR NZL AUS GBR SWE DNK CAN NLD USA NOR SGP UGA .8 MTEF Use 1990-2008; in % MWI FIN MLI KGZ .6 TZA GIN BEN MOZ ZAF BGR PRT GRC ESP ITA FRABEL IRL DEU AUT LUX SVK GHA MLT CHE SVN CZE KEN RWA PER VENCHL MDA EST CMR BFAPHLALB THA BRA NAM HUN JPN .4 NPL MRT ETH HND NIC ROM LCAARG LTU LVA ISR ARM KAZ INDSWZ COL KHM SRB MUSMEX BIH MKD BGD NER ZMB SEN PNG CPV LKA POL ISL HRV TCD LSO VNM MDGPAK MAR JOR DMA TUN URY HKG .2 ZAR SLE COMGEO PRY SLB TUR BHR TJK NGA YEM MNG BTN TON COG VUT DZA RUSGNQ MYS BRB UZB BDI LBR IDN UKR LBN FJI KWT 0 GNB CAF ERI TMP TGO MMR GMB HTI SDN LAO STP CIV DJI GUY CHN AGO WSM BOL TKM SYR AZE MDV GTM EGY SLV DOM SUR BLZ ECU VCT GRD JAM BLR CRI IRN PAN KNAMNE GAB TTO ATGSYC CYP OMN SAU ARE QAT 0 20000 40000 60000 Average Real GDP Per Capita 1990-2008; in PPP$ Notes: Authors’ calculations based on the sample of 181 countries during 1990-2008 described in the Data Appendix. Country codes adjacent to each scatter point. Scatter plot quadratic fit shown. Figure 4: MTEF and Development 1 BWA SGP KOR GBRDNK SWE NLD NZL CAN USANOR AUS UGA .8 MTEF Use 1990-2008; in % MWI FIN MLI KGZ .6 MOZ GIN TZA BEN ZAF BGR PRT GRC ESP ITA AUT LUX FRAIRL DEU BEL SVK GHA MLT CHE CZE SVN RWA KEN VEN PER CHL MDA EST BFA CMR PHLTHA BRA ALB NAM HUN JPN .4 ETH MRT NPL NIC HND ROMARG LTU LVA ISR KAZ ARM IND SWZ KHM COLMUS MEXSRB BIH MKD NER SEN ZMB PNG BGD CPV LKA POL ISL HRV TCD AFG PAK LSOMDG MAR VNM JOR TUN DMAURY HKG .2 ZAR SLE COM SLB IRQ PRY TUR GEO BHR YEM NGA COG GNQ TJKMNG DZA TON RUS MYS BRB BDI LBR IDN UZB FJI UKR LBN KWT WBG 0 SDN CAF GNBGMB ZWE CIV MMR TGO DJI HTIAGO LAO TMPSTP GTM EGY SYR CHN GUY BOL MDV SLV GAB DOM IRNSUR WSM BLZ TKM JAM ECU OMN CRI TTO PAN BLR MNE SAU ARE SYC CYP QAT .2 .4 .6 .8 1 Average HDI (Human Development Index) 1990-2008 Notes: Authors’ calculations based on the sample of 181 countries during 1990-2008 described in the Data Appendix. Country codes adjacent to each scatter point. Scatter plot quadratic fit shown. 28 A2. Tables Table 1: Summary Statistics Variable Obs. Mean Std. Deviation Min Max Across Within Government Balance 2,991 –2.24 11.87 9.59 –151.33 384.15 Spending Volatility 2,282 6.69 7.62 6.35 0.002 83.19 Technical Efficiency 2,359 86.48 12.11 1.67 39.21 99.00 MTFF 3,378 0.17 0.38 0.32 0 1 MTBF 3,378 0.07 0.26 0.22 0 1 MTPF 3,378 0.04 0.20 0.13 0 1 MTFF Regional Penetration 3,359 0.17 0.21 0.18 0 1 MTBF Regional Penetration 3,359 0.07 0.14 0.12 0 1 MTPF Regional Penetration 3,359 0.04 0.14 0.08 0 1 Openness 3,069 85.28 48.72 16.89 0.31 456.64 Conflict 3,439 0.05 0.21 0.17 0 1 Health Spending per Capita 2,460 669.04 983.71 284.64 7.09 7,536.2 Life Expectancy 3,331 66.09 10.39 2.13 26.41 82.58 Population Density 3,304 188.89 650.61 60.86 1.43 6,943.2 OECD Membership 3,439 0.16 0.36 0.07 0 1 Notes: The summary statistics are based on the sample of 181 countries during 1990-2008 described in the Data Appendix. The appendix contains details on the data sources, units of measurement and construction of variables. The differences in number of observations across variables reflect data availability in the different data sources. 29 Table 2: MTEF Growth, 1990-2008 1990 2008 Adoptions Transitions Reversals MTFF 9 71 104 –41 –1 MTBF 1 42 21 23 –3 MTPF 1 19 0 18 0 Total MTEF 11 132 125 0 –4 Notes: The summary statistics are based on the sample of 181 countries during 1990-2008 described in the Data Appendix. Of the eighteen transitions to MTPF nine are from MTFF and nine from MTBF. The MTFF reversal is Argentina. The MTBF reversals are Argentina, Estonia, and the United States. Table 3: Correlations Matrix Gov Spend Tech MTFF MTBF MTPF Openn. Confl. Bal Volat Eff Gov_Bal 1 Spend_Volat .01 1 Tech_Eff .02 –.05** 1 MTFF .06*** .04* .10*** 1 MTBF .01 .05** –.16*** –.13*** 1 MTPF .06*** –.05** .14*** –.09*** –.06*** 1 Openn. .09*** –.02 .17*** .04** –.05*** .05*** 1 Confl. –.09*** .05** –.17*** –.05*** –.00 –.05*** –.13*** 1 Notes: The correlations are based on the sample of 181 countries during 1990-2008 described in the Data Appendix. The number of observations varies between 2,038 and 3,378 depending on the pair of variables. ***, **, * indicate statistical significance at the 1%, 5%, 10% levels, respectively. 30 Table 4: Stochastic Frontier Model for Public Health Delivery Dependent Variable: log(Life_Expi,t) (Log of Life Expectancy at Birth) Coefficients Model Statistics log(Health_Spendi,t) 0.00348*** (0.00099) Pop_Densityi,t 0.00002*** Mean Efficiency: µ = 86.4783*** (0.00000) OECDi,t 0.07829*** Std. Deviations: σu = 0.1567*** (0.00402) σv = 0.0247*** Constant 4.28722*** Ratio: λ = 6.3398*** (0.00747) Year Effects Yes Log Likelihood: log(L)=1686.66 Sample Period 1995-2008 Countries 177 Observations 2,359 Notes: The unit of observation is a country-year from the sample described in the Data Appendix. Standard errors in parentheses. The table reports maximum likelihood (ML) estimates of a stochastic frontier model for life expectancy with time-varying inefficiency term uit. The model assumes an exponential distribution for the inefficiency term. ***, **, * indicate statistical significance at the 1%, 5%, 10% levels, respectively. Table 5: Unit Root Tests Government Spending Technical Balance Volatility Efficiency IPS Test Statistic Ztbar ! –9.348*** –17.066*** –2.047** p-value 0.000 0.000 0.020 Average Panel Length 18.55 12.94 13.83 Countries 159 175 167 Notes: The tests are performed on the sample of 181 countries during 1990-2008 described in the Data Appendix. The table reports Im-Pesaran-Shin (IPS) unit root test results for the dependent variables. The null hypothesis is H0: All panels contain unit roots. The varying number of countries reflects the unbalanced nature of our panel and the requirement that the minimum length of each individual panel has to be at least ten. Panel means included. Time trends or lags not included. 31 Table 6: MTEFs and Fiscal Discipline: Baseline Dependent Variable: Gov_Balancei,t (Central Government Balance, % of GDP) Model: OLS FE D-GMM D-GMM-IV (1) (2) (3) (4) Gov_Balancei,t–1 0.481*** 0.379*** 0.421*** 0.423*** (0.045) (0.041) (0.040) (0.040) Gov_Balancei,t–2 0.174*** 0.116*** 0.101*** 0.101*** (0.051) (0.038) (0.037) (0.038) MTFFi,t 0.070 0.018 1.936** 1.305** (0.268) (0.318) (0.929) (0.605) MTBFi,t –0.187 0.274 2.068* 2.427** (0.352) (0.448) (1.133) (1.147) MTPFi,t 0.897*** 1.066* 4.577** 3.375** (0.305) (0.609) (2.103) (1.359) Opennessi,t 0.002 0.005 0.007 0.009 (0.001) (0.007) (0.008) (0.008) Conflicti,t –1.296* –1.736** –1.193* –1.055 (0.685) (0.838) (0.693) (0.670) Year Effects Yes Yes Yes Yes Internal Instruments No No Yes Yes External Instruments No No No Yes AR(1) Test p-val. 0.000 0.000 AR(2) Test p-val. 0.902 0.911 Hansen J Test p-val. 0.681 0.792 Sample Period 1990–2008 1990–2008 1990–2008 1990–2008 Countries 162 162 161 161 Observations 2,478 2,478 2,316 2,316 Notes: The unit of observation is a country-year from the sample described in the Data Appendix. Columns (1) and (2) report standard errors clustered at the country level. Columns (3) and (4) report two-step estimates and standard errors with the Windmeijer correction. GMM models use the orthogonal deviations transform. The internal instruments are: the second and third lags of Gov_Balancei,t, and of MTFFi,t, MTBFi,t, MTPFi,t. The external instruments are: MTFF_Regional_Penetrationi,t, MTBF_Regional_Penetrationi,t, and MTPF_Regional_Penetrationi,t. The internal instruments enter uncollapsed. ***, **, * indicate statistical significance at the 1%, 5%, 10% levels, respectively. 32 Table 7: MTEFs and Fiscal Discipline: Alternative Specifications Dependent Variable: Gov_Balancei,t (Central Government Balance, as % of GDP) Model: D-GMM-IV D-GMM-IV D-GMM-IV D-GMM D-GMM-IV (1) (2) (3) (4) (5) Gov_Balancei,t–1 0.400*** 0.405*** 0.321*** 0.420*** 0.414*** (0.037) (0.035) (0.041) (0.043) (0.040) Gov_Balancei,t–2 0.110*** 0.055 0.033 0.099*** 0.104*** (0.038) (0.038) (0.047) (0.038) (0.039) MTFFi,t 1.204* 1.149* 1.539* 2.484*** 1.045* [MTFFi,t–1 in col. (4)] (0.643) (0.681) (0.918) (0.881) (0.571) MTBFi,t 1.965** 2.103* 3.679** 3.192** 2.162** [MTBFi,t–1 in col. (4)] (0.997) (1.235) (1.676) (1.495) (1.028) MTPFi,t 3.600** 2.783* 4.757** 6.106*** 3.227** [MTPFi,t–1 in col. (4)] (1.757) (1.598) (2.156) (2.210) (1.555) Opennessi,t 0.003 0.003 0.005 0.006 0.007 (0.007) (0.008) (0.014) (0.008) (0.008) Conflicti,t –1.804** –1.436 –1.469 –1.369** –0.807 (0.776) (0.889) (0.972) (0.653) (0.623) Change from Baseline One-Step Collapsed Difference MTEFs MTEFs Robust Instrum. Transform Lagged Predeterm. Year Effects Yes Yes Yes Yes Yes Internal Instruments Yes Yes Yes Yes Yes External Instruments Yes Yes Yes No Yes AR(1) Test p-val. 0.000 0.000 0.000 0.000 0.000 AR(2) Test p-val. 0.923 0.517 0.597 0.829 0.970 Hansen J Test p-val. 0.792 0.991 0.484 0.867 0.367 Sample Period 1990–2008 1990–2008 1990–2008 1990–2008 1990–2008 Countries 161 161 161 161 161 Observations 2,316 2,316 2,313 2,316 2,316 Notes: The unit of observation is a country-year from the sample described in the Data Appendix. Each table column reports the estimates of a variation on the baseline specification in Table 6 column (4). Column (1) reports one-step estimates with robust standard errors instead of two-step estimates with Windmeijer corrected standard errors. Column (2) collapses the instruments for the lagged dependent variable. Column (3) uses the difference transform instead of the orthogonal deviations transform. Column (4) enters the three MTEF variables lagged one period and drops the external instruments. The bracketing range for this specification is 0.379–0.482. Column (5) treats the three MTEF variables as predetermined. The internal instruments are now the first and second lags of MTFFi,t, MTBFi,t, MTPFi,t. ***, **, * indicate statistical significance at the 1%, 5%, 10% levels, respectively. 33 Table 8: MTEFs and Allocative Efficiency: Baseline Dependent Variable: Spend_Volatilityi,t (Per Cap. Health Spending Volatility) Model: OLS FE D-GMM D-GMM-IV (1) (2) (3) (4) Spend_Volatilityi,t–1 0.308*** 0.009 0.087 0.105 (0.055) (0.050) (0.067) (0.065) MTFFi,t 0.342 0.277 –1.935 –0.092 (0.620) (0.647) (1.946) (1.889) MTBFi,t 0.338 0.196 –6.439** –5.823** (0.577) (1.120) (2.829) (2.415) MTPFi,t –0.950 –3.386 –14.879* –10.530 (1.122) (2.078) (8.053) (6.699) Opennessi,t –0.002 0.006 0.003 0.005 (0.003) (0.012) (0.013) (0.013) Conflicti,t –0.556 –0.898 –0.668 –0.918 (0.769) (1.103) (1.445) (1.492) Year Effects Yes Yes Yes Yes Internal Instruments No No Yes Yes External Instruments No No No Yes AR(1) Test p-val. 0.000 0.000 AR(2) Test p-val. 0.888 0.769 Hansen J Test p-val. 0.688 0.582 Sample Period 1996–2008 1996–2008 1996–2008 1996–2008 Countries 172 172 170 170 Observations 1,870 1,870 1,698 1,698 Notes: The unit of observation is a country-year from the sample described in the Data Appendix. Columns (1) and (2) report standard errors clustered at the country level. Columns (3) and (4) report two-step estimates and standard errors with the Windmeijer correction. GMM models use the orthogonal deviations transform. The internal instruments are: the second lag of Spend_Volatilityi,t, and the second and third lags of MTFFi,t, MTBFi,t, MTPFi,t. The external instruments are: MTFF_Regional_Penetrationi,t, MTBF_Regional_Penetrationi,t, and MTPF_Regional_Penetrationi,t The internal instruments enter uncollapsed. ***, **, * indicate statistical significance at the 1%, 5%, 10% levels, respectively. 34 Table 9: MTEFs and Allocative Efficiency: Alternative Specifications Dependent Variable: Spend_Volatilityi,t (Per Cap. Health Spending Volatility) Model: D-GMM-IV D-GMM-IV D-GMM-IV D-GMM D-GMM-IV (1) (2) (3) (4) (5) Spend_Volatilityi,t–1 0.079 0.115 0.044 0.072 0.046 (0.068) (0.072) (0.065) (0.043) (0.063) MTFFi,t 0.897 0.425 0.047 –1.764 –0.112 [MTFFi,t–1 in col. (4)] (1.156) (2.390) (1.816) (1.638) (1.189) MTBFi,t –5.328* –5.913** –2.904 –6.977* –2.671 [MTBFi,t–1 in col. (4)] (2.846) (2.914) (3.546) (3.914) (2.344) MTPFi,t –11.086* –7.583 –8.651 –14.697** –12.578** [MTPFi,t–1 in col. (4)] (6.615) (7.957) (5.870) (7.041) (5.699) Opennessi,t 0.004 0.005 –0.000 0.004 0.004 (0.012) (0.013) (0.027) (0.013) (0.015) Conflicti,t –0.353 –1.306 0.187 –0.203 0.294 (1.136) (1.462) (1.578) (1.431) (1.141) Change from Baseline One-Step Collapsed Difference MTEFs MTEFs Robust Instrum. Transform Lagged Predeterm. Year Effects Yes Yes Yes Yes Yes Internal Instruments Yes Yes Yes Yes Yes External Instruments Yes Yes Yes No Yes AR(1) Test p-val. 0.000 0.000 0.000 0.000 0.000 AR(2) Test p-val. 0.808 0.721 0.939 0.944 0.956 Hansen J Test p-val. 0.582 0.743 0.762 0.732 0.386 Sample Period 1996–2008 1996–2008 1996–2008 1996–2008 1996–2008 Countries 170 170 170 170 170 Observations 1,698 1,698 1,694 1,697 1,698 Notes: The unit of observation is a country-year from the sample described in the Data Appendix. Each table column reports the estimates of a variation on the baseline specification in Table 8 column (4). Column (1) reports one-step estimates with robust standard errors instead of two-step estimates with Windmeijer corrected standard errors. Column (2) collapses the instruments for the lagged dependent variable. Column (3) uses the difference transform instead of the orthogonal deviations transform. Column (4) enters the three MTEF variables lagged one period and drops the external instruments. The bracketing range for this specification is 0.011–0.309. Column (5) treats the three MTEF variables as predetermined. The internal instruments are now the first and second lags of MTFFi,t, MTBFi,t, MTPFi,t. ***, **, * indicate statistical significance at the 1%, 5%, 10% levels, respectively. 35 Table 10: MTEFs and Technical Efficiency: Baseline Dependent Variable: Tech_Efficiencyi,t (Estimated, see Stoch. Frontier Table 4) Model: OLS FE D-GMM D-GMM-IV (1) (2) (3) (4) Tech_Efficiencyi,t–1 0.999*** 0.858*** 0.920*** 0.934*** (0.004) (0.015) (0.066) (0.047) MTFFi,t –0.072 –0.091 –0.049 0.009 (0.052) (0.060) (0.137) (0.119) MTBFi,t –0.048 –0.071 0.303 0.427 (0.079) (0.080) (0.423) (0.281) MTPFi,t –0.065 –0.060 0.616 1.015* (0.079) (0.216) (0.693) (0.588) Opennessi,t –0.002*** –0.002 –0.001 –0.001 (0.001) (0.001) (0.002) (0.001) Conflicti,t 0.176** –0.028 –0.011 –0.000 (0.080) (0.044) (0.070) (0.048) Year Effects Yes Yes Yes Yes Internal Instruments No No Yes Yes External Instruments No No No Yes AR(1) Test p-val. 0.040 0.039 AR(2) Test p-val. 0.375 0.384 Hansen J Test p-val. 0.182 0.647 Sample Period 1995–2008 1995–2008 1995–2008 1995–2008 Countries 169 169 169 169 Observations 1,970 1,970 1,801 1,801 Notes: The unit of observation is a country-year from the sample described in the Data Appendix. Columns (1) and (2) report standard errors clustered at the country level. Columns (3) and (4) report two-step estimates and standard errors with the Windmeijer correction. GMM models use the orthogonal deviations transform. The internal instruments are: the second lag of Tech_Efficiencyi,t, and the second and third lags of MTFFi,t, MTBFi,t, MTPFi,t. The external instruments are: MTFF_Regional_Penetrationi,t, MTBF_Regional_Penetrationi,t, and MTPF_Regional_Penetrationi,t The internal instruments enter uncollapsed. ***, **, * indicate statistical significance at the 1%, 5%, 10% levels, respectively. 36 Table 11: MTEFs and Technical Efficiency: Alternative Specifications Dependent Variable: Tech_Efficiencyi,t (Estimated, see Stoch. Frontier Table 4) Model: D-GMM-IV D-GMM-IV D-GMM-IV D-GMM D-GMM-IV (1) (2) (3) (4) (5) Tech_Efficiencyi,t–1 0.905*** 0.912*** 0.916*** 0.931*** 0.919*** (0.043) (0.060) (0.072) (0.069) (0.061) MTFFi,t –0.021 0.035 –0.105 –0.087 –0.131 [MTFFi,t–1 in col. (4)] (0.147) (0.133) (0.129) (0.105) (0.121) MTBFi,t 0.366 0.440* 0.095 0.136 0.153 [MTBFi,t–1 in col. (4)] (0.255) (0.258) (0.296) (0.283) (0.158) MTPFi,t 0.791 1.163* –0.529 0.615 0.276 [MTPFi,t–1 in col. (4)] (0.565) (0.617) (0.553) (0.700) (0.363) Opennessi,t –0.002 –0.001 –0.002* –0.001 –0.001 (0.001) (0.001) (0.001) (0.001) (0.002) Conflicti,t –0.036 –0.035 –0.031 0.054 –0.028 (0.047) (0.053) (0.044) (0.043) (0.046) Change from Baseline One-Step Collapsed Difference MTEFs MTEFs Robust Instrum. Transform Lagged Predeterm. Year Effects Yes Yes Yes Yes Yes Internal Instruments Yes Yes Yes Yes Yes External Instruments Yes Yes Yes No Yes AR(1) Test p-val. 0.038 0.041 0.040 0.042 0.039 AR(2) Test p-val. 0.376 0.381 0.364 0.408 0.381 Hansen J Test p-val. 0.647 0.735 0.007 0.300 0.143 Sample Period 1995–2008 1995–2008 1995–2008 1995–2008 1995–2008 Countries 169 169 169 169 169 Observations 1,801 1,801 1,798 1,801 1,801 Notes: The unit of observation is a country-year from the sample described in the Data Appendix. Each table column reports the estimates of a variation on the baseline specification in Table 10 column (4). Column (1) reports one-step estimates with robust standard errors instead of two-step estimates with Windmeijer corrected standard errors. Column (2) collapses the instruments for the lagged dependent variable. Column (3) uses the difference transform instead of the orthogonal deviations transform. Column (4) enters the three MTEF variables lagged one period and drops the external instruments. The bracketing range for this specification remains the same 0.858–0.999. Column (5) treats the three MTEF variables as predetermined. The internal instruments are now the first and second lags of MTFFi,t, MTBFi,t, MTPFi,t. ***, **, * indicate statistical significance at the 1%, 5%, 10% levels, respectively. 37 Table 12: MTEFs and Fiscal Performance: Excluding Highly Autocratic Countries Dep. Variable (yi,t): Gov_Balancei,t Spend_Volatilityi,t Tech_Efficiencyi,t Model: (1) (2) (3) (4) (5) (6) yi,t–1 0.415*** 0.418*** 0.103 0.117* 0.962*** 0.981*** (0.038) (0.037) (0.063) (0.061) (0.102) (0.062) yi,t–2 0.101*** 0.100*** (0.037) (0.037) MTFFi,t 1.905** 1.875*** –1.852 –0.520 –0.089 –0.081 (0.898) (0.676) (1.855) (1.791) (0.130) (0.129) MTBFi,t 1.867* 3.139*** –6.289** –5.708** 0.201 0.286 (1.042) (1.152) (3.019) (2.409) (0.366) (0.353) MTPFi,t 4.853** 4.894*** –14.77* –11.14* 0.618 0.888 (2.156) (1.709) (8.225) (6.630) (0.623) (0.665) Opennessi,t 0.005 0.007 0.006 0.007 –0.000 –0.001 (0.007) (0.008) (0.013) (0.013) (0.001) (0.001) Conflicti,t –1.351* –1.071* –0.770 –0.907 0.010 0.022 (0.697) (0.599) (1.428) (1.495) (0.073) (0.047) Year Effects Yes Yes Yes Yes Yes Yes Internal Instrum. Yes Yes Yes Yes Yes Yes External Instrum. No Yes No Yes No Yes AR(1) Test p-val. 0.000 0.000 0.000 0.000 0.042 0.039 AR(2) Test p-val. 0.777 0.788 0.904 0.815 0.399 0.391 Hansen Test p-val. 0.644 0.934 0.702 0.663 0.308 0.300 Sample Period 1990–2008 1996–2008 1995–2008 Countries 155 155 164 164 163 163 Observations 2,228 2,228 1,638 1,638 1,735 1,735 Notes: The unit of observation is a country-year from the sample described in the Data Appendix. The table reports estimation results from running the baseline D-GMM specifications, columns (3)-(4) from Tables 6, 8, and 10, on a restricted sample that excludes the most autocratic countries, see table 14. Two-step estimates with Windmeijer corrected standard errors in parentheses. The models use the orthogonal deviations transform. The internal instruments are: the second and third lags of yi,t in columns (1)-(2), the second lag of yi,t in columns (3)-(6), and the second and third lags of MTFFi,t, MTBFi,t, MTPFi,t. in all columns. The external instruments are: MTFF_Regional_Penetrationi,t, MTBF_Regional_Penetrationi,t, and MTPF_Regional_Penetrationi,t The internal instruments enter uncollapsed. ***, **, * indicate statistical significance at the 1%, 5%, 10% levels, respectively. 38 Table 13: MTEFs and Fiscal Performance: Excluding Highly Developed Countries Dep. Variable (yi,t): Gov_Balancei,t Spend_Volatilityi,t Tech_Efficiencyi,t Model: (1) (2) (3) (4) (5) (6) yi,t–1 0.422*** 0.425*** 0.070 0.087 0.924*** 0.937*** (0.039) (0.036) (0.062) (0.056) (0.071) (0.048) yi,t–2 0.103*** 0.096*** (0.037) (0.036) MTFFi,t 1.443 1.273* –2.632 0.187 –0.035 0.047 (1.063) (0.691) (2.364) (1.774) (0.116) (0.158) MTBFi,t 2.042 2.685** –5.915 –5.193* 0.286 0.531 (1.583) (1.270) (3.906) (2.738) (0.333) (0.372) MTPFi,t 3.939* 3.685** –23.63* –15.70* 0.939 1.321* (2.195) (1.507) (10.652) (9.352) (0.726) (0.801) Opennessi,t 0.012* 0.011 0.002 0.003 –0.001 –0.001 (0.007) (0.008) (0.013) (0.013) (0.001) (0.001) Conflicti,t –1.013 –0.787 –0.288 –0.246 0.003 0.007 (0.809) (0.654) (1.434) (1.357) (0.046) (0.055) Year Effects Yes Yes Yes Yes Yes Yes Internal Instrum. Yes Yes Yes Yes Yes Yes External Instrum. No Yes No Yes No Yes AR(1) Test p-val. 0.000 0.000 0.000 0.000 0.046 0.043 AR(2) Test p-val. 0.915 0.856 0.959 0.887 0.387 0.379 Hansen Test p-val. 0.785 0.946 0.889 0.821 0.677 0.791 Sample Period 1990–2008 1996–2008 1995–2008 Countries 151 151 160 160 159 159 Observations 2,173 2,173 1,602 1,602 1,700 1,700 Notes: The unit of observation is a country-year from the sample described in the Data Appendix. The table reports estimation results from running the baseline D-GMM specifications, columns (3)-(4) from Tables 6, 8, and 10, on a restricted sample that excludes the most developed countries, see table 14. Two-step estimates with Windmeijer corrected standard errors in parentheses. The models use the orthogonal deviations transform. The internal instruments are: the second and third lags of yi,t in columns (1)-(2), the second lag of yi,t in columns (3)-(6), and the second and third lags of MTFFi,t, MTBFi,t, MTPFi,t. in all columns. The external instruments are: MTFF_Regional_Penetrationi,t, MTBF_Regional_Penetrationi,t, and MTPF_Regional_Penetrationi,t The internal instruments enter uncollapsed. ***, **, * indicate statistical significance at the 1%, 5%, 10% levels, respectively. 39 Table 14: Country Extremes, 1990 Highly Autocratic Highly Developed Countries Countries Bahrain Australia Bhutan Belgium Oman Canada Qatar Japan Saudi Arabia Netherlands Swaziland New Zealand Norway Sweden Switzerland United States Notes: Authors’ calculations using the sample of 181 countries described in the Data Appendix. A country is defined as highly autocratic if its Polity IV score in 1990 takes the extreme value –10. A country is defined as highly developed if classified by UNDP in 1990 as having “very high human development� based on the country’s Human Development Index. 40 A3. Data Appendix This appendix contains the complete list of variables used in the paper, together with details on measurement and sources. Fiscal Performance Government Balance: Ratio of the overall central government Öscal balance to GDP, in percent. Sources: IMF World Economic Outlook. Spending Volatility: Absolute growth rate in health spending per capita, in PPP$. Sources: Authorsí calculations. See equation (1) in the paper. Technical E¢ciency: Estimations of e¢ciency scores from a stochastic frontier model that shows life expectancy as output and health spending per capita in PPP$ as input. Sources: Authorsí calculations. See equations (2) and (3), and Table 4 in the paper. Budget Institutions MTFF: Dummy variable that takes the value one if MTFF is the highest MTEF phase adopted, zero otherwise. Sources: World Bank and IMF documents and country specialists, case studies. MTBF: Dummy variable that takes the value one if MTBF is the highest MTEF phase adopted, zero otherwise. Sources : World Bank and IMF documents and country specialists, case studies. MTPF: Dummy variable that takes the value one if MTPF is the highest MTEF phase adopted, zero otherwise. Sources: World Bank and IMF documents and country specialists, case studies. MTFF Regional Penetration: The percentage of MTFF adopters in the countryís geographic region. We use the twenty-two geographic regions deÖned by the United Nations Statistics Division. See equation (4) in the text. Sources: Authorsí calculations. MTBF Regional Penetration: The percentage of MTBF adopters in the countryís geographic region. We use the twenty-two geographic regions deÖned by the United Nations Statistics Division. See equation (4) in the text. Sources: Authorsí calculations. MTPF Regional Penetration: The percentage of MTPF adopters in the countryís geographic region. We use the twenty-two geographic regions deÖned by the United Nations Statistics Division. See equation (4) in the text. Sources: Authorsí calculations. 41 Country Characteristics Conáict: Dummy variable that takes the value one if there are at least 1,000 battle- related casualties, zero otherwise. Sources: UCDP/PRIO Armed Conáict Dataset, Uppsala University. Health Spending per Capita: Health Expenditure per capita in PPP$ terms. Sources: World Health Organization. HDI: Human Development Index, a composite index measuring average achievement in three basic dimensions of human development: a long and healthy life, knowledge, and a decent standard of living. Sources: United Nations Development Programme. Life Expectancy: Life expectancy at birth, in years. Sources: World Bank World Development Indicators. OECD Membership: Dummy variable that takes the value one if the country belongs to the OECD, zero otherwise. Sources: OECD. Openness: Trade openness measured as the ratio of the sum of imports plus exports to GDP. Sources: World Bank World Development Indicators. Polity Score: Composite score ranging from -10 (strongly autocratic) to 10 (strongly democratic) based on the Polity IV methodology. Sources: Marshall, Jaggers, and Gurr (2010). Population Density: Residents per square kilometer. Sources: World Bank World Development Indicators. Region: One of twenty-two geographical regions. Sources: United Nations Statistics Division. 42