Growth Dynamics: The Myth of Economic Recovery
  • 1 0000000404811396https://isni.org/isni/0000000404811396International Monetary Fund

Contributor Notes

Author(s) E-Mail Address: vcerra@imf.org; ssaxena@gspia.pitt.edu

Using panel data for a large number of countries, we find that economic contractions are not followed by offsetting fast recoveries. Trend output lost is not regained, on average. Wars, crises, and other negative shocks lead to absolute divergence and lower long-run growth, whereas we find absolute convergence in expansions. The output costs of political and financial crises are permanent on average and long-term growth is negatively linked to volatility. These results also imply that panel data studies can help identify the sources of growth and that economic models should be capable of explaining growth and fluctuations within the same framework.

Abstract

Using panel data for a large number of countries, we find that economic contractions are not followed by offsetting fast recoveries. Trend output lost is not regained, on average. Wars, crises, and other negative shocks lead to absolute divergence and lower long-run growth, whereas we find absolute convergence in expansions. The output costs of political and financial crises are permanent on average and long-term growth is negatively linked to volatility. These results also imply that panel data studies can help identify the sources of growth and that economic models should be capable of explaining growth and fluctuations within the same framework.

I. Introduction

The recent vast empirical literature on economic growth has focused primarily on cross-sectional growth regressions, with few studies exploiting the variation of growth over time. Therefore, very little attention has been paid to whether countries recover from their crises, shocks, and downturns. Although it is known that many crises are associated with recessions (Kaminsky and Reinhart, 1999), there is little research on whether the output losses are fully reversed.

Three examples illustrate the different types of recoveries from shocks that countries may experience. Following a steep recession from 1965–68, output in Nigeria recovered rapidly to its former trend line (passing through the pre-recession peak). Conversely, only a tiny fraction of the output loss from Sweden’s banking crisis in the early 1990s was recuperated. At the opposite extreme, growth surged in Swaziland after a shallow recession in 1987. This paper investigates which of these recovery experiences is typical across countries and whether the type of recession or shock impacts the magnitude of recovery.

Do crises derail growth? If output doesn’t fully recover from a contraction, as in the Swedish example, the proclivity to shocks may be responsible for the absolute divergence of incomes across countries. That is, if poor countries are hit by more shocks than rich countries, the output losses could accumulate over long periods, causing incomes to diverge. Poverty traps would be different in character than currently envisaged by some economists, with vastly different policy implications. Sachs and others (2004) argue that African countries are stuck in a poverty trap because their low incomes prevent them from generating sufficient savings and investment. According to this view, the trap operates even when economic conditions are favorable and governance is sound. Due to savings traps, capital thresholds, and demographic traps, a poor country is unable to generate high growth, and thus cannot pull itself out of poverty unless it is rescued by large infusions of aid from wealthy nations. If, on the other hand, growth is thwarted by political and financial crises rather than poverty traps, the implication is that institutions and policies would need to be transformed so as to achieve stability and set the conditions for strong long-term growth.

Understanding how countries recover from shocks is also important for several other reasons and linked to different debates in the literature. First, the welfare costs of volatility are much larger when the output loss is not recovered. They are, in fact, permanent. Lucas (1987) showed that the welfare costs of economic instability in the postwar United States were only a minor problem relative to the costs of modestly reduced rates of economic growth. However, his calculation assumed that deviations from trend were temporary. If shocks have permanent output effects, the welfare costs of volatility and growth are intrinsically connected. Also, emerging market and developing countries have considerably higher volatility than the United States. Moreover, sharp and permanent drops in the level of output may lead to steep increases in unemployment and civil unrest, even when output drops to a level that was associated with stable and prosperous times only a few years earlier.

Second, the inference about the determinants of growth can be deceptive if crises and shocks have permanent effects and if recoveries resemble Sweden’s pattern rather than that of Nigeria or Swaziland. For example, many critics of Sweden’s welfare state policies have argued that such policies led to a drop in Sweden’s GDP per capita relative to other OECD countries. Yet, a simple examination of the timing of the output loss relative to the OECD average makes it very clear that the slippage was associated with Sweden’s severe banking crisis and recession in the early 1990s. Prior to this recession, Sweden’s output had been growing roughly parallel to the OECD average. The banking crisis resulted in a sharp decline in output below the prior trend and below the OECD average and this wedge persisted in the subsequent “recovery.” By comparing only the 30-year or 40-year growth rate to that of the OECD average, the critics misinterpret the slippage as due to welfare state policies as the long-term growth correlate.2

Third, if output losses associated with crises and shocks don’t dissipate, it becomes critical to link time-varying growth rates with time-varying policies and country characteristics. Crises, shocks, and changes in policies can be ignored only if their output consequences are transitory. The copious body of growth literature searching for policies to boost income levels3 has largely used a single average growth rate over a span of several decades for each country and a set of average or initial policies and country characteristics. Growth is not, however, a steady process. The variation across time is about as large as the variation across countries. Easterly and others (1993) note that, “with a few famous exceptions…countries are success stories one period and disappointments the next.” Rodrik (1999) and Hausmann, Pritchett, and Rodrik (2004) document a large number of shifts in trend growth and try to explain episodes of growth collapses and growth accelerations, respectively. Pritchett (2000) provides examples of patterns ranging from rapid steady growth to plateaus and valleys. Cross-section regressions ignore the considerable variation in the data across time. As illustrated above for the case of Sweden, the inference from such regressions can be questionable. Moreover, explanatory variables that are constant or change very little over the sample are unlikely to be the sources of the shifts in trend growth.

The optimal data frequency for growth regressions is controversial, but permanent output loss would tip the balance in favor of higher frequency panels. Potential growth correlates are abundant, possibly larger than the number of countries. Cross-country regressions have been fragile, with the coefficients subject to change depending on the presence of other variables in the equation (Levine and Renelt, 1992). Some recent papers contend with sample size limitations and multicolinearity by sampling potential explanatory variables to find the set of variables with coefficients that are reasonably robust when combinations of other variables are present in the regression. For example, Sala-i-Martin (1997) ran two million cross-country regressions and Doppelhofer, Miller, and Sala-i-Martin (2000) used a Bayesian approach to test the robustness of growth correlates. However, many policy and other variables change over time. Ignoring this source of variation throws away valuable information. Exploiting a panel of time series, cross-country growth rates can take advantage of temporal and cross-sectional variation, thereby increasing the sample size and reducing multicolinearity. Yet, this approach has received much less attention than cross-section regressions, with the few extant panel studies (e.g., Islam 1995) focusing mostly on tests of convergence. Pritchett (2002) argues against using annual panel data with fixed effects for a variety of technical reasons. But panel data techniques are rapidly improving and can combine constant and time-varying regressors. If shocks have permanent output effects, these new techniques become indispensable.

The optimal frequency in panel data studies is linked to a fourth controversy: whether output follows a deterministic or stochastic trend. The emerging literature that studies growth using panel data tends to use five- or ten-year averages of growth based on an apparent view that shocks are temporary and one can separate “business cycles” from deterministic trends.4 However, the evidence that fluctuations are transitory has yet to be demonstrated. Nelson and Plosser (1982) challenged the prevailing assumption that U.S. output follows a deterministic trend, and the nature of U.S. business cycles continues to attract an active debate. In emerging markets, some recent studies (e.g., Cerra and Saxena, 2005a) document that financial crises lead to long-run losses in the level of output of the affected countries. If recessions or growth collapses are not followed by faster-than-average growth, there may be no sense of “averaging out” over a business cycle. The recessions or growth collapses may themselves be responsible for lower average long-run growth, and the determinants of contractions may differ from those of expansions. The conditional expectation for the level of income at a distant future horizon would be unaffected by a current recession only if output reverts to trend. If output follows a stochastic trend, then every shock changes the conditional expectation of the future income level, one for one.

Finally, the paper contributes to research on the relationship between growth and volatility, upon which the extant empirical evidence is mixed. Some cross-section studies using international data find a positive link between mean growth and its standard deviation (Kormendi and Meguire, 1985; Grier and Tullock, 1989), while others (Ramey and Ramey, 1995) find the opposite relationship. Siegler (2004) reports a negative correlation between the standard deviation of real GDP and growth for a panel of 12 countries using decade averages over the period 1870 to 1929. Dawson and Stephenson (1997) find no association between output volatility and growth across U.S. states, and dispute the findings of Ramey and Ramey. They argue that countries with low growth tend to be the countries whose data contain large measurement errors, resulting in a spurious negative relationship between volatility and growth. By examining the recovery from shocks, this paper makes use of the within-country variation.

The paper proceeds as follows. In the next section, we discuss the predictions of theoretical models of growth and crises. Section III describes the data. In Section IV, we test whether negative shocks have a long-term impact on income levels in a broad sample of countries. We also ask if negative shocks contribute to unconditional divergence across countries (Section V). We investigate sources of recessions (Section VI), and test whether the speed of recovery depends on the type of shock (Section VII).

II. Theories of Crises and Growth

Crises and other negative shocks may, in theory, impose only a temporary restraint on output, but lead to strong future growth that offsets the initial decline. First, crises may facilitate beneficial political or economic reforms. Corrective policies could spur an economic recovery above the original trend line if they reduce inefficiencies. Second, following on the idea of Schumpeter’s (1942) creative destruction, recessions may cleanse the economy of inefficient firms, leading to higher productivity and output growth (see Caballero and Hammour, 1994). If either of these theories holds, crises or contractions may benefit long-term growth and we should be able to find evidence of strong recoveries following downturns. Using a structural vector auto regression for U.S. data, Gali and Hammour (1993) find evidence that recessions lead to higher productivity growth in the medium to long term. However, Caballero and Hammour (2005) find evidence that recessions reduce rather than increase the cumulative amount of restructuring in U.S. manufacturing firms. In general, if the “business cycle” implies that output reverts to trend, then there should be a fast growth recovery phase following a contraction. Reverting to the original path could occur, for instance, if a recession leads to a temporary disruption to economic conditions or a temporary fall in capacity utilization or employment, which is reversed as good times return.

Output loss from a balance of payments crisis is more difficult to generate from theoretical models. Chari, Kehoe, and McGrattan (2005) show that in a standard general equilibrium model, a decline in output cannot be accounted for by a tightening of a country’s collateral constraint on external borrowing. The sudden stop of capital inflows leads to an increase in output, not a drop. The sudden stop must be coupled with other frictions, such as an advance-payment constraint on intermediate goods, in order to generate output loss.

Some endogenous growth theories would support a negative relationship between volatility and growth. Martin and Rogers (1997) show that if future benefits of learning by doing are not fully internalized by workers, then recessions are periods in which opportunities for acquiring experience are foregone. Permanent output loss could also be characteristic of a reduction in productivity. Even if productivity growth resumes after a recession, there would be a permanent wedge in the level compared to a pre-recession forecast.

As a counterpart to the theoretical debate, there are opposing statistical views of economic fluctuations that have vastly different implications for welfare. In Hamilton (1989), output is modeled as a stochastic trend that undergoes Markov switching between positive and negative drift rates. Since the regime switch occurs in the growth rate of the permanent component of output, a negative state results in output loss that is persistent. After a recession, output resumes growth with positive drift, but remains on a parallel path below the original trend. Thus, a country hit by a negative shock would be worse off in the long-run relative to one that was never hit by the shock. In the Friedman (1993) plucking model, output springs back to the original trend during a fast growth recovery phase. Mean reversion implies the shock has no long-run impact. Both models involve “V-shape” growth recoveries, except that the Friedman model suggests that growth would be temporarily higher during the recovery than during a normal expansion. In this paper, we investigate whether output returns to the original trend path following a recession. Thus, we are testing whether there is a reversion to trend in the aftermath of economic contractions, including those related to financial and political crises.

The statistical response of output to shocks can also be compared to predictions of endogenous and exogenous growth models to distinguish between these models.

Exogenous growth models produce a steady-state relationship between capital per effective worker (k), output per effective worker (f(k)), the savings rate (s), population growth rate (n), depreciation (d), and exogenous technological progress (x) as described by the following equation:

sf(k*)=(n+d+x)k*

The steady-state level of capital and output per effective worker are determined by s, n, d, and x. The growth rate of aggregate capital and output depend on population growth and labor-augmenting technological change.5

The key feature of an exogenous growth model is that the production function exhibits diminishing returns to capital. If a negative shock reduces capital (per effective worker) below its steady-state level, the boost in the marginal product would lead to a high investment spurt. Growth in capital and output would be strong but diminishing as capital approaches the steady state. This theoretical response to a shock to capital would imply reversion of capital and output to trend, as in the Friedman model, and positive serial correlation in growth rates. Other sources of shock in an exogenous growth model include a change in the savings rate, population growth rate, depreciation rate, and growth of labor-augmenting productivity. For example, a decline in the savings rate would reduce the steady-state level of capital and output per effective worker. If the initial capital stock was above the new steady state, capital would decline until it reached the new steady state, implying that growth would be serially correlated. A sharp fall in capital per effective worker induced by a substantial decline in the steady state could more than offset the positive population and productivity growth rates, and aggregate capital and output could contract. Growth would return to normal (at population and productivity growth rates) once the steady state was reached. Thus, this type of negative shock would induce a recession that gradually petered out, turning into a weak recovery.

Endogenous growth models have different implications for transition and steady-state growth in response to shocks. The long-term growth rate is determined by deep structural parameters of preferences and technology and by policy variables. In these models, the production function exhibits constant returns to capital per effective worker. Thus, shocks to capital will have a lasting influence on the level of output. Shocks to the structural parameters or policy variables would have a lasting impact on the growth rate.

In both types of growth models, the level of productivity is a summary measure of all non-rival inputs in production. Given that productivity enters linearly in the production functions of both models, the persistence of a productivity shock on output will mirror the process of persistence for productivity itself.

This paper thus tests the statistical properties of recoveries and compares them to the theoretical models of growth and volatility discussed above. We also test whether volatility has an impact on the convergence of income levels across countries.

III. Data Sources and Descriptive Statistics

The main focus of the paper is the empirical behavior of aggregate GDP and its growth rate. As our primary data source, we use GDP growth rates from the World Bank’s World Development Indicators (WDI). This dataset contains the largest sample of countries. As a secondary analysis, we study growth per capita and we also test for the convergence in the level of income per capita. For this analysis, we use the Penn World Tables database (Heston, Summers, and Aten, 2002), which is a mainstay of the growth literature due to its comparable levels of GDP per capita. The WDI contains separate series with comparable levels of GDP per capita, but these are available only from 1980. Thus, our data consists of unbalanced panels of annual observations spanning 192 countries from 1960 to 2001 for the World Bank dataset, and 154 countries from 1960 to 2000 for Penn World Tables dataset. These are the two broadest datasets available and common in the literature.

We also investigate the sources of negative shocks to growth. Potential explanations for recessions include financial crises, civil wars, changes in government regimes, and changes in the terms of trade. We also expect that the extent of a country’s trade and financial openness may impact its growth and recovery from recession, although the sign (positive or negative) of the impact is uncertain.

Crises are defined in two ways. First, we obtain banking and currency crisis dates from Kaminsky and Reinhart’s (1999) influential study on twin crises. However, the drawback of this source is that there are only 23 countries included in the study. Second, we construct an exchange market pressure index (EMPI) for each country as the percentage depreciation in the exchange rate plus the percentage loss in foreign exchange reserves. This formulation makes indices comparable across countries.6 A dummy variable for a crisis is formed for a specific year and country if the EMPI is in the upper quartile of all observations across the panel. The construction of the dummy variable allows us to identify a recession episode that coincides with a currency crisis for comparison to a recession corresponding with a non-crisis episode. We obtain banking crisis dates on a large set of countries from Caprio and Klingebiel (2003).

The data for civil war is obtained from Sarkees (2000) Correlates of War Intra-State War Data, 1816–1997 (v3.0) (at www.correlatesofwar.com), which updates the work of Singer and Small (1994). The dataset identifies the participants of intrastate wars. We form a dummy variable for internal conflict by assigning a value of unity for a country in the years of conflict and zero otherwise.

The dummy for trade liberalization is formed from the dates of trade liberalization available in Wacziarg and Welch (2003). We assign a value of zero to pre-liberalization years and unity to the year of trade liberalization and subsequent years. We obtain measures of financial liberalization from the Financial Reform database compiled by the IMF’s Research Department (Omori, 2005). We use both the overall index of financial market liberalization and one of its components, capital account liberalization. The overall index also includes directed credit/reserve requirements, interest rate controls, entry barriers/pro-competition measures, banking supervision, privatization, and security markets. Along these various dimensions, countries can and sometimes do backtrack.

The terms of trade data comes from two sources: the IMF’s International Financial Statistics (IFS) unit price of exports and imports, and world commodity prices from COMTRADE. The index from the IFS source is a ratio of unit price of exports to unit price of imports. For the other source, we use the weights of the top three exports for 60 countries provided by Cashin, Cespedes, and Sahay (2002). We weight the world price of commodities to their share in exports of these countries to construct the TOT index. If the data for some country is available from both sources, we use the IFS unit prices.

The data on change in government comes from Polity International. The regime durability variable measures the number of years since the most recent regime change, defined by a three point change in the polity score over a period of three years or less or the end of transition period defined by the lack of stable political institutions. We construct a dummy which takes a unitary value when the durability variable becomes zero. The polity score is derived from codings of the competitiveness of political participation, the openness and competitiveness of executive recruitment, and constraints on the chief executive. The polity scale ranges from +10 (strongly democratic) to -10 (strongly autocratic).

Table 1 shows average and median growth rates in the two datasets. Growth averages 3.7 percent per year in the World Bank dataset, and falls to 2.1 percent in Penn World Tables due to population growth. Average growth for expansion years in the World Bank dataset is 5.7 percent, very similar to expansion growth rates immediately preceding and subsequent to a recession. Median growth rates in expansions are lower-than-average rates due to some positive skewness. Median growth rates in the year and three years immediately following a trough are about ½ percentage point lower than in a typical expansion year. In the Penn World Tables data, the expansion years surrounding a recession average slightly higher growth, but the median growth rates are the same or lower. In comparing the two datasets, it should be noted that for a common country and sample period, the Penn World Tables data could include more episodes of “recession” to the extent that output growth in a particular year, although positive, was insufficient to outpace population growth. In Table 1, the three-year average growth rate immediately before and after a recession is higher than the average growth rate because, by definition, a peak year and the year after a trough are expansion years.

Table 1.

Average Growth Rates

article image

For the sample of World Bank and Penn World Tables growth rates, we calculated the average cumulative loss and average length of recessions (defined as years of negative growth). The descriptive statistics and total number of recessions for each sample are shown in Table 2. The recession statistics are further broken into income groups and regions, as well as recessions corresponding with crises, wars, new governments, and countries that have liberalized their trade or financial systems.

Table 2.

Characteristics of Recessions

article image

The cumulative output loss in a recession averages 7½ percent for the full sample of countries in both datasets and recessions last 1.6–1.8 years on average. Recessions are much shallower and shorter for high-income countries—about 3–4 percent cumulative loss over 1.4 years—than for any other group. Civil wars, changes in government regime, and banking crises correspond with the deepest and most prolonged recessions. Countries that are most open to international capital flows and, more generally, countries that have the most liberal financial systems tend to have shallower and shorter recessions. Countries that have partially liberalized financial systems have somewhat deeper and longer recessions than those with more complete liberalization, although the loss and duration are lower than the average of all countries. Transition countries have the deepest recessions of the regional groups, partly because data is available only from the early 1990s when most began their transitions.

IV. Down and Out

Following an economic downturn, is there a high-growth recovery phase?

This section searches for a high-growth recovery using a variety of analytical tools and econometric tests. Through the use of a “timeline,” we illustrate the typical behavior of output in the years leading up to a peak, through the recession, and for several years after the trough. Turning to econometric analysis, we test for a fast-growth recovery following a recession for the full sample of countries and for samples containing different regions and income levels. We test whether the amplitude and duration of an expansion is influenced by the amplitude or duration of its preceding recession, and we also test whether booms impact the depth and length of subsequent recessions. Finally, we examine if the statistical properties of output in the panel more closely follows stochastic or deterministic trends.

Figures 1 and 2 present timelines depicting the average behavior of the level (in logs) of output for a boom-bust-recovery episode in the full World Bank and Penn World Tables datasets as well as various subsamples. A peak (trough) year is defined as a year of positive (negative) growth followed by a year of negative (positive) growth. To construct the timeline, we align the peak years for all episodes in a given sample and show the level of output in a 12-year window around the peak. We compute the average growth rate of expansion years occurring prior to the peak, with the peak year defined as time t = 0. We set the level of output equal to 100 in the peak year of the timeline and use the average growth rate to construct the level of output in years t-1, …, t-6. We use the average length and average cumulative output loss of the recessions in the data to determine the level and date of the trough on the timeline. The output levels following the trough in the timeline are constructed using the average growth rate of expansions occurring in the countries following their troughs. The timelines are intended to illustrate the typical behavior of all peak-recession-recovery episodes, thus a country’s growth rates will necessarily be double counted to the extent that it has multiple peaks and troughs within a twelve year window.7

Figure 1.
Figure 1.

Timeline of Recession: World Bank Data

Citation: IMF Working Papers 2005, 147; 10.5089/9781451861662.001.A001

Source: World Bank, World Development Indicators; and authors’ calculations.
Figure 2.
Figure 2.

Timeline of Recession: Penn World Tables Data

Citation: IMF Working Papers 2005, 147; 10.5089/9781451861662.001.A001

Source: Penn World Tables; and authors’ calculations.

Timelines are shown for boom-bust-recoveries in the full sample of countries (using both World Bank and Penn World Tables datasets) and in countries classified by income groups. Timelines are also shown for recessions corresponding to currency crises, banking crises, civil wars, and new governments and for subsets of recession episodes in which the country maintained a liberal trade regime, capital account, or financial system.

The timelines illustrate that output declines with the recession, but in the ensuing recovery, it does not recoup the level associated with the linear extrapolation of the original trend. In the World Bank dataset, a few percentage points of the output lost during the recession is recuperated in the recovery for episodes associated with civil wars and banking crises, but the gap widens for all other samples. Recessions lead to permanent losses in the level of output for all samples, at least through the end of five years after a trough.

Individual country episodes disproportionately resemble Hamilton recoveries (Figures 35). The Asian crisis countries, OECD countries, and many other episodes display output losses that appear to persist. Very few cases of complete Friedman recoveries can be found. One exception is the 1995 Mexican crisis, in which output appears to revert to the original pre-crisis trend line. The Mexican crisis is one of the better known and explored case studies, and this may explain the general unfamiliarity in the literature with the phenomenon of permanent output loss during a financial crisis. Several cases of partial Friedman recovery are shown in Figure 5, including several African countries. These cases illustrate that even if growth is rapid immediately following a contraction, output may not fully revert to trend. We also test formally for the difference in growth rates during an average expansion year and an expansion year occurring in the immediate aftermath of recessions. We define a “trough” nonparametrically as a year of negative growth that is followed immediately by a year of positive growth. That is, troughs are dated according to the calculus rule so as to be consistent with turning points in output:8

Figure 3.
Figure 3.

Episodes of No Recovery in Selected Crisis and Asian Countries

Citation: IMF Working Papers 2005, 147; 10.5089/9781451861662.001.A001

Figure 4.
Figure 4.

Episodes of No Recovery in Developed and Other Countries

Citation: IMF Working Papers 2005, 147; 10.5089/9781451861662.001.A001

Figure 5.
Figure 5.

Episodes of Complete and Partial Recovery

Citation: IMF Working Papers 2005, 147; 10.5089/9781451861662.001.A001

Troughit={1gt0 and gt+1>00gt0 and gt+100gt>0}

The “recovery phase” is the one or more years of positive growth after the trough. Dummy variables are constructed corresponding to the years in the recovery phase.

The econometric specifications exploit the time variation of growth within each country. Restricting the sample to expansion years, we test the magnitude of the growth rate (g) in a recovery phase following a trough. We allow the average rate of growth to differ across countries by imposing heterogeneous intercepts (using fixed effects) in the panel of annual growth rates. Although each country is allowed a different growth rate, we pool information on growth in the recovery phase by imposing a homogenous slope. Pooling has two advantages. First, several countries have insufficient data to estimate an individual slope coefficient, and the information could not be used unless pooled with other country episodes. Second, the pooled estimate provides summary information about the typical response, even if we can expect a variation around it. Since the growth rate in the year following a trough is a positive by definition, we condition the dependent variable on positive growth to compare the strength of recovery to the average growth during expansions. Thus, our basic equation has the following specification:

(gi,t/gi,t>0)=αi+β*Troughi,t1+εi,t

where Troughi,t-1 is a lag of the indicator dummy variable described above. In alternative specifications, the variable Trough(-x,-y) indicates that the dummy variable will take the value of unity for any year in a given country if a trough occurred either x or y years prior. The regressions are corrected for heteroskedasticity using feasible generalized least squares, which weights equations by the inverse of their variances. This procedure should also help reduce the potential for measurement error in driving the results. We also use the White (1980) heteroskedasticity consistent covariance matrix estimator.

Proof of “recovery” or reversion to trend, as in Friedman’s (1993) model, would require a significant positive coefficient on the dummy variables for the year or years following the trough. A coefficient of zero would be consistent with the Hamilton (1989) model, implying that output drifts up at its normal expansion rate.9

Following a recession, growth rebounds at a rate significantly below that of an averageexpansion year. Given the failure of output to revert to trend line, countries experiencing many shocks tend to fall behind.

The regression results shown in Table 3 are not consistent with mean reversion (or the Friedman model of recovery). Indeed, in the year after a trough, growth is significantly lower than in a typical expansion year.10 Recoveries are sluggish at the beginning. The next few years are indistinguishable from a normal expansion. Therefore, there is no evidence of a fast rebound that would bring output back to its pre-contraction trend. Contractions seem to have permanent effects on the level of output in each country.

Table 3.

Strength of Recoveries

article image
Note: T-stats are below the coefficients; GLS panel regression with fixed effects and robust standard errors.

Falling behind means staying behind unless the trend growth rate is high enough to offset the impact of the negative shocks or if there are compensating positive shocks. The panel regressions use a fixed effect estimator, which allows a different average growth rate for each country. If countries with more frequent contractions have stronger expansions (through the fixed effect) or shallower recessions, then they may still have higher long-run growth. For example, neoclassical theory would predict that countries open to international capital flows would have a high potential for growth as external savings would facilitate investment in projects with high marginal returns. On the other hand, international capital flows have been volatile, particularly in the last decade. These arguments would suggest a positive relationship between volatility and growth.

We tested the hypothesis that the average growth rate is different for countries with varying proportions of recessions. We estimated cross-section regressions relating the average long-run growth rate to the proportion of years in recession. Table 4 illustrates that countries experiencing more frequent years of negative growth have sharply lower average growth rates than stable countries. Breaking out average growth into expansions and contractions, we find statistically significant evidence that expansions are stronger for volatile countries in the Penn World Tables dataset, but this advantage is more than offset by sharply deeper contractions. In the World Bank dataset, recession-prone countries have marginally weaker expansions (not significant) and severe contractions. Thus, the results show that countries experiencing more frequent contractions lose considerable ground. For example, a country experiencing contractions 20 percent of the time will have an average annual growth rate approximately 1 percentage point lower than a country experiencing contractions only 10 percent of the time. Part of the explanation is the direct consequence of forgone growth, but the loss goes beyond the difference in average growth during an expansion year and a recession year.

Table 4.

Trend Growth and Volatility

article image

On average, countries with frequent troughs lag behind. By way of illustration, compare hypothetical countries A and B that begin at the same level of output. Country A experiences only one recession during the sample period, while country B experiences several downturns. The evidence above suggests that on average, country B’s average growth over its history would be considerably lower than A, and its level of income would trail that of country A.

Heterogeneity

In addition to the results based on the panel restriction of homogeneous slope, we also allow for heterogeneity in the slope estimates across countries. We tally the number of negative coefficients when allowing heterogeneous slopes on recovery. In the year after the trough, we find negative slope coefficients on 107 out of the 172 countries for which there was sufficient data for this regression in the World Bank data (Table 5). The probability of this result occurring with a “fair coin toss” is less than 1 percent.11 The proportion of negative slopes among those that are individually significant is even higher and is statistically significant for both datasets. The number of negative slopes two years after a trough appears to be a fair coin toss, but the individually significant coefficients remain disproportionately negative (Table 6). Group mean estimators allow each country to have a different slope coefficient, but can be tested for whether the slopes tend to be significantly different from zero. In the full sample of World Bank data, the distribution of slopes is centered at a value of −0.4, and the null hypothesis that the mean of the distribution is zero can be rejected at a 1 percent confidence level. The null of zero cannot be rejected for the Penn World Tables dataset for the full sample. However, the group mean estimate of the slope coefficient becomes considerably more negative for the recent sample from 1990, and the null hypothesis of zero can be rejected for both datasets.

Table 5.

Heterogeneous Slopes

article image
Note: In sign tests, asterisks denote significance of number of negative coefficients in total relative to a fair coin toss.
Table 6.

Individual Slopes Table

article image

We also parse the data into high, upper-middle, lower-middle, and low income groups. Income groups are classified according to the World Bank Atlas methodology. We test whether growth in recovery is significantly positive. Using the World Bank data, we find that there is a sharp difference between recoveries in higher and lower income groups (Table 7). In high- and upper-middle-income countries, the recovery is weaker than an average expansion year by a negative margin that is significant even at the 1 percent confidence level. Only in low-income countries can we find statistically significant evidence of a partial rebound, although the magnitude of the growth differential is well below that required to return to the original trend line.12 Interestingly, recoveries become even weaker in the recent sample period from 1990–2001, with all income groups having a negative and highly significant coefficient.

Table 7.

Income Groups

article image
Note: T-stats are below the coefficients; GLS panel regression with fixed effects and robust standard errors.

In order to check if the results are being driven solely by developing countries or by a specific region, we tested the strength of recovery for each of several regions. Regions were designated as industrial countries and continental or island blocs, as identified by a country’s World Bank numerical code. We find a negative coefficient on the year after the trough for each region, except for African countries (Table 8). The African results are consistent with Collier (1999), who finds rebounds in some African countries following a civil war. The industrial countries show a negative and significant coefficient, and transition countries have the largest negative coefficient of all the regions. The recent period from 1990 to 2001 shows generally weaker recoveries, similar to the results of the income breakdown. Asian, Middle-Eastern, and Transitions countries have highly significant negative coefficients on the recovery, and the coefficient on African countries turns negative and significant at the 1 percent level.

Table 8.

Regional Tests

article image
Note: T-stats are below the coefficients; GLS panel regression with fixed effects and robust standard errors.

Does the positive coefficient for Africa over the full period indicate an African recovery miracle? Unfortunately not. The average output loss in a recession is 6½ percent, whereas the rebound is less than ½ percent. Moreover, the results seem to be sensitive to a few outliers. Several African outliers with strong rebounds are displayed in Figure 5, such as Rwanda, Chad, Togo, Liberia, and Guinea-Bissau. However, even among the outliers with strong rebounds, there appear to be very few cases of full recovery.

What if it takes longer than a few years to regain lost output?

The regressions above test the magnitude of growth during recovery, in the year immediately after a trough and for several years after a trough. We also look at complete recessions (from peak to trough) and expansions (from trough to peak). For each of the datasets, we select all episodes of a complete recession (bounded at the beginning and end by at least one year of positive growth) followed by a complete expansion (bounded at the beginning and end by at least one year of negative growth). The World Bank dataset contains 469 pairs of complete recession and expansion episodes, and the Penn World Tables dataset contains 747 pairs (Table 9). We calculate the amplitude (absolute value of the cumulative output change) and the number of years in duration of the recession and subsequent expansion. We test whether the amplitude or duration of the recession influences the amplitude or duration of the subsequent expansion. If, after a deep recession, output were to rebound back to the original trend path, we would expect to find that the amplitude of a recession is positively associated with the amplitude of the following expansion.

Table 9.

Tests of Amplitude, Duration, and Steepness of Complete Expansions

article image
Note: T-stats are below the coefficients, ST refers to stochastic trend and DT refers to deterministic trend

For each pair of complete recession and expansion, we also verify whether the amplitude or duration of the recession leads to a significant change in the steepness (amplitude/duration) of the subsequent expansion compared with the steepness of the prior expansion. A significant change in the steepness of an expansion following a deep or prolonged recession would provide some evidence of a rebound in output or a growth takeoff following a negative shock, even if the rebound occurs over several years.

The results are not consistent with rebound or growth takeoff following severe recessions. Contrary to Friedman’s plucking model, Table 9 shows that expansions are weaker when preceding recessions are longer, and expansions are shorter when preceding recessions are longer and deeper. The depth of the recession is not significantly related to the strength of the subsequent expansion. The change in the steepness of expansions is not significantly related to the amplitude and duration of the prior recession. Thus, a deep or prolonged recession does not lead to a rebound or growth takeoff.

Are deep recessions preceded by strong, unsustainable booms?

If a recession is triggered as an adjustment to an excessively strong economic boom, then there may be no need for a strong recovery following the recession. This hypothesis could explain the scarcity of strong recoveries in the data. We tested this hypothesis in two ways. First, we checked whether recessions were in fact preceded by stronger than average growth. We formed dummy variables that picked up expansion years for the three years immediately preceding a recession and tested whether growth was higher than in a typical expansion year. The evidence, shown in Table 10, suggests that the years immediately prior to recessions tend to experience significantly lower growth. This result contradicts the hypothesis that strong booms trigger recessions. Second, we calculated the amplitude (cumulative growth) and duration of each complete expansion, and tested whether either of these factors could explain the amplitude or duration of the subsequent recession. As shown in Table 10, the amplitude of an expansion has a negative, but insignificant, impact on the amplitude (depth) of the subsequent recession. Strong expansions precede significantly shorter recessions. Prolonged expansions precede significantly shorter and shallower recessions. Thus, this evidence is contrary to the hypothesis that strong booms are the cause of deep or long-lasting recessions.

Table 10.

Tests of Strong Boom Prior to Recession

article image
Note: T-stats are below the coefficients, P3 refers to a dummy for peak year and previous two years. The first column of regression results is for available data in the period 1960-2001, using GLS and fixed effects with robust standard errors.

Persistence of shocks

An alternative way to evaluate the persistence of shocks to output is to test between stochastic and deterministic trend models as the data generating process. These tests do not specifically focus on the recovery phase of the growth process, but provide information about the overall behavior of output. We test for unit roots in output and per capita output using the Hadri panel unit root test (Table 11). For both datasets, we can reject the null hypothesis of no unit root, in favor of the alternative hypothesis of a unit root. A problem with Hadri and many other unit root tests is that they fail to account for dependence among the cross-sections. Pesaran (2003) proposes a simple panel unit root test in the presence of cross-section dependence. The test adds the cross section averages of lagged levels and first-differences of the individual series to a standard Dickey-Fuller regression (cross section augmented Dickey-Fuller or CADF). The average of the t-statistics (CIPS) in individual CADF regressions is compared to the critical values tabulated for the three main specifications of the deterministic variables: no intercepts or trends, individual-specific intercepts, and intercepts and incidental linear trends. In the World Bank and Penn World Tables datasets, all three specifications produced average test statistics smaller (in absolute value) than the critical values. The test statistics for the World Bank dataset are −1.09, −1.33, and −1.86 for models with no intercepts or trends, with an intercepts only, and with intercepts and trends, respectively compared to critical values of approximately −1.43, −2.0, and −2.5 at the 10 percent level for a panel with T between 30 and 50 and N between 100 and 200. The test statistics for the Penn World Tables dataset are −1.22, −1.56, −1.99, respectively. Thus, the null of a unit root cannot be rejected.

Table 11.

Panel Unit Root Tests

article image

Probabilities are computed assuming asympotic normality

Automatic selection of maximum lags and automatic selection of lags based on SIC: 0 to 9Newey-West bandwidth selection using Bartlett kernel

We find that growth and growth per capita exhibit statistically significant positive serial correlation. Table 12 presents estimates for a fixed effects specification. The lagged dependent variable is significantly persistent, particularly for the World Bank data. Although the fixed effects estimation eliminates endogeneity between the lagged dependent variable regressor and the country effect, the specification introduces correlation between the lagged dependent variable and the averaged error term used for the fixed effect. The coefficient estimate would be biased downward on the order of 1/T. Therefore, we also estimate a differenced equation and use the second lagged level of the dependent variable as the instrument.13 The magnitude of serial correlation in the growth rate increases somewhat for each sample.

Table 12.

Serial Correlation of Growth Rates

article image
Note: T-stats are below the coefficients; sample uses available data from 1962-2001

We conduct Monte Carlo experiments for different output processes to verify their conformity with the properties of the data found above. In particular, we generate artificial samples approximately matched to the first two moments of the actual data and the dimensions of the panel observations. We base the experiments on a data generating process (DGP) that follows a stochastic trend with positive serial correlation in the growth rates, and on a DGP that follows a deterministic trend with serial correlation in the growth rates.

  • The stochastic trend model is given by the following equation:

  • yit=yit1+μ+ϕΔyit1+εitεti.i.d.N(0,σ2)

  • The deterministic trend model can be represented as:

  • yit=yi0+at+ϕΔyit1+εitεti.i.d.N(0,σ2)

To broadly match moments of the data, we set µ = a = 0.03, ϕ = 0.2 and σ = 0.07.

We test the artificial panels for the magnitude of growth in a recovery following a trough. The Monte Carlo study shows that data generated as a stochastic trend yield a significantly negative coefficient on the dummy for an expansion year after a trough, whereas the data generated as a deterministic trend yield a significantly positive coefficient (Table 13). In addition, estimates for serial correlation in the growth rate are positive and significant for the stochastic trend data, regardless of whether it is estimated in levels using fixed effects or in differences with the second lag of growth as the instrument. Estimates of serial correlation in growth for data generated with a deterministic trend are negative regardless of estimation specification. Thus, the stochastic trend DGP is consistent with the properties of the data, but the deterministic trend DGP shows results opposite to the data.

Table 13.

Monte Carlo Results

article image
Note: T-stats are below the coefficients, ST refers to stochastic trend and DT refers to deterministic trend

Which aspects of business cycle and growth models are consistent with the empirical results?

The main properties of the data are: (1) negative shocks to the level of output don’t dissipate, and (2) growth is positively serially correlated. Which theories of business cycles and exogenous or endogenous growth models are consistent with these properties of the data?

Business cycles driven by temporary shocks to capacity utilization and the unemployment rate, stemming from demand innovations for instance, would not be consistent with the data. Such temporary shocks would be trend-stationary, and growth would exhibit negative serial correlation. At the other extreme, productivity shocks could generate the results in the data provided they are persistent and productivity growth is serially correlated.

In an exogenous growth model, a reduction in the capital stock, holding constant the determinants of the steady state, should elevate growth. Yet, the lackluster recoveries in the data are not consistent with this source of disturbance, except for low-income or African countries, where there is some evidence of fast recovery. On the other hand, shocks leading to a decline in the steady state below the level of the capital stock prevailing at the time would match the patterns observed in the data. Capital per effective worker would decline until the new steady state was attained. If the shock is large enough, aggregate capital and output would also decline until offset by population and productivity growth. This pattern would produce a weak recovery, consistent with the data. A financial crisis could be a source of reduction in the savings rate, one determinant of the steady state, if it increases the risk premium from borrowing externally to finance investment. The crisis could also have a more immediate impact on output if a loss in external financing makes it difficult to purchase intermediate inputs from abroad, thus directly depressing the production of output.

Turning to endogenous growth models, the balanced growth rate is a function of preference and production parameters and policy variables. Permanent innovations in these variables would raise or depress the rate of growth. Although growth rates are serially correlated, they do not appear to have a unit root. Thus, innovations to the variables influencing the growth rate could only be consistent with the data if they are temporary. In contrast, innovations to capital would be persistent since production is linear in capital (broadly defined). Nevertheless, empirical literature’s findings of conditional convergence would cast some doubt on the linearity assumption.

V. Divergence: Its the Crises, Stupid

Are contractions partly responsible for absolute divergence?

Convergence is a property of exogenous growth models resulting from diminishing private returns to the factors that can be accumulated (various forms of capital).14 Countries with low levels of capital stock would be expected to grow faster than capital-rich countries holding constant the factors that determine the steady state level of capital. However, empirical studies have found that countries with low initial income (e.g., in 1960) have grown more slowly on average than countries that started off rich (see Barro and Sala-i-Martin, 2004). Absolute convergence holds only for countries or regions that started off at somewhat similar levels, such as the OECD countries, U.S. states, and Canadian provinces. Convergence across a broad set of countries has been obtained only when conditioning on factors that could influence the steady state (such as the savings rate).

Given the striking result that a negative shock permanently reduces the level of output on average, we wondered if susceptibility to shocks could be contributing to the absolute divergence found in the cross-country growth literature. An exogenous growth model predicts that the capital stock per effective worker in capital-poor countries should rise rapidly relative to that of capital-rich countries. From a position below its steady state level, capital per effective worker should continue to grow toward the steady state. Aggregate capital would further expand due to the exogenous increases in population and technology.

How would we account for years of negative growth in output? If output reverted to a deterministic trend in response to demand-based business cycles, it might be unnecessary to worry about recessions in applying convergence tests. However, the persistent impact of negative shocks on output casts doubt on this portrayal of recessions. Explanations within the growth framework could include reductions in the country’s population or technological knowledge, destruction of the capital stock for an exogenous reason or unusually high depreciation experience, or an investment collapse induced by a fall in the steady state. We would only expect to find convergence between countries experiencing these negative shocks and others that aren’t experiencing them if the recessions are caused by destruction of capital with a sufficient period of adjustment to regain lost ground. We find evidence for a growth rebound only in the low-income and African countries. Even for this sample, the rebound does not appear to be strong enough to completely regain lost ground. Thus, the ceteris paribus conditions for observing absolute convergence do not appear to hold. That is, we would not expect to find convergence if, for example, capital-poor countries experienced shocks that reduced their steady states during the sample period.

To explore the impact of shocks on convergence, we compare a standard convergence regression to one in which convergence is conditioned on expansion phases rather than on average long-run growth rates, which include episodes of negative shocks.

There is statistically significant evidence of convergence if conditioning on the expansion rate of growth for each country.

We first test the standard hypothesis of beta convergence in a cross-country regression and find the standard results of absolute divergence across countries. However, we then calculate the average expansion growth rate for each country. When the test is confined to the expansion phase only, we find strong evidence of convergence (Table 14). The divergence of incomes across countries is thus a function of the crises and shocks, which have two effects on poor country incomes. First, an average recession year is more severe for countries that start out initially poor, leading to the highly significant divergence when the sample is confined to recession years. Second, the poor countries receive a significantly higher proportion of recession years. These results alone do not necessarily distinguish between trend-stationary and stochastic trend processes, as convergence conditional on expansion could be consistent with either data process given appropriate assumptions. However, in the previous section, we found that the data are not consistent with volatility around a deterministic trend.

Table 14.

Phase Conditional Convergence

article image

The results provide evidence of conditional convergence complementary to that found in the literature. Standard convergence regressions attempt to condition on factors that would lead to different steady states across countries. Typically, the factors will be measured at their initial values. However, if the steady state changes during the sample period, there is no reason to expect convergence. Our test conditions on shocks that could change the steady state or level of productivity, as identified by their impact on output. The findings further demonstrate the consequences of negative shocks and point to a higher frequency of these occurring in poor countries.

The figure at the right provides a stylized illustration of the main findings. In the Penn World Tables data, countries that are poor at the beginning of the sample period experience stronger expansions than the initially rich countries. If not for the recession years, the poor countries would converge toward the rich. However, the countries that are initially poor suffer from more frequent and deeper recessions than the initially rich, which leads to divergence.

This evidence argues that shocks do indeed derail growth. Poor countries show rather respectable expansions, which by themselves would propel their incomes to converge with those of rich countries. This finding casts doubt on the savings trap view, which is based on the inability to save and invest, even in good times. The possibility that political and financial crises disrupt growth would be consistent with the literature that finds deep-rooted institutional quality as an important determinant of both the frequency of crises (Acemoglu, Johnson, Robinson, and Thaicharon, 2003) and long-run growth (Acemoglu, Johnson, and Robinson, 2004), although there could be other causes of the instability. The results here point to the link between crises and long-term growth through the negative persistent impact of crises on the level of output.

VI. Wars, Crises, and Regime Change, Oh My!

Given the detrimental impact of contractions on the long-term level of output, what are some potential sources of such disturbances?

We investigate some of the likely suspects for economic contractions: civil wars, financial crises, political instability as measured by regime change, and negative terms-of-trade shocks. Given the richness of the panel dataset, we also look at the impact of trade liberalization on subsequent growth. We regress growth on dummy variables for wars, crises, and regime change.

Civil wars, financial crises, and changes in government regime have large and significant negative impacts on the growth rate.

In Table 15, we estimate the impact of financial crises on growth. We use two different dating conventions to check for robustness. The first set of currency crises and banking crises dates provides a large number of countries and observations. We construct the currency crises dates and take banking crises dates from Caprio and Klingebiel (2003). We use Kaminsky and Reinhart’s (1999) sample for the second set of dates. Table 16 shows the detrimental impact of civil wars, and changes in government regime on growth.

Table 15.

Crises and Growth

article image
Note: T-stats are below the coefficients; GLS panel regression with fixed effects and robust standard errors.C-S and C-K dates refer to the authors’ dates for currency crises and Caprio-Klingebiel dates for banking crises
Table 16.

Other Shocks and Growth

article image
Note: T-stats are below the coefficients; GLS panel regression with fixed effects and robust standard errors.

Financial and political instability have strong associations with recessions. Financial crises coincide with more than one-third of years of negative growth. Half of all recession years coincide with crises, regime change, civil war, or some combination of these variables.

article image

We also include measures of openness to trade in the list of variables as openness may influence vulnerability to external shocks. We find that trade liberalization has a significant positive effect on growth. Terms of trade movements have the expected positive correlation with growth, but the magnitude is negligible. The explanation likely owes to the poor quality of data on terms of trade. Data are difficult to obtain, and although we have patched together information from several sources, large gaps remain for many countries.

VII. How Shocks Interact with Recoveries

Does the strength of the recovery depend on the source of the negative shock?

We interact the disturbance variables that are associated with recessions, as discussed above, with the dummies for the expansions immediately following troughs, controlling for the lower-than-average growth in the year after the trough. A negative coefficient on an interaction variable would imply a weaker-than-normal recovery even after taking into account that growth in the recovery phase of an expansion is lower than growth in an average year of expansion. To maximize the number of observations, we use all available countries and time periods, and we focus on pooled estimates. However, some of the interaction regressors are available for only a subset of the countries.

Financial crises lead to weak recoveries.

Table 17 shows that recoveries are weak when the output contraction is associated with a financial crisis. Both banking crises and currency crises lead to significantly lower growth in the aftermath of a recession linked with them. Two lags of currency crises are shown because these crises often occur just before or during the early stage of a recession and therefore may not always coincide with the trough. We also interact the recovery year with measures of the extent of liberalization of the capital account regime and overall financial markets, and find that recoveries are weak in countries with more liberalized capital accounts and financial markets. This result suggests that lack of access to financing may be one of the mechanisms that prevents the recovery of output to its prior trend. Higher inflows of aid as a share of gross national income boost growth in the recovery.

Table 17.

Recovery Conditional on Financial Crises and Liberalization

article image
Note: T-stats are below the coefficients; GLS panel regression with fixed effects and robust standard errors.Sample is all available data from 1960-2001

The impact of openness and trade liberalization on the speed of recovery is mixed. Openness, measured as exports plus imports relative to GDP, could facilitate adjustment by promoting rapid export growth after a depreciation. However, the coefficient is negative and statistically significant in the World Bank data. The results indicate that recoveries are marginally weaker for open countries, although the economic significance is trivial. As an alternative measure of openness, we used dates on trade liberalization available from Wacziarg and Welch (2003). Recessions that occur after a country has liberalized its trade regime (measured as a dummy variable) display significantly weaker recoveries. However, in the World Bank dataset, liberalized trade regimes contribute to strong recoveries after controlling for their interaction with liberalized capital account regimes. Thus, an open trade regime helps countries to adjust, except in countries with liberalized capital accounts, possibly due to financing constraints on imported intermediate inputs to trade.

Political change toward autocracy leads to weak recoveries.

Political crises also tend to dampen growth in recovery, except when associated with a change in government toward a more democratic system. The Polity International dataset provides a summary variable (Polity2) for each country that measures the extent to which a government is democratic or autocratic, with positive values indicating a more democratic system. A recession associated with a change in the Polity2 score toward more democracy is followed by a faster-than-average recovery (using World Bank data). Therefore, a political crisis involving a change in government regime could be expected to exhibit a weak recovery. However, the change in government regime (three or more point change in the Polity2 score) is insignificant. One reason for this result may be that the regime change variable does not distinguish between the installation of a more democratic government and a more autocratic government, each with potentially different impacts on growth. Thus, we disaggregate government change into those that move in the democratic versus autocratic directions. Recoveries following change to more autocracy are significantly weak with large (in absolute value) negative coefficients, whereas the recoveries following democratic change are indistinguishable from typical recoveries. Although the interaction of civil wars with recovery is not significant, it seems to be due to some African outliers such as Liberia, Rwanda, and Nigeria. We thus distinguish civil wars in the full sample of countries from those in Africa, and from civil wars in Africa coinciding with change to autocratic governments. We find that civil wars lead to weak recoveries in general, and to very weak recoveries in Africa when the government regime becomes more autocratic. The strong recoveries from civil war appear to reflect African cases in which the regime change was neutral or more democratic.

Results for Africa

The strong recoveries in Africa appear to reflect high aid inflows and democratic change. We separately examine determinants of strong recovery in Africa, since it is the region with the strongest evidence of recovery. As shown in the bottom of Table 18, recoveries are positive and significant when aid inflow is high relative to gross national income. Indeed, the average recovery is no longer positive after separately controlling for aid inflows. There is also markedly different recovery behavior following a change to a more autocratic versus democratic government system. The weak recoveries following autocratic change are highly significant and contrast with the positive and highly significant recovery following democratic change in the Penn World Tables dataset.

Table 18.

Recovery Conditional on Political Crises and Africa

article image
Note: T-stats are below the coefficients; GLS panel regression with fixed effects and robust standard errors.Sample is all available data from 1960-2001, except for Africa sample, which includes 51 countries in World Bank data and 44 countries in Penn World Tables data.

VIII. Conclusions

Using panel data for broad datasets of countries, this paper documents that recessions are typically not followed by high-growth recovery phases, either immediately following the trough, over several years of the subsequent expansion, or even over the complete subsequent expansion that follows a complete recession. Indeed, for most countries, growth is significantly lower in the recovery phase than in an average expansion year. Thus, when output drops, it tends to remain well below its previous trend.

The results can help distinguish between different theoretical models of growth and business cycles, and suggest directions for future research that link volatility and growth. For example, recoveries in the data are not consistent with creative destruction, at least at the level of aggregate GDP. These results would also cast some doubt on the importance of shocks to capital stocks as the source of recessions, but would be consistent with productivity shocks or changes in the determinants of steady-state levels.

The results of this paper highlight the importance of panel studies for identifying for the timing of growth and taking advantage of the temporal variation in the data. Cross-section regressions (even two million of them) that average growth across decades can mask determinants of turning points and make it more difficult to distinguish between explanatory variables. Thus, panel regressions that control for the timing of shocks or policy changes offer a better chance of explaining sources of growth than cross-section regressions.

The paper also highlights that political and financial crises are costly at all horizons. Crises contribute to half of the episodes of negative growth, and there is no evidence that they typically lead to economic reforms or policy adjustments that restore output to trend. Change to a more democratic government system, on the other hand, improves the rebound from a recession. We also find evidence that while trade liberalization increases the long-run growth rate, it can weaken recovery from recession. However, such weak recoveries tend to occur in combination with liberalized capital account regimes, possibly as a result of restricted access to financing for imported intermediate inputs as the confidence of international investors is slow to return following a recession.

When shocks derail growth, incomes diverge. Poor countries have respectable expansions, and therefore do not appear to be stuck in a poverty trap. However, many poor countries do appear to be mired in a crisis trap. Countries that experience many negative shocks to output tend to get left behind and their long-term growth suffers. Thus, while standard growth theory may work well in explaining expansions, a fruitful direction for future research would be to explain the proclivity to wars, crises, and other negative shocks.

References

  • Acemoglu, Daron, Simon Johnson, and James Robinson, 2001, “Colonial Origins of Comparative Development: An Empirical Investigation,” American Economic Review, Vol. 91, pp. 13691401.

    • Search Google Scholar
    • Export Citation
  • Acemoglu, Daron, Simon Johnson, James Robinson, and Yunyong Thaicharoen, 2003, “Institutional Causes, Macroeconomic Symptoms: Volatility, Crises and Growth,” Journal of Monetary Economics, Vol. 50, pp. 49123.

    • Search Google Scholar
    • Export Citation
  • Aghion, Philippe, and Gilles Saint-Paul, 1991, “On the Virtue of Bad Times: An Analysis of the Interaction Between Economic Fluctuations and Productivity Growth,” CEPR Working Paper, No. 578.

    • Search Google Scholar
    • Export Citation
  • Arellano, M., and S.R. Bond, 1991, “Some Tests of Specification for Panel Data: Monte Carlo Evidence and an Application to Employment Equations,” Review of Economic Studies, No. 58, pp. 27797.

    • Search Google Scholar
    • Export Citation
  • Barro, Robert J., 2003, “Determinants of Economic Growth in a Panel of Countries,” Annals of Economics and Finance, Vol. 4, No. 2, pp. 23174

    • Search Google Scholar
    • Export Citation
  • Barro, Robert J., 1997, Determinants of Economic Growth: A Cross-Country Empirical Study (Cambridge, Massachusetts: MIT Press).

  • Barro, Robert J., 1991, “Economic Growth in a Cross Section of Countries,” Quarterly Journal of Economics Vol. 106, No. 2, pp. 40743.

    • Search Google Scholar
    • Export Citation
  • Barro, Robert J., and Xavier Sala-i-Martin, 2004, Economic Growth, Second Edition, (New York: McGraw Hill).

  • Caballero, Ricardo J., and Mohamad Hammour, 2005, “The Cost of Recessions Revisited: A Reverse-Liquidationist View,” Review of Economic Studies 72, (March), pp. 31341.

    • Search Google Scholar
    • Export Citation
  • Caballero, Ricardo J., and Mohamad Hammour, 1994, “The Cleansing Effect of Recession,” American Economic Review, Vol. 84, No. 5, pp. 135068.

    • Search Google Scholar
    • Export Citation
  • Caprio, Gerard, and Daniela Klingebiel, 2003, “Episodes of Systemic and Borderline Financial Crises,” World Bank data set available via the internet: http://econ.worldbank.org/view.php?id=23456.

    • Search Google Scholar
    • Export Citation
  • Caselli, Francesco, Gerardo Esquivel and Fernando Lefort, 1996, “Reopening the Convergence Debate: A New Look at Cross-Country Growth Empirics,” Journal of Economic Growth, Vol. 1, No. 3, pp. 36389.

    • Search Google Scholar
    • Export Citation
  • Cashin, Paul A., Luis Cespedes, and Ratna Sahay, 2002. “Keynes, Cocoa, and Copper: In Search of Commodity Currencies,” IMF Working Paper No. 02/223 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation
  • Cerra, Valerie, and Sweta Saxena, 2005a, “Did Ouput Recovery from the Asian Crisis?,” IMF Staff Papers, Vol. 54 (1), pp. 123.

  • Cerra, Valerie, and Sweta Saxena, 2005b, “Eurosclerosis or Financial Collapse: Why Did Swedish Incomes Fall Behind?,” IMF Working Paper 05/29 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation
  • Chari, V.V., Patrick Kehoe, and Ellen McGrattan, 2005, “Sudden Stops and Output Drops,” Federal Reserve Bank of Minneapolis, Research Department Staff Report 353 (Minneapolis).

    • Search Google Scholar
    • Export Citation
  • Collier, Paul, 1999, “On the Economic Consequences of Civil War,” Oxford Economic Papers, Vol. 51, pp. 16883.

  • Dawson, J. W., and F.E. Stephenson, 1997, “The Link Between Volatility and Growth: Evidence from the States,” Economic Letters, Vol. 55, pp. 36569.

    • Search Google Scholar
    • Export Citation
  • Doppelhofer, Gernot, Ronald Miller, and Xavier Sala-i-Martin, 2000, “Determinants of Long-Term Growth: A Bayesian Averaging of Classical Estimates (BACE) Approach,” NBER Working Paper 7750.

    • Search Google Scholar
    • Export Citation
  • Durlauf, Steven, 2003, “The Convergence Hypothesis After Ten Years,” University of Wisconsin at Madison, manuscript.

  • Easterly, William, Michael Kremer, Lant Pritchett, and Lawrence H. Summers, 1993, “Good Policy or Good Luck: Country Growth Performance and Temporary Shocks,” Journal of Monetary Economics, Vol. 32 (December), pp. 45983.

    • Search Google Scholar
    • Export Citation
  • Friedman, Milton, 1993, “The ‘Plucking Model’ of Business Fluctuations Revisited,” Economic Inquiry, Vol. 31 (April), pp. 17177.

    • Search Google Scholar
    • Export Citation
  • Gali, Jordi, and Mohamad Hammour, 1993, “Long-Run Effects of Business Cycles” (New York: Columbia University Graduate School of Business), unpublished manuscript.

    • Search Google Scholar
    • Export Citation
  • Grier, K.B. and G. Tullock, 1989, “An Empirical Analysis of Cross-National Economic Growth, 1951-80,” Journal of Monetary Economics, Vol. 24, pp. 25976.

    • Search Google Scholar
    • Export Citation
  • Hamilton, James D., 1989, “A New Approach to the Economic Analysis of Nonstationary Times Series and the Business Cycle,” Econometrica, Vol. 57 (March), pp. 35784.

    • Search Google Scholar
    • Export Citation
  • Harding Don, and Adrian Pagan, 2002, ‘Dissecting the Cycle: A Methodological Investigation’, Journal of Monetary Economics, Vol. 49, pp. 36581.

    • Search Google Scholar
    • Export Citation
  • Hausmann, Ricardo, Lant Pritchett, and Dani Rodrik, 2004, “Growth Accelerations,” Harvard University, unpublished manuscript.

  • Heston, Alan, Robert Summers and Bettina Aten, 2002, Penn World Table Version 6.1 Center for International Comparisons at the University of Pennsylvania (CICUP), October.

    • Search Google Scholar
    • Export Citation
  • Islam, Nazrul, 1995, “Growth Empirics: A Panel Data Approach,” Quarterly Journal of Economics, Vol. 110 (November), pp. 112770.

  • Kaminsky, Graciela, and Carmen Reinhart, 1999, “The Twin Crises: The Causes of Banking and Balance of Payments Problems,” American Economic Review, Vol. 89, No. 3 (June), pp. 473500.

    • Search Google Scholar
    • Export Citation
  • Kormendi, R., and Meguire, P., 1985, “Macroeconomic Determinants of Growth: Crosscountry Evidence,” Journal of Monetary Economics, Vol. 16, pp. 14163.

    • Search Google Scholar
    • Export Citation
  • Levine, Ross, and David Renelt, 1992, “A Sensitivity Analysis of Cross-Country Growth Regressions,” American Economics Review, Vol. 82, No. 4, pp. 94263.

    • Search Google Scholar
    • Export Citation
  • Lucas, Robert E., Jr. 1988, “On the Mechanics of Economic Development,” Journal of Monetary Economics, Vol. 22, pp. 242.

  • Lucas, Robert E., Jr., 1987, Models of Business Cycles, (Oxford and New York: Basil Blackwell).