Economic policy analysis in international development

Transcription

1 Economic policy analysis in international development D. Weinhold, A.S. Rigterink DV2169, Undergraduate study in Economics, Management, Finance and the Social Sciences This is an extract from a subject guide for an undergraduate course offered as part of the University of London International Programmes in Economics, Management, Finance and the Social Sciences. Materials for these programmes are developed by academics at the London School of Economics and Political Science (LSE). For more information, see:

2 This guide was prepared for the University of London International Programmes by: Dr Diana Weinhold, Department of Development Studies, London School of Economics and Political Science Anouk S. Rigterink, Doctoral Candidate, Department of International Development, London School of Economics and Political Science This is one of a series of subject guides published by the University. We regret that due to pressure of work the authors are unable to enter into any correspondence relating to, or arising from, the guide. If you have any comments on this subject guide, favourable or unfavourable, please use the form at the back of this guide. The University of London International Programmes Publications Office Stewart House 32 Russell Square London WC1B 5DN United Kingdom Website: Published by: University of London University of London 2007 Reprinted with minor revisions 2011 The University of London asserts copyright over all material in this subject guide except where otherwise indicated. All rights reserved. No part of this work may be reproduced in any form, or by any means, without permission in writing from the publisher. We make every effort to contact copyright holders. If you think we have inadvertently used your copyright material, please let us know

3 Contents Contents Chapter 1: Introduction... 1 Aims of the course... 3 Learning outcomes... 3 How to use this subject guide... 4 Essential reading... 5 Further reading... 7 Online study resources Examination advice Syllabus Chapter 2: Introduction to quantitative methodology Aims of the chapter Learning outcomes Essential reading Further reading Introduction Sampling Normal distribution and the central limit theorem Hypothesis testing Basics of regression Endogeneity Cross-section, time series and panel data Difference-in-difference estimation Summary Now read A reminder of your learning outcomes Sample examination questions Chapter 3: Economic growth: basic concepts, ideas and theories Aims of the chapter Learning outcomes Essential reading Further reading Introduction What is growth? How important is growth? Economic growth in the far distant past Modern growth theories: Harrod-Domar and Solow growth model Modern growth theories: endogenous growth theory Solow growth model versus endogenous growth theory A reminder of your learning outcomes Sample examination questions Chapter 4: New directions in growth theory Aims of the chapter Learning outcomes Essential reading i

4 169 Economic policy analysis in international development ii Further reading Introduction Heterogeneous firms, misallocation and intermediate goods Economic geography A reminder of your learning outcomes Sample examination questions Chapter 5: Institutions and (very) long-run growth Aims of the chapter Learning outcomes Essential reading Further reading Introduction What are institutions? Investigating how institutions could matter for long-run growth: IV regression revisited Property rights: Engerman and Sokoloff and Acemoglu, Johnson and Robinson Critiques: geography and human capital A reminder of your learning outcomes Sample examination questions Chapter 6: Globalisation and trade theory Aims of the chapter Learning outcomes Essential reading Further reading Introduction Classical trade theory: Ricardo and the Heckscher-Ohlin model New trade theory New new trade theory: heterogeneous firms A reminder of your learning outcomes Sample examination questions Chapter 7: Finance and financial crises Aims of the chapter Learning outcomes Essential reading Further reading Introduction Inflation Policy response to inflation Old style Balance of Payments crises Structural adjustment programmes (SAPs) New style currency crises Policy options for new style crises A reminder of your learning outcomes Sample examination questions Chapter 8: Microfinance Aims of the chapter Learning outcomes Essential reading Further reading Introduction

5 Contents The credit market Further reading Microfinance programmes The success of microfinance: empirical evidence A reminder of your learning outcomes Sample examination questions Chapter 9: Aid effectiveness Aims of the chapter Learning outcomes Essential reading Further reading Introduction Defining aid Whether and how to give aid? The current big debate on aid effectiveness Is aid related to growth? Empirical evidence A reminder of your learning outcomes Sample examination questions Appendix A: Sample examination paper Appendix B: Tables for Questions 1 and Table for Question Table for Question iii

6 169 Economic policy analysis in international development Notes iv

7 Chapter 1: Introduction Chapter 1: Introduction Welcome to 169 Economic policy analysis in international development. The purpose of this course is to provide an accessible, yet rigorous, introduction to evidence-based policy analysis for development that is suitable for students from a wide variety of backgrounds. You must have passed 171 Introduction in international development before you study this course. The design of development policies must be derived from development theory. There is no single theory of economic development, but we strive to provide a background review of competing theories of how many complex processes, both at the macroeconomic (country and international) level and the microeconomic (individual, household and firm) level, may lead to development. Given the large number of topics in development economics, this course will focus on macroeconomic theories, although the chapter on microcredit in particular, and the chapter on aid to a lesser extent, include some microeconomics. If these topics are particularly interesting to you, we suggest you consider taking a course in microeconomics of development. Theoretical questions we address in this course include: through what mechanisms are trade and growth related? Will providing microcredit enable farmers to improve their living standards? Does international aid promote growth? How can we explain large differences in living standards between developed and developing countries? For each of these topics there may be several theoretical answers, each theory corresponding to a unique, but not necessarily mutually incompatible, mechanism through which the variables of interest are related. Which policy design will be most successful will in turn depend on which underlying mechanism turns out to be most true. If all the competing theories are on first inspection logical and internally consistent, which policy should we choose? Even if one theory sounds more appealing (or plausible to our ears), how can we assess the validity of our intuition? For many years most debates surrounding development policies were characterised by rhetorical arguments about which theoretical approach was more intrinsically persuasive, something determined at least as much by ideological and political position as by anything else. However in the last 30 years or so, with access to comparable cross-national, household, and individual level data dramatically increasing, the discipline has completely shifted its approach. Today, we strive to evaluate theories and policies on the basis of empirical data; this is what we mean by evidence-based analysis. Broadly speaking, there are three possible approaches to evidencebased analysis that have been used in social science disciplines such as economics, political science and sociology. The first is to derive some testable implications from theory, and then use statistical methods to examine whether these implications are consistent with the observed data. Ideally we look for testable implications that differentiate between theories, or some relationship that is highly unlikely unless the associated theory had strong explanatory power. The second approach arises because sometimes it is difficult to differentiate between many possible theoretical mechanisms, so as a first pass to the problem we focus on a reduced form relationship (in which the 1

8 169 Economic policy analysis in international development 2 underlying mechanism is something of a black box ) and use observed data empirically to test directly whether some policies have had the causal effect that they were supposed to generate. For example, did trade opening increase growth? Did microfinance reduce poverty? Did building more schools increase enrolment? However, one of the major methodological challenges facing both of these approaches is to convincingly draw out causal conclusions from nonexperimental data. Many social and economic variables are correlated, but how can we discern whether there really is a causal relationship? In this course we will explore several different empirical strategies that have been utilised to identify causal relationships, but these statistical methods may not be applicable, or may be unconvincing, in some cases. As a result in the past decade it has become very popular, when feasible, to design and conduct a randomised controlled experiment to test the effects of a given policy. For example, if new textbooks are assigned randomly to different schools, then the differences in learning outcomes between those schools with and without the new textbooks can be considered to be the impact of the textbook policy. Despite the recent enthusiasm for randomised controlled experiments, it is important to keep in mind that no single methodology provides the most convincing answers to all possible research questions, and that sometimes choosing between particular techniques involves trade-offs. Consider the following: a randomised controlled trial may provide the most convincing evidence on a causal relationship between textbooks and pupil performance. However, designing such an experiment can be very costly and time-consuming, requiring significant financial resources. Non-experimental research could possibly be done using already existing data. Furthermore, the costs of running an experiment will likely force researchers to focus on a small subset of children that may benefit from receiving textbooks children in a particular province in Kenya, for example. If we find that textbooks increase pupil performance in Kenya, can we say with confidence that they will have the same effect on children in the whole of Kenya? All children in Africa? All children in developing countries? The experiment may have conclusively proven that the results hold for this particular subgroup (it has high internal validity), but it may not generalise to other cases (its external validity may be doubtful). In the case of a different research question, researchers may be unable or unwilling to randomise. It seems highly unethical to withhold life-saving drugs from people for the sake of an experiment. It seems infeasible to ask countries to randomise their trade policies. Imagine if policymakers exclusively focus on policies that could be randomised; this would limit their options to a specific category of policies, and these types of policies are not necessarily the most effective. As policymakers depend more and more heavily on evidence-based analyses to guide their decisions, it has become ever more important for a wide range of policymakers and other interested individuals to be able to read and critically assess the ever increasing quantity of quantitative studies. Many politicians, decision-makers, and other policy stakeholders (including the general public) have not had the several years of postgraduate training in statistics and econometrics (the statistical analysis of economic data) that it takes to master the sophisticated methodological tool kit used by modern-day social scientists. In sum, many more people really need the skills to critically read and consume the information conveyed in the academic quantitative literature than possess the time or desire to finish a PhD required to produce those studies.

9 Chapter 1: Introduction Thus for over more than a decade of teaching in the Department of International Development at the London School of Economics and Political Science, we have focused on conveying the underlying intuition behind the structure of quantitative work, teaching students not how to actually do econometrics, but rather how to read and consume quantitative analyses critically and rigorously. Each year our starting group of students ranges from many with no formal economics or mathematics backgrounds at all to those with extensive statistical educations, from English literature majors to lawyers to geologists to central bankers. By the end of the course they can all pick up any World Bank technical working paper or a good quality academic article in economics or political science and be able to read, critically assess, and come to their own conclusions about the validity of the empirical strategy and policy conclusions. A key lesson that we teach is that good quality quantitative policy analysis is not a purely technical or statistical exercise, but is firmly embedded within the theoretical framework of the processes under study. With a few basic concepts and some new vocabulary, we show you how econometric equations map back to the economic theories we have been discussing in each chapter. Good econometric analysis simply reflects good analysis, full stop, so critically engaging with the quantitative material is more a matter of learning how to think rigorously and thoroughly through a problem (whether it be quantitative or qualitative), and then being able to translate that intuition into the quantitative framework. The authors of this subject guide include one academic with many years of quantitative training and practice and a young researcher, who prior to taking a similar MSc course to this course, had not had any formal training in quantitative methods at all. We are both passionate about passing on the lessons of rigorous, analytical thinking in the context of economic development policy to students from all backgrounds. The process of integrating these new ways of thinking can be time consuming and, at times, frustrating, but we hope you will find that the rewards are well worth the effort. And this subject guide will be there to assist you along your way. Aims of the course The aims of this course are: to provide a critical overview of current growth and welfare policies in developing countries to demonstrate how the underlying theories that inform development policies are evolving in light of continuous empirical testing to provide a comprehensive introduction to evidence-based policy analysis, including a non-technical but operational ability to read and comprehend regression analyses used in quantitative policy evaluation. Learning outcomes On completion of this course, you are expected to be able to: describe the main theories, debates and concepts in development economics demonstrate a clear understanding of the major economic policy issues in developing countries read, comprehend and critique empirical analysis in the context of development policy evaluations at a non-technical level 3

10 169 Economic policy analysis in international development 4 demonstrate an understanding of how theories of development economics have evolved and shaped policy over the past 50 years. How to use this subject guide The aim of this subject guide is to help you understand the required material, highlight the main points and explain how various readings are connected. Each chapter will start by outlining aims and learning outcomes of the chapter and by listing essential and further reading. It will then take you through the relevant material. To enable you to test yourself, the learning outcomes are restated at the end of each chapter and sample examination questions are provided. We recommend that from Chapter 3 onwards you work through the topics in the order they are given in the subject guide. Read the essential reading at the points at which you are asked to do so. The subject guide will then explain the main points you should take away from what you have just read, and how these relate to other material. If you did not pick up these main points from the reading, you might find it useful to revisit it. When you are particularly interested in a specific topic and feel you have mastered the required material, you can read some of the further reading. Keep in mind that understanding the basics is the sure way to pass the course, but that showing a deeper understanding of how various pieces of research relate to each other and referring to extra material can earn you a higher grade. An essential element of this course is empirical analysis: the quantitative testing of the predictions of theories and the evaluation of policies using real-world data. Understanding regression analysis, an often-used technique for doing so, is therefore essential. In every chapter we will include at least one reading that uses quantitative analysis so that as you progress through the course you will get plenty of practice reading and engaging with regression tables. An outline of the essential knowledge you will need for understanding empirical analysis to the level required in this course is provided in Chapter 2. This chapter contains all methodological concepts necessary to understand the remainder of the subject guide, so it is very important to read it carefully (even if you have already taken or are taking another Statistics course). However, it is not necessary to understand all methodology after the first read: you will practise with the concepts throughout the subject guide and should refer back to Chapter 2 whenever necessary. Please note that you are not required to know how to do regression yourself or to remember numerical results. However, it is important to understand the intuition behind empirical analysis and why any given analysis was set up the way it was. For each reading the particulars may be slightly different, but all the quantitative analyses have the following points in common: The specification should reflect one or more testable implications of some underlying theories. Given the theory, the specification should take into account and try to control for possible omitted variables. Given the theory, the specification should take into account and try to control for possible reverse causality or simultaneity (or endogeneity). The assumptions required to ensure that the results are free from omitted variable and endogeneity bias introduce caveats and limitations to the results interpretation and policy implications. As

11 Chapter 1: Introduction all empirical work requires some assumptions along these lines, each analysis will be limited in its own way. When reading empirical analyses, it is important to revisit each of these points and be able to show whether (and how well) the analysis handles possible biases and, as a result, how comfortable you are with the resulting policy implications. The subject guide will give you guidance in this as well on a reading-by-reading basis. Empirical analysis plays dual roles in development policy; on the one hand, quantitative analysis is often used to differentiate between competing theoretical explanations of economic phenomena, and empirical policy evaluation is used to help judge the effectiveness of policies derived from those underlying theories. On the other hand, the basic theoretical evolution itself is highly influenced by the outcomes of empirical results. Quantitative analysis may find one set of theories more consistent with observed data and reject another set of theories, which then fall by the historical wayside. Often careful empirical work generates new stylised facts that inspire a new generation of academics to develop new theories. The amount of real-world data and empirical evidence is constantly increasing, and at least partially as a direct result, theory and policy designs are also evolving. You will therefore see that the subject guide does not proclaim one true theory or the best policy in an absolute sense. It will, however, explain from what evidence or observations a theory followed, and how the theory is or is not able to explain what we observe in the real world. We want you to think creatively and critically, but also rigorously and in a disciplined manner. In the end, you should have the tools not only to explain the current state of development policy, but also to track progress in this area going forward through critical reading of academic and popular literature. Essential reading You need to purchase one book: Ray, Debraj Development Economics. (Chichester: Princeton University Press, 1998) [ISBN ]. This textbook is an overview of theories in development economics. We will assign relevant parts of this textbook to read throughout the subject guide. There is no need to read specific parts of the book that are not assigned. In some assigned pages and articles there may be some technical material presented that is beyond the scope of this course. This subject guide will point out the sections that you will not be responsible for reproducing, and help you to understand the non-technical intuition. Textbooks often cannot keep up with recent developments in theory. Equally, empirical research is mostly published in academic journal articles. Therefore, the remainder of the essential reading consists of journal articles. These are available through the University of London Online Library. All International Programmes students have free access to this resource and you will be able to find the full-text articles here. You will need a username and password to get access. Alternatively, Google Scholar is a useful searching device. Finding literature can be timeconsuming, but remember that finding literature yourself is an important skill you will need to develop throughout your studies. For each topic, we also provide an extensive list of further reading. You can read extra material if you are particularly interested in the topic at 5

12 169 Economic policy analysis in international development 6 hand. Read extra material selectively: we by no means expect you to read all, or even most of it. If you are truly struggling with the course, we do not expect you to do any extra reading. Take into account that a student who has mastered the essential reading fully and has done no extra reading will generally do much better on the exam than a student who moved on to further reading before having understood the essential reading. However, when you can show yourself to be proficient in the essential material, showing knowledge of additional readings can take your grade up that extra notch. Full list of Essential reading Acemoglu, Daron, Simon Johnson and James A. Robinson The colonial origins of comparative development. An empirical investigation, American Economic Review 2001, 91(5), pp Amiti, Mary and Josef Konings Trade liberalization, intermediate inputs and productivity: Evidence from Indonesia, American Economic Review 2007, 97(5), pp Banerjee, Abhijit, Esther Duflo, Rachel Glennerster and Cynthia Kinnan The miracle of microfinance? Evidence from a randomized evaluation, Working Paper. (Boston: Massachusetts Institute of Technology, 2009). Available at: Miracle%20of%20Microfinance.pdf Bernard, Andrew et al. Firms in international trade, Journal of Economic Perspectives 2007, 21(3), pp Bhagwati, Jagdish N. and Anne O. Krueger Exchange control, liberalization and economic development, The American Economic Review 1973, 63(2), pp Bleakley, Hoyt Disease and development: Evidence from hookworm eradication in the American South, The Quarterly Journal of Economics 2007, 122(1), pp Burnside, Craig and David Dollar Aid, policies and growth, The American Economic Review 2000, 90(4), pp Dollar, David and Aart Kraay Growth is good for the poor, Journal of Economic Growth 2002, 7(3), pp Dornbusch, Rudi A primer on emerging market crises, NBER Working Paper series No. 8326, Available at: Easterly, William Foreign aid goes military! Review of The bottom billion by Paul Collier, The New York Review of Books 2008, 55(19). Available at: Easterly, William The Big Push déjà vu. Review of The end of poverty by Jeffrey Sachs, Journal of Economic Literature 2006, 44(1), pp Easterly, William, Ross Levine and David Roodman New data, new doubts: A comment on Burnside and Dollar s Aid, policies and growth (2000), NBER Working Paper No. 9846, Available at: w9846 Engerman, Stanley L. and Kenneth L. Sokoloff Factor endowments, institutions and differential paths of growth among new world economies, NBER Working Paper No. 9259, Available at: Galor, Oded and David N. Weil Population, technology and growth. From Malthusian stagnation to the demographic transition and beyond, The American Economic Review 2000, 90(4), pp Glaeser, Edward L., Rafael La Porta, Florencio Lopez-De-Silanez and Andrei Shleifer Do institutions cause growth?, Journal of Economic Growth 2004, 9(3), pp Hsieh, Chang-Tai and Peter J. Klenow Misallocation and manufacturing TFP in China and India, The Quarterly Journal of Economics 2009, 124(4), pp

13 Chapter 1: Introduction Jones, Charles I. Intermediate goods, weak links and superstars. A theory of economic development, NBER Working Paper series No , Available at: Kaplan, Ethan and Dani Rodrik Did the Malaysian capital controls work?, NBER Working Paper series No. 8142, Available at: papers/w8142 Morales, Juan Antonio and Jeffrey Sachs Bolivia s economic crisis, NBER Working Paper series No. 2620, Available at: w2620 Morduch, Jonathan The microfinance promise, Journal of Economic Literature 1999, 37(4), pp Puga, Diego and Anthony J. Venables The spread of industry. Spatial agglomeration in economic development, Centre for Economic Policy Research Discussion Paper No. 1354, Available at: Sachs, Jeffrey How to help the poor: piecemeal progress or strategic plans? Review of The white man s burden by William Easterly, The Lancet 2006, 376, pp Available at: article/piis (06) /fulltext Sachs, Jeffrey D. Institutions don t rule. Direct effects of geography on per capita income, NBER Working Paper No. 9490, Available at: www. nber.org/papers/w9490 Detailed reading references in this subject guide refer to the editions of the set textbooks listed below. New editions of one or more of these textbooks may have been published by the time you study this course. You can use a more recent edition of any of the books; use the detailed chapter and section headings and the index to identify relevant readings. Also check the virtual learning environment (VLE) regularly for updated guidance on readings. Further reading Please note that as long as you read the Essential reading you are then free to read around the subject area in any text, paper or online resource. You will need to support your learning by reading as widely as possible and by thinking about how these principles apply in the real world. To help you read extensively, you have free access to the VLE and University of London Online Library (see below). For ease of reference, here is a full list of all Further reading referred to in this subject guide: Acemoglu, Daron and Simon Johnson Unbundling institutions, Journal of Political Economy 2004, 113(5), pp Banerjee, Abhijit V. and Esther Duflo Inequality and growth. What can the data say?, Journal of Economic Growth 2003, 8(3), pp Berkowitz, Daniel and Karen Clay American civil law origins. Implications for state constitutions, American Law and Economics Review 2003, 7(1), pp Burnside, Craig, Martin Eichenbaum and Sergio Rebelo Prospective deficits and the East Asian currency crisis, World Bank Policy Research Working Paper No. 2174, Chari, Varadarajan, V., Patrick J. Kehoe and Ellen R. McGrattan Business cycle accounting, Econometrica 2007, 75(3), pp Clemens, Michael A. and Gabriel Demombynes When does rigorous impact evaluation make a difference? The case of the Millennium Villages, World Bank Policy Research Paper #5477, Available at: com/sol3/papers.cfm?abstract_id=

14 169 Economic policy analysis in international development Coleman, Brett E. The impact of group lending in Northeast Thailand, Journal of Development Economics 2000, 60(1), pp Collier, Paul The bottom billion. Why the poorest countries are failing and what can be done about it. (Oxford: Oxford University Press, 2007) [ISBN ]. Dell, Melissa The mining Mita: Explaining institutional persistence, MIT Mimeo, Diamond, Jared Guns, germs and steel. (New York: W.W. Norton & Company, 1997) [ISBN ]. Easterly, William The cartel of good intentions: the problem of bureaucracy in foreign aid, The Journal of Policy Reform 2002, 5(4), pp Easterly, William The elusive quest for growth. Economists adventures and misadventures in the Tropics. (Cambridge MA: MIT Press, 2001) [ISBN ]. Easterly, William The White Man s Burden. Why the West s efforts to aid the Rest have done so much ill and so little good. (Oxford: Oxford University Press, 2006) [ISBN ]. Easterly, William What did structural adjustment adjust? The association of policies and growth with repeated IMF and World Bank adjustment loans, Journal of Development Economics 2005, 76(1), pp Easterly, William When is fiscal adjustment an illusion?, World Bank draft working paper, Available at www-wds.worldbank.org/ external/default/wdscontentserver/iw3p/ib/1999/09/14/ _ /additional/ _ pdf Easterly, William and Ross Levine What have we learned from a decade of empirical research on growth? It s not factor accumulation: stylized facts and growth models, World Bank Economic Review 2001, (15)2, pp Edwards, Sebastian How effective are capital controls? The Journal of Economic Perspectives 1999, 13(4), pp Eichengreen, Barry Capital account liberalization. What do cross-country studies tell us?, The World Bank Economic Review 2001, 15(3), pp Eichengreen, Barry, Ricardo Hausmann and Ugo Panizza Currency mismatches, debt intolerance and original sin. Why they are not the same and why it matters, NBER Working Paper series No. w10036, Fernald, John G. and Brent Neiman Measuring the miracle. Market imperfections and Asia s growth experience, FRB of San Francisco Working Paper No , Grossman, Gene M. and Elbahan Helpman Trade, knowledge spillovers and growth, European Economic Review 1991, 35(2), pp Helpman, E. and P. Krugman Market Structure and Foreign Trade. (Cambridge, Massachusetts: MIT Press, 1985) [ISBN ]. Henderson, Vernon, Todd Lee and Yung Joon Lee Scale externalities in Korea, Journal of Urban Economics 2001, 49(3), pp Henderson, Vernon J., Adam Storeygard and David N. Weil Measuring economic growth from outer space, NBER Working Paper No. w15199, Kanbur, Ravi The economics of international aid in Kolm, Serge-Christophe and Jean Mercier Ythier (eds) Handbook of the economics of giving, altruism and reciprocity. (Amsterdam: Elsevier B.V., 2006) [ISBN ] pp Klenow, Peter J. and Andres Rodriguez-Clare The Neoclassical revival in growth economics. Has it gone too far? NBER Macroeconomics Annual 1997, 12, pp Kopits, George and Steve Symansky Fiscal policy rules, International Monetary Fund Occasional Paper. (Washington D.C.: International Monetary Fund, 1998). 8

15 Chapter 1: Introduction Krugman, Paul A model of innovation, technology transfer and the world distribution of income, The Journal of Political Economy 1979, 87(2), pp La Porta, Rafael, Florencio Lopez-de-Silanes, Andrei Shleifer and Robert Vishny Law and finance, Journal of Political Economy 1998, 106(6), pp Levitt, Steven D. Using electoral cycles in police hiring to estimate the effect of police on crime, The American Economic Review 1997, 87(3), pp Limao, Nuno and Anthony J. Venables Infrastructure, geographical disadvantage, transport costs and trade, World Bank Economic Review 2001, 15(3), pp Melitz, Marc J. The impact of trade on intra-industry reallocations and aggregate industry productivity, Econometrica 2003, 71(6), pp Moretti, Enrico Social returns to education and human capital externalities. Evidence from cities, Unpublished working paper, Available at: Moyo, Dambisa Dead Aid. Why aid is not working and how there is another way for Africa. (New York: Allen Lane, 2008) [ISBN ]. North, Douglas Institutions, institutional change and economic performance. (Cambridge UK: Cambridge University Press, 1990) [ISBN ]. Nunn, Nathan and Diego Puga Ruggedness: The blessing of bad geography in Africa, Mimeo, Pavcnik, Nina Trade liberalization, exit and productivity improvements: Evidence from Chilean plants, The Review of Economic Studies 2002, 69(1), pp Pritchett, Lant Divergence, big time, Journal of Economic Perspectives 1997, 11(3), pp Quah, Danny Empirics for growth and distributions. Stratification, polarization and convergence clubs, Journal of Economic Growth 1997, 2(1), pp Quah, Danny Some simple arithmetic on how income inequality and economic growth matter, Working Paper, Available at: edu/viewdoc/download?doi= &rep=rep1&type=pdf Rajan, Raghuram G. and Arvind Subramanian Aid and growth: What does the cross country evidence really show?, The Review of Economics and Statistics 2008, 90(4), pp Reinhart, Carmen M. and Kenneth Rogoff Banking crises. An equal opportunity menace, NBER Working Paper series No. w14587, Rodrik, Dani Why we learn nothing from regressing economic growth on policies, Harvard University Working Paper, Available at: international.ucla.edu/media/files/rodrik.pdf Sachs, Jeffrey The end of poverty: Economic possibilities for our time. (New York: Penguin Press, 2005) [ISBN ]. Sala-i-Martin The disturbing rise of global income inequality, NBER Working Paper No. w8904, 2002a. Stiglitz, Joseph E. and Andrew Weiss Credit rationing in markets with imperfect information, The American Economic Review 1981, 71(3), pp Williamson, Jeffrey G. Five centuries of Latin American inequality, NBER Working Paper No. w15305, Young, Alwyn The African growth miracle, Working Paper, Available at: Young, Alwyn The tyranny of numbers. Confronting the statistical realities of the East Asian growth experience, The Quarterly Journal of Economics 1995, 110(3), pp

16 169 Economic policy analysis in international development Journals These academic journals are particularly relevant to (development) economists: American Economic Journal: Applied Economics American Economic Journal: Economic Policy American Economic Review Econometrica Journal of Development Economics Journal of Economics Perspectives Journal of International Economics Journal of Political Economy Journal of Public Economics NBER working paper series Review of Economic Studies Quarterly Journal of Economics World Bank Economic Review. Unless otherwise stated, all websites in this subject guide were accessed in April We cannot guarantee, however, that they will stay current and you may need to perform an internet search to find the relevant pages. Online study resources In addition to the subject guide and the Essential reading, it is crucial that you take advantage of the study resources that are available online for this course, including the VLE and the Online Library. You can access the VLE, the Online Library and your University of London account via the Student Portal at: You should have received your login details for the Student Portal with your official offer, which was ed to the address that you gave on your application form. You have probably already logged in to the Student Portal in order to register! As soon as you registered, you will automatically have been granted access to the VLE, Online Library and your fully functional University of London account. If you forget your login details at any point, please london.ac.uk quoting your student number. The VLE The VLE, which complements this subject guide, has been designed to enhance your learning experience, providing additional support and a sense of community. It forms an important part of your study experience with the University of London and you should access it regularly. The VLE provides a range of resources for EMFSS courses: Self-testing activities: Doing these allows you to test your own understanding of subject material. Electronic study materials: The printed materials that you receive from the University of London are available to download, including updated reading lists and references. 10

17 Chapter 1: Introduction Past examination papers and Examiners commentaries: These provide advice on how each examination question might best be answered. A student discussion forum: This is an open space for you to discuss interests and experiences, seek support from your peers, work collaboratively to solve problems and discuss subject material. Videos: There are recorded academic introductions to the subject, interviews and debates and, for some courses, audio-visual tutorials and conclusions. Recorded lectures: For some courses, where appropriate, the sessions from previous years Study Weekends have been recorded and made available. Study skills: Expert advice on preparing for examinations and developing your digital literacy skills. Feedback forms. Some of these resources are available for certain courses only, but we are expanding our provision all the time and you should check the VLE regularly for updates. Making use of the Online Library The Online Library contains a huge array of journal articles and other resources to help you read widely and extensively. To access the majority of resources via the Online Library you will either need to use your University of London Student Portal login details, or you will be required to register and use an Athens login: The easiest way to locate relevant content and journal articles in the Online Library is to use the Summon search engine. If you are having trouble finding an article listed in a reading list, try removing any punctuation from the title, such as single quotation marks, question marks and colons. For further advice, please see the online help pages: Unless otherwise stated, all websites in this subject guide were accessed in We cannot guarantee, however, that they will stay current and you may need to perform an internet search to find the relevant pages. Examination advice Important: the information and advice given here are based on the examination structure used at the time this guide was written. Please note that subject guides may be used for several years. Because of this we strongly advise you to always check both the current Regulations for relevant information about the examination, and the VLE where you should be advised of any forthcoming changes. You should also carefully check the rubric/instructions on the paper you actually sit and follow those instructions. The examination paper for this course is three hours in duration and you are expected to answer three questions, from a choice of 10. The Examiners attempt to ensure that all the topics covered in the syllabus and subject guide are examined. Some questions could cover more than one topic from the syllabus since the different topics are not self-contained. 11

18 169 Economic policy analysis in international development 12 The most important thing to do in the exam is first to read the question carefully and fully understand what is asked. In answering the question, provide a clear argument, using the theories and empirical analyses you have learned. Remember that it is very important to provide evidence. Answering that institutions lead to growth because the theory says it will is not sufficient. Provide examples of empirical analyses or real-world observations that support this claim. A Sample examination paper appears as an appendix to this guide. The Examiners commentaries contain valuable information about how to approach the examination, so you are strongly advised to read them carefully. Past examination papers and the associated commentaries are valuable resources in preparing for the examination. Remember, it is important to check the VLE for: up-to-date information on examination and assessment arrangements for this course Syllabus where available, past examination papers and Examiners commentaries for the course which give advice on how each question might best be answered. Chapter 2: Introduction to quantitative analysis Sampling Hypothesis testing Normal distribution and central limit theorem Regression: basics, control variables, dummy variables, interaction terms Endogeneity Regression: IV regression, difference-in-difference estimation Cross-section, time series and panel data Chapter 3: Economic growth: basic concepts, ideas and theories Introduction to growth and GDP Growth and inequality Historical growth experience Neoclassical or Solow growth model Endogenous growth models Chapter 4: New directions in growth theory Extensions to the neoclassical model: heterogeneous firms, misallocation and intermediate goods Economic geography Chapter 5: Institutions and (very) long run growth Introduction to institutions and growth Property rights: Acemoglu, Johnson and Robinson Factor endowments: Engerman and Sokolof Geography and growth Human capital and growth

19 Chapter 1: Introduction Chapter 6: Globalisation and trade theory Ricardian comparative advantage Heckscher-Ohlin model New trade theory New new trade theory Chapter 7: Finance and financial crises Causes and consequences of inflation Old style financial crises: balance of payments crises and structural adjustment programmes New style financial crises and currency crises Policy options during financial crises Chapter 8: Microfinance The credit market: incomplete information and limited liability Consequences of imperfect credit markets for developing countries and the poor Introduction to microfinance Evaluating the success of microfinance Chapter 9: Aid effectiveness Introduction to ODA Big debate on aid: Jeffrey Sachs, William Easterly, Paul Collier and Dambisa Moyo Empirical evidence on aid and growth on country and programme level 13

20 169 Economic policy analysis in international development Notes 14

21 Chapter 2: Introduction to quantitative methodology Chapter 2: Introduction to quantitative methodology Aims of the chapter This chapter aims to familiarise you with some quantitative methodology, mainly regression analysis, to the point where you are able to consume (understand and critically assess) empirical academic papers in development economics. It is meant as a reference chapter that you can return to when reading the remaining chapters of this subject guide. Learning outcomes By the end of this chapter, and having completed the Essential reading and Activities, you should be able to: recognise when the topics introduced in this chapter are applied in subsequent chapters and reading. When you have completed the course you should be able to: read and critically assess empirical research employing the aspects of quantitative methodology covered in this chapter. One purpose of this chapter is to familiarise you with quantitative techniques used in the papers you will read throughout the course. Therefore, we urge you to read this chapter thoroughly now (even if you have taken another Statistics course) and review this chapter when reading these papers. Essential reading Ray, Debraj Development Economics. (Chichester: Princeton University Press, 1998). Appendix 2, pp Further reading Levitt, Steven D. Using electoral cycles in police hiring to estimate the effect of police on crime, The American Economic Review 1997, 87(3), pp Introduction The second chapter of this course is about quantitative methodology, mainly regression analysis. The techniques reviewed here are used extensively, but by no means exclusively, in research in development economics and development policy analysis. Scholars use regression analysis in their attempts to answer questions such as: What are the drivers of economic growth? Is openness to trade related to growth? Can a particular aid programme (a microfinance programme, a programme providing books to school children, budget support for developing country governments) be considered successful? As you can see from these questions, the aspects of quantitative methodology covered in this chapter can be applied to every topic in the remainder of the subject guide: economic growth, international trade, aid, credit, etc. 15

22 169 Economic policy analysis in international development There is a reason why we start with methodology and not with any of these topics specific to international development. The knowledge about methodology that you will need is similar for all these topics, and we wanted to provide you with a chapter where it is all neatly put together. In fact, the basics of all the methodology you will need to know are in this chapter. In the subsequent chapters, you will see these basics being applied to different areas again and again. This is to say that while the subject matter of this chapter is very important for understanding the rest of the course, you do not need to have fully grasped it all by the end of this chapter. You will practise with it again and again throughout the whole course. You can revisit relevant sections of this chapter every time you are asked to read an empirical paper and will probably understand the material better every time you do so. A disadvantage of providing you with a methodology reference chapter is that you may feel a bit overwhelmed reading it for the first time. Again, please remember that reading empirical papers and understanding their methodology is a skill that will be built up with time and practice. And as with many skills (like playing the piano, or public speaking), once you learn them, they stick with you and they can be applied in other contexts. If you learn the piano playing Mozart, you will probably do pretty okay playing Bach. Like piano-playing skills, the skills you learn reading an empirical paper about trade can be employed reading empirical work on child labour, the environment or even topics that are completely unrelated to development. A final thing to remember is that this course aims to teach you how to be a good consumer of empirical research. After working through the subject guide, we hope that you can read an empirical paper, understand most of the quantitative methodology being employed and make an informed argument about how convincing you find the empirical strategy. We do not ask you to be a producer of empirical research. You are not expected to be able to do regression analysis, use statistical programmes or do any calculations whatsoever. You will also never be asked to reproduce numerical results of an empirical paper. In terms of our earlier example, an examination question will never be: By what percentage does the productivity of firms go up, if trade barriers are lowered by 10 per cent, according to paper X? but it may be similar to: Why would we expect lower trade barriers to increase firm productivity? How does paper X test this expectation? Do you feel that the authors of paper X have successfully avoided [potential problems with this analysis]?. Sampling When carrying out research, we often want to draw conclusions about a large number of entities: a population. We may want to know the average income of all people in a country, the productivity of all firms in a country, the effect of having new schoolbooks on the attendance of all school children in a region. We can go and collect data on the whole population, but this is likely to be time consuming and expensive. When done correctly, sampling can provide a much cheaper alternative to a complete census at a small accuracy cost. Sampling is the process by which we select a subset of observations from all possible observations in a population. When taking a sample, the first step is to define the target population. This consists of all the observations about which the research aims to draw conclusions. The sampled population consists of all observations 16

23 Chapter 2: Introduction to quantitative methodology that have some chance of ending up in the sample. Ideally, the target population and the sampled population should be the same. An example: you want to investigate the average number of rooms in London houses. You take a sample from all houses in a London landline phone register. In this case the target population is all houses in London. The sampled population, however, is all houses in London with a landline. Houses without a landline have no chance of ending up in the sample: they are in the target population, but not in the sampled population. Once the target population is determined and we have a suitable sample frame, we take a sample. Various sampling techniques exist, but here we will focus on a simple random sample. In a simple random sample, every member of the population has an equal, known probability of ending up in the sample. You can think of this as throwing all population members into a hat and randomly picking a number of them. Proper random samples of a sufficient size can deliver strikingly accurate predictions about the total population. If it is not the case that each possible observation has an equal, known probability of ending up in the sample, and/or if the target population is different from the sampled population, the sample is biased. It is important to think about the impact that bias can have on the conclusions of the research. When observations that are more or less likely to end up in the sample are systematically different from each other on a characteristic relevant for the outcome, this impact is especially serious. Return to our earlier example: you want to draw conclusions about the average number of rooms in all houses in London, but take a sample of all houses in London with landlines. Houses with landlines may be significantly different from houses without landlines. Maybe older inhabitants are more likely to have a landline, whereas younger ones only have a mobile phone. Age of the inhabitants of a house may be relevant for the number of rooms in the house, as older people may be able to afford bigger houses. Then, the average number of rooms in your sample may be higher than that in the total population of all London houses. There are many types of sampling biases of which consumers of research have to be aware. We will review a number of them. First, self-selection bias. This occurs when subjects themselves have some say over whether or not they are included in the sample. An example: when mailing out surveys, receivers can choose whether or not to send them back. People interested enough to send the survey back may be systematically different from non-respondents. Secondly, strategic bias occurs when subjects feel they can manipulate policy choices by participating in the research, impacting the composition of the sample. An example: imagine a microfinance organisation doing a survey of poverty in a village where the organisation will set up a programme. If villagers feel that they can increase their chances of obtaining credit by participating in the survey, the eventual sample may be very different compared to a situation where villagers believe participating will not benefit them (for example when a government statistics bureau is doing a regular poverty survey). Finally, interviewer bias. This occurs when the subject responds to characteristics of the interviewer in such a way that the composition of the sample is influenced. When the subject of research is of a sensitive nature this may be especially serious. Consider doing a survey on condom use for example. Whether the interviewers are male or female is likely to impact the willingness of men and women to participate, biasing the sample. 17

24 169 Economic policy analysis in international development Normal distribution and the central limit theorem A probability distribution is a graphical or mathematical representation that tells us what the relative probabilities are of any of the possible outcomes of a process. For example, we can plot all possible outcomes of a process on the horizontal axis of a graph, and the percentage of the time that these outcomes occur on the vertical axis. Consider the following example: we observe the number of absent children in a 5-child village school, for 45 days. This may have the following result: No. of children absent Table 2.1 No. of days this is observed Relative frequency Probability 0 1 1/ / / / / / The probability distribution can be represented by the following graph: Figure 2.1 The normal distribution is a special kind of probability distribution. It is bell-shaped and symmetric, as in the graph below. 18

25 Chapter 2: Introduction to quantitative methodology Figure 2.2 As it is symmetric, the mean of the normal distribution lies exactly in the middle of the distribution. In the graph above, the mean is 0. Here we will call the mean of this normal distribution μ ( mu ). The bulk of observations is concentrated around the mean. This means that extreme outcomes (such as 3 in the graph above) are very unlikely, whereas outcomes close to the mean are more likely. The standard deviation of a normal distribution is a number that represents how widely observations are spread around the mean. A large standard deviation means a wide spread, a small standard deviation means a small spread. In the graph below, the dashed curve has a smaller standard deviation than the solid curve. x Figure 2.3 x 19

26 169 Economic policy analysis in international development Regardless of how wide or narrow they are, all normal distributions have the following properties in common: 68 per cent of the probability density of a normal distribution is located within 1 standard deviation around its mean. In Figure 2.4 below: 68 per cent of observations lie between the dashed lines at ±1. 95 per cent of the probability density of a normal distribution is located within 2 standard deviations around its mean. In Figure 2.5 below: 95 per cent of observations lie between the dashed lines at ±2. For example: suppose we know that the height of UK men is normally distributed, with a mean of 175cm and a standard deviation of 5cm. Using this information, we know that 68 per cent of UK men are between 170 (175 5) cm and 180 (175+5) cm tall. We also know that 95 per cent of UK men are between 165 (175 2*5) cm and 185 (175+2*5) cm tall. Figure 2.4 and Figure 2.5 x Imagine that you draw a large number of random samples from the same target population. You calculate the sample mean of each of these samples. You then draw a probability distribution of the sample mean: what shape would it have? The central limit theorem states that as you draw an increasing number of successively random samples from a population, the distribution of sample means will approach a normal distribution. As the number of samples tends to infinity, then the sampling distribution converges to N(µ, σ 2 /n). The distribution of the population itself is irrelevant for this statement. Activity Experiment with the central limit theorem on this website: Instructions: 1. When you arrive on this website, press Begin. You see four graphs. The first graph shows the distribution of the population from which a sample is taken (the parent population). 2. You can choose between normal, uniform, skewed and customised distributions. Choose the uniform distribution first. A black block will appear because each outcome is equally likely to appear in a (continuous) uniform distribution. x 20

27 Chapter 2: Introduction to quantitative methodology 3. The second graph is entitled Sample Data. Press the Animated button and you will see how a sample is drawn from the distribution above. Press Animated a few times. Each time a new blue column is added to the third graph representing the distribution of the sample means. It is the mean of each sample that you obtained by pressing Animated. 4. You can choose to draw 5, 1,000, or 10,000 samples at a time. Accordingly, 5, 1,000, or 10,000 blue sample means are added to the distribution of sample means. Press these buttons a few times. Soon you will see a normal distribution appear. Note that this happens even though the target population is not normally distributed. 5. On the right-hand side of the Distribution of Means graph you can choose the sample size of the samples you draw. When you choose N=2 a sample mean distribution with a large standard error ( standard error is the technical name for the standard deviation of the sample mean) will appear. If you choose a higher value, e.g. N=25 a more narrow distribution will appear. 6. Now repeat the exercise for a skewed distribution and you will see that again the distribution of means will become normal. Experiment with the simulation as much as you like! Hypothesis testing As explained in the introduction, theories or observations of the real world can give rise to certain predictions. To test these predictions, we formulate them as hypotheses. For each prediction, we formulate two hypotheses: the null hypothesis (H 0 ) and the alternative hypothesis (H 1 ). Together, these hypotheses should cover all possible outcomes. Usually, H 0 concludes that nothing happened, whereas H 1 concludes that something happened. For example: H 0 H 1 A treatment has had no effect on an outcome variable A treatment has had an effect on an outcome variable Group A is not different from group B Group A is different from group B Variable X is unrelated to variable Y Variable X is related to variable Y Table 2.2 H 0 is the hypothesis to be tested effectively the working hypothesis. We observe something in the real world and then calculate how likely it would be for us to observe what we did, assuming that H 0 is correct. If this is sufficiently unlikely, we reject H 0 in favour of the alternative hypothesis. Consider the following example: we may think that UK men aged are taller on average than all UK men. We know that UK men are 175cm tall on average. We formulate the null and alternative hypotheses: a. H 0 : UK men aged are not taller on average than all UK men. b. H 1 : UK men aged are taller on average than all UK men. We take a random sample of all UK men aged The men in our sample are 179cm tall on average. Now we ask ourselves: what is the probability that we would get a sample with a mean of 179cm, when the true population mean height of men aged is 175cm (assuming that H 0 is true)? We calculate (do not worry how for now) that this probability is 4 per cent. 21

28 169 Economic policy analysis in international development 22 So we now know that getting this sample with an average of 179cm is unlikely if H 0 were true. But is it sufficiently unlikely to reject H 0? The researcher chooses the probability level at which to reject H 0. This is called the significance level. Common significance levels are: 10 per cent, 5 per cent and 1 per cent. When the probability that H 0 is true given our observations is below 10 per cent, 5 per cent or 1 per cent respectively, H 0 is rejected at the respective significance level. In our example, H 0 is rejected when we choose a significance level of 10 per cent or 5 per cent, but we fail to reject H 0 when we choose a significance level of 1 per cent. (Note the phrase fail to reject we never accept H 0 merely we state that we have insufficient evidence to reject it.) When deciding whether or not to reject the null hypothesis we can make two types of error: A Type I error occurs if we reject H 0 even though it is in fact correct. A Type II error occurs if we fail to reject H 0 even though it is in fact wrong. The researcher controls the probability of making a Type I error by setting the significance level. If we choose a decision rule to reject H 0 when the probability that it is correct is below 5 per cent, there is a maximum 5 per cent chance that we reject H 0 when it is in fact correct. (Basically this would occur if we obtained an unlucky or unrepresentative sample.) If we set our significance level at 1 per cent, the probability of making a Type I error gets smaller. As may be evident from this example: the smaller the chance of a Type I error, the larger the chance of a Type II error hence a trade-off between the two error types exists (for a given sample size). Because we do not know the truth we cannot calculate the probability of a Type II error. So how do we calculate the probability that H 0 is true given our observations? We have an observed sample mean (x) and hypothetical population mean (µ 0 ). Because of the central limit theorem, we know that the sample mean is (approximately) normally distributed. We also have a good approximation of the standard error of the sampling distribution (s/ n, which approximates σ/ n, where σ is the population standard deviation). We calculate by how many standard errors the observed sample mean and the hypothetical population mean are apart. This number is called the t-statistic. For example, our calculated sample mean and the mean under H 0 are two standard errors apart: the t-statistic is 2. Written down as a formula: t = x μ 0 s/ n The t-distribution is very similar to the standard normal distribution N(0,1), but for small samples it has slightly fatter tails, as we had to estimate the standard error of the sampling distribution and therefore introduced a bit more uncertainty. However, these distributions are very similar when we take a sample of reasonable size. (As a rule of thumb, we may use the standard normal distribution as an approximation for the t-distribution for large sample sizes, for example n 50.) An overview of critical values of when the t-statistic converges to the standard normal (i.e. when the sample size is (very) large): If the observed sample mean and hypothetical population mean are further apart than 1.96 standard errors we can reject H 0 at the 5 per cent level.

29 Chapter 2: Introduction to quantitative methodology If the observed sample mean and hypothetical population mean are further apart than 1.64 standard errors we can reject H 0 at the 10 per cent level. If the observed sample mean and hypothetical population mean are further apart than 2.58 standard errors we can reject H 0 at the 1 per cent level. When sample sizes are a bit smaller, the tails of the t-distribution are a bit fatter and the critical values are a bit larger. Critical values for different degrees of freedom are included in the back of most statistics textbooks. For the most part, using a rule of thumb value that you will reject the null hypothesis for any t-statistic that is 2 or larger in absolute value (for statistical significance at the 5 per cent level) is a pretty common and relatively benign practice. The p-value reports the probability mass in the sampling distribution that lies outside of your estimate. For example, a p-value of.035 means that 3.5 per cent of the probability mass of the sampling distribution under the null lies outside your sample estimate (technically this definition is for a onesided test but you do not need to worry about one-sided and two-sided tests). Thus if you want to test the null at 5 per cent you will look for a p-value of.05 or lower. From the example above, our p-value of.035 means we can reject the null hypothesis at 5 per cent, but not at 1 per cent. Basics of regression In this section, we are going to bring together the material we have covered so far to learn the basics of regression. Regression analysis in essence does two things: it establishes a conditional correlation between two variables and it determines how likely it is that this correlation is due to chance. A correlation is a systematic relationship between two variables. For example: aid spending is related to poverty reduction, or: trade openness is related to growth. Note that a correlation may be positive or negative: more aid spending is related to less poverty (a negative correlation) and more trade openness is related to more growth (positive correlation). Note that a correlation between X and Y does not necessarily mean that X causes Y. Correlation does not necessarily imply causation. You will see why in the next section. Establishing that two variables are correlated in a sample might tell us something, but is of limited policy relevance by itself. Between most variables, there will be some correlation due to pure chance. If you let 100 people roll a die three times, you will probably be able to find people whose rolls seem to be correlated. However, as rolling a die is random, we can comfortably say that this correlation is due to chance. We need to apply what we learned about hypothesis testing here: if the probability of observing a particular relationship between X and Y, when they are in fact unrelated, is below a certain significance level (say there is less than a 5 per cent chance that we would observe a particular relationship in a sample if in fact there was none in reality in the population), we conclude that there is a significant relation between X and Y. For example: we observe a positive correlation between education and wage: the more education the higher the wage. From the central limit theorem we know there is less than a 3 per cent chance of observing this relationship if education and wage were unrelated. If our significance level was set at 5 per cent, we can now say that the positive relationship between education and wage is statistically significant, at the 5 per cent level. Note that the relationship is not significant at the 1 per cent level. 23

30 169 Economic policy analysis in international development How does establishing a correlation work more technically? Although we do not go into the full technical details, we will illustrate this graphically. Imagine that you have two variables that you think are related: X and Y. We can plot these on a graph. Figure 2.6 Now we think about a single line that will best depict the relationship between X and Y, given these points (see below). In this case, the line is upward-sloping suggesting a positive relationship. A negative relationship can be depicted as a downward-sloping line. Figure 2.7 How can we determine which is the best line? Statisticians have defined the best line as the line that minimises the sum of squared errors. The error is the vertical distance of each individual point to the regression line. Statistical software can single out the one line for which the sum of these squared distances is smallest. Since the sum of squared errors is central in this type of regression, it is also termed Ordinary Least Squares (OLS). 24

31 Chapter 2: Introduction to quantitative methodology Figure 2.8 The graph illustrates how a regression line is derived. You may remember that lines are usually described using a slope, a, and an intercept, b: y = ax + b Analogously, the regression equation is: y = α + β x + ε This simply says that each point in our graph can be described by the regression line (with an intercept, α, and a slope, β) plus the error, ε (the vertical deviation of a given point from the regression line). We call y the dependent variable or outcome variable, and x the independent variable or explanatory variable. β is a regression coefficient (or slope coefficient). Note that a positive β indicates a positive relationship between x and y, whereas a negative β indicates a negative relationship. No linear relationship at all would mean β = 0. In this equation, there is one explanatory variable. In practice, we will mostly see regressions with more than one explanatory variable, resulting in a multiple linear regression equation: y = α + β 1 x 1 + β 2 x 2 + β 3 x β k x k + ε Activity On this website you will see a regression graph. Add, delete or move points in the graph and see how the regression line changes. In particular, try the following: When the regression model opens, you see an example of a negative correlation. Add points to turn it into a positive correlation. Click on show residuals to see how large the error for each point is. Clear all points by clicking on clear all points. Add two points that display a clear positive relationship. Now add one more point that does not confirm this relationship (having a large error; being very far from the line). Observe the impact of adding this one point on the position of the regression line. Clear all points. Add twenty points that display a clear positive relationship. Again add one more point that does not confirm this relationship. How does the impact on the position of the regression compare to the previously observed impact? 25

32 169 Economic policy analysis in international development Clear all points and temporarily remove the regression line by clicking on no regression line. Attempt to add twenty-five points that are completely randomly distributed over the graph. Bring back the regression line by clicking on show regression line. Did you succeed in creating no correlation between x and y whatsoever? Clear all points and temporarily remove the regression line. Add twenty points that have a U-shaped relationship: high levels of y for both very low and very high levels of x and low levels of y for medium levels of x. Perhaps this represents the level of health care an individual needs at various ages: high levels of health care as a more vulnerable child or senior and low levels of health care as a resilient adult. Bring back the regression line. How well do you feel that the regression line reflects the relationship created? At the bottom of the graph, you can see the regression equation that describes the regression line you can create by adding, deleting and moving points. The values of and when the regression model opens are α=70.08 and β= respectively. Try the following, clicking clear points each time you start a new item: Create a line with a large negative value of β. Create a line with a small negative value of β. Create a line with a positive value of β. Attempt to create a line with β = 0. Attempt to create a line with α = 0. Returning to the one explanatory variable model, we reformulate our question: how likely is it that we observe a given relationship between X and Y in our sample when in reality, in the whole population, X and Y are unrelated? Alternatively, if we estimate a slope coefficient to be 5.46, we want to know, how likely is it that we could derive a slope estimate of 5.46 from a sample when the true underlying relationship (true slope coefficient) in the population was 0?' Or, more simply: How likely is it that we observe a given β^ (where β^ ( beta hat )is our point estimate of the true β), whereas in reality β=0? To distinguish the relationship between X and Y observed in a sample from the true, unknown relationship, we call the observed relationship β^. We now need to apply what we learned in previous sections. From hypothesis testing: We set H 0 : β=0. There is no linear relationship between X and Y. H 1 : β 0. There is a linear relationship between X and Y. From the central limit theorem we can analytically derive the shape of the sampling distribution of β^ and derive a t-statistic. We use the t-statistic to determine the level of statistical significance of our estimate. If we observe a β^ that is sufficiently unlikely given H 0, we reject H 0 and say the slope coefficient is statistically significantly different from 0. Note that we are using the term significantly very specifically here! The researcher can determine what is sufficiently unlikely his/herself by choosing a significance level. The sample distribution of β^ under H 0 : β = 0 can be visualised as in the graph below. 26

33 Chapter 2: Introduction to quantitative methodology 5% β = 0 2σ Figure 2.9 We omit technical details, but for β^ to be significant at the 5 per cent level the criteria listed below hold. If one of these is true, this automatically means they are all true, so in practice it is only necessary to check one: The chance of observing β^ when the true β=0 is smaller than 5 per cent. This 5 per cent is represented by the blue shaded area in the graph. The p-value is smaller than The observed β^ is at least two times the standard error away from 0. The distance between β^ and 0 (which is simply β^) divided by the sample standard error is also called the t-statistic. The t-statistic of β^ is bigger than 2 or smaller than 2. Note that the actual critical value associated with the 5 per cent level may be slightly smaller or slightly larger than 2, depending on the sample size. But we use 2 as a rule of thumb. Now, we have all the information we need to read a regression table. When reading a regression table, pay attention to these features: The dependent variable y is usually mentioned at the top or bottom of the table. All explanatory variables included in the regression (the different x s), are listed in the rows of the table. The resulting β^ s, or coefficients, are in the corresponding rows. In parentheses under these coefficients, we are provided with information on the significance of the coefficient. This can be provided in 3 forms, analogous to the three criteria that hold when β^ is significant. A p-value, indicating the chance that β^ is observed when β=0. A p-value lower than 5 per cent indicates significance at the 5 per cent level. A standard error. A coefficient that is larger than twice the standard error is significant at the 5 per cent level. A t-statistic. A t-statistic larger than 2 indicates significance at 5 per cent level. Often, the author indicates significance using asterisks (*). 27

34 169 Economic policy analysis in international development An example of a regression table is given below. The authors attempt to investigate the impact of independent media on public food distribution and calamity relief expenditure. Figure removed due to copyright restrictions These variables (newspaper circulation, turnout, political competition, and election dummy) are not included in regression 4 Figure 2.10 From: Besley, Timothy and Robert Burgess The political economy of government responsiveness. Theory and evidence from India. The Quarterly Journal of Economics 117(4), 2002, pp Used with kind permission. Endogeneity In the previous section we have said correlation does not imply causation. Although there may be a (significant) correlation between A and B, this does not necessarily mean that A has some causal effect on B. A number of endogeneity problems may inhibit us from drawing such a conclusion. We will discuss omitted variable bias and reversed causality. Before going into this, this section will give an introduction to endogeneity. 28

35 Chapter 2: Introduction to quantitative methodology Imagine we want to investigate the effect of a treatment on an outcome. Ideally, we would like all observations to be identical and to differ only in whether or not they get the treatment (as in panel A in the figure below). This way, we can have confidence that the observed effect on the outcome variable is due to the treatment, and not due to anything else. For example: We want to examine the effect of taking vitamins on health. We can administer vitamins to a group of genetically identical mice in a laboratory and observe any changes in their health. Figure 2.11 Please see the PDF on the VLE for a colour version of this Figure. However, ideal experiments like A will rarely occur in the social sciences. Most analyses must be done on non-experimental data. Observations are likely to be different across many characteristics, not only whether or not they get the treatment. We want to make sure that any effect we observe on the outcome variable is solely due to the treatment, not due to these other characteristics. One way to ensure this would be for the treatment to be distributed at random, which makes it likely that the observations receiving the treatment will not be systematically different from the observations not receiving the treatment (as in panel B). To go back to our example: we can bring in a group of participants, randomly assign half of them to take vitamins, give the rest a placebo and observe any health effects. Conducting an experiment as in B is not always possible. We may have to rely on observed data on treatment and outcome. It is possible that individuals who received the treatment are systematically different with respect to some characteristics, from those who did not receive the treatment, as in panel C. We cannot be completely sure that any observed effect on the outcome is due to the treatment, rather than these characteristics. It may also be difficult to observe what came first : the outcome or the treatment. Go back to our example once more: we observe the health status of a number of people and ask them about their vitamin use. People taking vitamins are on average healthier than people who don t. However, people choosing to take vitamins may be different with respect to a number of characteristics: they may have a healthier lifestyle in general, be better informed about health issues, etc. These characteristics may influence both their health status and their decision to take vitamins. We cannot conclude that taking vitamins improves health necessarily. 29

36 169 Economic policy analysis in international development These examples illustrate the difference between an exogenous and an endogenous variable: Variation in an exogenous variable is determined outside the model. Variation in an endogenous variable is determined within the model. When we allocate vitamins at random, variation in vitamin use is not determined by some process relevant to health. The variation can be considered exogenous. However, when we observe vitamin use, variation is a function of processes within the system (people s choice to take vitamins is related to all the other choices they make, their preferences), that may be relevant to health. This variation can be considered endogenous. One endogeneity problem is omitted variable bias. This section will explain what omitted variable bias is and how it can impact conclusions drawn from a regression. An omitted variable is correlated with both an explanatory variable and the outcome variable in a regression, but is itself not included in the regression. If such an omitted variable exists, the relationship between an explanatory variable and the outcome variable observed in a regression, may be different from the true relationship between these variables. This is called omitted variable bias. An example We run a regression with the number of babies born in a particular area as a dependent variable and the number of storks in that area as an explanatory variable. The regression shows that the number of storks is significantly correlated to the number of babies born. We may conclude that babies are delivered by storks. Schematically: Observed relationship Regression equation: #babies = α + β #storks + ε #storks #babies born Figure 2.12 You may already suspect that the observed relationship is different from the actual relationship. The omitted variable in this case may be whether or not the area is in the countryside. Countryside is correlated to the number of babies born, as people may prefer raising their children in a greener, safer environment. We can also expect storks to prefer to live in the countryside. Hence, the omitted variable causes variation in both the dependent and independent variables: areas in the countryside have both a higher number of storks and a higher number of children born. There is no actual causal relationship between the number of children born and the number of storks. Schematically: 30

37 Chapter 2: Introduction to quantitative methodology Actual relationship Regression equation: #babies = α + β #storks + ε #storks #babies born More storks in countryside Families move to countryside Countryside Figure 2.13 The previous example is an example of upward bias: because of the omitted variable, the relationship between the modelled explanatory variable and the outcome variable appears more strongly positive than the true relationship. An omitted variable can also result in downward bias: because of the omitted variable, the relationship between the explanatory variable and the outcome variable appears more strongly negative than it is in reality. Schematically: Downward bias Regression equation: y = α + β x 1 + ε X 1 Y + X 2 Figure 2.14 An example We want to investigate whether government financial support to schools increases pupil performance. We find a negative relationship between the amount of money a school received from the government and the average performance of its pupils. We may conclude that the more money a school receives from the government, the worse its students do. However, an omitted variable may be school wealth. The government may target financially needy schools, so wealthier schools are less likely to 31

38 169 Economic policy analysis in international development receive government support. Pupils in wealthier schools can be expected to perform better, as the school can afford better teachers, materials etc. In this case, the omitted variable makes the impact of government spending on pupil performance appear more negative than it is. In reality, government spending may still improve student performance, but the effect of the omitted variable may be stronger, and the regression coefficient may have a negative sign. Reversed causality is another problem associated with endogeneity. Reversed causality occurs when variation in the outcome variable causes variation in the explanatory variable, rather than the other way round. Reversed causality can make an explanatory variable appear a stronger cause of the outcome variable than it is in reality. For example, we may observe a positive correlation between the number of policemen and the crime rate. This may mean that policemen cause people to commit crimes. It may also mean that in cities with a higher crime rate, more policemen get hired (which sounds more likely). Activity Consider the following explanatory and outcome variables. For each of these cases, answer the following questions: What is a possible omitted variable? Which mechanisms connect this omitted variable to both the explanatory and outcome variable? What is the direction of the bias (upward or downward)? Explanatory variable A microfinance programme Trade openness Taking a statistics course Educational attainment Distance of individuals' houses to the town centre Farm size Outcome variable Poverty GDP growth Knowing that correlation does not equal causation Earnings Expenses made for travel ( ) Farm productivity Table 2.3 Control variables, dummy variables and interaction terms This section will cover different types of variables that are commonly used in regression analysis, and their interpretation. Including a control variable in a regression is one way to attempt to deal with omitted variable bias. Take the following example: we are interested in the question of whether a lower student-teacher ratio (str) improves students performance. To assess this, we regress average test scores of students in a district on the average student teacher ratio: test score = α + β 1 str + ε The regression shows a negative relationship between the average studentteacher ratio in a district and the average test scores. Students in districts with lower student-teacher ratios achieved higher test scores on average. 32

39 Chapter 2: Introduction to quantitative methodology We may suspect that this result is biased downward, as we can think of several potential omitted variables: Perhaps district wealth is both related to higher test scores (students in richer districts are in a more conducive environment to learn) and a lower student-teacher ratio (schools in richer districts can afford to hire more teachers). Perhaps the value that parents put on education is correlated with test scores (parents valuing education more require their children to work harder in school) and with student-teacher ratios (parents putting high value on education push the school to hire more teachers). To address the first problem, we include district wealth in our regression: we control for district wealth. test score = α + β 1 str + β 2 wealth + ε You can think about this as artificially holding wealth across districts constant; we compare student-teacher ratios and test scores of districts that artificially have the same wealth. If our suspicion that district wealth is an omitted variable was correct, we would see that ^ β 1 increases (it becomes less negative ) compared to our regression without a control variable. We expect that omitting district wealth makes the relationship between student-teacher ratios and student test scores seem more negative than it is in reality and controlling for district wealth should bring ^ β 1 closer to its true value. Variables that you include in a regression to exclude a source of omitted variable bias, while not being primarily interested in its causal effect on the outcome variable, are sometimes called control variables. The distinction between explanatory variable and control variable, however, is very loose and is not always made. The name control variable has to do with the function of the variable in relation to the purpose of the regression analysis, not because there is anything intrinsically different about them compared to other variables. Including control variables cannot solve all omitted variable problems. Consider our second potential omitted variable: parents valuation of education. This is very hard to observe, let alone to measure and to include in a regression. Some sources of omitted variable bias can be dealt with by including control variables, others cannot. Some variables we may want to include in a regression are dichotomous: they can take only one of two values: a person s gender (male/female) whether a person has a college degree (yes/no) whether a person has been to jail (yes/no) This information could be captured using a dummy variable. A dummy variable is a variable that can take on only one of two values: coded either zero or one. We may, for example, be interested in investigating the effect of education on wage. We suspect that whether an individual is male or female also has an impact on wage. We have a database with information on individuals, including their level of education, gender and the wage they earn. We construct a dummy variable ( male ) equalling 1 if a person is male and 0 if a person is not. Ignoring omitted variable problems for now, we estimate the following regression: 33

40 169 Economic policy analysis in international development wage = α + β 1 years of schooling + β 2 male + ε We can interpret β 2 as the effect of being male on wage. β 1 can be interpreted as the effect of schooling, controlling for gender. Graphically, including the dummy variable allows for two regression lines representing the relationship between schooling and wage, one for men and one for women, with different intercepts (but the same slope, β 1 ). In the graph below: α captures the wage of a woman without schooling. α + β 2 captures the wage of a man without schooling. For both genders, the effect of schooling on wage is captured by β 1. The wage of a man with one year of schooling, is reflected by α + β 2 + β 1 * 1. The wage of a woman with one year of schooling is reflected by α + β 1 * 1. Figure 2.15 Note that we did not include dummy variables for being female and male plus an intercept term in our regression. Doing so would violate one of the basic assumptions of OLS regression. It would become impossible to estimate the regression equation. All you have to remember about this is that a researcher will omit one category of the dummy variable from the equation if an intercept term is included. In our example, we assumed that the effect of education is the same for men and women. We did not consider the possibility that the effect of education on wages might be different for men and women. Say we expect that one year of education will steeply increase a man s wage, whilst it has a weaker effect on a woman s wage. An interaction term allows us to incorporate this idea into a regression. An interaction term consists of two explanatory variables multiplied together. A regression with interaction term can take this form: y = α + β 1 x 1 + β 2 x 2 + β 3 (x 1 * x 2 ) + ε In this equation, β 1 captures the effect of x 1, β 2 captures the effect of x 2 and β 3 captures the additional effect of the combination of x 1 and x 2. For our example, the regression equation would be: wage = α + β 1 years of schooling + β 2 male + β 3 (years of schooling * male) + ε 34

41 Chapter 2: Introduction to quantitative methodology where β 3 captures the additional effect of being male and having a certain amount of schooling on wage. Graphically, an interaction term allows for different slopes (as well as intercepts) of the regression line for different categories of observations. Remember that a dummy variable allows for different intercepts only. Figure 2.16 In the graph above: α captures the wage of a woman without schooling. The effect of schooling on wage for women is captured by β 1. As before, the wage of a woman with one year of schooling, is reflected by α + β * 1. 1 α + β 2 captures the wage of a man without schooling. The additional effect on wage of the combination of having schooling and being male is captured by β 3. The wage of a man with one year of schooling, is reflected by α + β 2 + β 1 * 1 + β 3 * 1. Cross-section, time series and panel data This section will consider three kinds of data that could be used to run a regression: cross-section, time series and panel data. Different types of data may be more or less suitable to answer different research questions. Cross-section data contains information on different cross-section units at the same point in time (or an average over time). Cross-section units may be countries, people or any unit of observation arrayed across space. The cross-section dataset only contains information on different values of the cross-section units, and no information on variation within these cross-section units over time. Thus, cross-section data may contain characteristics (wealth for example) of different countries, provinces, individuals, households, etc. at a single point in time or an average over a single time period. It may give information about the differences in wealth between countries or households, but no information on how the wealth of individual countries or households developed over time. 35

42 169 Economic policy analysis in international development An example Imagine that we want to investigate whether a lower national tax rate attracts more investment from abroad. Cross-section data we could use to try to answer this question could look as follows: Tax rates and FDI inflows in 1998 Country Tax rate (%) FDI inflow (million$) Argentina China France Nigeria United States Zambia Table 2.4 Time series data contains information on a single unit of analysis at different points in time. It contains variation over time only. Thus, time series data may contain information on how the wealth of an individual country or household developed over time. We could use only one time series, using information about past wealth in a country to predict future development of its wealth. Alternatively, we could investigate the relationship between multiple time series, for example research the relationship between changes in trade policy and changes in wealth in one country. Returning to the earlier example, investigating the potential relationship between taxes and foreign investment, time series data could look like this: Tax rate and FDI inflow in China Year Tax rate (%) FDI inflow (million$) Table 2.5 Panel data contains information on different cross-sectional units at different points in time. In other words, it combines cross-section and time series data. Panel data contains both variation between units of observation and within units of observation over time. Thus, panel data may contain information on how the wealth of an individual country or household developed over time, and information on differences in wealth between various countries or households. 36

43 Chapter 2: Introduction to quantitative methodology Panel data that could be used to do our example research into the relationship between taxes and FDI could look like this: Tax rates and FDI inflows Country Year Tax rate (%) FDI inflow (million$) China China China Nigeria Nigeria Nigeria Zambia Zambia Zambia Table 2.6 We have identified three kinds of data: cross-section, time series and panel data. None of these types of data is intrinsically better than others. However, when we have a particular research question in mind, a specific type of data may be better suited to answer it. Here is a (non-exhaustive) list of examples: Cross-section data can be useful when the research question concerns differences between observations, for example in the following cases: When one or more of the variables we wish to research changes very little over time. Do countries with a legal system of Anglo-Saxon origin provide better protection for investors than countries with a legal system of a different origin? In a country with low household mobility: do households that live in an area with TV reception have a different attitude towards the ruling government than households in areas without such reception? Even when we do expect the variables of interest to change over time, data may only be available at one point in time. A research question that concerns only one cross-section unit (there is only one world price of oil, exchange rate between the euro and the dollar, Dow Jones index) will necessarily use time series data, for example: How does the world price of grain respond to changes in the production of bio diesel? Given past US interest rates and other variables, what is the most likely US interest rate one month from now? Panel data can be useful when the research question concerns both differences between entities and over time, for example: When we want to investigate the effect of some treatment. Do women that participate in a microfinance scheme experience a greater increase in their sense of empowerment over time than nonparticipating women? 37

44 169 Economic policy analysis in international development In the context of panel data, you may come across the term fixed effects. To understand this concept, let s go back once more to our example, investigating whether low tax rates attract foreign investment with a panel data set. There are a number of possible omitted variables that might bias the analysis if we run a regression with FDI as a dependent variable and tax rates as an independent variable. Perhaps countries with a low tax rate also have other policies that attract investors. Or perhaps some attribute of Chinese culture both favours low taxes and draws investors. We could introduce numerous control variables to address the omitted variable problems. However, there could still be something about China we do not observe or that we cannot measure, that may make it both prone to have a low interest rate and high levels of FDI. To (partially) address omitted variable problem(s), we can introduce country-fixed effects. We add a dummy variable that equals one if the observation concerns China and zero otherwise. We construct such dummy variables for all countries in the sample, which will make our data look like this: Tax rates and FDI inflows Country Year Tax rate (%) FDI inflow (million$) Dummy China Dummy Nigeria Dummy Zambia China China China Nigeria Nigeria Nigeria Zambia Zambia Zambia Table 2.7 You can think of the variable dummy China as a control for being China. It controls for both observed and unobserved characteristics of China that do not change over time. In our example, fixed effects might control for some attribute of Chinese culture, if we expect this to be stable over the research period. Including fixed effects would not adequately control for GDP differences between research countries: we would expect GDP to change substantially over time. When estimating a model using panel data and including fixed effects, it includes dummy variables for each individual entity, controlling for characteristics of these entities that do not change over time. We can thus introduce country-fixed effects, province-fixed effects, householdfixed effects, individual-fixed effects, etc. depending on the data at hand. Another type of fixed effects we can introduce is time-fixed effects. When including time-fixed effects, one includes a dummy for each relevant time period (a year dummy, or decade dummy). This controls for characteristics of a period that do not vary between cross-section units. 38

45 Chapter 2: Introduction to quantitative methodology Instrumental variable (IV) regression Problems of endogeneity make it hard for researchers clearly to identify causal relationships. Using instrumental variable regression can potentially solve, or at least mitigate, problems of endogeneity and identify causal relationships. The challenge is to find a truly valid instrument. This section will provide an introduction to instrumental variable regression. Consider a regression equation: y = β 0 + β 1 x + ε We suspect that x is an endogenous variable. We can thus not interpret β 1 as the causal effect of x on y. Attempting to solve this problem, we search for a valid instrumental variable or instrument. Let us call our instrument z. For z to be a valid instrument, it has to be exogenous and satisfy two conditions: Relevance condition: z is correlated with x. Exclusion restriction: z is not related to y, except through its correlation with x. Using the instrument z we can extract out some exogenous variation in x. If we then use this extracted exogenous variation in x to estimate the effect of x on y, we can measure the unbiased causal effect of x on y. This is all rather abstract, so let us consider an example (after Levitt 1997, further reading). We want to assess the effect of police presence in a city on crime. Variation in police presence, however, is endogenous. Reversed causality is an especially pressing concern: in cities that have a high crime rate, more police officers get hired. Running a regression that looks like this: crime rate = α + β # police officers + ε is unlikely to give us unbiased estimates of the effect of police presence on crime. Now say that we think that a dummy indicating whether a given year is an election year is a valid instrument for the number of police officers. Would this satisfy our two conditions? Is being in an election year correlated to the number of police officers? Maybe, if we expect that officials aiming to get re-elected would hire more police officers to satisfy voters. It is easy to check this condition: we can run a regression with the number of policy officers as a dependent variable and the election year dummy as an explanatory variable and see if the two are significantly correlated. Is being in an election year unrelated to crime except for its effect on crime through the number of police officers? It may seem reasonable to believe that an election year does not have an independent effect on crime. There is no way to prove this definitively, however, and we have to use our common sense to assess whether the exclusion restriction is satisfied. For example, if we believe that there are more protests in an election year, which may be related to crime (throwing rocks at the police, smashing cars or shop windows), our instrument is not valid. Equally, if we think that an increase in crime will lead citizens to force officials out of office, leading to new elections, our instrument is not valid. In the USA (where the study was done) municipal elections occur 39

46 169 Economic policy analysis in international development on a fixed schedule, and not too many riots occur surrounding mayoral races. So we are probably okay. Authors using instrumental variables often spend a large part of their paper arguing that their instrument is indeed exogenous and trying to disprove counterarguments like the ones above. If both conditions are met, we use our instrument to isolate exogenous variation in police presence. In this case, exogenous variation is change in police presence as a result of being in an election year. Variation in police presence due to other factors may be endogenous to the model. We would like to filter out the first type of variation. Schematically: Figure 2.17 Note that an election year does not have to be the only, or even the most important determinant of the number of police officers. It just has to be a source of exogenous variation. How does this work more technically? Instrumental variable estimation consists of two stages: the First Stage Regression and the Second Stage Regression, together known as an IV regression or often a Two-Stage Least Squares (2SLS) regression. In the first stage, we regress the number of police officers (our endogenous variable) on a dummy for election year (our instrument): # police officers = α + β election year dummy + ε If our instrument satisfies the relevance condition, the election year dummy and the number of police officers will be significantly correlated. We now know the size of the effect of being an election year on the number of police officers. We also know which year in our dataset was an election year. We can use this information to predict how many police officers there are in any given year. In this case: α for a non-election year and α + β for an election year. Call this predicted number of police officers # of ^. police officers Because # of ^ police officers only contains variation in police presence due to the election year, we have isolated some exogenous variation in police presence. 40

47 Chapter 2: Introduction to quantitative methodology Now we use this exogenous variation in police presence to estimate the IV regression. We estimate: crime rate = α + β # of ^ police officers + ε Note that we use the predicted number of police officers, not the actual number of police officers (which is endogenous). If the exclusion restriction is satisfied, this should allow us to estimate the causal effect of police presence on crime. In terms of the abstract model, the two stages of a two-stage least squares model are: Stage 1: x = α + β z + ε. We use this regression to construct predicted x or, isolating exogenous variation in x. Stage 2: y = α + βx^ + ε. Estimating the causal effect of x on y. To sum up, the instrumental variable approach can potentially solve endogeneity issues due to reverse causality and establish a causal relationship between two variables. We must be convinced, however, that the instrument is exogenous, reasonably correlated with the original (endogenous) variable, and that the exclusion restriction holds. Assessing these assumptions is key to assessing the credibility of a paper using instrumental variable estimation. The most important tool for doing so is your common sense! Difference-in-difference estimation Difference-in-difference estimation may be employed when we want to investigate the effect of some treatment on an outcome and we have information both over time and across space (or cross-section units) for both the treated and the untreated. In the context of development, this is often the effect of a particular aid programme (providing schools with textbooks, providing microcredit, etc.) on the well-being of the poor (pupil performance, income). Call the cross-sectional units receiving the treatment the treated group. We often know the score of the treated group on the outcome variable before and after the treatment. For example, we know that pupils on average scored 210 points on a test before they received schoolbooks and 220 points on average after. The difference between pupil s scores before and after they received schoolbooks is 10 points. We might say that the schoolbook programme caused the pupils scores on tests to increase by 10 points. However, we would be wrong in drawing such a causal conclusion: maybe something else changed during the time that the programme was running, causing performance of pupils to go up irrespective of the programme. Maybe the government has increased its budget on education and all children in the country (those receiving books and those not receiving books through the programme) score better on the test. To draw credible causal conclusions, we need a control group. Ideally, the treatment and control group should have identical characteristics and differ only in whether or not they receive the treatment. The control group is our counterfactual: we believe that, in absence of the treatment, the outcome variable for the treatment and control group would follow the same trend. Any difference between the trends of the control group and the treatment group can then be attributed to the treatment. 41

48 169 Economic policy analysis in international development In terms of our example, say that the schoolbook programme has been randomised. A group of potential recipient schools was selected, a lottery was held and only half of these schools received books. Pupils in the schools not receiving books form our control group. We also measure the performance of control group pupils before and after the programme: they score on average 220 points before and 228 points after the programme is executed. The difference between control group pupils scores before and after programme execution is 8 points. The difference between the differences in test scores in the control and treatment group before and after the programme is 2 (10 8). (You may understand now why this method is called difference-in-difference estimation.) We conclude that the programme caused pupils scores to rise by 2 points. For this conclusion to hold, we must credibly argue that the trends in test scores would be the same for the pupils receiving books and those not receiving books, in absence of the programme. More technically, a difference-in-difference regression set-up will look like this: Outcome ti = β 1 + β 2 treatment i + β 3 post t + β 4 (treatment * post) ti + ε ti We observe the outcome variable for an entity (i) in two time periods (t), before and after the treatment could have been received. Treatment is a dummy indicating whether an individual entity has received the treatment and post is a dummy indicating whether the observation is made before or after the treatment period. Graphically, we could represent the interpretation of the coefficients in the following series of graphs: Figure 2.18 Note that the regression equation includes a dummy variable (treated). This allows for two separate regression lines, each with their own intercept. β 1 captures the effect of being in the non-treated group. It is the intercept term for the non-treated group. 42

49 Chapter 2: Introduction to quantitative methodology Figure 2.19 β 3 captures the effect of being in the post-treatment period. In absence of the treatment, the treatment and the control group should display the same trend in outcome. This allows us to draw a (hypothetical) line showing how outcome in the treatment group would have developed, in absence of the treatment. Figure 2.20 β 2 captures the effect of being in the treated group. It is the intercept for the regression line for the treatment group, minus the intercept of the regression line for the control group (β 1 ). 43

50 169 Economic policy analysis in international development Figure 2.21 Finally, β 4 is our variable of interest. Note that this is the coefficient on an interaction term. It captures the additional effect of the combination of being in the treatment group and being in the post-treatment period. In other words, it captures the effect of having received the treatment. Summary Difference-in-difference estimation is an approach to estimate the causal effect of a policy on a treatment group. Its credibility, however, depends entirely on the validity of the key assumption and chosen control group: In the absence of treatment the difference in outcomes between treatment and control group would not have changed. Is it reasonable to assume that this assumption is true? An important way of assessing the validity of this assumption is to use your common sense. The researcher can also supply evidence that suggests that the key assumption is valid, for example, if the researcher can show a graph of the trend in the outcome before the treatment. If trends look similar, we may be inclined to think the research strategy is valid. Furthermore, a researcher may show that treatment and control group do not differ significantly in characteristics that may be relevant to the research, using certain statistical tests. Now read Ray, Debraj Development Economics. (Chichester: Princeton University Press, 1998) Appendix 2, pp Ray treats some of the material above in Appendix 2. Generally, he does so in a more technical way. If it helps you to see the same thing explained using a different approach, please use the material in Ray. Again, if you do not fully understand the material yet, do not worry. You have just laid the basics of all the methodology you will need to learn for the course. It is only natural for you to need more practice to grasp it all. More practice will follow in each chapter from now on. 44

51 Chapter 2: Introduction to quantitative methodology A reminder of your learning outcomes Having completed this chapter, and the Essential reading and Activities, you should now be able to: recognise when the topics introduced in this chapter are applied in subsequent chapters and reading. When you have completed the course you should be able to: read and critically assess empirical research employing the aspects of quantitative methodology covered in this chapter. Sample examination questions There will be no examination questions exclusively testing your knowledge of methodology. However, if you look at the examination questions in the subsequent chapters, you will realise that being able to apply the material in this chapter to the topics in subsequent chapters is necessary to get a good score in the examination. 45

52 169 Economic policy analysis in international development Notes 46

53 Chapter 3: Economic growth: basic concepts, ideas and theories Chapter 3: Economic growth: basic concepts, ideas and theories Aims of the chapter In this chapter we will: (a) explain how economic growth is measured and discuss some associated advantages and disadvantages of different approaches; (b) examine the empirical relationship between economic growth, poverty and inequality, and (c) review the two primary theoretical models of economic growth showing how they can be used to shed light on observed patterns of development. Learning outcomes By the end of this chapter, and having completed the relevant reading and activities, you should be able to: explain how we conventionally measure economic growth and the problems associated with this illustrate some alternative ways of measuring growth explain potential relationships between economic growth and both poverty reduction and between/within country inequality discuss empirical evidence regarding these relationships roughly sketch the pattern of economic growth from the far distant past to today and understand the processes behind it know the crucial assumptions and implications of the following growth models: Harrod-Domar model Crucial assumptions: output is proportional to the stock of capital, investment is proportional to output. Implications: indefinite growth is possible by increasing the stock of capital. Model is unstable except under very particular parameterisations. Solow or neoclassical growth model Crucial assumptions: output is a function of capital, labour and technology, with diminishing returns to capital. Implications: long-run growth solely depends on the rate of technological progress; increase in capital stock, savings rate or oneoff increase in efficiency will only lead to a higher level of output, not to an increased long-term growth rate. Limited scope for policy. Conditional convergence, flow of capital to capital-scarce regions. Endogenous growth theory Crucial assumptions: human capital displays increasing returns at an aggregate level due to externalities. Ideas have non-rival characteristics. Implications: one-off increase in human capital can lead to permanent growth increases. Scope for policy intervention. Human capital flows to places where human capital is abundant. Poverty traps. discuss empirical evidence on how well both the neo-classical model and the endogenous growth model explain observed growth patterns. 47

54 169 Economic policy analysis in international development Essential reading Dollar, David and Aart Kraay Growth is good for the poor, Journal of Economic Growth 2002, 7(3), pp Galor, Oded and David N. Weil Population, technology and growth. From Malthusian stagnation to the demographic transition and beyond, The American Economic Review 2000, 90(4), pp Ray, Debraj. Development Economics. (Chichester: Princeton University Press, 1998). Paragraph 2.2.1, pp.10 16, Chapter 3, pp.47 97, Chapter 4, pp Further reading Banerjee, Abhijit V. and Esther Duflo Inequality and growth. What can the data say?, Journal of Economic Growth 2003, 8(3), pp Easterly, William The elusive quest for growth. Economists adventures and misadventures in the Tropics. (Cambridge MA: MIT Press, 2001). Fernald, John G. and Brent Neiman Measuring the miracle. Market imperfections and Asia s growth experience, FRB of San Francisco Working Paper 2006, No Henderson, Vernon J., Adam Storeygard and David N. Weil Measuring economic growth from outer space, NBER Working Paper No. w15199, Klenow, Peter J. and Andres Rodriguez-Clare The Neoclassical revival in growth economics. Has it gone too far?, NBER Macroeconomics Annual 1997, 12, pp Pritchett, Lant Divergence, big time, Journal of Economic Perspectives 1997, 11(3), pp Quah, Danny Some simple arithmetic on how income inequality and economic growth matter. Working Paper, Available at: edu/viewdoc/download?doi= &rep=rep1&type=pdf Sala-i-Martin The disturbing rise of global income inequality, NBER Working Paper No. w8904, 2002a. Young, Alwyn The African growth miracle, Working Paper, Available at: Young, Alwyn The tyranny of numbers. Confronting the statistical realities of the East Asian growth experience, The Quarterly Journal of Economics 1995, 110(3), pp Introduction Economic growth is at the core of development; for poor countries, no long-term, significant and sustained social progress can occur without it. As co-features of the dynamics of the income distribution, the evolution of poverty and inequality is inexorably linked to the rate of economic growth. Theoretically, academic economists have made considerable progress in the past 50 years in understanding some of the fundamental causes of economic growth. However, despite its centrality and the recent theoretical progress, in practice economists have very little idea how actually to achieve sustained high growth rates (hence the title of one of our core readings, The elusive quest for growth (Easterly 2001)). This chapter will introduce basic concepts, ideas and theories regarding growth. We will ask questions such as: What is economic growth? Why is growth important for development? We will introduce two main theories of economic growth and discuss the insights produced by each. We will then build upon these fundamental insights in further chapters as we explore more recent alternative theoretical frameworks and complementary empirical work.

55 Chapter 3: Economic growth: basic concepts, ideas and theories What is growth? Now read Further reading Ray, Debraj Development Economics. (Chichester: Princeton University Press, 1998). Paragraph 2.2.1, pp Henderson, Vernon J., Adam Storeygard and David N. Weil Measuring economic growth from outer space. NBER Working Paper No. w15199, Young, Alwyn The African growth miracle, Working Paper, Available at: When we say development we often think of economic growth. Indeed both terms are sometimes used interchangeably. However, we will see shortly that activities we do not generally think of as contributing to development (such as market-based child labour) do contribute to economic growth as we conventionally measure it, as growth of GDP. Thus, a good start of any course in development is to ask: What is economic growth theoretically, and how is this distinct both from how we measure growth empirically, and from broader conceptual notions of development? One component of economic development that development economists pay a lot of attention to is poverty reduction. Although we expect (and have seen in the past) that poverty will generally be reduced as the economy grows, the extent to which this occurs may vary considerably from country to country, and in some cases growth may occur without much poverty alleviation at all, or may actually be associated with increased poverty, so we will want to examine this relationship more closely. Another important and related question is this: What is the relationship between growth and inequality? Or, how important is growth for poverty reduction? This chapter and the associated readings will go some way towards answering these questions. The most common measure of economic growth is an increase in Gross Domestic Product (GDP). Nominal GDP is the market value of all final goods and services made within the borders of a country in a year. Nominal economic growth is thus the change from one year to the next in the value of all final goods and services; real economic growth is nominal economic growth adjusted for inflation. This approach to measuring economic wealth has the advantage that it is objective, not easily influenced by politicians, and can be consistently applied across countries and through time. However, the approach does have some limitations, and for some uses of GDP statistics it is important to be aware of these caveats. For example, when comparing the GDP figures for developing countries we want to keep in mind that: Companies in the formal sector are required to keep records of their income, sales and transactions. However, a significant part of economic activity in many developing countries takes place in the informal sector, made up of small, unlicensed firms and individuals who do not report their activity to the government and generally do not pay taxes. As this activity is underground it is hard to measure and often is very inaccurately estimated, or not included at all, in GDP estimates. GDP counts every transaction as positive, even if it is associated with misery: a child working, sales of weapons/drugs, a company receiving a contract to dump toxic waste, etc. 47

56 169 Economic policy analysis in international development GDP fails to take into account non-market activity, such as household work and childcare. GDP estimates do not take into account the reduction of finite stocks of resources or pollution of the environment. These problems are aggravated by the limited capacity of many developing countries to collect and report national statistics, increasing measurement error. Furthermore, to correct nominal GDP estimates for inflation to create real GDP measures, it is necessary to construct price indices based on a set bundle of consumption goods. For many poor countries it is difficult to ascertain what the appropriate bundle of goods would be, and in many countries these price indices are not created at all (see Young, 2009). So given these problems, are there any viable alternatives to the GDP statistics? Green accounting is one proposed approach in which researchers adjust GDP statistics to take into account the depletion of natural stocks and the negative costs of environmental damage of production. While these alternative figures can be very interesting, producing them requires a considerable number of assumptions and valuejudgments, and researchers can disagree about the underlying theoretical models used to generate some of the adjustments. These statistics thus lack the robust simplicity, universality and objectivity of the plain (but flawed) GDP and thus can be useful complements to, but probably not universal substitutes for, the standard national accounts. Other alternatives address the measurement problems of poor data availability and large informal sectors in developing countries. Young (2009) reports measures of real consumption based on the more easily available statistics on ownership of durable goods, the quality of housing, the health and mortality of children, the education of youth and the allocation of female time in the household. This indicates that sub-saharan living standards have, for the past two decades, been growing at more than three times the rate indicated in international data sets! Henderson et al. (2009) use changes in the amount of electrical lighting at night to measure economic growth from outer space. Take a look at the results presented in their tables to see that they conclude there are large underand over-estimations in official growth rates. Summary You are expected to be able to explain what GDP growth is, why measuring GDP is problematic, especially in developing country contexts, and how economic growth can be distinct from development. In addition, you should be able to identify alternative ways of measuring economic growth and contrast their outcomes to conventional GDP estimates. How important is growth? Now read Further reading Dollar, David and Aart Kraay Growth is good for the poor, Journal of Economic Growth 2002, 7(3): Banerjee, Abhijit V. and Esther Duflo A reassessment of the relationship between inequality and growth: Comment. Manuscript, MIT, Quah, Danny Some simple arithmetic on how income inequality and economic growth matter, Working Paper, Available at: edu/viewdoc/download?doi= &rep=rep1&type=pdf 48

57 Chapter 3: Economic growth: basic concepts, ideas and theories Sala-i-Martin The disturbing rise of global income inequality, NBER Working Paper No. w8904, 2002a. Even if measuring GDP was unproblematic, we are still left with the question of how important is economic growth for poverty reduction? Is it possible that the poor do not benefit from economic growth? Or perhaps the poor may benefit in absolute terms, but relatively less than the richer parts of society, increasing inequality. We can certainly think of situations in which certain groups of people have not benefited from growth, so the argument should be taken seriously. What does the empirical evidence on growth and inequality suggest? It is tempting to conclude that if income inequality increases with economic growth, growth is bad for the poor. However, the relationship is much more subtle than that. In particular, inequality within a country must increase at a sufficiently high rate relative to economic growth if the number of people in poverty is to increase; historically we rarely observe within-country inequality changing at the rates that would be required to observe increasing poverty given the historically observed growth rates. Time t=1 Time t=2 Time t=1 Time t=2 Figure 3.1 It is also important to distinguish between income inequality between countries and income inequality within a country. Consider the following examples. There are two countries: a rich one and a poor one. In these countries, there are six people that can be of four types: very rich (VR) rich (R), poor (P) and very poor (VP). For simplicity, assume that the same nominal income increase is required to jump from each group to the next. There are two time periods, in between which growth may occur at different rates in the rich and the poor country. In example 1, pro-poor growth occurs in the rich country: some people that were poor in t=1 are rich in t=2. There is no growth in the poor country. Even though growth is pro-poor in the sense that it has increased poor individuals incomes more than rich individuals, the inequality between the poor and the rich country increases, as the rich as a whole country grows richer and the poor country s income is stagnant. Note that the inequality within the rich country decreases in contrast. 47

58 169 Economic policy analysis in international development 48 The inequality within the poor country is unchanged. In example 2 the rich economy encounters a decline in GDP, that hits the poor individuals disproportionately. Again, the poor economy is stagnant. Even though GDP decline is anti-poor in the sense that it decreased poor individuals incomes more than rich individuals, the inequality between the poor and the rich country decreases, as the difference between the average income in the rich and the poor country is smaller in t=2 than in t=1. Within the rich country however, inequality increases. These examples do not tell us that growth will do any of the things depicted. However, they illustrate that we cannot draw the simple conclusion that growth is anti-poor/pro-poor if we observe increased/ decreased inequality between countries. It also tells us that both withinand between-country inequality can change (potentially in opposite directions). It would be interesting to see which has the largest influence on changes in global inequality. Activity Draw some diagrams like Figure 3.1 above (two countries, with six inhabitants each of four potential types, in two time periods in between which growth or decline occurs). Illustrate some of the following situations where pro- (anti-) poor growth is defined as growth disproportionately increasing the incomes of poor (rich) individuals respectively: Pro-poor growth decreasing both within- and between-country inequality. Pro-poor growth increasing within country inequality and decreasing between-country inequality. Anti-poor growth increasing within-country inequality and increasing between country inequality. Anti-poor growth increasing within-country inequality and decreasing betweencountry inequality. Can you show any other combinations of these? You are required to understand potential relationships between growth and poverty reduction and between- and within-country inequality. Now let us turn to the empirical evidence on these relationships. On the question of what determines global inequality, within-country or across-country inequality, there is very strong evidence from many authors that within-country income inequality is relatively stable over time. Thus much of the change in the global income distribution of individuals seems to be coming from changes in the relative distribution of income across countries (see Quah, 2001, for an example). On the question whether global inequalities have increased or decreased the evidence is very mixed: Quah (2001) has found a slight increase in global inequality. However, using a different methodological approach, Xavier Sala-i-Martin (2002a, 2002b) finds that global inequality has decreased. He also concludes that both poverty rates and absolute head counts of people in absolute poverty have declined significantly. When assessing these results it is important to remember that they are largely driven by what happens in China and India. Because these countries are so populous, when we look at the global distribution of income at the individual level, whatever happens in these two countries takes on enormous importance. Both of these countries have grown dramatically over the past decades, and although income inequality within each country (especially China) has increased, this increase has not been dramatic enough to cancel out the dramatic forward shift of the entire distribution.

59 Chapter 3: Economic growth: basic concepts, ideas and theories The paper by Dollar and Kraay in the reading list draws the following conclusions: The incomes of the poor increased, approximately one for one, with the overall growth in mean income. This can be interpreted as poor- neutral growth; the poor neither benefit nor suffer disproportionately from growth. (Note this is an average result; other studies since have found much more heterogeneity in the poverty elasticity of growth.) The incomes of the poor do not fall disproportionately during periods of crisis or negative overall growth. Aggregate GDP growth is not associated with increases in income inequality within countries. When you read the Dollar and Kraay paper, do not worry about the more technical aspects of the estimation techniques (primarily discussed on pages and which you can ignore). For example, they use an instrumental variables estimator in some equations, and run some regressions with levels data, others with differenced data, and one estimator that combines the two. You do not need to understand the difference between these estimators, and you do not need to understand the technical estimation method. We will study instrumental variables estimators in later papers, but for now there is no need to understand the technical methodology. Instead, focus on the basic specification of the regressions and the results. In order to explore the relationship between poverty and growth, they start with a basic estimating equation (1) on page 202 that looks like: y p = α + α y + ct 0 1 ct α X +μ + ε 2 ct c ct They show that this is equivalent to estimating equation (2), but we can still interpret the coefficient estimate α 1 in the same way in either model. Since equation (1) is more intuitive, we will stick with that for our purposes. The variable y p is the log of per capita income of the poor, and y is the log overall average income. The subscripts c and t correspond to countries and year, so y p means the log of the per capita income of the ct poor in country c in year t. The variable X is actually a vector of variables i.e. it corresponds to a set of control variables. The variable μ c corresponds to time invariant country characteristics that are not controlled for in the regression (and are thus part of the error term). In the original regression (1) in Table 3 (with no control variables X), it is because the presence of these country-specific time invariant characteristics (or country-fixed effects) in the errors could bias the estimates. Therefore in regressions (3) and (4) they use differences instead. What this means is that instead of looking y ct in each time period and comparing to the level of y p to see ct whether when y ct is higher (or lower) there is a systematic relationship with y p being higher (or lower), they look at the relationship between ct changes over time, or the correlation of (y ct y ct-1 ) and (y p ct yp ). ct-1 Because both y p and ct yp will incorporate any country specific time ct-1 invariant characteristic, by subtracting one from the other we will cancel out ( difference out ) these factors. Thus regressions (3) and (4) control for the country-fixed effects, or μ c. However, the authors are worried that by only looking at change over time within a country they will not have enough data to see an effect, so regression (5) combines the two approaches. In Table 4 the authors introduce some additional control variables, X ct, into the analysis. Each set of columns presents the coefficient estimates and corresponding t-statistic for a different set of control variables. For 47

60 169 Economic policy analysis in international development example, in the first regression they control for regional dummies. Thus, if there are systematic differences between average incomes and incomes of the poor across different regions that might have been driving the results from Table (3), these are now taken into account. In other words, the regression is now asking whether, within each region, you see a systematic relationship between average incomes and incomes of the poor. The coefficient estimate of interest is still α 1, the coefficient on the log of per capita GDP, and as we can see it is 0.905, which is statistically significantly different from zero (with a standard error of 0.094), but not statistically different from one (with a p-value of 0.313). On page 208 the authors further describe how to interpret the results for each set of control variables. At this point, you do not need to understand fully each of these regressions, but notice that they are exploring how robust their conclusions are to different ways of looking at the relationship. Although you do not need to understand all the equations in the remainder of the paper, do the following activity. Activity Choose one or two variables that interest you, such as trade (globalisation) or social spending. Locate the table in which the authors explore the relationship between your chosen variable and the incomes of the poor. What is the coefficient of interest? What does it mean? Interpret the coefficient in the following way: as [variable of interest] increases/decreases, the incomes of the poor increase/decrease. Increase [variable of interest] seems to be good/bad for the poor. The text of the article walks you through the results so you may use the discussion as a guide. Although we observe growth and income inequality both evolving over time, is there a causal relationship between them? Does increasing growth cause increased (or decreased) income inequality? As the Dollar and Kraay article discusses, the evidence here is very mixed. Some papers have found a negative relationship, but others have found a zero, nonlinear or even positive relationship between inequality and growth. Banerjee and Duflo (2000) come to the conclusion that there is no robust relationship between these two variables. While all these results should be interpreted with some caution as this is an ongoing research question, it is important to recognise that the data do not show evidence of the kinds of anti-poor effects of growth that some people fear. There seems to be no systematic relationship between growth and inequality over the time horizon of our recent data sets (i.e. 20 or 30 years). In absolute terms, there is solid evidence that in fastgrowing economies, such as India and China, millions have been lifted out of poverty. The evidence suggests that great scope for massive poverty alleviation lies in encouraging aggregate growth. It seems growth is indeed very important for development. Summary You are not expected to reproduce exact or numerical results from the outlined papers. However, you should be able to discuss the overall picture that emerges from empirical research on growth and inequality/poverty reduction. 48

61 Chapter 3: Economic growth: basic concepts, ideas and theories Economic growth in the far distant past Essential reading Galor, Oded and David N. Weil Population, technology and growth. From Malthusian stagnation to the demographic transition and beyond, The American Economic Review (4), pp We have seen that when we speak of development, a lot of attention is paid to economic growth per capita. Over the past decades, per capita growth has been the norm to the extent that when developing countries fail to grow, we start to wonder what they are doing wrong. However, it is important to realise that per capita economic growth is a very new phenomenon if we take into account the far distant past. As you can see in the table on page 4 of Galor and Weil s article, per capita income in Western Europe did not grow at all in the period and only after 1820 did per capita growth exceed 1 per cent. When reading the article, it is important to realise that people s living standards (per capita GDP) can only increase when population growth is smaller than the growth of total output. The total wealth in the economy may increase, but if this increase is matched or outpaced by population growth, wealth per person will remain equal or decrease. Galor and Weil distinguish three stages of growth and reconcile them in their model. You should be able to describe this pattern of economic growth and understand the processes behind it. Stage 1: Malthusian growth Some factor of production is fixed (it cannot increase). Usually this is land. This implies decreasing returns to scale for all other factors of production. Any increase in the standard of living will lead to increased population growth. Any increase in the available resources (more land, increases in the quality of technology) will be offset by an increase in population. The net effect is that per capita income does not increase. Increases in the availability of resources are rare in this stage and population growth is very slow. A testable hypothesis that follows is that differences in technology across countries will be reflected in different population densities but not in different living standards. There is historical archival and archaeological evidence that the Malthusian model was consistent with reality for a long time: per capita GDP did not grow for long periods of time, population growth was very small, increased living standards were strongly related to fertility and differences in living standards between countries were relatively small prior to Stage 2: Post-malthusian growth As the population slowly grows, the rate of technological progress increases (a standard result of the endogenous growth model we will encounter shortly). This progress translates into a faster increase in output compared to Stage 1. Population growth speeds up as a result, as in the previous stage. The crucial difference from Stage 1 is that total output now increases faster than the population and per capita income starts to increase. This increase is small by today s standards, estimated to be well under 1 per cent per year in Western Europe. 47

62 169 Economic policy analysis in international development Stage 3: Demographic transition Due to the faster population growth in the previous stage, technological progress continues at increased pace, as does growth in total output. However, in this stage, the rate of population growth begins to fall. The Malthusian link between output growth and population growth is broken and only now per capita income starts to increase at the levels that we consider normal today. At the same time, the level of education increases dramatically. As more educated children are more likely to invent new things as adults, technological progress (and per capita growth) continues. Why do people decide to have fewer, better-educated children in this stage? Galor and Weil argue that this is a logical choice in the face of technological progress. If technology (for example farming practices) is constant, farmers over many generations will have learned from experience (trial and error) how to farm efficiently. Children can learn how to do this by observing their parents and schooling will not necessarily lead to better farming outcomes. However, when farming technology changes quickly, there is little time to figure out how to farm by trial and error and the knowledge of parents is less relevant. Children now need to be able to analyse and evaluate production possibilities and there is a value in education. Modern growth theories: Harrod-Domar and Solow growth model Now read Further reading Ray, Debraj Development Economics. (Chichester: Princeton University Press, 1998) Chapter 3, pp Easterly, William The elusive quest for growth. Economists adventures and misadventures in the Tropics. (Cambridge MA: MIT Press, 2001) Chapter 2 and 3, pp Pritchett, Lant Divergence, big time, Journal of Economic Perspectives 1997, 11(3), pp We now move on to growth models that attempt to help us understand growth in modern times. You are required to be able to explain crucial assumptions and implications of these models. The first model Ray discusses is the Harrod-Domar model. This model (published by Eversy Domar in 1946) was originally intended to explain the relationships between short-term recessions and investment in the USA. It was not meant to be a model of long-run growth, nor was it meant to be applied in developing countries. Essentially, the idea is that output is proportional to the stock of capital. Some fixed fraction of output is invested and investment adds to the capital stock. This in turn will make output grow and the whole process starts over. The implication of this model is that you just have to increase the amount of physical capital in order to grow. Early development economists thus thought that insufficient physical capital was a binding constraint to growth in developing countries. The Harrod-Domar model was used to calculate the financing gap, the gap between actual investment in a developing country and the investment needed to attain some desired growth rate. This financing gap could then be filled with foreign aid. 48

63 Chapter 3: Economic growth: basic concepts, ideas and theories There is very little to no evidence that supports this model, as Easterly humorously illustrates with the following example of Zambia: given the amount of aid that Zambia has received over the last 40 years, the Harrod- Domar model predicts that the Zambian GDP would be over 40 times larger than it actually is. Even so, a number of the International Financial Institutions (IFIs) (such as the World Bank and the IMF) still occasionally use the model. The second model described in Ray is the Solow growth model, or Neoclassical growth model. Easterly (2001) provides a non-technical understanding of both the Harrod-Domar and Solow model, you may find useful to read. According to the Solow, or neoclassical, model, the total output of an economy is a function of the basic inputs, capital and labour. This function, called a production function, describes how capital and labour are transformed into output, and can be denoted: Y = A*F(K, L) Where Y is total output, K and L denote capital and labour respectively, and A is a technology constant. This neoclassical production function has two essential properties: If you double capital and labour, output will double. This is called constant returns to scale. Note however, that when doing so, the amount of output per worker and thus total wealth per worker remains the same. Also, the amount of capital per worker does not change. However, if you hold labour constant and only increase capital, eventually capital will display decreasing marginal returns. For example, imagine that you start out with 10 workers and no sewing machines (capital). If you add one sewing machine you will probably increase production by a lot. Adding further sewing machines will continue to increase total output, but at some point after labour is fully utilised, adding yet more sewing machines will not increase total output as much as earlier additions. Furthermore, some sewing machines will break down and there won t be enough people to fix them. Thus at some point additional sewing machines will not increase production any more at all. In other words at some level of capital, increasing the amount of capital per worker will give you a progressively smaller return in terms of production. This is called diminishing returns to capital. In order to increase the amount of capital in the economy, you have to invest using savings. Saving or investment (they are assumed to always be equal) in the simplest version of the Solow model is some proportion (say 20 per cent for example) of the total output. This assumption is made for convenience and can be relaxed. The main implications of the model are not changed by specifying another form of savings function. Another parameter to notice in the Solow model is technological progress, or the growth of A in the production function above. Growth in A causes the production possibilities frontier to expand so that it is possible to produce more and more goods (Y) using the same amount of capital and labour (K and L). However the Solow model says nothing about how technology ( A ) might increase. It is exogenous, a set parameter that can be arbitrarily set by a researcher, and is not endogenously determined within the model. 47

64 169 Economic policy analysis in international development Graphically, this is illustrated by Figure 3.2, below. Because it is difficult to graph three variables on 2-dimensional paper, we normalise by dividing both Y and K by L, denoting our new per capita variables with lower case letters. Thus on the horizontal axis is k, amount of capital per worker. The curve F(k), is the output per worker. We can see that this increases as the amount of capital per worker (k) increases. However, we can also see that it increases ever more slowly as k increases. These are the diminishing returns to capital. An increase in technology A will shift the entire curve upwards: more output per worker can be attained with the same amount of capital per worker. In the graph, we can also see savings or investment s*f(k). Some fixed proportion of output is saved, so this curve has approximately the same shape as F(k). The last curve represents depreciation d(k). Over time the value of capital goods decreases, things such as machines or computers are worn down or get outdated. So the capital stock per worker loses value at some fixed rate, hence d(k) is a straight line. Again, these assumptions about savings and depreciation rates are purely for convenience you can draw them in any shape you like in the graph below and the basic conclusion of the neoclassical model will still hold. In other words, the insights generated by the neoclassical growth model are invariant to the savings behaviour and depreciation rates of capital. Figure 3.2 Solow growth model What are those insights? What does the Solow model tell us about growth? We can see that at any point to the left of k* saving is higher than depreciation. Investment in that period is greater than depreciation, so the capital stock increases and in the next period capital per worker, k, is increased and thus we move along the F(k) curve and increase per capita output, y. However at points to the right of k*, depreciation is greater than investment, total capital stock falls in the next period and k decreases, as does per capita output, y. Only at k*, where savings and depreciation are equal, is k, the amount of capital, and hence output, per worker, stable. k* and y* are thus the equilibrium levels of capital and output per worker. If all the curves in the graph stay where they are, y and k will automatically return to these equilibrium levels. In other words per capita growth will eventually cease. 48

65 Chapter 3: Economic growth: basic concepts, ideas and theories Note that it is possible for per capita output to grow temporarily by increasing the savings rate, moving the sf(k) curve upwards. This would move point A to the right, and an equilibrium value of k* would be attained at a higher level of output per worker. In the period before this point is reached, growth will occur. However, it is important to see this is only temporary: once the economy has reached a new point like A, growth will stop. A new equilibrium, with higher levels of capital per worker and income per worker, will be reached. A one-off increase in efficiency of production will have a similar effect. It will shift the F(k) curve upwards and the equilibrium amount of capital per worker will result in higher output per worker. In other words point B will shift upwards. If this increase in efficiency is one-off, an equilibrium will again be reached and growth will only be temporary. However, if we picture technological progress as a continuous upward shift of the F(k) curve, it is possible to have growth in the long term. As technological progress will make it possible to attain ever more output per worker with the same amount of capital and workers, growth does not cease. In fact, as technological progress is the only source of permanent, long-term growth in this model, the long-run rate of growth will equal the rate of technological progress. The Solow model s primary insight was that capital accumulation alone cannot drive long-term per capita growth (unlike in the Harrod Domar model). Instead, you need technological progress for permanent growth. But after fingering technological growth as the key to sustained economic progress, the Solow model leaves this main factor in long-run economic growth unexplained. For policymakers, it follows from the above that unless their policy succeeds in increasing the rate of technological progress (and how this could be done is left unexplored), any policy will only result in a temporary increase in growth, not a permanent one. Scope for policy is thus relatively limited in the Solow model. Activity Draw a copy of the Solow graph and: Increase the savings rate. Convince yourself that a new equilibrium will be reached. What does this mean for growth? What does this mean for total output? Imagine a one-off large inflow of aid that increases the amount of capital per worker. Convince yourself that in the long run, the model will return to the same equilibrium. What does this mean for growth and total output? Draw a one-off efficiency increase. Convince yourself that a new equilibrium will be reached. What does this mean for growth? What does this mean for total output? Draw a number of new curves representing continuous technological progress. Convince yourself that output per worker keeps increasing. What does this mean for growth? The Solow model also has some implications for global differences in wealth. Imagine that the world consists of countries that are all in equilibrium. This does not mean that they are in the same equilibrium (that they all have the same production structure, hence the same graph and the same equilibrium values of y and k). It means that they are all in their own, country-specific equilibrium (they all have their own, country-specific graph and equilibrium values). We further assume that technological progress occurs at a global scale and thus that the rate of progress is the same for all countries. This would mean that all countries are growing at the same long-run rate, the rate of technological progress. 47

66 169 Economic policy analysis in international development 48 Now let s relax some of these assumptions and imagine that some countries, the rich ones, have already reached their equilibrium but the poor ones have not done so yet (they are still at some point to the left of k* in the graph). The rich countries grow only at the rate of technological progress and have the equilibrium amount of capital; they are at the technological frontier. Technological progress equally moves the technological frontier of poor countries outwards but in addition, these countries grow because they move towards the technological frontier. This implies that poor countries grow faster than rich ones and that the gap in wealth between them will shrink. As long as the underlying parameter values (such as the savings rate, the production function, etc.) of rich and poor countries are different they will not converge to the same level, but they will converge to a level that is similar to that of other similar countries. This is called conditional convergence. It does not mean that the rich-poor gap will disappear (absolute convergence). The equilibrium level of output specific to the poor countries might be lower than that of the rich ones. Ray (1998) gives an overview of empirical evidence on the convergence hypotheses. Another relevant implication of the Solow model follows from its decreasing returns to capital. If capital can flow freely across the world (i.e. if we increase globalisation and openness), we expect capital to go to the place where it earns the highest returns. The Solow model predicts that this will be in a country where there is initially comparatively little capital as returns to capital decrease when you have more of it. Thus the model predicts a flow of investment to regions that do not have a lot of capital yet, increasing the rate of the transitional growth towards equilibrium. Therefore the Solow model predicts that increased trade and openness will accelerate convergence (or conditional convergence), and thus medium term growth in poor countries, but will not permanently increase per capita growth. As you will see in the chapter on trade theory, however, empirical evidence does not support this prediction. You are expected to know the crucial assumptions and the implications of both the Harrod-Domar and Solow growth model. As emphasised, these are: Harrod-Domar Crucial assumptions: output is proportional to the stock of capital, investment is proportional to output. Implications: indefinite growth is possible by increasing the stock of capital. Solow Crucial assumptions: output is a function of capital, labour and technology, with diminishing returns to capital; saving and investment are proportional to output, constant rate of depreciation. Implications: long-run growth solely depends on the rate of technological progress; increase in stock of capital, savings rate or one-off increase in efficiency will only lead to a higher level of output, not to an increased growth rate. Limited scope for policy. Conditional convergence, flow of capital to capital-scarce regions. Modern growth theories: endogenous growth theory Now read Ray, Debraj Development Economics. (Chichester: Princeton University Press, 1998) Chapter 4, pp

67 Chapter 3: Economic growth: basic concepts, ideas and theories Further reading Fernald, John G. and Brent Neiman Measuring the miracle. Market imperfections and Asia s growth experience, FRB of San Francisco Working Paper 2006, No Klenow, Peter J. and Andres Rodriguez-Clare The Neoclassical revival in growth economics. Has it gone too far? NBER Macroeconomics Annual 1997, 12, pp Young, Alwyn The tyranny of numbers. Confronting the statistical realities of the East Asian growth experience, The Quarterly Journal of Economics 1995, 110(3), pp We have seen that the Solow model explains the growth effects of the accumulation of capital, but leaves changes in the rate of technological increase unexplained (or exogenous). Endogenous growth models, on the other hand, endogenise technological change, or make technological change an outcome of internal processes within the model. Endogenous growth models endeavour to illustrate the underlying mechanisms through which technological change is generated within an economy, just as the neoclassical models endeavoured to illustrate the underlying mechanisms behind the relationship between capital accumulation and growth. They each focus on a completely different aspect of the growth process, and by simplifying away the other complicating factors, give us a much deeper and clearer picture of the particular process under study. Our understanding of the whole growth process must incorporate insights from both of these types of models (and a few more as well, as we shall see later). Chapter 4 in Ray (1998) deals with endogenous growth models. You can ignore any equations, just focus on the following main ideas. There are many versions of endogenous growth models (as there are multiple versions of neoclassical models). Thus it is best to think of these as two families of models. Although the specifics vary, there are a few primary mechanisms that these models have in common, and in this course we will focus on one of the primary engines of endogenous growth: spillovers or externalities to knowledge. A certain good is characterised by a positive/negative externality if either the consumption, production or exchange of this good has positive or negative consequences for the wellbeing of a party not involved in the transaction. In the case of a positive consequence, we say that the benefits associated with the good spill over to a third party. Knowledge and ideas are usually considered to be characterised by externalities, as the below will illustrate. A key insight from endogenous growth theory (and reflected in the Galor and Weil story from above) is that ideas, or knowledge, have a large component that is non-rival. This means that if an idea is used in one production process, it is not diminished but can be used by others in their production as well. Sometimes firms or people will try to keep certain knowledge private, through patents or secret formulas, but this only partially or temporarily works, and if the ideas escape, their use by the original firm in no way diminishes their usefulness to subsequent users. In many, if not most cases, knowledge and ideas can flow freely across people. Thus knowledge is broadly a non-rival input. Take the example of producing a bike: if you allocate steel, labour etc. to producing a bike, they cannot be used to produce something else. They are rival inputs. However, the idea that a wheel with spokes is superior to a solid wheel is non-rival, i.e. it can be used in many bike-producing processes at the same time and indeed for simultaneously producing many other goods. Others are now also able to make superior bikes etc. without going through the trouble of reinventing the wheel. 47

68 169 Economic policy analysis in international development While ideas are mostly non-rival, human capital of a person, or the ability of the individual to assimilate and generate ideas, is rival. If you hire a scientist to work for your firm, she cannot work for another firm. Therefore she gains some private return (her wage) from her human capital (if her human capital were free for everyone to use, nobody would pay her for it!). However the ideas generated by the scientist may leak out to other firms. Furthermore, your scientist will engage socially and professionally with other people. Her interesting ideas, good feedback and analytically insightful social interactions will augment the ability of other people around her to come up with good ideas themselves. Thus although an individual s human capital is rival, it displays spillovers or externalities. Your scientist cannot herself capture all the benefits of her education (human capital) other people around her benefit too. In turn, she benefits from her interactions with other people of high human capital. Indeed, two heads are not only better than one, they are better than the sum of one head alone and another head alone. A key implication of the presence of rival human capital with externalities is that the private returns to human capital do not equal the benefits for society. For a young woman calculating how many years to go to school, she will observe that after some point, increasing the years of schooling will not increase her earnings (if you don t believe this, ask your professor!). So, for her there are diminishing returns to human capital accumulation and at some point it will not make economic sense to continue in school. However, the social benefit of her education will be greater than the increase in wages, because of the positive externalities discussed above. Her education makes her more productive (and thus earning higher wages), but it also increases the productivity of other people around her. Thus the optimal level of education for her will be lower than the optimal level of education that society would like her to have; this is one of the primary theoretical justifications for government subsidies to education. Furthermore, because of human capital spillovers, the gain in productivity (and hence the increased wage) she will get from an extra year of education will be greater if there are more educated people around her. In other words, at the aggregate level, the return to an additional unit of human capital (in this case an additional year of schooling) will be higher the higher the aggregate level of human capital is. This means that human capital displays increasing returns to scale at the aggregate level (although still diminishing returns at the individual level). Note this aggregate relationship is the exact opposite of that of physical capital in the neoclassical model. A number of interesting policy implications can be derived from these insights from endogenous growth models. Given the increasing returns to human capital at the aggregate level, a one-off increase in the stock of human capital can permanently increase the rate of growth. This is because an increase in the aggregate stock of human capital increases the return to further increments of human capital (i.e. schooling), which in turn increases investment in human capital, which increases the size of the aggregate stock of human capital, and so on This leaves far greater scope for big long-term impacts from policy intervention than were implied under the Solow growth model: in a world characterised by endogenous growth, a one-off government policy increasing the human capital stock can translate in permanent long-term per capita growth. Furthermore, because individuals on their own would invest less in human capital than the societal optimum, a subsidy to education will improve societal welfare as well as growth. 48

69 Chapter 3: Economic growth: basic concepts, ideas and theories Another implication of the model is that human capital is expected to flow to places where a lot of human capital is already available (the opposite of the Solow prediction for physical capital). This is indeed what we observe in the real world most educated people generally want to work/study in places with other educated people. Internationally, educated migrants tend to move from less educated countries to more educated places. Even within the same country, say the USA, more educated people are more likely to move to New York or Boston than to Goodland, Kansas or Gering, Nebraska. They might not understand the concept of externalities, but they do expect to earn more for their education in real terms. Another key insight (also mirrored in the Galor and Weil story) of endogenous growth theory is the importance of population size. With more people, there is a higher chance that someone will come up with a new (non-rival) idea. If there are more people around to use the new idea, it has greater benefit and it is also more likely that someone else will improve or add to that idea, and so on. So the bigger the population size, the faster the rate of idea generation. Also, to the extent that part of the idea can be held privately (by patents, for example) at least for some time, then the larger population implies greater returns to the individual for generating ideas, and increases the probability that ideas will be generated. This is an example of a virtuous circle. However, the flip side of this is that a vicious cycles (or poverty trap) can also occur when there are increasing returns. For example, if the total stock of human capital is low, the returns to investment in human capital (education) are also low, leading people to invest very little in human capital and keeping the total stock low. Summary You are again required to know the crucial assumptions and implications of this endogenous growth model: Crucial assumptions: human capital has increasing returns at an aggregate level, due to externalities. Implications: A one-off increase in human capital can lead to permanent per capita growth increase. Larger scope for policy intervention compared to the Solow growth model. Human capital flows to places where human capital is abundant, rather than to areas where capital is scarce (as the Solow model would predict). There is a possibility of vicious circles, where poorer countries are likely to grow more slowly than richer countries: poverty traps. Solow growth model versus endogenous growth theory The final question of this chapter is: What model does a better job of describing observed growth patterns, the Solow model or the endogenous growth model? Remember, it needn t be one or the other; they can both shed important insight into different components of the growth process. However, it is important to know, for example, whether recently observed growth miracles such as those that occurred in Korea and Taiwan from the 1960s to the 1990s can be explained primarily by neoclassical models i.e. by the increase in factor accumulation (increases in factor 47

70 169 Economic policy analysis in international development inputs of production like physical capital and labour), or by endogenous growth processes i.e. by increases in total factor productivity or technological innovation. Researchers investigate how much of the observed growth is due to increases in factor accumulation and how much is due to an increase in total factor productivity (TFP) using a technique called growth accounting. Essentially, the technique involves estimating the extent to which changes in factor inputs can explain changes in output, then any output growth that is not explained by factor accumulation is assumed to be due to an increase in TFP. Ray quotes a paper by Alwyn Young (1995), which concludes that factor accumulation alone largely explains East Asian growth. He concludes that neoclassical growth theory does a reasonable job explaining their growth miracle. However, others have debated this conclusion based on methodological grounds. Klenow and Rodriguez-Clare (1997) argue that increased productivity growth will increase returns to investment, increasing investment rates. The increased output due to this increase in investment should be attributed to TFP, not factor accumulation. They find that the explanatory power of factor accumulation decreases somewhat once this is taken into account. More recently Fernald and Neiman (2006) use an alternative approach to growth accounting that takes into account that firms face differential costs of factor inputs, due to varying levels of access to subsidised credit and other political wedges. When taking this into account, they even find negative growth of technology in those firms receiving the most subsidies. Note that the idea that not all firms are similar, that they are heterogeneous, is also sparking new ideas in growth theory that will be discussed in the next chapter. In any case, although the evidence seems to be tipping towards neoclassical models lately, research into the underlying explanations of the Asian Miracle is mixed and ongoing, with no definitive answers available yet. Summary You are expected to discuss empirical evidence on how well both the neoclassical model and the endogenous growth model explain observed growth patterns. A reminder of your learning outcomes Having completed this chapter, and the relevant reading and activities, you should now be able to: explain how we conventionally measure economic growth and the problems associated with this illustrate some alternative ways of measuring growth explain potential relationships between economic growth and both poverty reduction and between/within country inequality discuss empirical evidence regarding these relationships roughly sketch the pattern of economic growth from the far distant past to today and understand the processes behind it 48

71 Chapter 3: Economic growth: basic concepts, ideas and theories know the crucial assumptions and implications of the following growth models: Harrod-Domar model Crucial assumptions: output is proportional to the stock of capital, investment is proportional to output. Implications: indefinite growth is possible by increasing the stock of capital. Model is unstable except under very particular parameterisations. Solow or neoclassical growth model Crucial assumptions: output is a function of capital, labour and technology, with diminishing returns to capital. Implications: long-run growth solely depends on the rate of technological progress; increase in capital stock, savings rate or oneoff increase in efficiency will only lead to a higher level of output, not to an increased long-term growth rate. Limited scope for policy. Conditional convergence, flow of capital to capital-scarce regions. Endogenous growth theory Crucial assumptions: human capital displays increasing returns at an aggregate level due to externalities. Ideas have non-rival characteristics. Implications: one-off increase in human capital can lead to permanent growth increases. Scope for policy intervention. Human capital flows to places where human capital is abundant. Poverty traps. discuss empirical evidence on how well both the neoclassical model and the endogenous growth model explain observed growth patterns. Sample examination questions 1. Given the historical experience of developed countries, it is unrealistic to expect that all developing countries will be able to create growth in the short term. Discuss. 2. What is the difference between the role of human and physical capital accumulation in the Solow growth model and the neoclassical growth model respectively? 3. It is sometimes said that the rising tide [of economic growth] will lift all boats. Discuss this statement with reference to theoretical and empirical evidence on the impact of growth on inequality. 47

72 169 Economic policy analysis in international development Notes 66