A quarterly magazine of the IMF

March 2006, Volume 43, Number 1

The Quest Continues
Lant Pritchett

After decades of growth research, what can academic economists really say about policy?

So here I am, in the middle way, having had twenty years—
Twenty years largely wasted, the years of l'entre deux guerres—
Trying to learn to use words, and every attempt
Is a wholly new start, and a different kind of failure
Because one has only learnt to get the better of words
For the thing one no longer has to say, or the way in which
One is no longer disposed to say it.
—T. S. Eliot, Four Quartets

Modern growth theory, which built on the Harrod-Domar model, was born in 1956 with Robert Solow's famous papers and will turn 50 this year. Even the "new" growth theory, born with Paul Romer's papers, is now in its 20s. Why is it that with aptness, if poetic ineptness, many economists feel they could replace "words" with "growth research" in T. S. Eliot's refrain above about his "middle way?" This article is a brief retrospective and prospective on growth research in three parts: growth theory (old and new), empirical growth research (short and long), and the way forward. The theme that runs through all three parts is the tension between the logics of academic interest and the needs of the policy practitioner. The typical policymaker or advisor—whether politico or technocrat—wants to know the likely consequences of concrete public sector actions (not necessarily limited to policies) over their relevant time horizon. If growth research is a quest to satisfy this need, the journey is far from over.

Growth theory: old and new

Growth Theory (as we shall understand it) has no
particular bearing on underdevelopment economics,
nor has the underdevelopment interest played
any essential part in its development.
—John Hicks, Capital and Growth, 1965

Dozens of brilliant (and thousands of amateurish) "new growth theory" papers later, it is hard to re-create the enthusiasm with which "endogenous growth" was first greeted in academic and policy circles in the early 1980s. At that time, growth theory was still dominated by Robert Solow's model, now known as the "exogenous growth" model, and was stuck, badly stuck. By 1971, Solow could write that everything that could be said with his growth model had been said. And what was said was not that nothing could be said with certainty but that, with certainty, nothing could be said by the model. In the model (and all its variants), equilibrium or steady-state growth rates of per capita output were driven by technical progress—but these models, as constructed, could say nothing about the determinants of technical progress. This point is often misunderstood but is important to understanding just how badly cornered the profession was and, hence, how excited it was about new growth.

The Solow model did not assume that technical progress was exogenous—that is, determined outside the model. Rather, the model made the assumptions necessary to produce a model of an economy with a dynamic equilibrium, a path to which, in the long run, the economy would settle down. The implication of those assumptions was that technical progress had to be exogenous to the model. The technical problem is that the assumptions necessary to produce a model with an equilibrium implies that payments to the standard factors of production—labor and capital—exhaust the total product. Nothing is left over to pay entrepreneurs or innovators. But if innovators cannot be paid at all in equilibrium (given the assumptions of the model), then nothing—including policy of any kind—can affect their incentives to innovate. This problem of reconciling purposive economic innovation that results in greater productivity of all factors (such innovation obviously exists and is a key to the success of capitalism) with formal economic models (which could adequately model this phenomenon) was long standing. It was clearly recognized by Joseph Schumpeter and Frank Knight, but Solow's clean algebra and exposition just made it stark.

The best macroeconomic models circa 1980 said national policies could affect the level of incomes but they could not affect steady-state growth. Meanwhile, microeconomic analysis of sector reforms—for example, in trade, privatization, or the financial sector—produced estimates of the level effects of policy changes that were typically only small fractions of GDP. This marriage—of a macroeconomics in which long-run growth rates were driven by technical progress that was independent of national policies to a microeconomics in which policy reform could produce only small efficiency gains in levels—was a stable but increasingly unhappy one. The marriage was stable because both the macro and micro outcomes were not "assumptions" but were technically driven features of their respective analytical approaches.

Into this unhappy marriage of macro and micro came new growth theory models. One of the reasons the marriage was unhappy was entirely technical and internal to the discipline of economics: couldn't economists produce a model that had both purposive innovation and a steady state? Romer (1983) used advances in techniques of modeling "noncompetitive" equilibrium (in which factor payments do not exhaust product) to produce new models that did just that. In this new class of models, since innovation could be driven by incentives, national policies could influence not only the level of income but also countries' steady-state growth rates—that is, the very long run rates.

A second reason the marriage of macro and micro was unhappy was external: its inability to explain obvious facts about the world (for instance, that levels of income did, in fact, differ by orders of magnitude) or to provide policy guidance. Therefore, the spirit (if not the actual technical rigor) of new growth models was quickly put to use in the policy world to justify policy reform based on its impact on growth. Structural adjustment as a response to negative economic shocks emphasized the need to restore short-run macroeconomic stability and raise growth rates, with the promise that postshock and poststabilization levels of income would be much higher.

While the contribution of the new growth models to the internal logic of the economics discipline has been lasting, the bloom came off the rose of the explicit use of new growth models for policy purposes in developing countries relatively quickly. Nearly everything about the first-generation growth models was at odds with the needs of developing country policymakers. The new growth literature focused on the very long run and on incentives for expanding the technological frontier—not particularly useful for most developing countries, whose primary interest was in restoring short- to medium-term growth and accelerating technological catchup by adopting known innovations. In the end, Hicks was not only right but also prescient about the irrelevance of new growth to the "underdevelopment interest."

Empirical growth research: short and long

I just ran four million growth regressions.
—Xavier Sala-i-Martin

New growth theory and empirical research into growth have the same relationship as lightning and a forest fire. Although lightning may start the fire, where the fire goes and how much it burns are determined by the condition of the forest, not its origins. New growth theory legitimized a research agenda—including growth regressions—focused on the correlates of national growth rates that has continued long past the relevance of the original spark. In fact, nearly all growth regressions today, while they place growth of output on the left-hand side as the variable to be explained, are arithmetically equivalent to "levels" regressions that do not depend on new growth theory at all. Of course, from a policy viewpoint, this is irrelevant. The relevant question is what actions can be taken to spur more rapid economic growth over the relevant horizon. The answer to this question does not depend at all on whether the growth impact is a steady-state gain (and, hence, an infinitely large-level impact) or just a big enough impact on levels to accelerate growth rates.

As they have evolved, growth regressions have moved toward the "short" (using panel regression techniques) or the "long" (which goes back to regressions on the level of income, which is, after all, the same as the longest possible growth rate). Let us review these two strands separately.

Growth regressions: the short. Growth regressions have proved a very useful extended investigation into the correlates of medium- to long-run growth. From the original growth regressions, which used single cross sections with 20- to 30-year growth rates and ordinary least squares, to the latest use of panel data with 5- and 10-year growth rates and more sophisticated estimation techniques, the growth regression literature has added to the stock of knowledge about the correlates of growth with various policy and nonpolicy variables. However, in the end, as an instrument for informing policy, regressions have suffered five exhausting, if not mortal, wounds.

First, growth regressions never satisfactorily resolved the "symptoms versus syndromes" problem. They could establish the relationship between particular observable empirical variables ("symptoms") and growth but could only rarely move convincingly from those observed associations to specific recommendations on how to tackle the underlying causes of the emergence of the symptoms. For example, suppose one finds that a high black market premium is associated with lower growth. Is this an indicator that trade policy was awry, that there were macroeconomic imbalances, or that the country was unable to respond to adverse trade shocks? Or was it just an indicator of general dysfunction? Hence, although it was clear that there was something about a country's external policies that affected growth and triggered a syndrome of misguided inward orientation, even two decades into trade liberalization, prominent scholars are still debating exactly what that something is and how best to address it.

Second, growth regressions were widely seen as producing estimates of gains from policy reform that were orders of magnitude larger than the microeconomic estimates of those gains—without any particularly convincing economic explanation. As trade theorists like T.N. Srinivasan and Jagdish Bhagwati consistently pointed out, the justification of trade reform based on growth regressions was a dubious exercise indeed, compared with the firmer groundings in theory and microeconomic empirics such reforms already had. The growth regression literature in field after field (trade, tax and expenditure, and finance) has had a very difficult time reconciling micro and macro evidence.

Third, growth regressions have a tough time dealing with the huge differences in country experiences. Establishing a correlation between economic output and some variable, say, corruption, on average across countries is only the very beginning of wisdom. Is the effect proportional to corruption as measured? Or does it follow a different relationship? Is the effect stronger in democracies than in nondemocracies? In poorer than in richer countries? In more open than in less open countries? In reality, the growth regression is only a crude approximation that indicates the average impact of corruption, but it does not provide the information policymakers really want—the specific impact in a particular country. This inability to tackle country differences head-on led to differing results about one set of growth determinants (for example, external orientation) depending on sample, period, technique, and the other variables included—creating a general skepticism about the reliability of all growth regressions.

Fourth, growth regressions cannot predict turning points— either accelerations or decelerations—and we know that developing country growth rates have these turning points. The growth regression approach is much better at identifying the longer-run (30 years or more) determinants of growth than at predicting the very short run ones. In part, this is because the short-run volatility of many growth determinants is simply too small to play much of a role in explaining growth changes. For instance, while few doubt that human capital is vital for long-run development, it evolves so smoothly that it can explain, at best, 1 percent of cross-national differences in 5-year growth rates. As a recent study of episodes of growth acceleration shows, very few acceleration turning points are associated with reform episodes, and few reform episodes are associated with accelerations (Hausmann, Rodrik, and Pritchett, 2004).

Fifth—and perhaps most disillusioning—it was felt (perhaps a bit unfairly but nevertheless widely) that growth regressions did not help policymakers anticipate either the disappointments or the surprises of the experiences of the 1990s (World Bank, 2005). What were the disappointments of the 1990s?

• The severity and intensity of the international and domestic financial crises that rolled through East Asia, affecting countries that had been growth "miracles"—Indonesia, Korea, and Thailand. Also, Argentina's financial and economic implosion after the collapse of its currency convertibility regime. (While it is perhaps unfair to ask growth theory to predict crises, the fact that countries performing well on the fundamentals of growth could undergo such massive shocks was an unpleasant surprise.)
• The length, depth, and severity of the transition recessions in the reforming socialist countries in Eastern Europe and the former Soviet Union.
• The weakness of the response of growth to reform, especially in Latin America.
• The continued stagnation in sub-Saharan Africa, the paucity of success stories, and the apparent wilting of optimism around the "African Renaissance"—resulting in reform fatigue, whether or not reform had taken place.

There were pleasant surprises as well. These include bright spots of sustained rapid growth, especially in China, India, and Vietnam; the strong progress in noneconomic indicators of well-being (especially basic education and children's health) in spite of low growth in some cases; and the resilience of the world economy to stresses.

Institutions and levels of output: the long. One important strand of the literature on economic growth went back to investigating the empirical determinants of the level of output. After all, since at some sufficiently distant point in (pre)history, output per capita in all countries was roughly equal, the current level of their output is thus also an estimate of their very long run growth rates. This literature—starting with a series of seminal contributions by Hall and Jones (1999), Engerman and Sokoloff (2002) and Acemoglu, Johnson, and Robinson (2001)—emphasizes that the primary determinant of a country's level of income is the quality of its institutions, or what Hall and Jones call "social infrastructure." The phrase "institutions rule" (Rodrik, Subramanian, and Trebbi, 2004) suggests the importance given to institutions as the determinants of growth.

That the quality of institutions affects the level of income and current growth is interesting. But this insight is not necessarily good news for policymakers, given that no one has a very clear idea about how to improve the quality of institutions and many believe that the determinants of institutional quality are very long run historical processes. As one economist (who will remain anonymous) puts it: "the definitions seem to be that policies are what can be purposively changed and institutions are what cannot."

The way forward: states and transitions

You can't always get what you want,
But if you try sometimes you might find,
You get what you need.
—The Rolling Stones

Given the dramatic failures in growth research, I would like to conclude by pointing to a new direction for future research. As neither I (Pritchett, 2003) nor anyone else has produced a complete and adequate version of what I propose (although Michal Jerzmanoski has made a promising start), I am vulnerable to Wolfgang Pauli's critique when Werner Heisenberg announced he had a new theory with only the technical details missing. Pauli sent a letter to a friend, saying, "This is to prove to the world I can paint like Titian," followed by a scrawled blank rectangle with the conclusion: "Only the technical details are missing." The key idea in my proposal is that economies are in different "states," and, therefore, the dynamics of output are different for economies in different states, and the dynamics of transitions between states are different from the dynamics within states.

Suppose you have a pot of water and you pick it up and turn it over. Where will the water go? The answer, that it will spill out onto the ground, is so obvious that the astute reader already realizes it is a trick question. If the water is frozen, it may stay right in the pot. If the water is vapor, then turning the pot over will trap the steam in the pot. The obvious point is that the equations of motion of water (or any other substance) depend on the state—solid, liquid, or vapor—it is in. What determines the transitions of water between states? Well, applying heat will cause water to change states, but only in a discontinuous way—water at 35° F and water at 95° F behave almost the same, while water at 32° F and at 102° F behave nothing like each other. The equations of motion of water in one state do not work at all when water is in another state, and the response of water to heat applied within a state does not work at all well when applied to transitions from one phase to another.

Growth theory can be viewed as the attempt to formalize the equations of motion of aggregate output, and growth empirics can be viewed as the related effort to estimate these equations of motion, including the "impulse response functions" showing how output responds to various shocks (for example, policy, institutions, geography, prices, and technical change). The answer to the question "what happens to output growth when trade barriers or inflation are reduced" could, in principle, have a single theoretical and empirical answer. If France and Nepal can both be treated as water in a liquid state, then it is conceivable that a theory and empirics of growth that treat France and Nepal as both generic countries is adequate. I regard it as much more likely that growth dynamics are characterized as equations of motions within states and equations that determine transitions across states.

All of the empirical stylized facts about cross-national growth can be matched with an encompassing model with five states (Pritchett, 2003). Each state is characterized by the typical rate of growth and level of income: advanced industrial (high income, steady growth) and poverty trap (low income, no growth) are two extremes, with low- and middle-income countries having three other possible states: rapid convergence, nonconverging growth, and collapse. Within the limits of each of these states, there can be more or less rapid growth (for example, France can grow more or less rapidly within the advanced industrial growth state, Colombia can grow more or less rapidly within the nonconverging growth state), and the equations of motion appropriate to each state would apply.

The transitions between these states are determined by a distinct dynamic, and the probabilities of transitions between the states are neither equal (the probability of transition from nonconverging growth to collapse is not the same as to rapid convergence) nor necessarily symmetric (the probability of transition to advanced industrial is not the same as the transition away from it).

This is an ambitious agenda for growth research because rather than a single growth theory, there is a collection of growth theories (for each state) and transition theories (potentially one theory from each state to each other state). The debate between exogenous and endogenous growth has mostly been about identifying the dynamics of motion of a country like France or the United States in a state of advanced industrial growth, and not about escaping a poverty trap or even initiating a growth boom.

This agenda also requires empirics that do not begin by assuming one is estimating the parameters of a single set of equations of motion but rather by thinking of growth regressions as a first, descriptive set of partial correlations across various horizons that reveal patterns of relationships that smooth over a more complex dynamic (which accounts for the differences across countries and lack of robustness of growth regressions). In other words, if a measure of corruption is associated with lower growth, is that because in each state corruption lowers the growth rate or because corruption raises the probability of the transition away from a good state (for example, rapid convergence into collapse) and lowers the probability of transition to a good state?

By pursuing this ambitious agenda, the profession will move closer to giving policymakers what they want: a growth strategy that is relevant to their particular circumstances. It would focus on the actions that attack the binding constraints to moving a country from the state it is in to a better state (for example, from collapse to nonconverging growth, or from poverty trap to rapid convergence). Constructing policy advice from such a growth strategy is going to be much more complicated than reading off a list of policies that are partial correlates in a growth regression and saying "do that." After all, the advice will need to be tailored to a country's circumstances and potential and be specific to its current state, with limited empirical evidence to draw upon. This means that crafting the policy advice will remain as much art as science, but it will be well worth the effort if policymakers succeed in getting what they need.