The New Growth Evidence

Journal of Economic Literature Vol. XXXVII (March 1999),pp. 112-156

The New Growth Evidence

JONATHAN TEMPLE'

1.Introduction ultimately driven by technical advances, and it Gas widel; felt that these ad- HY HAVE SOME countries grown vances would be resistant to further rich while others remain poor? It W analysis. Only rarely did anyone con- is hard to think of a more fundamental sider that even these early models, with question for economists to answer. Ac- their simple assumptions about the cording to the definition used by the exogeneity of technical progress, might World Bank's 1996 World Development be used to help understand the wide Report, over 4.5 billion of the world's 5.6 international variation in levels of billion people live in developing coun- income and growth rates. tries, and so a better understanding of Thirty years later, macroeconomists what generates economic growth could returned to growth issues, spurred by make a huge contribution to human wel- the availability of the Summers-Heston fare. Yet macroeconomists have tradi- data set, and by the work of William tionally shown little interest in the gulf Baumol (1986), Robert Lucas (1988) between rich and poor. The study of and Paul Romer (1986). Although it was growth at the aggregate level has often new growth theory that initially drew been something of a backwater, rele- the most attention, the last ten years gated to a brief last chapter in main- have also seen an outpouring of empiri- stream textbooks, and rarely taken on by cal work intended to explain post-1960 anyone outside development economics. growth experiences. This work, the There are at least two reasons for this "new growth evidence," is the subject of state of affairs. One is that, until re- what follows. cently, lack of data made it hard to A common view of this research is compare income levels across a wide that, somewhat ironically, it has mostly range of countries. A second reason is demonstrated the explanatory power of more subtle, and turns on a common in- the original 1950s-style neoclassical terpretation of the early theoretical lit- models. I believe that this interpreta- erature. The theoretical work of the tion misses a great deal, and one aim of mid-1950s suggested that growth was this essay is to argue- for a different 1 Hertford College, and Institute of Economics view. Another aim, perhaps no less cen- and Statistics, Oxford University. I am grateful to tral, is to examine why this crucial area Philippe yhion, me Aron, Tony Atkinson, John McMi Ian, Jo n Muellbauer, Steve Nickell, is regarded

Steve Redding. and the referees for hel~fulcontri- with such widemreadI distrust. Em~iri- butions I duld es ecially like to thank John cal work on growth has often Pencavel, who PlayeSC%a major role in shaping the deen paper. The ESRC and Hertford College, Oxford and there is a wide- provided financial support. spread feeling that growth theory and 112 Temple: The New Growth Euidence 113 econometrics are best kept apart. Their (45)Why do growth rates differ over offspring, cross-country regressions, are long periods? not greatly loved by either parent. (Q6) What happens in the long run? It is not difficult to see why this might be the case. The early papers The first question covers such widely- present some easy targets, and basic discussed issues as whether poor coun- points are often misunderstood. How- tries are catching up with the rich. I will ever, some frequent criticisms are easily discuss possible answers in Section 2. To answered. Too many people have dis- place the discussion in context, I also missed the prospects for cross-country discuss the purpose and nature of the research, ignoring recent improvements Summers-Heston data set, and some and greater awareness of econometric subtleties in measuring output levels and issues. I will argue that we have already growth rates. learned some interesting things. In answering the other five questions, The focus of the paper is predomi- the new growth evidence makes use of a nantly the cross-country empirical work variety of methods, many of them con- carried out by macroeconomists. This troversial. Much of the first half of the work has three distinguishing features. survey is taken up by a discussion of the First, researchers have often tried to in- possible approaches. Section 3 contrasts tegrate developing and developed coun- cross-country growth regressions with tries in a single empirical framework; as the older methods, historical case stud- we shall see, this endeavour is not with- ies and growth accounting, and builds out its problems. Second, the research the case for using the cross-country makes intensive use of the cross-section variation in the data. Next, I turn to variation in growth rates and other vari- the many econometric problems with ables. Finally, the research questions growth regressions (Section 4) and to are often inspired at least in part by re- approaches that may overcome some of cent growth theory. There is a renewed these problems (Section 5). An under- emphasis on human capital, and to a lying argument will be that the use of lesser extent, research and develop- panels is often the best way forward, ment (R&D), as important variables in though not without problems of its explaining differing growth experiences. own. Work sharing these features has been The second half of the survey is more used to address a wide variety of directly concerned with what we have questions. Among the many issues of learned from this research. Section 6 interest, I choose to emphasize six: gathers together some of the issues often discussed under the heading "con- (Q1) How is the world income distri- vergence." It makes good sense to treat bution evolving? these issues together, as answers to one (42) Do countries converge to steady question are usually relevant to the state paths and, if so, how next. For instance, measurement of the quickly? rate of convergence is often used to assess the extent of diminishing returns, (43) How rapidly do returns to inputs which then feeds into the debate about like physical capital diminish? the role of technology differences, (Q4) Are poor countries poor mainly which then affects our view of growth because they lack inputs, or be- differentials and the likely nature of cause of technology differences? long-run growth. 114 Journal of Economic Literature, Vol. XXXVIl (March 1999)

For practical ends, the main question monetary unit using the relevant to answer is (45):why do growth rates exchange rates. differ over long periods? Our knowl- However, observant travellers know edge is incomplete, but a synthesis of that a given sum of, say, dollars will buy recent contributions is provided in very different amounts of goods in dif- Sections 7 and 8. Although growth different countries. In other words, there ferentials have been a central focus of are often large and systematic depar- the literature, it is clear that there is tures of exchange rates from "purchas- more to explain, and issues needing fur- ing power parity" (PPP). Instead of ther study are the subject of Section 9. using exchange rates, incomes should Finally, the conclusions give a personal be converted using special currency in- view of the most reliable findings. dexes which are calculated so that one One aspect of my approach should unit will purchase the same bundle of already be clear. Since the recent litera- goods across countries. ture has partly been spurred by theo- The calculation and use of these in- retical developments, where relevant I dexes, called purchasing power parities will relate findings to growth models. or PPPs, is essential for accurate cross- However, I will not take the attitude country comparisons of real incomes that distinguishing between specific and expenditures. To give some idea of growth models should be the main aim the difference this makes, consider a of the empirical literature. Anything comparison between the incomes of the that might contribute to better policy- US and India. If we convert Indian per making should be of interest, and the capita GDP into US dollars at the offi- coverage here reflects that. cial exchange rate, this tends to suggest that India's average income is just 2 2. Prelinainaries percent of the USA's. If we use PPPs, India's relative position is improved by In thinking about growth, it helps a factor of two and a half, so that aver- to begin by trying to establish some age income is in fact 5 percent of the stylized facts. This section draws on the USA's. It turns out that comparison us- work of many researchers to high- ing exchange rates tends to overstate light some of the most interesting regu- the magnitude of income disparities.2 larities in the data. One advantage of The construction of a world table of this approach is that important mea- national accounts with figures that are surement issues can be introduced in comparable across space and time, what should be a relatively painless based on PPPs, is clearly a difficult and way. research-intensive exercise. It relies on 2.1 lncome Disparities obtaining price data for a wide range of goods, and then devising suitable aggre- The most striking aspect of the world gation procedures to obtain a national income distribution is the very large PPP. The United Nations International and persistent disparities in per capita Comparison Project (ICP), launched in income. However, one has to be rather the late 1960s, was designed to make careful in making such comparisons. such comparisons possible. Over 90 The obvious method is to value each country's quantities of final goods and 2Another way of saying this is that price levels are higher in richer countries. For a very clear services at domestic prices, and then survey of possible explanations, see Kenneth convert these figures into a common Rogoff (1996). Temple: The New Growth Euidence

TABLE 1 GDP RELATIYE TO THE US, SELECTEDCOUNTRIES,1960 AND 1990

Proportion of US GDP per capita Population (at current prices) 1990 1960 1990 Zaire 36m 0.05 0.02 Nigeria 96m 0.05 0.05 India 850m 0.07 0.07 China 1134m 0.05 0.07 Bangladesh 108m 0.09 0.08 Pakistan 112m 0.07 0.08 The Philippines 61m 0.12 0.10 Indonesia 178m 0.06 0.11 Iran 56m 0.23 0.16 South Africa 38m 0.21 0.18 Thailand 56m 0.10 0.20 Brazil 149m 0.18 0.22 Argentina 32m 0.44 0.25 Mexico 81m 0.28 0.32 Korea 43m 0.09 0.38 Japan 124m 0.30 0.81

Note: Most recent figures for Zaire from 1989.

countries have participated in bench- in 1960, the table shows their incomes mark studies, which are then used relative to the US for 1960 and 1990, to derive an aggregate PPP for each ranked by the 1990 figure. The table participating economy. makes clear that, even after making These estimates have then been com- PPP adjustments, there are enormously bined with national accounts data to large disparities in average living stan- form the Penn World Table (PWT). This dards. Zaire's per capita income in 1990 table, often known as the Summers- was less than a fortieth of the USA's, Heston data set, has been used by while the huge populations of India and empirical growth researchers since the China have average incomes rather less mid-1980s. The data set has been dis- than a tenth of the USA's. cussed most recently by Alan Heston Remember, too, that these figures and Robert Summers (1996), and is are the average income per head: many described in more detail in Summers in Zaire have much less than 2 percent and Heston (1988, 1991).3 of the average US income to live on. Table 1 displays some useful informa- Supplementing the raw figures with tion calculated from version 5.6 of the other pieces of evidence gives some Penn World Table. For a selection of idea of the human cost. The typical Af- countries with large populations that rican mother has only a 30 percent might have been considered developing chance of seeing all her children survive to age 5. The life expectancy of a 3 For a more detailed exposition of the methods person born in Sub-Saharan Africa in of the ICP, includin an assessment of its achieve- 1980 is just 48 years (William East- ment, see the artic7e by one of its originators, Irving Kravis (1984). erly and Ross Levine 1997a). In the 116 Journal of Econonaic Literatur -e, Vol. XXXVlI (March 1999)

dency for poor countries to grow faster TABLE 2 GROWTHI\.IIR~ICLESAND Dlb lSTEKb, 1960-90 tlian rich ones, which would allow them ANNUALGROFITHRATES OF OUTPUTPERWORKER to catch up. Figure 1 shows the growth of real GDP per worker over 1960-90 Miracles Growth Disasters Growth against its initial level, a familiar dia- Korea Ghana gram in the growth literature. If coun- Botswana Venezuela tries are converging, one would expect Hong Kong I\.Iozamblque to see a negative slope."n fact, like Taiwan Nicaragua Table 1, tlie figure shows that there is Singapore 5launtani2 no general tendency for countries to Japan Zambia converge to a common level of per Malta Mali capita income. Cyprus hladsgascar Some interesting patterns emerge Seychelles Chad from Figure 1. Tl'ithin the group of Lesotho Gu) ana countries that were relatively well off by 1960, those with 1960 GDP per Note: Figures for Botswana and Malta based on 1960- worker greater than $10,000 in 1985 89. "international dollars," there has been some convergence.5 \Vithin the group of poorer countries, there has been a circumstances, it is not surprising that greater variety of experience: some we should want to know as much as pos- have done very well and others very sible about how countries can grow badly. more quickly. Table 2 lists the ten fastest and the We can extract some other useful in- ten slowest growers among the coun- formation from Table 1. One interest- tries in the Sumiilers-Heston data set, ing fact is that most of these large de- using data on output per worker. The veloping countries roughly maintained table demonstrates a regional pattern their positions relative to the USA be- familiar to growth researchers: many of tween 1960 and 1990. Since the USA's the fastest growing countries are in East per capita income grew at around 2 per- Asia, many of the slomrest in Sub- cent a year over this period, tliis indi- Saharan Africa. The strength of this cates that these countries have grown at pattern suggests that growth rates are a similar rate. There does not appear to quite persistent over time, but this is be an absolute poverty trap. misleading. For the sample as a whole, 2.2 Growth Miracles and Disasters the correlation between growth in GDP per worker over 1960-75 aiid that over Although there is considerable persis- 1975-90 is just 0.17. Relatively few tence in relative positions, Table 1 also countries have done consistently well, hints at the possible variety of growth and averaging over long periods tends experience. There have been "growth to obscure tlie episodic nature of disasters," countries in which per capita growth, particularly in Sub-Saharan income has fallen since 1960, and Africa. Frequently countries have done "growth miracles" like Japan and Korea, well for short periods, only for growth countries which have risen rapidly up the world income distribution. Xavier Sala-i-Martin (1996) discusses convergence concepts in more depth. One of the first questions we might 5 The "international dollar" is the currency unit ask is whether there is a systematic ten- used in the Penn World Table. Temple: The New Growth Evidence

Korea *Hang. Kong . Japan . 0. :. CYF~~S Spain *Indonesia* . Italy ~urtey ...... a . **o *. . ** Gennany • ** . . Mexico *UK Canada . ....** . . -0. . *. *Al;tralia USA .* . . a Jamaica \ Argentina Trinidad New Zealancl 8 •

%* Ghana . Venezuela • • Nicaragua Mali Madagascar Chad Guyana

I I 1 1 I I

1960 real income per worker

Figure 1. Growth and initial income, 1960-90

to collapse later on. This has important middle income around 1960; some of implications for growth studies, which I these have joined the relatively well off. will return to later on. Most of the countries in Sub-Saharan Overall, the figure and tables suggest Africa have done very badly. that many of the most important poor These stylized facts might lead one to countries do not seem to be catching up think about theories that emphasize to the USA's level of income. Instead, relative development traps and multiple countries have roughly maintained their equilibria, but it is always worth replace within the world income distri- membering that there are often simpler bution over the last thirty years or so, explanations, not least low levels of with little tendency for reduced income factor accumulation among some coun- dispersion, and perhaps even some tries. To distinguish between theories divergen~e.~The exception to this sta- we clearly need more sophisticated bility is some of the countries that were methods. Sections 3, 4, and 5 will consider the pros and cons of those usually

6s ace precludes a more detailed picture, adopted, before Section 6 returns to whicR can be found in the recent work of Charles convergence issues in depth. Jones (1997a,b) and Danny Quah (1993, 1997). Before reviewing more complex em- The presentation here follows most of the literature in analysing post-1960 experience; for a pirical work, however, it is worth assess- longer view see Lant Pritchett (1997). ing two of its foundations. These are 118 Journal of Economic Literature, Vol. XXXVII (March 1999) the measurement of output and the cal- Japan's GDP per capita is around 80 culation of growth rates. Both involve percent of the USA's, according to the subtleties that some growth researchers PWT, but its GDP per worker is just 60 have missed. Lack of space precludes a percent. full treatment, but the coverage here at Growth theories are essentially con- least points towards useful further reading. cerned with output per worker hour, but for developing countries in particu- 2.3 The Quality of the Output Data lar, the available data on worker hours are weak. Holger Wolf (1994) discusses Nobody disputes that there are very the implications for the convergence large income disparities, but the accu- literature, and argues that the problem racy of precise comparisons is often should not be neglected. It is clear that questioned. Given the difficulty of the we need either better data on working task facing the compilers of the PWT, it hours and participation or, at the very is not surprising that worries have least, a framework for relating them to sometimes been expressed about some more easily observable variables. of the methods used and the reliability Other problems include the likeli- of the final data (for instance, Robin hood that measured output per worker Marris 1984). hour will vary with changes in the age For those countries that have not par- structure of a country's population ticipated in ICP benchmarks, and in (Michael Sarel 1995) and with price which the PWT estimates are based on distortions (Jagdish Bhagwati and Bent extrapolations, the quality of the data is Hansen 1972). Finally, the presence of likely to be particularly low.7 Also wor- a sizeable nonmarket sector may lead to rying is the quality of the underlying the output of developing countries national accounts data for certain devel- being understated. oping countries. Heston (1994) quotes confidence intervals for estimated 2.4 Measuring Growth growth rates that are wide enough to be The Summers-Heston data set was somewhat dispiriting. primarily intended as a means of com- More recently, Heston and Summers paring income levels. It is widely used (1996) have pointed to two major con- to derive growth rates as well, but there cerns. One is that data on capital ser- is noticeable confusion about whether vice lives is hard to come by, which or not this is a good idea. Some re- makes it difficult to arrive at accurate searchers are unaware that growth rates figures for net investment and capital obtained from the Penn World Table stocks. A second problem is the quality will usually differ from those implicit in of available data on labor force partici- countries' own national accounts. pation and working hours. Heston and Both sets of growth rates are Summers write that participation rates weighted averages of growth in the "vary enough to make GDP per capita a components of GDP, but the PWT very unsatisfactory proxy for GDP per weights are the real GDP shares of the worker" (1996, p. 23).8 As an example, components-based on an average of

7 Summers and Heston use a grading system to international prices-rather than the highlight the worst cases. nominal shares used in domestic na- 8In an earlier article, Heston and Summers tional accounts. If a country's price (1988) note the low rank correlation between GDP per capita and GDP per worker hour for 12 structure changes much between the OECD countries. initial and final years, the growth rate Temple: The New Growth Evidence 119

based on domestic prices will differ 3. Cross-Country Research from the Summers-Heston growth rate. Daniel Nuxoll (1994) argues that us- Now that we have briefly covered ing domestic prices to measure growth data issues, and gained some impression rates is more reliable, because the do- of recent growth experiences, it is time mestic prices are the ones reflecting the to dig a little deeper, and start to iden- trade-offs that agents actually face. He tify reasons why those experiences have concludes that "probably the ideal is to been so widely divergent. The ultimate use Penn World Table numbers for aim of this literature is to understand levels and the usual national-accounts why growth rates differ, and which pol- data for growth rates" (Nuxoll 1994, p. icy measures will be effective in raising 1434). In general, the decision to use growth. domestic or international prices de- In this section, I turn to considera- pends on the hypothesis being tested. tion of the methods that might be used As Nuxoll points out, too little thought to address these questions. The most is given to this question at present. popular method has been cross-section Once one has decided on a set of out- growth regressions, combining data put measures, the next question is how from the Penn World Table with other to calculate the growth rate. The usual variables. Despite the popularity of this method uses only initial and final out- endeavour, many believe that it is fruit- put. Since either of these may be some less, partly because there are not likely distance from the trend path of output, to be any general answers. The appro- it may well be preferable to use the least priate research questions and policies squares growth rate, obtained by re- will depend on a country's particu- gressing the whole of the log output se- lar situation, for instance whether a ries on a constant and a time trend. This country is a technological leader, or a should be more robust to short-run in- developing country trying to catch up. stability, such as business cycle effects.9 Historical studies are likely to be far Talking about measuring "the trend more sensitive to these issues, and I growth rate" presupposes that the trend first consider their particular merits as is deterministic rather than stochastic. a way of understanding growth. This is For the most part, growth researchers followed by a comparison of growth ac- have been content to assume that out- counting and cross-country regressions. put is trend stationary. One justification In keeping with my underlying argu- might be this. When unit root tests are ment that careful growth regres- used country by country, the presence sions can play a valuable role, I then of a unit root is typically rejected for review various possible specifications. some of them. It seems unlikely that One or two problems are highlighted the process generating output is so dif- in passing, but detailed consideration ferent across countries that shocks have of the econometric problems will be a permanent effect in some and only a postponed until Section 4. temporary one in others.1° 3.1 The Role of Historical Studies 9 It mi ht be useful to go a step further, by including &oamics in the regression For more on For many people, statistical research growth rate illeasurement see Eugene Canjels and on growth seems rather cruder and less Mark Watson (1997) and Nanak Kakwani (1997). informative than historical case studies. lOThe use of panel data unit root tests should be increasingly informative on this kind of ques- Certainly it is important to remem- tion. ber that growth regressions will never 120 Journal of Economic Literature, Vol. XXXVII (March 1999) offer a complete account of the growth siastic, it is worth remembering that process, and that historical analyses econometric problems such as endo- must have an important complementary geneity and omitted variables have role. Historians bring to bear a much close equivalents in historical studies, deeper conception of the social, politi- usually with even less hope of a solu- cal, institutional and technological tion. In the conclusions, I consider the sources of growth than theoretical and role of historical case studies as one of empirical economists are usually able to several possible ways forward. incorporate in formal models. 3.2 Growth Accounting versus Growth Equally, the limitations of historical Regressions treatments should not be ignored. One of the pioneers of growth research, In some contexts, another alternative Simon Kuznets, argued that the isolated to cross-country regressions is growth study of just one or two countries accounting. When investigating why provides only a partial view (Kuznets growth rates differ, a common starting 1966, p. 32). It is also worth quoting point is to think about the relative another influential economic historian, contributions of growth in inputs, and

Alexander Gerschenkron: ugrowth in efficiency or total factor The historians' contribution consists in point- productivity (TFP). h hen one might ing at potentially relevant factors and at think about separately modelling input potentially significant contributions among growth and TFP growth. Traditionally, them which could not be easily perceived though, growth accountants have only within a more limited sphere of experience. made progress on the first issue. A fa- (Gerschenkron 1952, p. 4, emphasis in original) miliar response is that estimates of TFP growth are frequently no more than a This seems to me exactly right: histo- measure of our ignorance. They bring rians can usefully point to particular us little closer to an understanding of factors that others are likely to miss. An why TFP growth rates might differ unspoken corollary is that statistical across time and space. work, perhaps using cross-section vari- It is important to understand that one ation, is often necessary to quantify the aim of running growth regressions is to importance of the "potentially relevant" use the cross-country variation in TFP factors. Another way of saying the same growth to understand its determinants. thing is that case studies are a useful However, the TFP growth rates are de- means of generating hypotheses, but rived somewhat differently from growth econometrics is needed to test the va- accounting. The cross-country variation lidity of generalizations (Angus Deaton in the data is used to estimate the pa- and Ron Miller 1996). The two methods rameters of an "average" production are complementary in other ways: function, instead of imposing these Gustav Ranis (1984) points out that the parameters based on factor shares or "average pattern" is only the beginning micro evidence. of wisdom, but knowledge of this pat- If we are to understand the determi- tern can inform one's choices of the nants of TFP growth, use of its variation countries to study in greater detail. across disparate units (countries, re- The problems with statistical work gions) seems pretty much essential. that are discussed below might lead one Whatever method of deriving TFP to disillusion, and a general preference growth one prefers, at some point one for history. Before becoming too enthu- will need to think about cross-section or Temple: The New Growth Evidence panel data econometrics. In this re- investment. In a sense, however, this spect, the approaches used in the new does not get us very far. It merely growth evidence are simply reflecting pushes the demanding question down the inevitable. The idea of returning to a rung, from explaining growth to ex- single-country growth accounting as the plaining investment. As I will discuss main mode of investigation is not an in Section 9, we need better explana- enticing one, especially when one is tions of the cross-country variation in confronted with the sparse data of investment ratios. developing countries. In practice, whether or not we seek Where one may depart from the re- to decompose output growth into the cent literature is in the derivation of contributions of inputs and TFP should TFP growth. Should the elasticities of depend on the research question. Often output to inputs be imposed, or esti- we are interested in the overall growth mated? Nearly all of the work surveyed effect of a policy outcome like inflation below implicitly estimates them, by in- or the budget deficit; whether or not it cluding factor accumulation in growth acts through factor accumulation or regressions. One reason lies in recent TFP growth is of secondary importance, theory. When the parameters are im- at least initially. Hence the approach of posed, typically one assumes perfect some recent papers, though perhaps too competition, constant returns to scale, few, is to use a regression model which and the absence of externalities. All simply relates growth to policy out- three assumptions have been ques- comes, using a model sufficiently gen- tioned, often convincingly, by new eral that it should account for most growth theorists. influences on factor accumulation as Yet the remainder of this paper will well as on TFP. This approach often has reveal many difficulties in estimating much to recommend it. technology parameters. If we are willing 3.3 Specqying Growth Regressions to ignore externalities to physical and human capital, which are quite possibly Overall, then, cross-country empirical negligible, one might well ask why we work will continue to form a useful choose estimation. We could set these complement to more traditional ap- parameters based on micro data or proaches, depending on the particular factor shares, and thus avoid all the questions that are being asked. Cross- problems connected with estimating country work allows differences in pro- production functions. In time, this may ductivity growth to be explained, unlike turn out to be the best method of ob- conventional growth accounting. If well taining comparable TFP growth figures done, this work can help identify the for a large number of countries. The relative contributions of different influ- point remains that cross-country studies ences more precisely than historical will still be needed to understand the studies. Given that there is a clear place determinants of TFP growth. for cross-country work, in this section Finally, one might ask whether the I turn to the specification of growth initial decomposition into input and regressions. TFP contributions is always a useful First, though, a warning is necessary. one. Researchers may well be right to Many of the issues require technical attribute a high degree of the cross- discussion, and this discussion pre- country variation in growth rates to the sumes a good knowledge of econo- variation of inputs, like physical capital metrics and an eagerness to get to grips 122 Journal of Econonzic Literature, Vol. XXXVII (March 1999) with methodological issues. Those read- Romer and Weil show, by approximating ers who want only a view of the forest, around the steady state, that growth in without worrying too much about the this model is given by: trees, should cover the remainder of this section only briefly, and then skip to the summary of research findings, which begins with Section 6. Before then, I will discuss several approaches in turn, starting with the framework of N. Gregory Mankiw, David Romer and David N. Weil where 8= 1-e-kt and h is the rate of (1992), henceforth MRW. The exposi- convergence to a country's steady state, a tion should clarify some of the under- measure of how fast countries attain lying ideas that almost all growth their long-run equilibrium path.12 The regressions hold in common. It can also measure is defined by be used to understand the usual selec- d In (t) tion of variables in equations that are = h[ln y' - ln y(t)] not explicitly derived from a theoretical dt model, and this less formal approach I and the convergence rate is related to will discuss in Section 3.4. other variables by Hence the framework introduced by h = (n + g + @(I-a - p) (3) MRW is a useful starting point.11 They In implementing the equation empiri- start from a Cobb-Douglas production cally, there are one or two subtleties. In function with constant returns to scale, practice investment rates are not con- Y = K"HP(AL)l- - P (1) stant, so MRW average them over the where K is physical capital, H is human period. Another point to note is that n capital, L is labour supply and A is an and 6 are likely to vary across countries, index of technical efficiency. MRW as- and so too will the rate of convergence. sume that investment rates in physical Conventional estimation of the model and human capital are constant at sk and ignores this subtlety, and we shall sh respectively, and that both types of return to this later on.13 capital depreciate at a common rate 8. Why does initial income affect Technical efficiency grows at the same growth in the above equation? The exogenous rate g across countries, while negative coefficient implies that, if we the labour force grows at differing rates take two countries with the same rates n. This is clearly just the Solow model of investment and the same level of augmented with human capital and an efficiency, the poorer one will grow assumption that countries share the same more quickly for a transitional period. rate of efficiency growth. The initial level of efficiency, A(O), is assumed to '"ow ood is the approximation used to derive vary randomly across countries (due to these resAts? David Romer (1996, p 23) argues that the approximation is likely to be fairly reliable local factors like climate) and this can be for the changes in parameter values typically con- used to justify the error term. Mankiw, sidered. 13Note also that the convergence equation is obtained by taking a Taylor series approximation l1More general treatments, particularly useful around a deterministic steady state. Michael for s ecialists, can be found in Steven Durlauf and Binder and M. Hashem Pesaran (1996) argue that, (1998) and Jeffrey Sachs and Andrew M. when growth is stochastic, this a proach to deriv- Warner (1997). ing convergence paths is potentiaty misleading. Temple: The New Growth Evidence 123

The reason for these "transitional dy- A, the only way to obtain consistent namics" is that the relatively poor econ- estimates of a conditional convergence omy must have lower stocks of physical regression is to use panel data methods. and human capital. Hence the marginal Since initial efficiency is an omitted product of extra capital is higher in this variable that is constant over time, it economy, and for a given rate of invest- can be treated as a fixed effect, and the ment its growth will be faster. time dimension of a panel used to elimi- One consequence is that if we run a nate its influence. The advantages of regression which controls for the deter- panel data approaches will be discussed minants of steady states, like invest- in Section 5. ment ratios, then initial income will Finally, note that we could also use take a negative sign. This is the "condi- the MRW approach to explain the vari- tional convergence" result. The conver- ation of income levels, rather than gence may be conditional, in that it is growth rates. Robert Hall and Jones only apparent when we take into ac- (1997) have argued that this may be a count the variation across countries in more natural research question, since steady state levels of income. Hence the ultimately we are interested in growth result does not imply that poorer rates mainly because of their impact on countries will catch up with rich ones. levels. A crucial drawback is that, in the Indeed, the Solow model is perfectly likely absence of good instruments, we compatible with income divergence: cannot explain the cross-section vari- anything that drives apart investment ation in income using any variables rates in rich and poor countries will which are endogenous to the level of in- tend to lead to increased income come, like the development of physical dispersion.14 infrastructure or the financial system. Given the role of transitional dynam- Hence 'levels accounting' tends to be a ics, it is widely agreed that growth little short of implications for policy, regressions should control for the and growth regressions retain the ad- steady state level of income. Note, vantage that they can be used to study a though, an important consequence: in wider range of variables. principle the level of technology A 3.4 Informal Growth Regressions should be included in the regression." Since this variable is unobserved, it has MRW's work provides a simple to be omitted. As with any other omit- theoretical framework for growth re- ted variable problem, the other parame- gressions. The most common approach, ter estimates are biased if one or more though, is simply to use a more or less regressors are correlated with the level ad hoc regression, driven in its specifi- of technology. In practice, countries that cation mainly by previous results in the are relatively less efficient are also likely literature. Typically such a regression to have lower investment rates, and one will include the investment ratio, initial can easily imagine further correlations income, and measures of policy out- with other right-hand-side variables. comes like school enrollment and the In the absence of a suitable proxy for black market exchange rate premium.15

14 It is also worth noting that conditional conver- 15 Regressions of this type are sometimes known gence does not imply a rejection of "Ak" models of as "Barro regressions," after Robert Barro (1991). growth. On this see Narayana Kocherlakota and Precursors that were less influential, but are simi- Kei-Mu Yi (1995) and Charles Leung and Quah lar in spirit, include Sherman Robinson (1971) and (1996). Roger Kormendi and Philip Meguire (1985). 124 Journal of Economic Literature, Vol. XXXVII (March 1999)

Since these regressions usually in- policy. I will return to this issue in the clude the investment ratio and initial conclusions. income, they can be interpreted in 3.5 Cross-Country Growth Accounting terms of the MRW model. This interpretation is rarely made explicit, and The cross-country growth literature there are few attempts to calculate has sometimes been unnecessarily am- technology parameters from the coeffi- bitious in trying to infer whether vari- cient estimates. Yet the informal growth ables act through factor accumulation regressions can be seen as a crude ex- or TFP growth. Yet this remains a tension of MRW, in that they attempt question of great interest to many. In this to account simultaneously for input section, I consider other ways of learning growth and variation in TFP growth. about technology parameters through Hence g is seen as a function of observ- estimation. The central idea is that we able variables, g(X). In practice the ex- can avoid the omission of a term in in- tension is not perfect, since the term itial efficiency by making use of capital In(n + g(X) + 6) is difficult to estimate stock data. One way of doing this is to use even when using nonlinear methods. the initial capital-output ratio in place More importantly, the extension is of initial income on the right-hand-side also imperfect because the additional X of the regression (2),as suggested by variables, like financial depth or infla- Jess Benhabib and Jordi Gali (1995). tion, may be correlated with initial effi- However, if we have data on capital ciency. If countries that are relatively stocks, a more straightforward approach inefficient also tend to have higher in- is that adopted by Benhabib and Mark flation rates, inflation will take a nega- Spiegel (1994) among others. It might tive sign when entered into (2), even if be called cross-country growth account- it has no long-run effect on output. It is ing, or growth accounting with exter- also possible that the supposedly puz- nalities. The method starts from the zling continent dummies often used in framework of traditional single country convergence regressions just proxy for growth accounting, but with two dif- differences in initial efficiency. Hence, ferences. The output elasticities are although simple aggregate models are estimated rather than imposed; and always dubious, some important insights part of the model may be designed to are neglected in the absence of a formal capture the cross-country variation theoretical derivation. in TFP growth. Unlike the MRW Going back to the production func- and Benhabib-Gali approaches, no tion derivation of MRM7, another closely approximations are needed. related point should be noted. When a Remember that in the conditional variable enters an informal growth re- convergence regression, initial income gression, it is not clear whether it is used to control for the transitional affects the long-run growth rate, the dynamics induced by factor accumula- steady state level of income, or both. tion. If, however, we have data on fac- Disentangling the two is rarely possible, tor inputs, it is not difficult to set up a and the problem is not unique to this regression in which the change of out- method. It is not as devastating as it put is directly regressed on changes in may first appear. A large effect on the these inputs. Taking the simplest exam- steady state level of income may be as ple, we can take logs of a Cobb-Douglas important in practical terms as a growth production function and difference it to effect, and of just as much relevance for obtain: Tenzple: The New Growth Evidence

will be mismeasured for individual countries. One promising elaboration is + @Alog L + E (4) that of Gary Koop, Jacek Osiewalski, where the notation is the same as before. and Mark Steel (1995). They estimate This provides direct estimates of factor stochastic production frontiers which shares, and there is no term in initial ef- vary across regional country groups; the ficiency. Instead, in estimating the method nllo\+s them to decompose model we should replace the unobserv- growth into input changes, efficiency able AlogA by soine function of change, and technical progress. One observables, g(X). Otherwise, the esti- might be sceptical about the details of mates of factor shares will be biased.16 In this decomposition, but this research is any case, it should be clear by now that innovative in deri\ing rates of produc- building a good model for g(X) is often a tivity growth on a comparable basis for motivation for this kind of study. an unusudll> wide range of countries. One problem with estimating (4) is One remaining disadvantage with all well known from the microeconoinic lit- these approaches is that they require erature on productioil functions. Since information on the initial stocks of the factor inputs are decision variables, physical and human capital. For many agents may respond to shocks by alter- developing countries, investment series ing inputs of, say, ~h~sicalcapital. In are not available before the late 1950s, that case, the regressors will be corre- and hence any estimate of a stock for lated with the error term, and estimates this period is likely to be no rnore than of the parameters will be inconsistent. an educated guess. The implication is Varlous solutions to this are possible, as that these techniques will become rnore outlined by Zvi Griliches and Jacques useful as longer spans of data become Mairesse (1995). Perhaps the method available. most likely to find favor is using lagged levels of the factor inputs as instruments, given the likely absence of other sultable variables. hyhatever empirical framework is It is worth remembering that these adopted, there are usually substantial difficulties apply equally to the other problems in estimating and interpreting possible specifications of growth regres- growth regressions. The difficulty of un- sions. The framework advocated here observed fixed effects is just one. In has a key advantage: there is no longer a this section I consider other frequent danger of spurious correlations driven concerns, starting with the objection by the omission of initial efficiency ii that very different countries are un- second advantage is that it can be ~m- likely to be drawn from a common sur- plemented ill a panel wlthout the addi- face, as niultiple regression assumes. tional complexities of dynamic panel Section 4.2 discusses outliers, while data models, provided that TFP growth Section 4.3 esarnines model uncer- is unrelated to initial income. tainty, drawing on influential work by As it stands, the method will recover Levine and Da~idRenelt (1992). Then only the average output-capital elastic- the problems of measurement error, ity across countries, and so TFP growth endogerieity and error correlation are also discussed.17 lh Be~lhaLlb and Boyall Jovanovlc (1991) and Bellhabib dnd S iegel (1994) analyse the likely 1'Heteroscedasticity I leave aside, partly be- Jlrectlous of the tlases cause it is better uuderstood. Most researchers use 126 Journal of Economic Literature, Vol. XXXVII (March 1999)

The underlying aim is to ensure that acknowledge that problems of parame- the reader gains an impression of why ter heterogeneity are likely to be more some results are fragile, and how this severe in the cross-section growth con- fragility might be overcome in future text than in other areas of economics research. Again, I should emphasize (Levine and Renelt 1991). It is easy to that readers anxious for a summary of see, for instance, that the coefficient on findings should now skip to Section 6, the investment ratio is likely to be perhaps returning to this material on a lower in war-torn and unstable coun- second reading. tries than in peaceful ones. It is possible that parameters vary across coun- 4.1 Parameter Heterogeneity tries in ways which render conventional I start with a frequent objection to estimates inconsistent. empirical work on economic growth, Several papers have provided strong one which some mistakenly seem to re- evidence for widespread heterogeneity. gard as devastating. Countries differing Particularly important in this respect is widely in social, political and institu- the work of Durlauf and Paul A. tional characteristics are unlikely to fall Johnson (1995), who use regression on a common surface. One well known trees to identify multiple regimes, those comment along these lines is that of country groupings across which parame- Arnold Harberger (1987): ters differ widely. Together with Dur- lauf and Quah (1998), they have empha- What do Thailand, the Dominican Repub- sized that linear regressions assuming lic, Zimbabwe, Greece, and Bolivia have in colnmon that merits their being put in the common parameters are an inappro- same regression analysis? priate way to investigate new growth theories. Such regressions will often There is a cheap answer to this, and a fail to distinguish recent models with more thoughtful one. The cheap answer multiple equilibria from the traditional is that any statistical modelling requires Solow-Swan framework. assumptions about parameter con- One approach that has been advo- stancy, and we can see how closely the cated instead is to set up a transition countries fall on a common surface sim- matrix, showing the probabilities of ply by looking at a regression R2. This is moving between particular income typically about 0.5 in a cross-section ranges in a given time period. Quah growth regression, suggesting that these (1993, 1997) argues that this kind of regressions do have some explanatory study is preferable to growth regres- power, despite the heterogeneity diffi- sions in describing the evolution of the culty and, one might add, measurement world income distribution. The transi- error and omitted variables. Recent re- tion matrices form a very useful sum- gressions, such as those in Sachs and mary of income dynamics, but are less Warner (1997), can explain more than effective at revealing causal mecha- 80 percent of the variation in growth nisms, at least as yet. rates. Given the existence of heterogeneity, A more thoughtful answer is to a key question for traditional methods is heteroscedastic-consistent standard errors. Per- how much they tell us about parameter haps, though, alternative methods like wei hted averages. Given that the purpose of least squares might be referable. it shoulf also cross-country empirical work is often to be remembered that wKere heteroscedasticit is observed, it may reflect neglected parameter let- arrive at generalisations about growth, erogeneity. the averages are important. It turns out Temple: The New Growth Evidence 127 that when a dynamic panel data model model is only a crude approximation to is used, as is frequent in the growth reality. This means that it is likely to fit literature, the estimate of the average some observations particularly badly. A effect will be inconsistent, even when problem immediately arises here: obser- the length of the time series tends to vations that are "unrepresentative," due infinity. The reasoning is that if the re- to parameter heterogeneity, omitted gressors are serially correlated and one variables or measurement error, can neglects to model parameter heteroge- act as influential outliers or leverage neity, this will induce serial correlation points.18 in the disturbance. In a model with a Unfortunately, attempts to address lagged dependent variable the outcome this difficulty are often rather token. will be inconsistent estimates (Pesaran Sometimes observations are dropped and Ron Smith 1995). one at a time, or single-case diagnostics In contrast, cross-section regressions like Cook's distance are used; but these can produce consistent estimates of the methods are well known, at least to stat- average long-run coefficients even in a isticians, to be inadequate. Single-case dynamic model, provided that the vary- diagnostics are likely to miss groups of ing coefficients differ randomly and are outliers (the "masking" effect) and may distributed independently of the regres- even be misled into identifying repre- sors. This does not mean that cross- sentative observations as outlying (the section regressions are necessarily the "swamping" effect). best technique: we have already noted A better alternative is provided by ro- the omitted variable bias inherent in bust regression. As discussed by Peter conditional convergence regressions. It Rousseeuw and Annick Leroy (1987), does, however, reinforce the case for robust estimation procedures allow re- using static panel data models of the searchers to characterise the most co- kind proposed in Section 3.5. herent part of the data set, and use the Unfortunately, it will be difficult to resulting parameter estimates to iden- make much progress on questions of pa- tify possible outliers. These estimators rameter heterogeneity until we have have rarely been applied in the growth data over a longer time period. When literature, even though they have con- more data becomes available, it may be siderable advantages over least squares, possible to use panel data models with at least at the exploratory stage. The stochastic parameters. At present, more identification of possible outliers will imaginative use of methods for detect- not only help render generalizations ing heterogeneity, such as robust esti- more robust, but will also highlight mation, regression trees, sample splits, countries with atypical growth experi- interaction terms, and dummy variables, ences, ones that are particularly likely would be welcome, especially when the to reward further study. use of panel data allows more degrees of freedom. Recent research has been 4.3 Model Uncertainty far more careful about this. Harberger A central difficulty with empirical had a point. growth studies has been widely 4.2 Outliers 18 AS an illustration, in the context of growth re- Once we acknowledge the existence gressions analysing equipment investment, Brad De Long and Lawrence Summers (1991) discuss of parameter heterogeneity, we must the large effect that Botswana and Zambia have on also acknowledge that our regression the coefficient estimates and their precision. 128 Journal of Economic Literature, Vol. XXXVIZ (March 1999)

acknowledged by practitioners since the too much attention is paid to the statis- work of Levine and Renelt (1992). tical significance of variables. A worry- Many variables have been found to be ingly common mistake is to conclude significant in growth regressions; nearly that "there is no effect of variable X" as many have been found to be "frag- simply because its effect is imprecisely ile," in the sense that their statistical measured; one suspects that Type I1 er- significance disappears when a different rors are frequent.19 This is not a mis- group of right-hand-side variables is take of the original paper, which makes selected. clear that a lack of robustness should Problems of this kind have been often spur further investigation into given some discussion in the statistics causality and inter-relationships. Find- literature. The difficulty is that several ing that a result is fragile to a particular models may all seem reasonable, but conditioning variable could in itself be lead to different conclusions about the valuable information. parameters of interest. In these circum- Another illustration of the dangers of stances, presenting the results of a taking the Levine-Renelt analysis at single model is misleading. It ignores face value is useful. It is plausible that the researcher's inevitable uncertainty many of the variables in growth regres- about the form of the model, and so sions are symptoms of deeper problems leads the reader to underestimate the or "syndromes": high inflation reflects uncertainty actually present. Edward bad macroeconomic policy, while a high Leamer (1983, 1985) was one of the black market premium may reflect a first to emphasise the difficulty, and mixture of bad short-run policy and in- it is a variant of his proposed solu- ward orientation. There are many possi- tion, "extreme bounds analysis," which ble indicators for each syndrome, which underlies the work of Levine and will tend to render one another insig- Renelt. nificant when included at the same Since their paper, several writers time. If one treats individual variables have used the same framework to dem- as symptoms of underlying problems, onstrate that a correlation is robust to finding that some are "fragile" in the changes in specification. This is an ad- Levine-Renelt sense does not tell us vance, but not without problems of its much about their potential usefulness own. Several of the regressors used by (Pritchett 1998). One promising way Levine and Renelt are almost certainly forward is to combine indicators using endogenous. More fundamentally, the latent variable methods, and then ex- robustness study which simultaneously amine the robustness of these overall addresses fixed effects, measurement measures. error, endogeneity, outliers and model 4.4 Endogeneity uncertainty, is yet to be written, and one suspects that this will remain the One of the most frequently expressed case for some time. Hence the finding concerns about work in the growth that a variable is robust to changes in the model specification is not sufficient 19 Havin said that, the issues in statistical infer- for valid inference. ence raisef by growth regressions are interesting, It is also worth pointing out that since we are often looking at the population rather robustness is not a necessary condition than a sample from it. Even when the regressors are considered to be stochastic, arguably inference for useful information. One danger with should proceed conditional on their realized val- the Levine and Renelt approach is that ues. Temple: The New Growth Evidence 129 literature is the probable endogeneity tests can be used to examine the iden- of some regressors. To avoid simultane- tifying assumptions, it should be re- ity concerns, researchers often make membered that these are based on use of initial values, for instance re- the estimated residuals, not the true gressing growth over 1960-85 on the disturbances. 1960 secondary school enrollment rate. This will be a particular difficulty This is not quite as watertight as re- when researchers estimate conditional searchers seem to think: even if the en- convergence regressions but do not take dogeneity problem is solved, perhaps steps to eliminate the effect of the ini- some omitted variable, like the political tial level of technology. Since this level regime, affects both growth and the in- is persistent but omitted, there is likely itial level of variables like schooling. to be serial correlation in the error term Early papers often used the average for this reason alone. Serial correlation investment ratio as a regressor. It is tests may not detect this. The variables easy to construct arguments that causal- in the growth equation and the lagged ity could go from growth to investment, values used as instruments will both be as well as vice versa." Recent work uses correlated with the level of technology, instrumental variables to avoid these and so little of this level will remain in problems (Barro and Sala-i-Martin the estimated residuals. Hence the in- 1995; Barro 1997; and Francesco struments will appear to be valid, when Caselli, Gerald Esquivel, and Fernando a serial correlation test on the true Lefort 1996). There is a tendency for residuals would show them not to be.21 the coefficient on the investment To avoid endogeneity problems, the ratio to fall, as one might expect. In- best solution may be to use panel data cidentally. this suggests that some of to estimate the cross-country growth ac- the earlier results in the literature counting framework proposed above, may understate the effect of variables still using lags as instruments. Ad- like hurrlan capital that are positively mittedly the use of capital stock data correlated with investment and growth. will introduce measurement error. It In general there is a shortage of good should, though, be easier to find in- instl-uinents. So lnanv variables could be struments uncorrelated with this error used to explain that it is difficult than it is to find instruments uncor- to find variables that are not only highly related with the unobservable level of correlated wit11 the endogenous vari- technology. ables but can also plausibly be excluded Another fruitful line of research frorn the regression. One solution is to might be to adapt empirical growth use a panel. and employ lags of the models to allow explicitly for the possi- endogenous variables as instruments. bility of regressors that are endogenous Unfortunatelv their exogeneity is not al- to the growth rate or the level of in- ways clear. ~l~eremay be long delays in come. In an innovative paper De Long the effects of. say. human capital accu- (1997) extends a Solow model by adding mulation. Although serial correlation equations relating population growth and the investment share to the level "For instance, it seems to be a fairly robust of per capita income, and he shows st)rlized fact that lonth raises saving (Christopher how this affects the interpretation of Carroll md \Yei'f'l881). \\'ith imperfect capital mobility, this will induce a correlation between investment and growth even in the absence of any "The arguments in this and the next paragraph independent causal role for investment. are due to Christopher Sims (1996). 130 Journal of Economic Literature, Vol. XXXVIZ (March 1999) parameter estimates. A useful insight is method-of-moments estimators are two that increases in total factor produc- particularly useful techniques. Another tivity, in raising income, can also reduce possibility is the use of instrumental population growth and the relative variables, although I have noted above price of capital goods, which then feed the shortage of plausible candidates. through to further increases in the When a panel is used and the measure- capital-labour ratio and per capita income. ment error is not thought to be persistent, one can use lags of the variables as 4.5 Measurement Error instruments in GMM estimation, as As discussed earlier, the Penn World Caselli, Esquivel, and Lefort (1996) Table used in cross-section growth re- d'1scuss. gressions is inevitably flawed in some 4.6 Error Correlation and Regional important respects. Researchers also Spillouers frequently make use of additional data, such as that on income inequality, for Several researchers have suggested which problems of mismeasurement are that the disturbances in cross-section even more severe. Yet few studies at- growth regressions may not be indepen- tempt to assess sensitivity to rneasure- dently distributed. Although this can ment error, let alone correct for it. In often be seen as an omitted variable part, this may be due to the common problem, there are common shocks like misperception that measurement errors the climate which will always be diffi- simply bias coefficients towards zero. cult to model, and could lead to corre- When there are several badly measured lation in the disturbances. Although variables, or the errors depart from De Long and Summers (1991) found classical assumptions, then biases may little evidence for spatial correlation, go in either direction. it is often the case that regional That said, some regard the presence dummies add substantially to a growth of measurement error as pretty much regression's explanatory power. fatal for the whole exercise, and to hold These dummies partly reflect the fact this view is surely to go too far. Those that failure has been concentrated in instinctively suspicious of the literature sub-Saharan Africa, and success in East sometimes fall back on measurement Asia. It is also interesting that large error as an excuse to avoid engaging growth fluctuations seem to have been with it. Yet some results hold using dif- sychronized across many of the major ferent measures, samples and specifica- Latin American countries over the last tions. A more useful reaction to the three decades. Observations like these problem of measurement error is to try have led several researchers to think and identify which variables are particu- about spillovers between geographical larly badly measured, and hence isolate neighbors (Alberto Ades and Hak B. the findings that are least reliable. This Chua 1997; Easterly and Levine 1998; would also help researchers direct their Ramon Moreno and Bharat Trehan efforts towards the variables that are 1997). relatively well measured. Adjustment for spatial correlation Occasionally sensitivity to measure- raises formidable statistical problems, ment error is explicitly assessed, as in and the identification of genuine spill- Torsten Persson and Guido Tabellini overs is problematic (Charles Manski (1994) and Temple (1998a). Multivari- 1993). The spillover effects that have ate reverse regression and classical been found may simply reflect omitted Temple: The New Growth Evidence 131 variables. As it stands, these questions shares or micro evidence. However, probably need to be investigated fur- theory gives little guidance on a theo- ther. Without more evidence that the retical specification for TFP growth, or disturbances are independent, the stan- the micro parameters that could be used dard errors in most growth regressions to calibrate it. Hence the econometric is- should be treated with a certain degree sues raised here will need to be addressed of mistrust. whether or not calibration is used. 5.1 Panel Data Methods 5. Panel Data and Time Series Methods The advantages of using panel data Having absorbed the potential diffi- techniques to study growth are several. culties of cross-country studies, some First, and most fundamentally, they al- readers will be tempted to dismiss the low one to control for omitted variables usefulness of this style of research alto- that are persistent over time. As we gether. I think that would be a mistake, have seen, conditional convergence re- and I now complete the discussion of gressions rarely give unbiased esti- methods by explaining why. Techniques mates, because variations in technical that make more use of time series vari- efficiency across countries are likely to ation in the data might yet overcome be correlated with the regressors. By many of the objections often raised to moving to a panel data framework, we cross-country research. My argument can at least control for unobserved will be that although the various heterogeneity in the initial level of approaches are complementary, panel efficiency. data studies will increasingly offer the Another advantage is that several lags best way forward for many questions of of the regressors can be used as instru- interest, especially as longer spans of ments where required, thus alleviating data become available. measurement error and endogeneity This conclusion may be controversial, biases. A highly promising approach is especially since it is becoming fashion- that of Caselli, Esquivel, and Lefort able to argue that calibration tech- (1996), who use GMM to estimate a dy- niques can overcome some of the namic panel data model. The growth difficulties of the cross-country growth model is first differenced to eliminate literature.'Wne argument along these the effect of initial efficiency, and then lines, already encountered, is that we lags of the variables used as instru- should select the parameters of the ag- ments. Their work suggests that existing gregate production function based on cross-section estimates of convergence micro evidence. rates are severely biased, and so demon- As pointed out earlier, this kind of strates the potential usefulness of this approach ignores externalities. If we are technique. willing to do that, we do not need to However, this approach is still in its assume a common aggregate production infancy and there remain reasons to be function at all, but could probably de- sceptical about even the latest results. rive comparable TFP growth rates The particular GMM estimator used by across countries using data on factor Caselli, Esquivel, and Lefort is likely to "Unfortunately, shortage of space precludes a perform bidly when the variables .are full discussion. David Romer (1996) provides an highly persistent, because then lagged interesting critique of calibrating short-run macro models, much of which applies in this context as levels are weak instruments for first dif- well. ferences. More generally, the finite 132 Journal of Economic Literature, Vol. XXXVII (March 1999) sample properties of most dynamic largely unsettled. If annual data is used, panel data estimators are not yet well one must take great care in modelling understood. Different estimators do the short-run dynamics, especially as well in different circumstances, so as some heterogeneity is inevitable. Most yet it is difficult to advise on the best researchers have opted for five or ten course for growth researchers. year averages, but the latter in particu- There are some remaining worries lar mean that one is left with little time about the use of fixed effects specifica- series variation. It is clear that the pros- tions. Too often researchers use fixed pects for rigorous growth research will effects approaches to analyse the effects greatly improve as longer spans of data of variables that are fairly constant over become available. In the meantime, we time, or that will affect growth only need to investigate how well short time with a long lag. Standard transforma- averages eliminate business cycle effects. tions like first differences or "within Despite the problems, the use of groups" are likely to exacerbate the panel data methods, together with tests problem of measurement errors, at least for parameter heterogeneity, can prob- if these errors are not persistent. They ably succesfully address many of the typically lead to a large fall in precision, objections raised to cross-country em- since in effect the between-country pirical work on growth. These papers variation is thrown away. One might certainly present a harder target for the even prefer to ignore fixed effects when sceptics. At the very least, standard using something like a mean-square cross-section approaches should some- error criterion. times be supplemented with an assess- In other areas of economics, the ment of sensitivity to a fixed effects results from fixed effects estimation are specification. often found to be disappointing. However, my own view on the use of Griliches and Mairesse (1995) argue dynamic panel data methods is that that it may be better to proxy for omit- they may introduce an unnecessary de- ted variables, since this leaves more gree of sophistication. In conventional identifying variance in the independent specifications, they are needed because variables, and is informative in itself. of the presence of transitional dynamics One simple approach is to use carefully in output. It may be much simpler to specified regional dummies. The justifi- adopt cross-country growth accounting, cation for this is that much of the vari- either by estimating equations like (4) ation in efficiency levels occurs be- above or by using other data to impose tween rather than within continents, an the technology parameters. Panel data argument suggested by the estimates methods can then be used to study the in Koop et al. (1995). Along similar determinants of TFP growth, whilst lines, one could use panel data to esti- avoiding the extra complexities of mate growth regressions separately for dynamic models. different regional groupings. 5.2 Time Series Methods Given the presence of cyclical effects, researchers face some difficult Some econometricians, primarily choices in selecting the time intervals those from a time series background, over which to study growth. Whether argue that we should go further than us- one is best using annual data, or five- or ing panels. They argue that standard ten-year averages to avoid business cy- cross-section methods throw away cle effects, is a question that remains useful information, while panel data Temple: The New Growth Evidence 133 methods make unjustifiable assump- the cross-country literature by more tions about parameter homogeneity. In- careful attention to econometric issues. stead, we should estimate parameters for However, surveys that conclude that all countries individually, using separate the valuable work lies ahead are never time series regressions for each country. particularly useful, and are likely to One immediate problem is the qual- gloss over some useful contributions. ity of the data for developing countries. With that in mind, the second half of Many important variables, such as the survey reviews what we have population growth and school enroll- learned so far. ment, are often interpolated from just In reviewing what we know, I start in three or four census years. A second this section with issues at the heart of difficulty is the limited timespan of the the literature, those relating to conver- available data; even when observations gence and technology. Section 7 will are available for each of thirty years, it cover the contribution to per capita is difficult to discern the long-run ef- growth of the various inputs (physical fect of variables like inflation. To pre- and human capital, and research). Then vent short-run business cycle effects Section 8 will broaden the survey to in- driving apparent long-run correlations, clude wider influences, such as inequal- long lags of the independent variables ity, infrastructure, and macroeconomic need to be included, and one starts to management. run into a degrees-of-freedom problem In the convergence literature, some if the cross-section variation is ignored. confusion has been caused by the ten- As it stands, growth researchers have dency to see various studies as being only rarely confined themselves to the opposing answers to the same question. information in the time series proper- Instead, the different studies are best ties of the data. Among the exceptions seen as complementary approaches to are work by Jones (1995a) and several the six questions outlined in the intro- studies of inflation and growth, all of duction. Section 6.1 will discuss mea- which will be discussed later on. My surement of the rate at which countries own feeling is that time series econo- converge to their respective steady metricians are usually mistaken in be- states, and its implications. Sections 6.2 lieving that there is much to be gained and 6.3 focus on the role of technology. by moving to the use of annual data, Section 6.4 briefly considers the forces even in panels. Inevitably the short-run at work in convergence, paying especial variation in growth rates is dominated by attention to trade. business cycle effects, not by changes in 6.1 Measuring the Rate of Convergence fundamental long-run growth prospects. Only long time averages of growth Researchers have been keenly inter- rates, compared in the cross-section or ested in the proportionate rate at which using a panel, can allow us to address regions and countries close the gap be- the determinants of long-run growth tween their current positions and their with any degree of confidence. respective steady states. It is not un- common to read strong claims that this 6. Convergence and Technology: rate is remarkably stable, at around 2 What Ha~eWe Learned? percent a year (Barro and Sala-i-Martin 1992; Mankiw 1995; Sala-i-Martin The preceding sections argue that im- 1996). The significance of this lies in its portant improvements can be made to implication that returns to physical and 134 Journal of Econ,omic Literature, Vol. XXXVlI (March 1999) human capital diminish, but only very 6.2 Technology and International slowly. Dqferences in Growth Rates This kind of story about convergence and the extent of diminishing returns, The debate about convergence rates though popular, is not entirely reliable. is linked to another ongoing dispute, The 2 percent estimates are taken from the importance of differences in tech- rather crude cross-section regressions, nology across countries. At one extreme and there is no shortage of accompany- is the work of MRW, who claim that ing econometric problems. Fixed ef- income and growth differences can be fects are either ignored or dismissed a almost completely explained using a priori. Sensitivity to measurement error model in which technology is a public is rarely explicitly discussed or assessed. good, freely available to individuals in The possibility of heterogeneity biases all countries. Differences in incomes or outliers is not considered. More so- originate largely in differences in steady phisticated panel data and time series state levels of human and physical studies suggest that these flaws are cru- capital. cial to obtaining the 2 percent result, Near the other extreme are most de- and recent estimates of the rate of con- velopment economists, economic histo- vergence have varied between zero and rians, and theorists, who see "idea gaps" 30 percent a year.23 as central to the problems of developing The consensus now emerging is one countries. Capital accumulation is often of uncertainty. It is just not very easy to seen as a subsidiary part of the central disentangle the convergence rate from task, which is adopting best-practice other aspects of growth. Arguably this technology from abroad. Researchers in should not be surprising. Conditional this tradition emphasize that many im- convergence implies mean reversion, and portant ideas are protected or secret, so there is a close link between investi- and others can only be acquired by ex- gating convergence and testing for unit perience. An emphasis on the need to roots. We know from the persistence lit- transfer ideas and expertise points to erature just how hard it is to arrive at factors one instinctively thinks of as im- definitive conclusions in this field. portant, such as national culture and Progress is likely to require more institutions. careful thought about fixed effects, pa- Why, then, has the "public good" rameter heterogeneity and the ade- view of technology held such a sway quacy of unit root tests in short panels. over the recent empirical literature? The attendant danger is that re- One argument in its favour is that, at searchers will see technical sophistica- least ultimately, countries should have tion as an end in itself, and lose sight of access to any technical knowledge avail- the reasons for interest in conditional able elsewhere. Hence, in the long run, convergence. At present, too little the rate of technical progress will be thought is given to the value added in the same worldwide, even if levels of studying convergence rates. There are productivity differ because of local other, perhaps easier, ways of learning factors like climate. about the extent of diminishing returns. Unfortunately empirical researchers often jump from this view to a much more controversial position, without 23 See Caselli, Esquival, and Lefort (1996),Paul Evans (1997), Nazrul Islam (1995), Kevin Lee, much explanation. In some hypothetical Pesaran, and Smith (1996)and Temple (1998a). long-run equilibrium technical progress Temple: The New Growth Evidence 135 will be the same across countries. This them in assuming that TFP has grown does not mean, however, that one is jus- at a rate of 2 percent a year in all coun- tified in assuming constant rates of TFP tries since 1960, what does one make of growth in any sample we observe. The the many countries which have grown at key difficulty should be obvious. If de- less than 2 percent a year over this veloping countries often lag behind oth- period? Perhaps these countries have ers in technical efficiency, there is no received a large negative shock to reason to expect their efficiency to grow technology. Otherwise, the logical at the same rate. Furthermore, we mea- implication of the model is that these sure not technical progress but TFP countries are converging to their steady growth, and this may be affected by states from aboue. Any country which such things as instability and war. has grown at less than 2 percent a year Yet the early assumption of similar must have exceeded its steady state technologies and TFP growth rates ap- capital stock in the early 1960s, and peared quite successful. MRW draw at- have been running it down sub- tention to one particularly dramatic sequently. Perhaps this is right, but it finding. They claim that around 80 per- seems unlikely. cent of the international variation in per Other problems abound. Economic capita incomes can be explained using miracles, such as Japan's post-war just three variables: population growth, growth, are hard to explain by capital and investment rates for physical and accumulation alone. In most frame- human capital. The corollary is that dif- works, such a process would have to be ferences in technical efficiency can have accompanied by a steeply falling real in- only a small role to play in explaining terest rate, something that has not been cross-country income variation. observed. To paraphrase the words of However, their approach makes two Klenow and Rodriguez-Clare, the neo- controversial assumptions, that invest- classical revival in growth economics ment rates are exogenous to the level of went too far. Understanding the reasons income and uncorrelated with effi- for differences in technical efficiency ciency. More recently, Peter J. Klenow and TFP growth is essential to the and Andrks Rodriguez-Clare (1997) empirical growth project. have drawn attention to another under- 6.3 Technical Progress in the Long Run lying problem with the MRW finding. The human capital variable that MRW Although it is probably safe to dis- use only captures variation in secondary miss the idea that TFP growth rates are schooling. Since it ignores primary presently the same across countries, schooling, it tends to exaggerate the there are some interesting issues sur- variation in human capital across coun- rounding rates of technical progress in tries. When Klenow and Rodriguez- the advanced industrial countries, con- Clare correct for this, they find that nected to endogenous growth models. MRW's model only explains around half Theorists working in this field argue the variation in incomes, leaving a that the long run growth rate depends central role for technology differences. on the allocation of resources, and Consider some other difficulties with the MRW approach.24 If one follows and Sergio Rebelo (1993), Stephen Parente and Edward Prescott (1994), and Jonathan Eaton and 24Papers relevant to this discussion include Samuel Kortum (1995). For a broader treatment Andrew Bernard and Jones (1996b), Dongchul of technology and growth, see Jan Fagerberg Cho and Stephen Graham (1996), Robert King (1994). 136 Journal of Economic Literature, Vol. XXXVII (March 1999) particularly the amount of resources de- (1998). There is certainly evidence for voted to research. At least in the early such spillovers. David Coe and Elhanan models. one implication is that even in- Helpman (1995) find large effects of d~ist~ialcountries will grow at different foreign R&D on domestic total factor rates for long periods of time. productivity, while Eaton and Kortum One useful piece of evidence here is (1994) calibrate a model of interna- the path of the cross-country variance tional technology diffusion and find of incomes. Evans (1996) starts from that, even for the USA, around half its t!lc observation that, if countries differ productivity growth depends on foreign in their trend growth rates, the loga- technology improvements. This sug- rithms of their per capita incomes gests that the evidence for a common should wander away from each other long-run growth rate can be seen as over tirile. More precisely, it can be consistent with endogenous growth. shown that the cross-country variance 6.4 What Dri~esCon~ergence? of log incomes will be integrated of order one around an upward quadratic If we agree that there is more to in- trend. Evans presents evidence for thir- come variation than difference~in in- teen industrial countries that the vari- vestment rates, it is interesting to ask ance has not trended upward and that, why convergence might be taking place. if anything, the reverse has occurred. If we allow for technology differences, This clearly corresponds to the ear- the negative coefficient on initial in- lier finding of convergence among de- come requires careful interpretation. It veloped countries. It tends to suggest may indicate not just that countries are that at least the advanced industrial converging to their steady -states countries are growing at the same rate through capital accumulation (as MRW over the long run. Evans writes that proposed) but also that technology "either endogenous growth models transfer is taking place. are fundamentally flawed, or else the In this case, convergence studies will cffects they predict must be rela- be an especially poor way of learning ti1 ely unimportant for the countries about the extent of diminishing returns, considered here ." and there is no longer a theoretical rea- Since his paper, some work has sug- son to expect the relation between gested that rates of technical progress growth and initial income to be linear. may Lary even across industrial coun- Several writers, most recently Easterly tries.25 It is also worth pointing out and Levine (1997), have modelled that, even if countries do grow at the growth as a quadratic function of initial same rate in the long run, new growth income. A common finding is that mid- theories may still provide the explana- dle income countries have grown relation of the world growth rate. For tively fastest once controlling for policy empirical applications of endogenous differences. growth theory, one is likely to need a Some studies have looked directly at model in which domestic research is ac- convergence in TFP, particularly for companied by knowledge spillovers the OECD. Given links through trade, from other countries, as provided by one might expect to find TFP conver- Philippe Aghion and Peter Howitt gence between sectors in different countries, perhaps especially in manu-

25 See in particular Lee, Pesaran, and Smith facturing. In practice, there is only (1997). weak evidence that efficiency in OECD Temple: The New Growth Evidence 137 manufacturing sectors converged in a small one. It invites many questions, the 1970's, and some evidence of di- not least concerning the relati1.e impor- vergence in the 1980's. This suggests tance of particular inputs, and the role that the aggregate TFP convergence is in growth of wider influences, including driven by the performance of services.26 some that have long fascinated acadt'ln- The link between trade and conver- ics and policy-makers, like free trade gence is of great interest, partly be- and democracy. cause trade theory provides other mech- In this section I will assess the rolc of anisms, like factor price equalisation, various inputs, or what might be callecl through which countries might con- the 'proximate' sources of growth: in- verge. There is some suggestive evi- vestments in physical and human capi- dence of decreasing income dispersion tal, and in research and development, among countries linked by trade, which The next section will turn to the n.id~r can be partly attributed to the closing influences. of technology gaps. 7.1 In~estmentin Physical Capital Finally, it should be noted that trade links between countries may undermine As most undergraduates know, the many of the traditional beliefs about Solow-Swan model taught us that long- conditional convergence. Jaume Ven- run growth is likely to be independent tura (1997) points out that there is no of the investment rate, because the re- necessary link between diminishing returns to accumulable factors are prob- turns and conditional convergence in ably sufficiently diminishing for this to models of trading economies. If a weak be true. The same students were often form of factor price equalisation holds, taught that no correlation between in- the law of diminishing returns applies vestment rates and growth is observable only to the world's stock of capital, and in the data. We now know that was investment will be equally productive in wrong. There is a robust correlation be- each country. Hence one would not tween investment rates and growth, necessarily expect to find any link be- which sometimes survives the use of in- tween growth and initial income, at strumental variables. A key point is that least when controlling for investment. over a finite time period, this correla- Thus Ventura's work reinforces the tion is to be expected even in the idea that the usual conditional conver- Solow-Swan model, and its strength can gence result may partly be driven by be used to estimate the share of capital technology transfer. in the production function. However, some argue that if the endogeneity of investment could be 7. The Proximate Sources of Groulth correctly dealt with, the effect of investment would be close to zero in Overall, my reading of the literature developing countries, because much of is that growth differentials are partly it takes place in heavily distorted envi- explained by varying degrees of technol- ronments. As discussed in Section 4.4. ogy transfer, as well as variation in in- there are several reasons for belie\-ing put growth. It is a useful first step, but that endogeneity is important to the in- 26See the papers by Bernard and Jones (1996a) vestment and growth results. and also and Steve Dowrick and Duc-Tho Nguyen (1989) several problerns in overcorning its for more on TFP convergence. On trade, see Dan Ben-David (1993) and Ben-David and A. K. M. presence. Atiqur Rahman (1996). One interesting reason for believing 138 Journal of Economic Literature, Vol. XXXVZI (March 1999) that endogeneity is important is the lit- Given the apparent robustness of the erature's general ranking of estimated results, attention should perhaps shift returns to different types of investment. to their interpretation. Perhaps the re- For instance, the returns to foreign di- turns to equipment investment are rect investment are often found to be overestimated in the presence of indus- extremely high, but it also seems likely trialization. Another argument is that that here the endogeneity problem is equipment investment is an important greatest: foreign direct investment will part of technology transfer, and the be pulled into countries already doing returns to it then appear to be high be- well, or expected to do well in the fu- cause other costs of technology trans- ture. The usual ranking of returns to fer (technical adaptation, training, re- different types of investment tends to ,organization) are omitted. Both these suggest that endogeneity is driving the points may have something to them, estimates.27 and there remains considerable scope The strongest result in the invest- for further work. It is not as easy as it ment-growth literature is that the re- once looked to dismiss the findings of turns to physical capital are almost cer- De Long and Summers. tainly diminishing, in agreement with 7.2 Hurnan Capital the Solow-Swan growth model and most theoretical work since. This is the find- Harold Alderman et al. (1996) point ing of both convergence regressions and out that developing country govern- cross-country growth accounting (Ben- ments spend over $100 billion a year on habib and Spiegel 1994, King and education, health, and other human Levine 1994). capital investments. It is clearly impor- The finding that returns are diminish- tant to understand how these investing does not rule out important exter- ments contribute to growth. There is a nalities to investment. The work of De long tradition of microeconomic studies Long and Summers (1991) has drawn of the returns to schooling, but given attention to the potential importance of recent emphasis on externalities to investment in equipment, not without education, macroeconomic studies also controversy along the way. A consensus have a role to play. has emerged that the correlation be- The MRW paper seemed to provide a tween equipment investment and broad-brush answer to questions about growth appears to be weak in the the role of human capital in growth. Us- OECD, but the debate over this has dis- ing the proportion of the adult popula- tracted attention from its role in devel- tion enrolled in secondary school as a oping countries, for which the findings proxy for human capital investment, of De Long and Summers appear to be they argued that aggregate technology rather robust.28 It is difficult to explain is simply described by a Cobb-Douglas away their results by simply referring production function taking human capi- to influential outliers or measurement tal as one of its inputs (the augmented error concerns, although endogeneity al- Solow model). An estimated exponent most certainly accounts for some portion of 1/3 on human capital is, they of the very high estimated returns. claimed, consistent with both the estimated rate of convergence and sketchy "The arguments in this paragraph are due to evidence on the importance of human Lant Pritchett. 28For more on equipment investment, see capital to the United States. Temple (1998b) and the references there. Papers since then, however, have Temple: The New Growth Evidence 139 questioned whether the links between school enrollment rates are intended to human capital and growth are quite so represent a flow of investment in hu- simple. We have already seen that the man capital, or its stock. In practice MRW analysis perhaps overstates the these rates may be a poor proxy for variation in output per head that can be either, and given that data on average explained by variation in human capital. years of schooling in the population or Another problem to emerge is that labour force is now available, the con- changes in human capital appear to ex- tinuing use of school enrollment figures plain little of the variation in changes in has little to recommend it.30 output, casting doubt on the augmented An alternative approach is to con- Solow formulation.29 This macro- struct indices of human capital based on economic evidence conflicts with the the returns to schooling and experience finding of the micro literature that indicated by micro studies. The basic schooling has a significant return in idea is to combine schooling coeffi- terms of higher wages. The failure to cients from Mincerian wage regressions discern this effect at the macro level is with data on years of schooling, and worrying. then examine the contribution made to The literature uses somewhat dubious cross-country income differences by proxies for aggregate human capital. variation in this human capital measure. The focus is almost exclusively on This approach is useful, but it is neces- schooling rather than training. This is sarily silent on the extent of human mainly due to data limitations, but less capital externalities, and tends to as- excusable is the attention frequently sume that returns to schooling are not paid to school enrollment rates. These driven by signalling effects. rates were initially regarded as one of As this discussion indicates, there is the more robust and satisfactory vari- much work still to be done on the role ables in the growth literature, but it is of human capital. One problem is that worth remembering that they are likely the aggregate data seems to be too to be positively correlated with initial crude to ask some of the most interest- efficiency, so the results could often be ing questions: for instance, what mat- spurious. Certainly it has been much ters most, breadth of access or school harder to find an effect of human capi- quality? Other questions surround the tal in panel data studies, although it is role of health status. Variables like life also true that too few researchers think expectancy are often used in growth carefully about the specification. Rather regressions, but their role is never optimistically, they tend to expect a justified by a well-articulated theory. change in school enrollments to raise One interesting possibility is that growth almost instantly. good health status raises the return to Equally importantly, there are some education and training. conceptual difficulties with the use of Finally, there is widespread agree- school enrollment data. Only rarely do ment that human capital accumulation these rates correspond well to the is not a sufficient condition for growth. human capital variables highlighted in 30Those not convinced should consult Norman theoretical models. In many empirical Gemmell (1996) and Pritchett (1996b). Eric growth papers it is not clear whether Hanushek and Dongwook Kim (1995) consider another alternative to average years of schoolin 2gTwo important pa ers here are those by index of cognitive skills based on internationa&eas: Benhabib and Spiegef (1994) and Pritchett scores. Its determinants are investigated by Jong- (1996b). Wha Lee and Barro (1997). 140 Journal of Economic Literature, Vol. XXXVZZ (March 1999)

Most development economists are fa- effect, it might be thought that they miliar with examples of countries that typically overstate the social returns to have expanded education only to see research. This intuitive reasoning may high subsequent unemployment rates be incorrect. It is possible that distor- and falling returns; Korea before 1960 tions like creative destruction and mo- might be one example. In other devel- nopoly power affect the share of R&D oping countries, wage employment is in output, but do not affect the relation- sparse and the highly educated have ship between the social rate of return often found work only in the public sec- and the R&D share. Hence regressions tor, perhaps undermining the overall based on the R&D share may recover impact on growth. Hence a key chal- the social rate of return reasonably ac- lenge is to elicit the conditions in which curately, and perhaps even understate it expanding education is most beneficial. (Jones and John Williams 1997). Although the microeconomic evi- 7.3Research and Development dence suggests that theorists are right Of the "proximate" sources of growth, to emphasise R&D, some researchers the last to be considered is research and have criticised the research-driven development. Since the modelling of in- growth models on empirical grounds. dustrial research lies at the heart of Central here is the work of Jones some of the most promising new growth (1995a,b). He points out that growth theories, it is worth trying to clarify em- rates in OECD countries since World pirically the role of R&D, and I will War I1 have not shown any persistent take particular care here to relate such upwards trend in spite of policy findings to recent growth models. I will changes: trade liberalisation, an in- follow convention and common sense in crease in average years of schooling, in- assuming that research is of greatest creases in investment, and a substantial importance for advanced industrial increase in R&D efforts. economies. This apparent constancy of long-run As in the study of human capital, this growth, in the face of structural is a field in which there is already a changes, obviously calls into question wealth of microeconomic evidence. the validity of research-driven growth Early studies found private rates of re- models. Jones argues that such models turn to R&D as high as 30-50 percent are driven by a knife-edge assumption for the USA in the 1950's and 1960's. in the way that the benefits of research There are some well-known problems in are related to the existing stock of measuring the contribution of research knowledge. Although a more general re- to productivity growth. Overall, though, lationship can explain the evidence, so studies have concurred in finding high can other modifications, not all of them private returns. Estimates of R&D spill- ruling out research-driven models. overs are necessarily flawed and subject There is a well-known stylized fact to a variety of reservations, but the im- that the productivity of research, at pression remains that social returns are least as measured by realised patents, even higher.31 shows a long decline from the mid- Since empirical studies ignore such 1950's until the mid-1980's. This may externalities as the "business stealing" indicate some systematic relationship between the knowledge stock and 3lTwo particularly useful references on research and productivity growth are Griliches research benefits, or it may just repre- (1979, 1992). sent an exogenous fall in research Temple: The New Growth Evidence 141 productivity. Certainly there is little di- ferent to those in countries already near rect evidence of diminishing returns to the technological frontier, one should research. In general, one might ex- often be careful about extrapolating pect the progress of knowledge to be findings from the developing countries uneven, depending as it does on the to the more developed, and vice versa. emergence of new "fishing grounds." The influences considered here in- Since some of the underlying conclude population growth, the financial cepts (knowledge, ideas) are so hard to sector, the macroeconomic environ- pin down, the validity of research- ment, government spending, income driven growth will be a difficult debate distribution, and political and social to resolve.32 Taking sides is not as im- arrangements. Most of the findings portant as it might first seem. The discussed in this section are from cross- model Jones proposes allows research to country empirical work. Throughout, I have large level effects, and this is suffi- concentrate on findings which seem to cient for an interest in the policy effects have survived the attentions of a variety of research incentives. It is always of researchers, using different data sets worth remembering that "the long run and specifications, and where at least growth rate" is a theoretical abstrac- some thought has been given to prob- tion, never observable in practice. De- lems such as endogeneity, outliers and bates about whether or not policy af- model uncertainty. fects it will distract us from questions Since I concentrate on synthesizing that are ultimately of more practical previous work, there are some impor- importance. tant influences on growth that are not given the coverage they are perhaps 8. Wider InjTuences on Growth due. Chief among these are microeconomic structure and the fine detail So far, we have considered three of institutions. Although the new proximate sources of growth: physical growth evidence has covered trade pol- and human capital accumulation, and icy and political institutions, there are spending on research and development. other aspects which are harder to quan- In this section, I widen the survey, to tify. Few of the variables considered assess the contribution of macro- here would offer much insight into the economic evidence to our under- experience of China or the former com- standing of wider influences on growth. mand economies, for example. It may Sometimes these will work through the well be that these issues are better ad- proximate sources, explaining the cross- dressed by case studies, or by statistical country variation in input changes. work within countries at the firm or They may also have a direct effect sectoral level. through total factor productivity, for 8.1 Population Groujth instance by affecting technology transfer. In examining these effects, I tend to One area in which aggregate studies focus on developing countries. Since may make an especially useful contribu- institutions and the growth process in tion is the relation between population developing countries may be very dif- growth and economic performance.33

32 In an ambitious and important paper, Ricardo "The microeconomic evidence is often felt to Caballero and Adam Jaffe (1993) make some be unsatisfactory because the variables of interest, rogress in identifying proxies for ideas and like education and fertility, are endogenous within Enowledge spillovers, a household demand framework. Journal of Economic Literature, Vol. XXXVIZ (March 1999)

Most recent growth researchers have in- fully about causality. If people perceive cluded population growth as a variable that incomes are likely to rise, and pos- of subsidiary interest, perhaps in the sibly the returns to human capital, they manner suggested by MRW, and then may decide to have fewer children. An- noted a weak negative correlation be- other worry is that fertility decisions tween it and the growth of per capita appear to be quite strongly correlated income. This raises some endogeneity with various wider aspects of social de- concerns, although typically one might velopment. It may be that high fertil- think of causality running to population ity rates, rather than having a direct growth from the level of per capita negative effect, simply proxy for so- income rather than its rate of growth. cial arrangements that are relatively Most economists, asked to suggest "traditional" in other respects.35 negative consequences of population Overall, the popular belief that popu- growth, would probably argue that aver- lation growth is economically harmful is age human capital and capital-labour not yet well supported by statistical ratios are likely to fall. In the context of evidence, and overall the new growth MRW's work, there is a negative effect evidence has done little to modify the of population growth on the steady state conclusions of the influential survey by capital-labour ratio for a given invest- Allen Kelley (1988). He argues that ment rate, but it makes only a small some government policies are likely to contribution to the observed variation exacerbate the negative effects of popu- in levels of income and growth rates. lation growth, and this would be an in- There is some evidence that students in teresting area to explore. It should also countries with higher population growth be remembered that population growth record lower achievement (Hanushek may have adverse consequences be- 1992). There is also a weak negative yond those for per capita income, on relation between population growth and such things as access to safe water, and changes in total factor productivity, environmental quality more generally. perhaps something of a puzzle for 8.2 Trade and Growth theorists. More generally, it is clear that in- Understanding the interaction be- depth studies are needed to address the tween trade policies and macro- links between population growth and economic performance has traditionally macroeconomic outcomes. The work so been one of the central concerns of de- far has established one key finding: the velopment economics. Often it seems small negative effect of population that microeconomic studies allow much growth on per capita growth is partly sharper discrimination between hy- mediated through changes in labour potheses than the aggregate studies of force participation.34 These shifts in the trade and growth. However, researchers composition of the labor force may have have long been aware that micro studies only second-order welfare effects. often miss the economy-wide resource Some researchers have looked at the allocation effects that may be central link from fertility rates to subsequent to understanding the effects of trade growth, sometimes finding a negative policy. correlation. Again, one must think care- This awareness partly explains why 35Temple and Johnson (1998) discuss the possi- 34The key papers here are James Brander and ble role of social factors. See also Section 8.8 Dowrick (1994) and Pritchett (1996a). below. Temple: The New Growth Evidence the first cross-country studies of open- do not emerge as central except for ness and growth considerably pre-date Singapore and Malaysia. much of the rest of the empirical These articles are exceptions, and if growth literature. The early start does macroeconomic studies are to be genu- not seem to have conferred much ad- inely useful, the measurement and si- vantage, and in this area as in others, multaneity problenls often need to be the usual suspects-endogeneity, speci- better addressed. One way of doing this fication problems, outliers-are still at will be through finding natural experi- large. Since Sebastian Edwards (1993) ments. As an example, Ben-David covers the field in some detail, I will (1993) documents the observed conver- only briefly consider the issues. gence effect of trade liberalization by One of the most important difficul- the European Economic Community, ties is finding some way of quantifying and shows that it represents neither trade regimes. The measures which are a continuation of a long-run trend most defensible on theoretical grounds, nor a return to the pre-war extent of such as effective rates of protection, dispersion. can be difficult to calculate for a suffi- This kind of work may clarify the con- ciently large number of countries. Typi- ditions under u7hich openness is most cally, researchers fall back on simple beneficial. Recent endogenous growth proxies, such as trade shares in GDP or models have emphasised that open the black market exchange rate pre- economies, if they specialise according mium, meant to give some indication of to comparative advantage, may do less openness. Sachs and Warner (1995a) well than under autarky. There are create a binary dummy variable based several qualifications to these argu- on judging trade regimes against a ments, especially before translating set of such criteria. They find some them into activist trade policy, but re- striking results, including unconditional cent empirical work indicates that the convergence among the group of open ideas have some relevance for develop- economies. ing countries.37 The gains from open- Causality is another key problem in ness mav be greatest for countries al- the study of trade and growth. It is easy ready specialising in manufacturing to think of ways in which fast growth exports, like those in East Asia. may lead to, say, an increase in the 8.3 Finance and Growth trade share.36 One interesting possibility is to use trade shares predicted by a If cross-country work is going to be gravity model as instruments for actual useful, it should sometimes succeed in trade shares (Jeffrey Frankel, David shifting priors. One area in which this Romer and Teresa Cyrus 1996). They may have already happened is the em- find that this tends to raise the sug- pirical literature on finance and growth. gested importance of trade. However, Some prominent economists have the gravity model predicts high shares tended to dismiss the role of financial for most East Asian countries, so poli- factors, arguing (from rather little evi- cies for outward orientation (reflected dence) that financial development is in actual trade shares above predicted) simply a passive consequence of growth. There are now a number of papers 36Some researchers work with initial shares; a remaining problem is that these initial shares may 37 See for instance the evidence of Sachs and be correlated with unmeasured aspects of domes- Warner (1993b) that natural resource abundance tic policy. works against long-run growth. 144 Journal of Economic Literature, Vol. XXXVII (&larch1999) looking at the effect on growth of initial selection of other variables. A common financial development, including both conclusion from this literature is that banking system and stock market vari- although "policy matters" we do not yet ables. Taken together, these studies have any clear idea which elements of make those traditional arguments policy are crucial. harder to justify. Given the importance of this informa- The survey by Levine (1997) covers tion for policy-makers, our lack of prog- this area in some detail, so here the ress on this question is disappointing. issues are only briefly addressed. To be Yet it may be an inevitable consequence useful, regressions of growth on finan- of cross-country studies. The central cial systems must take into account the problem is that things tend to go wrong endogeneity of financial development. together: high inflation is accompanied Most studies, including the influential by political instability, exchange rate work of King and Levine (1993), have volatility, and so on, and disentangling done this. Hence perhaps the major the various contributions is not easy. worry with these results is the fixed Given that timing evidence would be effects difficulty discussed earlier. very useful, this is one field in which a As in the study of trade effects, cau- very strong case for historical case stud- sality problems can sometimes be ad- ies, in preference to regressions, can be dressed by natural experiments. In made. particular, it will be essential to investi- It is also worth pointing out that gate the consequences of financial many of these growth-and-policy stud- liberalizations across developing coun- ies control for the accumulation of tries, including the rise of stock mar- physical capital. It may be that the most kets. Work in Levine and Sara Zervos important and robust links between pol- (1997) suggests that active stock mar- icy and growth are mediated through kets play a role in subsequent growth, investment. Tbe focus in the policy and since they examine the value of literature could perhaps shift to ex- trades while controlling for capitaliza- plaining cross-country differences in tion, as well as looking at turnover, investment. their results are unlikely to reflect The links between inflation and only the forward-looking nature of stock growth are particularly controversial. prices. Most studies have looked at the cross- section but it is increasingly clear that 8.4 Short-Run Macroeconomics such results can be distorted by just one One of the most controversial areas is or two high-inflation outliers. Michael the link between short-run macro- Bruno and Easterly (1998) chose in- economic management and long-run stead to take a short run approach, and growth. It is not difficult to present re- examine the path of output during high gressions which show significant corre- inflation crises. They found that high lations, of the expected signs, between inflation crises are associated with out- the growth rate and variables such as put losses, but that output returns to budget deficits, inflation, and real ex- the same long-run growth path once in- change rate instability. It is rather flation has been reduced. This may be harder to isolate any particular policy the reason for the weak inflation and variable and demonstrate that it has a growth relation in the cross-section. robust correlation with growth, regard- From a methodological point of view, less of endogeneity concerns and the some of the most interesting studies are Temple: The New Growth Evidence 145 those making use of the time series ernment and fast growth leaps out fro111 properties of inflation and output. The the data. strategy is to emphasize that inflation is In particular it is worth noting that highly persistent, to the extent that it government consumption is one of the may be I(1).This finding implies that, variables whose correlation with growth should inflation affect growth, some is identified as fragile by Levine and shocks to the growth rate will be perma- Rerlelt (1992). As for the effect of social nent, a result that is rarely found. security transfers, Anthony Atkinson Clearly strong coriclusions here are (1995) points out that the detailed built on a somewhat shaky foundation: structure of welfare state institutions is the finding that inflation may have a likely to be crucial, and cross-section unit root inevitably uses tests with low studies will only succeed in obscuring power. the most important issues.38 Arguably more reliable is the re- Overall, this seenis to be one of those search on short-run output volatility fields in which stylized facts appear and long-run growth. Usually econo- hard to come by, at least at the aggre- mists have reserved separate compart- gate level, Microeconomic evidence on ments for the analysis of business cycles labour supply and investment responses and growth. Maybe that was wrong: to changes in tax rates is likely to be some recent theories and evidence sug- more fruitful. Perhaps the only promis- gest that there are deep connections being macro approach is that of Elall and tween the two. There appears to be a Jones (1997), who find that high gov- negative relationship between output ernment consumption lowers the level volatility and growth in the OECD, and of income, and also point out that in in sarrtples that combine the OECD and their framework, endogei~eit~is likely developing countries (Garey Ramey to mean that the effect is understated. and Valerie Ramey 1995). As yet, the interpretation of the findings is unclear, 8.6 Government Spending on but this is a fascinating area for further Infrastructure theoretical and empirical work, especially on the links between uncertainty More progress has been made on the and growth. question of the composition of government spending. There has been intense 8.5 Gouet-nment Size debate surrounding the role of public Anlong questions that are very impor- capital in the growth of developed tant but likely to remain largely unset- countries (Edward Grainlich 1994). tled, that of government size and This debate has a quieter counterpart in growth looms large. In political discus- the cross-country literature. Although sion it is common to hear claims that a there is a lack of data on the quantity high ratio of social security transfers to and quality of public infrastructure in GDP and a high level of government developing countries, measures of tele- consumption can be damaging to phone networks and electricity capacity growth prospects. The evidence is not have been found to have a significant strong. Some researchers find a nega- effect on subsequent growth. Easterly tive link between government consump- and Rebelo (1993) find that the share of tion and growth, but overall studies dis- public investment in transport and agree, and it would be wrong to argue 3sOther important discussions are Easterly and that a correlation between small gov- Rebelo (1993) and Joel Slemrod (1995). 146 Journal of Economic Literature, Vol. XXXVII (March 1999) communication is robustly correlated pressure to redistribute income in un- with growth. Infrastructure spending equal societies, because of the relatively appears to raise the social return to on- low income of the median voter. The ar- going investment, rather than spurring gument has the advantage of clarifying further efforts. a key point: just because high inequality There is an endogeneity problem lowers growth does not necessarily here, and one that instrumental vari- mean that one should redistribute. It ables usually cannot solve. Infrastruc- may be redistribution which drives the ture investments may be made in antici- correlation in the first place. pation of fast growth, rather than That said, the political economy argu- playing a causal role. Even so, it seems ments have not done particularly well easier to make headway on this issue empirically. There is not much evidence than on the more controversial one of that fiscal policy is strongly related to government size. inequality. Testing a central implication, that the inequality-growth relation 8.7 Inequality differs between democracies and non- democracies, is made difficult by the One of the most active areas of cross- fact that stable democracies are usually country research has been investigating found only among the richer countries. the consequences of inequality for Hence the literature seems to be mov- growth. Somewhat unusually for the ing away from the political economy growth literature, studies have tended line, towards an examination of the ef- to concur in finding a negative effect of fects of inequality on fertility rates, in- high inequality on subsequent growth.39 vestment in education, and political The evidence has not been accepted by stability. all: some writers I~oint out the concen- tration of richer countries at the lower 8.8 Social and Political Factors end of the inequality spectrum, the Some of the most interesting thinking poor quality of the distribution data, on economic growth is to be found on and the lack of robustness to fixed ef- the borders of political science and fects specifications.40 At least, though, sociology. Of the two areas, political it has become extremely difficult to factors have received more considera- build a case that inequality is good tion in the empirical growth literature, for growth. This in itself represents a probably because they lend themselves considerable advance. to measurement more easily, and be- Given the indications that inequality cause the lines of causation are better is harmful for growth, attention has understood. moved on to the likely mechanisms. The A common approach has been to re- best known line of argument is the po- late growth to indices of civil, political, litical economy one, in which demo- and economic rights. Studies of this cratic governments are under more kind tend to leave open the question of 39Roland Benabou (1996) surveys this work in which type of rights ire most important. depth. Roberto Perotti (1996) is another very good have argued that political and introduction, and the summary here relies heavily some on his evidence and ar uments. economic rights are complementary, 4ONote that fired efiects specifications may not but there are counter-examples of be particularly useful here. Inequalit is usually autocratic regimes that have extended found to be fairly stable over time, andthe lags in its effects may well be too long to be picked up by economic freedoms. the panel data analyses that have been carried out. At present, economic freedom seems Temple: The New Growth Evidence 147 to have a stronger observable link to mocracy; since, however, the indices growth than political rights. Re- devised by different methods are often searchers use indicators of property highly correlated, there does remain rights provided by country risk evalua- some hope for this research.4" tors to potential foreign investors.41 The Also interesting is the role of social indicators, which evaluate such things and political instability. The excellent as the ease of enforcing contracts and survey by Alberto Alesina and Perotti the risk of expropriation, are found (1994) concludes that it is here, rather to have a relatively strong connection than in the question of democratic free- to growth. Other studies of institu- doms, that the most important political tional factors tend to find that, even influences on growth are to be found. when good institutions do not raise Simple proxies for political instability growth through technical progress, they have been in use in cross-country re- often have an effect in encouraging gressions almost from the start. The key investment. problem with most results is that There is an emerging consensus that growth and stability are likely to be the link between democracy and growth jointly determined. However, papers is difficult to isolate. Casual empiricism that explicitly address this difficulty suggests a wide variety of experience tend to confirm the idea that instability under both authoritarian and demo- has a strong negative effect. One issue cratic regimes. Some studies have found for further work will be to draw a positive effects of political rights, but as clearer distinction between the effects many if not more fail to find a signifi- of civil war and those of collective cant correlation.. Barro (1997) suggests protest. that this may be because the relation is Also interesting is the impact of so- nonlinear: extensions of political rights cial arrangements on growth. Laymen appear to be beneficial only up to a are often willing to suggest that there certain point. Ultimately there is not a are sometimes fundamental social and great deal in the empirical literature to cultural barriers to growth, and there is support the positions either of instinc- support for this idea in Marx, Weber tive democrats or those who advocate and Kuznets, among others. More re- "strong states" isolated from the popu- cently, theoretical papers have started list and lobbying pressures sometimes to appear that draw out the relation associated with democratic freedoms. between social arrangements and vari- It may be that the type of regime is ous kinds of incentives. Some econo- far more important-whether, for in- mists distrust an emphasis on "social stance, those running autocratic re- capital" or "social capability" since, as gimes are bent on self-interest and Moses Abramovitz (1986) has pointed accumulating wealth, or whether they out, "no one knows just what it means are oriented to national economic goals. or how to measure it." In general, empirical assessment of Researchers are starting to make political regimes clearly presents some some progress on the measurement formidable measurement problems. issue. One promising approach is to There are many conceptual and mea- model social and political influences on surement traps just in quantifying de- growth as latent variables related to a

41 Evaluations provided b private consultancies have been used to good efLct b Stephen Knack 42Kenneth Bollen (1990) discusses measure- and Philip Keefer (1995)and Paoyo Mauro (1995). ment issues in more detail. 148 Journal of Economic Literature, Vol. XXXVZI (March 1999) variety of observable indicators. De- on the likely contributions of a set of spite some interest from development variables. economists, simple techniques for data For the last few years, commentators reduction like factor analysis and princi- have been claiming that the Summers- pal components have been largely ig- Heston data set has been mined to the nored by recent growth researchers.43 point of exhaustion, and implying that Their use seems to have a great deal of there is nothing new to be learned from potential, and the renewed interest in further work of this kind. From some social factors aligns well with recent published comments, it seems that "re- theoretical work, reinforcing the case gression fatigue" has set in, so that for further study. hearts sink when yet another dubious growth regression is presented. Despite 9. Where Next? this pessimism, researchers in the field continue to show some ingenuity in In summarising the recent literature, finding new and interesting variables to I have often indicated areas in which combine with the data set. Recent stud- there is much scope for further work, ies have looked at, among other things, and the direction it might take. In this the growth effects of natural resource section, I will consider some broader abundance, social factors, protection for questions about the future direction of intellectual property rights, the quality the empirical growth literature. There of infrastructure, and the composition seems to be general agreement that em- of government spending. pirical testing of research-driven growth The genuine worry is not that the models, perhaps building on earlier re- core data set has been exhausted, but search in applied microeconomics, has that this kind of relatively simple aggre- much to recommend it. There is far gate work may drive out other promis- less consensus on the future status of ing approaches, in a kind of Gresham's cross-country empirical work. Law of Growth Studies. Clearly, there is a danger of this happening. Yet those 9.1 The Prospects for Dfferent Methods who try to rule out cross-country em- The reasons for this lack of consensus pirical work rarely seem to have should be clear by now. Even the most thought much about the problems of al- enthusiastic proponent of cross-country ternatives. There is a now routine call regressions must acknowledge that we for more detailed studies of individual are a daunting distance from the ulti- countries' experience, but we have had mate goal, a model with high explana- studies of this kind for years, sometimes tory power which indicates with preci- without being able to draw many useful sion the relative contributions of generalizations. In particular, it is al- different influences. For the most part, most impossible, from the experience of researchers are satisfied with isolating just a few countries, to quantify the ef- relationships that are statistically sig- fect of different variables on long-run nificant and in some sense quantita- growth. tively 'strong'. Few attempt the more Even so, historical analysis must be a ambitious task of trying to place bounds major way forward. Amongst other things, historians may be able to iden- 43 Studies makin use of latent variables include tify some interesting natural experi- Irma Adelman an% Cynthia Taft Morris (1968), Eberhard Scholing and Vincenz Timmermann ments. There are important examples of (1988), and Temple and Johnson (1998). influential historical growth studies, Temple: The New Growth Evidence 149 notably those of experiences with dif- in more concrete terms, we know that fering trade policies. There are areas physical capital investment has been which have been largely neglected by important to growth in Singapore, but the new growth evidence, like the role why has Singapore's investment ratio of external debt, where the use of case been so high? We have learnt some- studies is essential. Historical analyses thing about the general question from may be particularly useful in identifying Levine and Renelt (1992) and Barro reasons for the instability of growth and Sala-i-Martin (1995), among others, within countries; as noted earlier, the but one persistent difficulty is that history of many developing countries explanations for differences in, say, has been marked by alternating booms investment ratios, tend to be rather and growth collapses (Pritchett 1998). ad hoc. Compared to growth regres- As the earlier quotation from Ger- sions, there is even less consensus on schenkron was meant to suggest, those the variables that are essential to with a knowledge of history can identify explaining differences in rates of factor factors of potential relevance and gen- accumulation. erate interesting hypotheses, and thus There is another issue that should be the two approaches have much to learn central to the literature, but which is from each other. So far there has been rarely even acknowledged. Develop- disappointingly little dialogue.44 ment economists have long distin- Traditional growth accounting can guished between different types of also be useful for certain questions. growth, and their varying consequences These methods have been put to good for welfare. Again putting the question use in describing the experience of in more concrete terms, we know that small groups of countries, as in the East Asian countries have managed to work of Alwyn Young (1992, 1995) on combine fast growth with low inequal- East Asian growth. The next task will be ity, but we lack a good understanding of to try and derive comparable TFP how this type of growth has been growth figures for a larger number of achieved. This kind of question is a cen- countries, a necessary first step in ex- tral challenge for the many develop- plaining the variation in productivity ment economists, economic historians, growth across countries. theorists and empirical macroeconomists currently working on growth. 9.2 Open Research Questions Connected to this is the fundamental question of whether growth will raise There will continue to be a place for welfare. Useful progress can be made cross-country work. With new and more here by relating indices of "human de- thoughtful methods, there is much left velopment" and quality of life to macro- to learn. As pointed out earlier, we may economic variables. The intuition of have a better idea of how factor in- most economists is that there will be a puts influence growth, but our under- strong connection between welfare indi- standing of the international variation ces and per capita income, but given the in factor accumulation is still weak. Put controversy surrounding certain poli- 44A notable exception is the book edited by cies, such as those for structural adjust- Nicholas Crafts and Gianni Toniolo (1996). This ment, there is plenty of room for fur- represents perhaps the first major attem t to inte- ther research. Easterly (1997) finds that grate the lessons of the new rowth tReory and evidence with the economic fistory of the ad- the effects of growth on a wide range of vanced countries. indicators are surprisingly uneven. 150 Journal of Economic Literature, Vol. XXXVII (March1999)

As well as classifying and analysing ized facts; more work on deriving and differing types of growth, development testing implications would clearly be economists have sometimes found it useful. At present, it is difficult to useful to draw distinctions between distinguish these models empirically types of countries. Ranis and John Fei from more orthodox explanations for (1988) argue that it is important to underdevelopment. move away from characterizing the Another issue that has not been suffi- "average" developing country, and ciently addressed in the macroeconomic work towards a deeper understanding literature is that of structural transfor- of differences. Along similar lines, mation. The study of sectoral shifts has Pritchett (1997) and Quah (1997) have a long history in development econom- emphasised the need to acknowledge ics.4"ew policy recommendations have heterogeneity, and move away from been forthcoming, which has perhaps techniques based on "representative" led some macroeconomists to neglect economies. structural change. Given that many of Quah's work makes clear that it is the countries under study go through possible to approach the data with more significant changes as they industrialize, imaginative methods than are usually it will sometimes be important to take adopted in the cross-country literature. this into account if conventional cross- In general, careful thought about how country work is not to give misleading countries and their experiences differ, results. For instance, Temple and Hans- drawing on historical studies, may sug- Joachim Voth (1998) argue that this gest new and useful ways of examining problem may have occurred in the study the data. One example of a thoughtful of equipment investment. approach is Easterly (1994). He points out that a surprisingly high number of 9.3 The Role of Aggregate Production countries have stagnated at some point Functions since the 1960s, and attempts to con- At present, there do remain some struct an empirical model explaining long-term concerns with the whole stagnation. cross-country research endeavor. The Similarly, Rodrik (1998) seeks to ex- mention of structural change leads plain why growth collapsed in many naturally to questioning the relevance countries in the 1980s. Given the insta- of aggregate production functions. Ar- bility of growth, noted in Section 2, in- guably the aggregate production func- vestigations of growth variation over tion is the least satisfactory element of time could have a high payoff. This is macroeconomics, yet many economists one area where case studies and panel seem to regard this clumsy device as es- data analyses may both have something sential to an understanding of national to contribute, allowing identification income levels and growth rates. of the key variables that determine One could debate the need to divide whether or not growth can be sustained. growth between the accumulation of in- Another challenge will be to identify puts and improvements in technology. the empirical relevance of theoretical For some questions of interest, this work on multiple equilibria, and exam- division is unnecessary, and should ine the possibility of "development traps." At present, theorists justify their 45This work is surveyed by Hollis Chenery (1988). A paper by Robinson (1971) estimates frequent emphasis on multiple equilib- cross-country regressions allowing for structural ria with some rather rudimentary styl- change. Temple: The New Growth Evidence probably be avoided. As was empha- some useful things about the study of sized earlier, if the focus of interest is a convergence, but some of the most in- policy variable like inflation or the fluential papers take an approach to in- budget position, it will often be prefera- ternational technology diffusion which ble to omit factor accumulation alto- has little support from elsewhere. The gether and concentrate on the overall most interesting findings are rarely con- growth effect of policy measures. If a vincing, while the more reliable ones sufficiently complete model is used, this hold few surprises. approach should be very useful. It is What then have we learned? I return often the overall effect that is of princi- to the six questions posed in the intro- pal interest and, as a side benefit, fewer duction, and answer them as follows. variables need to be handled. Poor countries are not catching up with Sometimes, though, we would like to the rich, and to some extent the inter- know whether a growth effect acts national income distribution is becom- through factor accumulation or effi- ing polarized. Countries do converge to ciency change. If answers using macro their own steady states, but at an uncer- data require us to make incredible as- tain rate. One reason for this uncer- sumptions, this naturally suggests that tainty is that countries catch up by we should be working at a more disag- adopting technologies from abroad, as gregate level. Useful data sets of this well as by investing in physical capital kind are hard to come by for developing and education. It is easy to envisage a countries, and further problems lie in hypothetical long-run equilibrium in synthesising results for different indus- which countries grow at the same rate, tries and countries to get an overall but over the last thirty years, rates of picture of what drives growth. Yet there efficiency growth have allnost certainly can be little doubt that this older varied widely. approach, in which a branch of applied How compatible is this view with ex- microeconomics meets development, is isting growth models? The Solow-Swan also an important way ahead. model is almost certainly correct in assuming that returns to physical capital 10. Conclusions are diminishing. However, it would be a long way off the mark to assert that the One of the themes of this paper has new growth evidence only succeeds in been that cross-country research pro- demonstrating the explanatory power of vides a useful complement to other in- the original neoclassical growth model. vestigative techniques, and for some A wide variety of variables have been questions of interest, it may even be the shown to affect growth over thirty year only way forward. I make this claim de- time spans, often through systematic spite the scepticism that is possibly in- differences in rates of efficiency growth vited by the discussion above. Those who across countries. These differences are instinctively distrust the new growth left unexplained by the Solow-Swan evidence may well have discovered model, but are surely important in un- further support for their position. derstanding the variation in experience It is certainly true that, taken as a over the last thirty years. whole, the growth literature can seem Another implication should be noted. something of a disappointment. Conclu- Either growth is endogenous, or it is sions that once seemed well established exogenous and level effects are large. have been overturned. We have learned Given the presence of large level 152 Journal of Economic Literature, Vol. XXXVII (March 1999) effects, distinguishing between exoge- ment and high taxation may have a nous and endogenous growth models is negative effect, but the evidence is still not as pressing as it might seem. The somewhat ambiguous. Government important point is that policy can have a spending on infrastructure is beneficial. major impact on a country's level of Openness to trade also appears to be a welfare. As pointed out earlier, the de- good thing, although we do not yet bate on whether policy affects the long- know enough about the conditions un- run growth rate or just the steady state der which this is true. level of income is almost impossible In my view, those are the main find- to resolve, and not much of practical ings to emerge. Some, like those on fi- importance will turn on it. nance and on inequality, tend to contra- Arcane discussions about the validity dict the earlier conventional wisdom. of endogenous growth theories are Others, like those on the importance of likely to frustrate the policymaker in infrastructure, reinforce it. In some ar- search of knowledge that can be put to eas, such as the size of government and good use. Those forming policies would the effect of democracy, the main point ideally like to find a set of instructions to note is the lack of support for posi- that is short, clear and easy to imple- tions that are often too fiercely held. At ment. It is a foolhardy researcher who least some knowledge of the average responds to this desire by condensing pattern is the beginning of wisdom, and many books and articles into a one para- although we have not learnt as much as graph summary, but somebody perhaps might be hoped, it is always worth re- ought to try. A quick overview of why membering how little we knew when we growth rates differ is a useful way to started. close. A key reason why growth rates differ across countries is that macroeconomic Abramovitz, Moses. 1986. "Catching Up, Forging stability differs across countries. This Ahead, and Falling Behind," J. Econ. Hist., 46:2, pp 385-406. effect partly acts through capital invest- Adelman, Irma and Cynthia Taft Morris. 1968. ment, and equipment investment may "An Econometric Model of Socioeconomic and have a special role. As yet the growth Political Change in Underdeveloped Coun- tries," Amer. Econ. Rev., 58, p 1184-218. benefits of education are imprecisely Ades, Alberto and Hak B Cfua 1997 "Thy measured. The social returns to R&D Neighbor's Curse: Regional Instability and Eco- are high, and even if the long-run nomic Growth," J. Econ. Growth, 2:3, pp. 279- 304. growth rate is independent of research Aghion, Philippe and Peter Howitt. 1998. En- efforts, the welfare effects of changes in dogenous Growth Theory. Cambridge MA: MIT R&D expenditure can be large. Popula- Press. Alderman, Harold; Jere R. Behrman, David R. tion growth does not seem to have the Ross, and Richard Sabot. 1996. "The Returns to large negative effects that are fre- Endogenous Human Capital in Pakistan's Rural quently conjectured. High inequality Wage Labour Market," Oxford Bul. Econ. Sta- tist.,58:1, pp. 29-55. lowers growth, perhaps by raising social Alesina, Alberto and Roberto Perotti. 1994. "The and political instability. The depth of fi- Political Economy of Growth: A Critical Survey nancial intermediation seems important of the Recent Literature," World Bank Econ. Rev., 8:3, p 351-71. to subsequent development. Democra- Atkinson, Anttony B 1995 "The Welfare State cies do not do noticeably better than and Economic Performance," Nat. Tax J., 48:2, autocratic regimes, but countries that pp. 171-98. Barro, Robert J. 1991. "Economic Growth in a extend economic freedoms and protect Cross-section of Countries," Quart. J. Econ. property rights grow faster. Big govern- 106:2, pp 407-43. Temple: The New Growth Evidence

-. -. 1997. Determinants of Economic Growth: Canjels, Eugene and Mark W. Watson. 1997. "Es- A Cross-Country ~mpiricdStudy. Cambridge, timating Deterministic Trends in the Presence MA: MIT Press. of Serially Correlated Errors," Rev. Econ. Sta- Barro, Robert J. and Xavier Sala-I-Martin. 1992. tist., 79:2, pp. 184-200. "Convergence," J. Polit. Econ., 100:2, pp. 223- Carroll, Christopher D. and David N. Weil. 1994. 51. "Saving and Growth: A Reinter retation," -. 1995. Economic Growth. Boston: Carnegie-Rochester Conf Ser PubPic Policy, McGraw-Hill. 40, pp. 133-92 . Baumol, William J. 1986. "Productivity Growth, Caselli, Francesco; Gerard Esquivel, and Fer- Convergence, and Welfare: What the Long-Run nando Lefort. 1996. "Reopening the Conver- Data Show," Amer. Econ. Rev., 76:5, pp. 1072- gence Debate: A New Look at Cross-Country 85. Growth Empirics," J. Econ. Growth, 1:3, pp. ~enabou,Roland. 1996. "Inequality and Growth," 363-90. in NBER Macroeconomics Annual. Ben Ber- Chenery, Hollis B. 1988. "Structural Transforma- nanke and Julio Rotemberg, eds. London and tion: A Program of Research," in The State of Cambridge: MIT Press, pp. 11-74. Develo ment Economics. Gustav Ranis and Ben-David, Dan. 1993. "E ualizing Exchange: T PaufSchultr, eds Oxford: Basil Blackwell. Trade Liberalization an1 Income Conver- Cho, Dongchul and Stephen Graham. "The Other gence," Quart. J. Econ., 108:3, pp. 653-79. Side of Conditional Convergence," Econ. Let- Ben-David, Dan and A. K. M. Atiqur Rahman. ters, 50, p 285 90 1996. "Technological Convergence and Interna- Coe, DaviBT and~ihananHelpman 1995 "In- tional Trade," CEPR Discussion Paper No. ternational R&D Spillovers," Europ. Econ. 1359. Rev., 39:5, p 859 87 Benhabib, ess and Jordi Gali. 1995. "On Growth Crafts, Nichoras' F. R. ond Gianni Toniolo, eds. and Indeterminacy: Some Theory and Evi- 1996. Economic Growth in Europe since 1945. dence," Carnegie-Rochester Conference Series Cambridge: Cambridge U. Press. on Public Policy, 43, pp. 163-211. Deaton, Angus and Ron Miller. 1996. "Interna- Benhabib, Jess and Boyan Jovanovic. 1991. "Exter- tional Commodit Prices Macroeconomic Per- nalities and Growth Accounting," Amer. Econ. formance and Po f..>-ltlcs In Sub Saharan Africa," J. Rev., 81:1, pp. 82-113. African Econ., pp. 99-191. Benhabib, Jess and Mark M. Spiegel. 1994. "The Deininger, Klaus and Lyn Squire. 1996. "A New Role of Human Capital in Economic Develop- Data Set Measuring Income Inequality," World ment: Evidence from Aggregate Cross-Country Bank Econ. Rev., 10:3, pp. 565-91. Data," J. Monet. Econ., 34:2, pp. 143-73. De Long, J. Bradford. 1997. "Cross-Country Vari- Bernard, Andrew B. and Charles I. Jones. 1996a. ations in National Economic Growth Rates: the "Productivity Across Industries and Countries: Role of 'Technology'," in Technolo y and Time Series Theory and Evidence," Rev. Econ. Growth Jeffery Fuhrer and ane Snedion Lit- Statist., 78:1, pp. 135-46. tle, eds. Federal Reserve Ban I of Boston. - . 199613. "Technology and Convergence," De Long, J. Bradford and Lawrence H. Summers. Econ. J.,106:437, pp. 1037-44. 1991. "Equipment Investment and Economic Bhagwati, Jagdish and Bent Hansen. 1972. Growth," Quart. J. Econ., 106:2, pp. 445-502. "Should Growth Rates Be Evaluated at Interna- Dowrick, Steve and Duc-Tho Nguyen. 1989. tional Prices?" in Development and Planning. "OECD Comparative Economic Growth 1950- Jagdish Bhagwati and Richard S. Eckaus, eds. 85: Catch-up and Convergence," Amer. Econ. London: George Allen and Unwin, p. 53-68. Rev., 79:5, pp. 1010-30. Binder, Michael and M. Hashem Pesaran. 1996. Durlauf, Steven N. and Paul A. Johnson. 1995. "Stochastic Growth," U. Cambridge, Dept. Ap- "Multiple Regimes and Cross-Country Growth flied Econ working paper no 9615. Behaviour," J. Appl. Econometrics, 10:4, pp. Bo len, Kenneth A. 1990. "Political Democracy: 365-84. Conceptual and Measurement Traps," Stud. Durlauf, Steven N. and Danny T. Quah. Forth- Comp. Int. Devel., 25:1, pp. 7-24. comin . "The New Empirics of Economic Brander, James and Steve Dowrick. 1994. "The Growtf," in Handbook of Macroeconomics. Role of Fertility and Population in Economic John Taylor and Michael Woodford, eds. Growth: Em irical Results from Aggregate Easterly, William. 1994. "Economic Sta nation, Cross-Nation3 Data," J Population Econ., 7:1, Fixed Factors, and Policy Threshofds," J. pp. 1-25. Monet. Econ., 33:3, pp. 525-57. Bruno, Michael and William Easterly. 1998. "In- -. 1997. "Life During Growth," World Bank flation Crises and Long-Run Growth," ].Monet. manuscript. Econ., 41:1, pp. 3-26. Easterly, William and Ross Levine. 1997. "Africa's Caballero, Ricardo J. and Adam B. Jaffe. 1993. Growth Tragedy: Policies and Ethnic Divi- "How High Are the Giants' Shoulders?" in sions,'' Quart J Econ, 112, ip120350, NBER Macroeconomics Annual. Olivier Jean . 1998. "Troubles with t e Nelghbours: Af- Blanchard and Stanley Fischer, eds. London rica's Problem, Africa's Opportunity," J. African and Cambridge: MIT press, pp. 15-74. Econ., 7:1,pp. 120-42. 154 Journal of Economic Literatur.e, Vol. XXXVII (March 1999)

Easterly, William and Sergio Rebelo. 1993. "Fiscal We Have Learned about Prices and Quantities Policy and Economic Growth," J. Monet. Econ., from International Comparisons," Amer. Econ. 32:3, pp. 417-58. Rev., 78:2, pp. 467-73. Eaton, Jonathan and Samuel Kortum. 1994. "In- -. 1996. "International Price and Quantity ternational Patenting and Technology Diffu- Comparisons: Potentials and Pitfalls," Amer. sion," NBER working paper no. 4931. Econ. Rev., 86:2, pp. 20-24. -. -. 1995. "Engines of Growth: Domestic and Islam, Nazrul. 1995. "Growth Empirics: A Panel Foreign Sources of Innovation," NBER working Data Approach," Quart. J. Econ., 110:4, pp. 1127-70. EcP,"~s,n&~~%in1993 "Openness, Trade Lib- Jones, Charles I. 1995a. "Time Series Tests of En- eralization, and Growth in Developing Coun- dogenous Growth Models," Quart. J. Econ., tries,"J. Econ. Lit., 31:3, pp. 1358-93. 110:2, pp 495-525. Evans, Paul. 1996. "Using Cross-Country Vari- . 1995b. "R&D-Based Models of Economic ances to Evaluate Growth Theories," J. Econ. Growth,"J. Polit. Econ., 103:4, pp. 759-84. Dynam. Control, 20, pp. 1027-49. . 1997a. "Convergence revisited," J. Econ. -. 1997. "How Fast Do Economies Con- Growth, 2:2, pp. 131-53. ver e?" Review Econ. Statist., 79:2, pp. 219-25. . 1997b. "On the Evolution of the World Fagergerg, Jan 1994 Technology and Interna- Income Distribution," J. Econ. Perspectives, tional Differences in Growth Rates," J. Econ. 11:3, p . 19 36 Lit., 32, pp 1147;75. Jones, CareI and ohn C Williams 1997. Frankel, Je frey A David Romer, and Teresa "Measuring the SociaI Return to R&D," Stan- Cyrus. 1996. "Trade and Growth in East Asian ford U., manuscript. Countries: Cause and Effect?" NBER working Kakwani, Nanak. 1997. "Growth Rates of Per-Cap- paper no. 5732. ita Income and Aggregate Welfare: An Interna- Gemmell, Norman. "Evaluating the Impacts of tional Comparison," Rev. Econ. Statist., 79:2, Human Capital Stocks and Accumulation on 201-11. Economic Growth: Some New Evidence," Ox- Kef&, Allen C 1988 "Economic Consequences ford Bul. Econ. Statist., 58:1, pp. 9-28. of Population Change in the Third World," J. Gerschenkron, Alexander. 1952. "Economic Back- Econ. Lit., 26:4, pp. 1685-728. wardness in Historical Perspective," in The King, Robert G. and Levine, Ross. 1993. "Finance Progress of Underdeveloped Areas. Bert F. and Growth: Schumpeter Might Be Right," Hoselitz, ed. Chicago: U. Chicago Press. Re- Quart. J. Econ., 108:3, pp. 717-37. printed in A. Gerschenkron, Economic Back- --. 1994. "Capital Fundamentalism, Eco- wardness in Historical Perspective, New York, nomic Development and Economic Growth," Frederick A. Praeger, 1962. Carnegie-Rochester Con$ Ser. Public Policy, Gramlich, Edward M. 1994. "Infrastructure In- 40, pp. 259-92. vestment: A Review Essay,", "/. Econ. Lit., 32:3, King, Robert G. and Sergio T. Rebelo. 1993. pp 1176-96. "Transitional Dynamics and Endogenous Griliches, Zvi 1979. "Issues in Assessing the Con- Growth in the Neoclassical Model," Amer. tribution of R&D to Productivity Growth." Bell Econ. Rev., 83:4, pp. 908-31. J. Econ., 10:1, p. 92-116. Klenow, Peter J. and Andres Rodriguez-Clare. - 1992. "TKe Search for R&D S~illovers." Forthcoming. "The Neoclassical Revival in Scand. J. Econ., Supplement, 94, ~29-'47. Growth Economics: Has It Gone Too Far?" Griliches, Zvi and Jacques Mairesse. 1995. "Pro- NBER Macroeconomics Annual. duction Functions: The Search for Identifica- Knack, Ste hen and Keefer, Philip. 1995. "Institu- tion," NBER manuscript. tions an$ Economic Performance: Cross-Coun- Hall, Robert E. and C. I. Jones. 1997. "Funda- try Tests Using Alternative Institutional Mea- mental Determinants of Output per Worker sures", Econ. Polit., 7:3, pp. 207-27. across Countries," Stanford U., manuscript. Kocherlakota, Narayana R. and Kei-Mu Yi. 1995. Hanushek, Eric A. 1992. "The Trade-off between "Can Convergence Regressions Distinguish Child Quantity and Quality," J. Polit. Econ., between Exogenous and Endogenous 100:1,pp 84-117. Growth Models?" Econ. Letters, 49:2, pp. 211- Hanushek, Eric A. and Dongwook Kim. 1995. 15. "Schooling, Labor Force Quality, and Economic Koop, Gary; Jacek Osiewalski, and Mark F. J. Growth," NBER working paper no. 5399. Steel. 1995. "Measuring the Sources of Output Harberger, Arnold C. 1987. "Comment," NBER Growth in a Panel of Countries," CORE discus- Macroeconomics Annual. Stanley Fischer, ed. sion paper no. 9542. London and Cambridge: MIT Press, pp. 255- Kormendi, Roger C. and Philip G. Meguire. 1985. 58. "Macroeconomic Determinants of Growth: Heston, Alan. 1994. "A Brief Review of Some Cross-Country Evidence," J. Monet. Econ., Problems in Using National Accounts Data in 16:2, pp 141-63. Level of Output Comparisons and Growth Kravis, Irving B. 1984. "Comparative Studies of Studies,"J. Devel. Econ., 44:1,29-52. National Incomes and Prices," J. Econ. Lit., Heston, Alan and Robert Summers. 1988. "What 22:1, pp 1-39. Temple: The New Growth Evidence 155

Kuznets, Simon. 1966. Modern Economic Growth. Heterogeneous Panels," J.Econometrics, 68:1,' New Haven: Yale U. Press. pp 79-113. Leamer, Edward E. 1983. "Let's Take the Con Pritchett, Lant. 1996a. "Population, Factor Accu- Out of Econometrics," Amer. Econ. Rev., 73:1, mulation, and Productivity," World Bank work- pp. 31-43. ing paper no. 1567. - . 1985. "Sensitivity Analyses ' would Help," -. 1996b. "Where Has All the Education Amer. Econ. Rev., 75:3, p. 308-13 Gone?" World Bank working paper no. 1581. Lee, Jong-Wha and Roiert J Barro 1997. -. 1997. "Divergence, Big Time," J. Econ. "Schooling Quali in a Cross Section of Coun- Perspectives, 11:3, pp. 3-17. tries." NBER wor2 ing paper no. 6198. -. 1998. "Patterns of Economic Growth: Lee, Kevin; M. Hashem Pesaran, and Ron Smith. Hills, Plateaus, Mountains, and Plains," manu- "Growth and Convergence in a Multi-Count script, World Bank. Empirical Stochastic Solow Model," J App? Quah, Danny T. 1993. "Empirical Cross-Section Econometrics, 12, pp. 357-92. Dynamics in Economic Growth," Europ. Econ. Leung, Charles and Quah, Danny T. "Conver- Rev., 37:2-3, pp. 426-34. , endogenous growth, and productivity - . 1997. "Empirics for Growth and Distri- $fs?:bances," J Monetary Econ,, 1996, 383, bution: Stratification, Polarization, and Conver- 535-547. gence Clubs," J. Econ. Growth, 2:1, pp. 27-59. Levine, Ross. "Financial Development and Eco- Ramey, Carey and Valerie A. Ramey. 1995. nomic Growth: Views and Agenda," J. Econ. "Cross-Country Evidence on the Link between Lit., 35, pp. 688-726. Volatility and Growth," Amer. Econ. Rev., 85:5, Levine, Ross and David Renelt. 1991. "Cross- pp. 1138-51. Country Studies of Growth and Policy," World Ranis, Gustav. 1984. "Typology in Development Bank working paper WPS no. 608. Theory: Retrospective and Prospects," in Eco- -. 1992. "A Sensitivity Analysis of Cross- nomic Structure and Performance. Moshe Syr- Country Growth Regressions," Amer. Econ. quin, Lance Taylor and Larry E. Westphal, eds. Reu., 82:4, pp. 942-63. London: Academic Press. Levine, Ross and Sara J. Zervos. Forthcoming. Ranis, Gustav and John C. H. Fei. 1988. "Devel- "Stock Markets, Banks, and Economic Growth," o ment Economics: What Next?" in The State Amer. Econ. Rev. OF Development Economics Gustav Ranis and Lucas, Robert E. "On the Mechanics of Economic T. Paul Schultz, eds. Oxford: Basil Blackwell. Development," J. Monet. Econ., 22, pp. 3-42. Robinson, Sherman. 1971. "Sources of Growth in Mankiw, N. Gregory. 1995. "The Growth of Na- Less Developed Countries: A Cross-Section tions," Brookings Pap. Econ. Act., 1, pp. 275- Study," Quart. J. Econ., 85:3, p 391-408. 310. Rodrik, Dani 1998. "Where DiB~llthe Growth Mankiw, N. Gregory; David Romer, and David N. Go? External Shocks, Social Conflict and Weil. 1992. "A Contribution to the Empirics of Growth Collapses," CEPR working paper Economic Growth," Quart.]. Econ., 107:2, 407- no.1789. 37. Rogoff, Kenneth. 1996. "The Purchasing Power Manksi, Charles F. "Identification of Endogenous Parity Puzzle," J. Econ. Lit., 34, pp. 647-68. Social Effects: The Reflection Problem," Reu. Romer, David. 1996. Advanced Macroeconomics. Econ. Stud., 60, pp. 531-42. Boston: McGraw-Hill. Marris, Robin. 1984. "Com aring the Incomes of Romer, Paul. 1986. "Increasing Returns and Long- Nations: A Critique of t K e International Com- Run Growth," ]. Polit. Econ., 94:5, pp. 1002-37. parison Project," J. Econ. Lit., 22:1, pp. 40-57. Rousseeuw, Peter J, and Annick M. Leroy. 1987. Mauro, Paolo. 1995. "Corruption and Growth," Robust Regression and Outlier Detection. New Quart. J. Econ., 110:3, pp. 681-712. York: Wile!. Moreno, Ramon and Bharat Trehan. 1997. "Loca- Sachs, Jeffrey D. and Andrew M. Warner. 1995a. tion and the Growth of Nations," J. Econ. "Economic Reform and the Process of Global Growth, 2:4, pp. 399-418. Integration," Brookings Pap. Econ. Act., pp. 1- Nuxoll, Daniel A. 1994. "Differences in Relative 118--- Prices and International Differences in Growth -. 1995b. "Natural Resource Abundance and Rates," Amer. Econ. Rev., 84:5, p 1423-36. Economic Growth," NBER working paper no. Parente, Stephen L and Edwarl'C Prescott. 5398. 1994. "Barriers to Technology Adoption and 1997. "Fundamental Sources of Long-Run Development,"J. Polit. Econ., 102:2, pp. 298-321. Growth," Amer. Econ. Rev., 87:2, pp. 184-88. Perotti, Roberto. 1996. "Growth, Income Distri- Sala-I-Martin, Xavier. 1996. "The Classical Ap- bution, and Democracy: What the Data Say," J. proach to Convergence Analysis," Econ. I., 106, Econ. Growth, 1,p . 149-87. p 1019-36. Persson, Torsten an$ Guido Tabellini 1994. "Is Sare?: Michael "Demographic Dynamics and the Inequality Harmful for Growth?" Amer. Econ. Empirics of Economic Growth," IMF Staff Pa- Reu., 84:3, pp. 600-21. ers, 1995, 42:2, 398-410. Pesaran, M. Hashem and Ron Smith. 1995. "Esti- ScKoling, Eberhard and Vincenz Timmermann. mating Long-Run Relationships from Dynamic "Why Growth Rates Differ: Measuring Journal of Economic Literature, Vol. XXXVIZ (March 1999)

'Unmeasurable' Influences," World Deuel., Temple, Jonathan R. W. and Paul A. Johnson. 16:11, pp 1271-94. 1998. "Social Capability and Economic Sims, Christopher A. "Commentary," Fed. Res. Growth," Quart. J. Econ., 113:3, pp. 965-90. Bank St. Louis Reu., Ma 78.3, pp. 173-78. Temple, Jonathan R. W. and Hans-Joachim Voth. Slemrod, Joel 1995 "Wiat DO Cross-Country 1998. "Human Capital, Equipment Investment, Studies Teach about Government Involvement, and Industrialization," Europ. Econ. Rev., 42:7, Prosperity, and Economic Growth?" Brookings pp. 1343-62. Pap. Econ. Act., pp. 373-431. Ventura, Jaume. 1997. "Growth and Interdepen- Summers, Robert and Alan Heston. 1988. "A New dence," Quart. J. Econ., 112, pp. 57-84. Set of International Comparisons of Real Prod- Wolf, Holger C. 1994. "Growth Convergence Re- uct and Price Levels Estimates for 130 Coun- considered," Weltwirtschaftliches Archiu, tries. 1950-1985." Reu. Income Wealth. 34:l. 130:4, p 747-59. pp. 1-25. Young, A?&, 1992 "A Tale of Two Cities: Factor - . 1991. "The Penn World Table: Mark 5: An Accumulation and Technical Change in Hong Expanded Set of International Comparisons, Kong and Sin apore," in NBER Macro- 1950-1988," Quart. J. Econ., 106:2, pp. 327-68. economics Annu$. Olivier Jean Blanchard and Tem le, Jonathan R. W. 1998a. "Robustness Tests Stanley Fischer, eds. London and Cambridge: o?the Augmented Solow Model," J Applied MIT Press, pp. 13-54. Econometrics, 13:4, pp. 361-75. -. 1995. "The Tyranny of Numbers: Con- -. 1998b. "Equipment Investment and the fronting the Statistical Realities of the East Solow Model," Oxford Econ. Papers, 50:1, pp. Asian Growth Experience," Quart. J. Econ., 39-62. 110:3, pp 641-80.