Chapter

CHAPTER 3. Growth Accelerations

Author(s):
Kevin Carey, Sanjeev Gupta, and Catherine Pattillo
Published Date:
February 2006
Share
  • ShareShare
Show Summary Details

Very large and sustained increases in growth rates are necessary if sub-Saharan Africa is to have a realistic prospect of halving income poverty by the year 2015. To meet the poverty Millennium Development Goals (MDGs), sub-Saharan Africa’s real GDP growth rates will have to double from a base scenario to about 7.5 percent.17 Although knowledge about what leads to sustainable, large accelerations of growth in sub-Saharan Africa is limited, it is instructive to look at some recent success stories within the framework of growth accelerations. A paper by Hausmann, Pritchett, and Rodrik (2004) (hereafter Hausmann, Pritchett, and Rodrik) has proposed that the traditional focus of empirical growth research on long-horizon or panel-data growth regressions can camouflage important medium-term patterns in a country’s growth. By looking at jumps in countries’ medium-term growth trends, they argue, one can gain insight into the sources of successful growth transitions. In addition, standard methods do not directly address a policymaker’s key question: how likely is it that a particular country will experience a growth acceleration that is sustained for a period of time?

The recent Regional Economic Outlook (IMF, 2005c), hereafter REO) for sub-Saharan Africa found that countries that have experienced jumps in their growth rates have registered improvements in broad measures of their policy stance and institutional quality. Accelerations appeared to operate via the trade channel, and were accompanied by increases in investment and productivity. Within this group of accelerations, those that were sustained for 10 years had stronger trade and investment, lower debt burdens and higher aid, and more democratic institutions than countries that did not sustain their accelerations.

While the REO’s findings were supported by bivariate analysis (reviewed in Chapter 2, Section A), it is shown here that the broad messages are maintained in a multivariate extension. Two types of investigation are undertaken. The first is a direct analog of the REO bivariate correlations that relate the event of being in an acceleration episode to a range of possible explanatory or associated factors (3.B). The second seeks to explain the timing of an acceleration in terms of a small set of potential triggering factors (3.C). The analysis then turns to comparing an important subset—sustained accelerations—with those that are not sustained, in terms of both associated and triggering factors (3.D). This section finds a robust association between growth accelerations and the trade channel, measures of policy and institutional quality, and productivity growth; in addition, sustained episodes are associated with lower debt burdens. Episodes are triggered by political transitions and economic liberalizations.

A. Identification and Bivariate Correlates of Accelerations

Growth accelerations are identified by a comparison of backward- and forward-looking per capita growth rates calculated over a moving window for each country. Hausmann, Pritchett and Rodrik compare seven-year forward- and backward-looking growth rates of per capita GDP from a given year. An acceleration is identified when the forward-looking rate exceeds the backward-looking rate by at least 2 percent, and the jump is to a level of at least 3.5 percent, with an additional proviso that the post-acceleration GDP level must the exceed pre-acceleration GDP level (to exclude crisis recovery periods). In the interest of focusing the analysis on more recent experiences, this study uses IMF World Economic Outlook real GDP per capita data from 1980–2004, rather than Penn World Tables (PWT) data (available only until 2000), and shortens the acceleration window to five years, allowing identification of acceleration episodes beginning up to 1999. Given lower average sub-Saharan Africa growth rates than in the full developing country sample of Hausmann, Pritchett, and Rodrik, while in our definition an acceleration still requires a jump in per capita growth over a five-year window of at least 2 percent a year, the cutoff for the post-jump growth rate is 2 percent (rather than 3.5 percent). The requirement that the level of GDP per capita must exceed the pre-acceleration level is retained.

This method identifies 34 growth acceleration episodes in the region since 1980, with more such episodes in the 1990s than in the 1980s, including several episodes currently under way. Episodes occur in countries at all levels of per capita income. The original Hausmann, Pritchett, and Rodrik cutoffs would produce 22 accelerations in total, of which 8 are in the 1980s (indicated by an asterisk in Table 1). In the benchmark set, 6 countries experience two accelerations over the full period, indicating nonetheless that accelerations are a surprisingly widespread phenomenon. Equally, however, the presence of accelerations in 28 separate countries over this period of sluggish overall growth points to the difficulty in sustaining them beyond the five year period. This motivates interest, pursued in the probit models below, in analyzing sustained accelerations as well as examining whether the same factors are associated with the accelerations identified by the tighter Hausmann, Pritchett, and Rodrik filter, as in the benchmark set.

Table 1.Acceleration Start Dates and Per Capita Growth Rates for 1980s and 1990s
1980s1990s
Start dateEpisode growthPost-episode growthStart dateEpisode growthPost-episode growth
Botswana*19867.71.2Angola*19934.92.6
Burkina Faso19833.32.9Benin19932.22.0
Burundi19832.4–0.1Botswana*19964.7
Chad*19833.31.4Burkina Faso*19944.73.2
Congo, Rep. of*19845.2–2.7Cape Verde*19924.55.1
Gabon19862.90.5Chad*19998.3
Ghana19832.92.0Côte d’Ivoire*19932.3–4.2
Kenya19842.5–1.6Equatorial Guinea*199429.718.5
Lesotho*19864.22.8Ethiopia*19923.81.4
Mauritius*19847.35.6Gambia, The19952.2
Mozambique*19866.02.4Guinea19942.30.0
Seychelles*19875.72.6Malawi*19944.8–3.5
Tanzania19852.3–1.6Mozambique*19947.15.1
Uganda*19863.94.1Rwanda*19962.6
Zimbabwe19862.6–1.2Senegal19942.21.5
Seychelles19957.5
Sierra Leone*199910.9
Tanzania*19994.0
Zambia19992.1
Source: IMF staff calculations from World Economic Outlook database, 2004.Notes: GDP per capita data in U.S. dollars. Acceleration episodes last five years and are identified as described in text. Post-episode growth rates cannot be calculated for accelerations after 1994. A sustained acceleration (shaded) is one where the average per capita growth was at least 2 percent for five years after the acceleration ended. All growth rates are calculated by a regression of per capita income on a constant trend. An asterisk indicates accelerations where the growth exceeds the 3.5 percent cutoff of Hausmann, Pritchett, and Rodrik (2004). For Chad, this cutoff dates are acceleration to 1981, whereas for Rwanda it begins in 1994.
Source: IMF staff calculations from World Economic Outlook database, 2004.Notes: GDP per capita data in U.S. dollars. Acceleration episodes last five years and are identified as described in text. Post-episode growth rates cannot be calculated for accelerations after 1994. A sustained acceleration (shaded) is one where the average per capita growth was at least 2 percent for five years after the acceleration ended. All growth rates are calculated by a regression of per capita income on a constant trend. An asterisk indicates accelerations where the growth exceeds the 3.5 percent cutoff of Hausmann, Pritchett, and Rodrik (2004). For Chad, this cutoff dates are acceleration to 1981, whereas for Rwanda it begins in 1994.

Empirical investigation sought to identify determinants of accelerations during the 1980s and 1990s using bivariate analysis. A broad range of explanatory variables covering macroeconomic stability, trade, debt, institutions, capital, and geography were examined, some of which can be thought of as triggering an acceleration, and some of which enable an acceleration to continue. We first examine the correlates of accelerations using bivariate analysis, which is useful to give an overall sense of the relationships in the data. However, given the limitations of bivariate analysis, robustness of the findings is verified using multivariate probit models in subsequent sections.

Findings in Table 2 are based on a comparison of average values of economic variables during the acceleration episodes with those during times when there was no acceleration, as well as relative to the period prior to an acceleration, augmented by formal tests of statistical significance. In interpreting the results, one should bear in mind that the analysis is limited to correlations, not causal determinants; it is difficult to distinguish between the causes and the consequences of accelerations.

Growth accelerations do not come at the expense of macroeconomic stability; inflation and budget deficits are either insignificantly different or better in acceleration episodes than in control groups. Inflation is slightly lower during the episodes of accelerated growth, but not significantly so, and the episodes of the 1980s also feature better central government budget balance, including grants. Furthermore, the results for trade variables (discussed further below) show a real exchange rate depreciation in acceleration episodes, which also suggests that inflation expectations are well contained. The most striking finding here is that policies improve for accelerating countries and are better than for countries that did not experience an acceleration of growth. The World Bank’s Country Policy and Institutional Assessment (CPIA), a broad measure of policy stance, shows a positive association with acceleration episodes in both decades.

There is a strong association between acceleration episodes and trade. Episodes are correlated with strong growth in the economies of a country’s trade partners, export growth, and a more competitive real exchange rate. Exports were also facilitated by real effective exchange rate (REER) depreciations, a result that is nearly as strong when countries in the CFA franc zone are excluded, pointing to the importance of careful management of competitiveness regardless of the exchange rate regime.

Measures of political and economic liberalization have a robust correlation with accelerations; some plausibly function as measures of reforms that trigger growth, such as trade liberalization and leadership transitions. The Sachs-Warner economic liberalization index displays a small but significant association with accelerations in both decades.18 Broader indices of democracy are likely to capture the enabling environment. The composite measure of the autocracy-democracy mix (polity) captures an association between alignment toward democratic institutions and accelerations.19 Consistent with recent research, the 1990s evidence also indicates an expansionary role for a transition to new leadership after the departure of a longtime incumbent.

Table 2.Differences Between Sample Averages for Acceleration Episodes: Own Past and Nonepisodes
1980s1990s
Accelerations vs. nonaccelerations: duringAccelerations: during vs. beforeAccelerations vs. nonaccelerations: duringAccelerations: during vs. before
Macroeconomic
Inflation–2.7–5.6*–1.9–2.3
Central government balance to GDP2.4*1.4*–0.90.5
REER, change1–6.0*–9.9*–1.8–2.0
REER, percent change, non-CFA–8.5*–14.3*–1.0–1.3
CPIA20.3*0.3*0.2*0.03
Trade
Partner growth0.31.1*0.3*0.3*
Sachs-Warner (updated)0.03*0.04*0.02
Real export growth10.2*14.4*5.8*6.5*
Debt
Debt service0.79.1*–2.4*–4.3*
Debt/GDP–39.327.6*5.68.8
NVP of debt growth30.8–9.44.0*–3.8*
NVP of debt/exports0.31.5*0.30.1
Institutions
Polity index1.1*–2.1*0.23.9*
Longtime leader change0.20.61.1*1.1
Capital and productivity
Investment to GDP1.8*–1.46.1*6.0*
TFP growth40.03*0.03*2.3*3.3*
Source: IMF staff calculations.

Note: Asterisk (*) indicates that the difference in means was significant in at least a one-tailed test at 10 percent.

REER: Real Effective Exchange Rate.

CPIA: Country Policy and Institutional Assessment.

NPV: Net Present Value.

TFP: Total Factor Productivity.

Source: IMF staff calculations.

Note: Asterisk (*) indicates that the difference in means was significant in at least a one-tailed test at 10 percent.

REER: Real Effective Exchange Rate.

CPIA: Country Policy and Institutional Assessment.

NPV: Net Present Value.

TFP: Total Factor Productivity.

Accelerations coincide with increases in investment and productivity improvements; both higher investment and TFP growth seem to be required for an acceleration to occur. The results support, in particular, an investment-productivity nexus operating for the more recent accelerations. The most important finding here is the role of TFP growth, which is statistically significant for both decades and of considerable economic magnitude for the 1990s.

The growth of the net present value (NPV) of debt falls significantly for 1990s accelerations, pointing to the important role of debt concessionality in supporting surges in growth in the region. Whereas accelerating countries in the 1980s had increased debt-service ratios, the 1990s episodes saw reduced debt-service ratios, as well as reduced growth in the NPV of debt levels. Although countries that experienced growth accelerations also experienced a general rise in the NPV of debt-to-export ratios in the 1980s, they avoided that problem in the 1990s. Concessionality is important for these results, as the face value of debt-to-GDP ratios increases for accelerating countries. It is plausible that relaxed claims on current fiscal revenues through debt relief and greater debt concessionality have facilitated the investment increases associated with growth accelerations.

When the focus is further narrowed to accelerations sustained over 10 years, the key correlates are robust trade and investment, lower debt burdens, and more democratic institutions. Half of the accelerations analyzed above can be considered sustained over the medium term, because per capita annual growth rates over five years following an acceleration episode were also above 2 percent (Table 1). Analysis of the 5- to 10-year growth rates reveals some disappointments, such as Kenya and Zimbabwe in the 1990s and Côte d’Ivoire more recently, but also accelerations that were sustained over the medium term in Uganda, Burkina Faso, and Ghana, among others. The methodology looks for statistically significant differences in averages for these sustained episodes compared with unsustained accelerations (Appendix Table A9). The key finding is a strengthened emphasis on favorable trade and debt alignment along with political institutions and investment as correlates of sustained growth. The analysis also shows that sustained accelerations are associated with increases in aid. In addition, aid combined with a good policy and institutional environment is shown to be a strongly significant correlate of the sustained accelerations.

The strong association between accelerations and trade is consistent with literature suggesting that a lack of openness to trade has substantially reduced Africa’s growth. Cross-country regressions indicate that Africa’s greater closure to international trade than the average developing country has cost the region 0.4–0.7 percentage point a year in growth. Indeed, being less open is more costly to Africa than to other developing countries. These findings are not surprising given the large body of empirical literature that shows that open economies grow faster than closed ones. While these econometric findings should be treated with caution as the debate on the interpretation of such results continues to evolve, research based on other methodologies also supports the view that trade openness promotes growth in Africa. In general, African countries with lower tariffs tend to have higher TFP growth (Figure 6).

B. What Is Different About Acceleration Episodes?

The empirical model estimates the probability of being in an episode, or the probability of an episode beginning in a given year, in a multivariate context, implemented using a probit regression. Since the event of beginning an episode or being in one corresponds to a discrete event with two outcomes, the dependent variable is equal to one for a year in an acceleration episode or a year when an episode begins and zero otherwise. Because the probit method estimates a linear equation within a cumulative normal density function, the regression coefficients are most easily understood when transformed to show the marginal change in the probability in response to an infinitesimal change in the explanatory variable, evaluated at the mean. This convention is followed for most of the tables in this chapter.20

Figure 6.Tariffs and Total Factor Productivity (TFP) Growth in Sub-Saharan Africa, 1997-2003

(Percent)

Source: IMF staff estimates.

Variables consistently and positively associated with a country that is in an acceleration episode are real exchange rate depreciation, investment, total factor productivity, debt burden, and the overall quality of institutions and policy as measured by the International Country Risk Guide (ICRG) index. In Table 3, results are presented separately for accelerations from the 1980s and 1990s, along with the combined sample for both decades and the corresponding set that would be obtained by the Hausmann, Pritchett, and Rodrik cutoff (a jump of 2 percent growth to a rate of at least 3.5 percent).21 The most robust correlates are those that emerge as significant in support of a strong association between trade-related factors and acceleration episodes—combined regression or in both decades separately—providing the five variables listed above.22

Although a trade interpretation of acceleration episodes is supported by all samples, the 1980s sample features particularly strong evidence that trade channels during an episode are active. It is notable that growth of terms of trade, exports, and trade partners, which might tend to diminish each other’s explanatory power because of collinearity, are all significantly positive in the regression for this decade. Using a general-to-specific estimation strategy, as warranted by the presence of multiple indicators of the same channel, the most robust of the three measures of trade buoyancy is terms of trade growth, which is significant in the combined regression and has borderline significance in the 1990s regression when the other measures of the role of trade are excluded. Finally, regardless of sample and specification, REER depreciation is always significantly correlated with acceleration episodes.

The regressions show a very strong association between institutional quality and policy stance and acceleration episodes. The significant associations between episodes and the ICRG index across all samples and specifications are a multivariate counterpart to the REO finding of a linkage between episodes and good rankings on the World Bank’s CPIA. As a residual variable picking up unmeasured improvements in the productive climate, TFP growth may also be capturing the institutional improvements that accompany accelerations, explaining the positive link found for this variable. One divergence between the results reported here and in Chapter 2, Section A, is that the ICRG is the empirically dominant measure of country institutions and policy in the multivariate context, whereas the CPIA had performed better in the bivariate analysis.

Checks for robustness using different estimation techniques confirm the importance of real depreciation, external conditions, TFP growth, and debt as correlates of accelerations. The final two columns in Table 3 show the correlates of episodes determined by the original Hausmann, Pritchett, and Rodrik cutoffs. Debt burden, real depreciation, and institutional quality remain significant and positive, and the economic liberalization index and budget balance are added to that list. Conversely, the coefficient on aid, which was negative for the larger set of accelerations, is now insignificantly different from zero. Appendix Table A10 reports the results from the use of alternative estimation methods (for efficiency, results are shown only for the full-period sample). A random effects probit is estimated, whereby the error term is allowed to have a country effect drawn from a distribution. The most noteworthy change from the earlier regressions is the loss in significance of the institutional quality variable (ICRG) and the emergence of significant coefficients on the dummies for coastal or resource-rich countries. The likely explanation is that the random effects absorb much of the explanatory cross-sectional country variation in the ICRG, leaving a role for the subregion dummies. This is consistent with the instrumental variable (IV) estimates,23 whereby the ICRG variable is instrumented by historical settler mortality (which has only cross-sectional variation) and returns to its typical strong level of significance.24 The IV regression also attaches significance to the economic liberalization measure while removing significance from aid.

Table 3.Probit Marginal Estimates for Probability of a Country Being in an Acceleration in a Year, 1980–2004
1980s1990sAll AccelerationsAll with 3.5 Percent Growth
Marginal coefficientp-valueMarginal coefficientp-valueMarginal coefficientp-valueMarginal coefficientp-value
Macroeconomic
Inflation0.0020.200.0010.670.0010.720.0010.34
Deficit0.0060.120.0070.230.0070.140.010.01
Debt and aid
Aid–0.020.010.0010.88–0.0050.04–0.0020.17
Debt service–0.0010.650.0010.650.0010.470.0010.73
Debt net present value
burden0.0300.010.0130.120.0200.010.020.01
Trade
Terms of trade0.0020.100.0030.120.0020.10
Real exchange rate–0.0020.01–0.0030.05–0.0030.01–0.0010.1
Partner growth0.0330.050.0040.69
Export growth0.0030.09–0.0010.33
Institutions
Country risk0.0400.170.0900.030.0600.050.0630.01
Sachs-Warner0.2500.310.1300.400.0800.310.0150.02
Geographic
Coastal–0.040.470.0800.260.0200.77–0.040.28
Resource rich–0.180.010.1100.20–0.020.77–0.030.41
Capital and productivity
Investment0.0110.01–0.0050.260.0050.090.0010.8
Total factor productivity0.470.251.820.011.490.010.660.03
Summary statistics and goodness of fit:
p-value for χ-squared test0.010.010.010.01
Pseudo R20.290.110.10.22
Percent of acceleration years predicted79545859
Percent of predicted acceleration years incorrect52616268
Source: IMF staff calculations.Notes: The indicated coefficient refers to the probability of an infinitesimal change in the independent variable x, evaluated at the mean value of x. The p-value is the analog of the usual regression test for the probit coefficient being zero. The cutoff probability for the goodness-of-fit calculations is 0.25. Country risk refers to the index from the International Country Risk Guide (ICRG). Results with a smaller set of accelerations are shown in the far right column. These require 3.5 percent growth, as in Hausmann, Pritchett, and Rodrik (2004).
Source: IMF staff calculations.Notes: The indicated coefficient refers to the probability of an infinitesimal change in the independent variable x, evaluated at the mean value of x. The p-value is the analog of the usual regression test for the probit coefficient being zero. The cutoff probability for the goodness-of-fit calculations is 0.25. Country risk refers to the index from the International Country Risk Guide (ICRG). Results with a smaller set of accelerations are shown in the far right column. These require 3.5 percent growth, as in Hausmann, Pritchett, and Rodrik (2004).

Despite the reasonable significance levels associated with certain variables, the empirical model should also be assessed in terms of its ability to predict acceleration episodes in the sample. The standard goodness-of-fit tests for probit models can be augmented by examination of how well the estimated model predicts positive outcomes. Because the dependent variable is binary and the estimated probabilities range continuously between zero and one, a threshold must be chosen to convert an estimated probability into a prediction. A conventional threshold is 50 percent (that is, estimated probabilities of 0.5 or greater are taken to predict that the dependent variable equals one). However, when the overall frequency of the event is less than 50 percent, researchers have suggested that this cutoff is not appropriate and recommend adjusting it toward the observed frequency in the sample or optimizing a loss function. The model can then be evaluated in terms of two types of prediction: the proportion of actual positive outcomes that were correctly predicted and the proportion of predicted positive outcomes that were incorrect (“false positives”).

When an estimated probability of 0.25 or greater is taken as predicting an acceleration, the model correctly predicts about half of episode years, but also incorrectly identifies a sizable number of nonepisode years as accelerations (Table 3). The model classifies at least half of episode years correctly. But the model tends to make a sizable proportion of incorrect predictions that an episode is occurring; for instance, in the regression for all accelerations (in the third column), 62 percent of all predicted positive outcomes are incorrect. The performance of the random effects and instrumental variables methods is similar to that of the benchmark estimates (Appendix Table A10). The instrumental variable model does somewhat better in matching observed positive outcomes, but both models fare poorly in terms of false positives.

C. What Triggers an Acceleration?

An empirical model that can predict the timing of accelerations is more challenging, given that determining when an acceleration began is imprecise, and the more difficult objective of linking a discontinuous event like an acceleration to big changes in other variables occurring around the same time. To deal with uncertainty about timing, we follow Hausmann, Pritchett, and Rodrik in labeling each episode as beginning in the originally determined year from Table 1 plus the two adjacent years. In addition, the dating of policy transitions includes three leading-in years (in the case of economic liberalization) or three lagging years (in the case of political transitions) to increase the chance that a transitional event overlaps with the beginning of an episode. Finally, to isolate large triggering events, terms of trade growth is transformed into a dummy variable equal to one whenever the three-year change is in the upper 75 percent of the sample.

The analysis finds that economic liberalization, as measured by the Sachs-Warner index, plays a significant role in explaining the timing of accelerations (Appendix Table A11). The variable is a significant factor in the timing of 1990s accelerations and in the combined sample. This is the major difference between our results and those of Hausmann, Pritchett, and Rodrik, who did not find a significant role for liberalization in their timing regressions.

Political transitions are also a significant determinant of acceleration timing. Using the polity indicator from the Polity IV database (which ranges from–10 for autocracy to +10 for democracy) identifies a regime change by a change of 3 or greater in the indicator. Regime changes can be in the direction of increased autocracy (negative change of 3 or greater) or democracy (positive change of 3 or greater). A second measure of political transitions is provided by the leader tenure variable from Olken and Jones (2004), but it does not emerge as significant in the probit framework.25

Similar to Hausmann, Pritchett, and Rodrik’s findings with a global sample, when regime changes are separated into movements toward democracy and those toward autocracy, the latter are a more robust determinant of acceleration timing than the former. Researchers have hesitated to take this implication too literally, suggesting that it may be a proxy for improved state capacity or restoration of control following a period of disorder or conversely, weakened state capacity following a shift to more participatory government. Among the acceleration episodes matched to the negative change in the polity score are Angola in 1993, Chad in 1984, Ethiopia in 1992, and Zambia in 1999. Nonetheless, when a more direct measure of recent civil conflict was included in the same regression, it was not significant.26

Theory suggests multiple channels linking political transitions to growth and generally implies that broad-based governments are most likely to supply the public good prerequisites for sustained growth. The theory is reinforced by consideration of the region’s relatively high number of resource-rich economies. Collier and O’Connell (2005), whose discussion is highly relevant, describe a complicated interaction of resource rents, political structure, and ethnic diversity. When an autocracy in a resource-rich economy is identified with a single ethnic group in an ethnically diverse country, the prioritization of transfers over growth becomes a dominant influence on economic outcomes. But democracy can also have counterproductive effects on growth if rents are dissipated by individuals seeking votes from groups when checks and balances are weak. Collier and O’Connell confirm an association between autocracy and weak, uneven growth in resource-rich countries in the region. The seemingly contradictory finding here of a link between moves toward autocracy and the timing of accelerations is not robust in the refinement of the criteria for an episode or in the consideration of sustained accelerations, as discussed below.

Tightening the criteria for an acceleration strengthens the link of timing to economic variables and weakens the link to the autocracy indicator. When the probit timing model is estimated for the sample that identifies an acceleration using the tighter Hausmann, Pritchett, and Rodrik cutoffs, coefficients are significant for the terms of trade variable and for the positive political regime change variable but not for its negative counterpart. Because the Hausmann, Pritchett, and Rodrik cutoff is a tighter filter of high-growth experiences, the results suggest that, whatever the positive growth effects of autocratic transitions, they arise for modest accelerations only. In results not reported, we find no econometric gain to interacting the polity changes with ethnic fractionalization. In a probit model with limited degrees of freedom, it is difficult, however, to fully explore the Collier and O’Connell (2005) channels relating autocratic political structures to rent transfers in ethnically diverse and resource-rich economies.

The timing model’s in-sample predictive power is poorer than that of the model for being in an acceleration; the rate of incorrect predictions of the start of an acceleration is particularly high. The overall frequency of acceleration initiations is low, so a 15 percent cutoff probability was used. At best, the model was able to match about one-third of positive outcomes, but this outcome is associated with a rate of incorrect predictions of positive outcomes of up to 80 percent (Appendix Table A11). This is consistent with Hausmann, Pritchett, and Rodrik’s message: “A lot of takeoffs take place when … conditions appear not to be particularly favorable. And growth takeoffs typically fail to materialize when the conditions are indeed favorable.” Indeed, this conclusion is reinforced when the episodes are based on Hausmann, Pritchett, and Rodrik’s stricter cutoff. Now, just one-fifth of episode starts are correctly called, while most predicted starts are incorrect.

D. What Sustains an Acceleration?

Separating accelerations that were sustained from those that were not reinforces the emphasis on the role of trade, while also pointing to explanatory roles for location and conflict recovery. First, a model that explains the probability of a country’s being in a sustained acceleration episode is estimated (Appendix Table A12) using a specification similar to that used earlier for being in an acceleration episode. The model includes only acceleration episodes and compares values of the variables (during the five-year initial acceleration episode) for sustained versus unsustained episodes. Noteworthy here is the continued role of real depreciations but also significant positive coefficients on dummies for coastal and recent civil conflict countries. This result is understandable given that Mauritius, Mozambique, Seychelles, and Uganda are included in the set of countries that experienced sustained accelerations. Goodness of fit for this model, assessed at a 50 percent probability threshold, is a little better than for the corresponding model from Table 3. However, the model here is for sustained accelerations within the set of all accelerations, so the underlying sample is quite different.

Sustained acceleration occurrence is now negatively correlated with debt burdens, pointing to debt as a constraining factor on growth. Two measures of debt—the ratios of net present value and of debt service to exports—are negatively associated with sustained episodes. This finding has two complementary implications: that debt relief has the potential to spur growth, and that growing economies need to avoid accumulating excessive debt. The finding of a negative association between debt and accelerations survives the narrowing of the set of accelerations to those determined by the Hausmann, Pritchett, and Rodrik cutoff, with aid and investment (both positive) being the only other significant variables. This result for debt is linked to a finding from the bivariate correlations in the REO that 1990s accelerations and sustained accelerations are negatively correlated with debt burdens.

Comparing explanations for the timing of sustained accelerations with those of unsustained accelerations directs attention toward economic variables and away from political transitions (Appendix Table A13). From our benchmark sample of accelerations, the economic liberalization variable is significant, as in the earlier timing regressions (Appendix Table A11), but so also are democratic transitions and the recent conflict dummy. This finding marks an alignment of our results with Hausmann, Pritchett, and Rodrik, who find that economic liberalization and moves toward democracy lead to sustained accelerations, but that big terms of trade shocks do not. When sustained accelerations are defined using the tighter Hausmann, Pritchett, and Rodrik cutoffs, of those variables in the original set, only the democratic transition indicator is significant.

E. Summary

Consistent themes from this chapter’s more detailed analysis and from the REO are the key roles of trade, investment, productivity, and policy and institutional soundness in supporting growth accelerations. Not unexpectedly, the significant variables within these broad categories change somewhat when the analysis shifts from the bivariate to the multivariate context. For example, the probit regressions tend to find the ICRG index rather than the CPIA index to be the stronger correlate of accelerations among the institutional quality variables. In addition, although real depreciation is a robust correlate of accelerations across all specifications, the effect of trade buoyancy is captured variously by trade partner growth, export growth, and the terms of trade.

Box 3.Risks to Sustainability of Growth Accelerations: HIV/AIDS and Poverty

The HIV/AIDS epidemic is jeopardizing the sustainability of growth in several sub-Saharan African countries. Although some countries have taken bold steps to slow the epidemic, and recent large increases in donor funds for prevention and treatment are encouraging, the HIV/AIDS epidemic is taking a serious toll on societies and economies in the region. Studies identify several channels through which the disease affects economic growth. In addition to reducing the labor supply, which translates into lower output, increased mortality and morbidity lower private and public sector productivity and lower the efficiency of labor by eroding human capital; at the same time, increased health expenditures tend to crowd out savings and reduce investment. For the worst-affected countries (those with HIV prevalence rates in over 20 percent in the working-age population), studies have projected that that the epidemic could reduce growth by 1 to 1.5 percentage points.1 These estimates omit an important concern of the business communities; namely, that an uncertain and deteriorating outlook could deter domestic and foreign investment. In addition, in the longer term, HIV/AIDS could discourage individuals and companies from investing in human capital, given significantly lower expected returns. It is these risks to the outlook for investment and productivity (important for growth accelerations) that raise concerns about the sustainability of growth in some countries.

Growth will not be sustainable unless it is shared by broad segments of the population. However, poverty outcomes in countries experiencing sustained accelerations have been varied. Given the infrequency of household surveys and the lack of data on the share of the population living below national poverty lines in the 1980s, it is difficult to trace the evolution of poverty rates in many sub-Saharan African countries. For the seven countries that experienced sustained accelerations, and for which some poverty data are available, poverty rates declined significantly during the 1990s in Ghana, Uganda, and (in the early 1990s) Seychelles.2 Burkina Faso and Benin report increases in poverty rates of less than 1 percent. In contrast, poverty rates increased significantly during the 1990s in Cape Verde and Lesotho.

1 See Haacker (2004), which draws on Joint United Nations Programme on HIV/AIDS studies. Note that data limitations prevent the formal consideration of the role of HIV/AIDS in the growth acceleration analysis.2 The percentage of households living below the poverty line in Seychelles fell to 19 percent from 30 percent between 1984 and 1992 (World Bank, 1994). This is a slightly different measure than that considered for other countries; that is, percentage of the population living below the poverty line. Inferences are based on poverty data from the World Bank’s World Development Indicators and on country Poverty Reduction Strategy Paper (PRSP) documents.

There are two important cases where findings from the REO reemerge in a different form in the multivariate analysis. First, although the REO found a strong positive link between the polity indicator and acceleration episodes in bivariate analysis, the probit regressions for being in an episode did not reproduce this finding. Recall, however, that the regressions did find a strong role for the ICRG index, which includes indicators of political freedom that overlap with those measured by the polity indicator.

Second, the REO found that debt rose in the 1980s growth acceleration episodes and fell in the 1990s episodes. In the multivariate analysis, debt tends to rise in episodes from both decades, but falling debt is found to be a characteristic of sustained accelerations. This is an important category because, although a relatively large number of countries in the region have been able to generate a single growth acceleration, only sustained growth will deliver significant progress toward the MDGs.

As others have found, the most difficult challenge is to discover what can predict the onset of an acceleration. In other words, the search for factors that change around the same time as the initiation of an episode is a much sharper filter than the search for factors that differ between episodes and nonepisodes. The data support a role for economic liberalization and political changes in triggering accelerations. However, the timing of accelerations is very poorly predicted, indicating that the models fall short of providing guidance on either necessary or sufficient conditions for an acceleration to occur.

The findings regarding acceleration triggers vary with the type of episode considered. For the benchmark set as used in the REO, initiation of accelerations is linked to economic liberalization and big changes in the polity indicator, with a move toward autocracy having a strong role in the latter effect. But when the criteria are refined—limited either to sustained episodes or to those with a higher growth rate—only democratic transitions are significant.

The most encouraging finding for policymakers is the link between the quality of policies across a range of dimensions and the propensity for acceleration. The empirical analysis uses indicators of macroeconomic stability, institutional quality, trade openness, and productivity and demonstrates their link to the occurrence of an episode. On the other hand, neither resource availability nor geography was consistently associated with episodes, suggesting that an acceleration is feasible for most countries in the region. Finally, the finding that sustained accelerations are associated with lower debt and higher investment along with trade indicates that Africa’s development partners need to continue to help promote growth in the region. (See Box 3 on particular risks to sustainability of growth accelerations.)

17

Most sub-Saharan African countries are not on track to meet the income poverty MDG. Already the highest in the world, sub-Saharan African poverty rates increased during the 1990s. Of the 28 countries for which household surveys are available, only 5 seem well positioned to meet the poverty goal (Cameroon, Ethiopia, Senegal, South Africa, and Swaziland), with two others relatively close (Mauritania and Mozambique) (World Bank and IMF, 2005).

18

The economic liberalization variable is the update of the original Sachs-Warner trade liberalization dating by Wacziarg and Welch (2003).

19

This measure is taken from the Polity IV database at http://www.cidcm.umd.edu/inscr/polity. It ranges between–10 and 10, with the bounds corresponding to autocracy and democracy.

20

The reported coefficients are the output of the dprobit command in Stata 8. For two methods, these transformed estimates are not available: random effects and instrumental variables.

21

In comparing the decade subsample results, one must consider both the possibility of structural change between the two decades and small-sample bias in deciding whether a coefficient difference between the two decades is meaningful.

22

The ICRG index is a proprietary index available at http://www.icrgonline.com. It is an average of component indicators that assess the quality of political and economic institutions and policies.

23

The IV method is implemented using the probitiv routine for Stata 8 made available by the World Bank’s Poverty Research Group.

24

Acemoglu, Johnson, and Robinson (2001) have argued that the quality of contemporary institutions in low-income countries was largely determined by the objectives of the colonizing powers, which, in turn, were related to the viability of settler communities. When measured by settler mortality, this provides the basis for the use of this variable in estimation. However, Sachs (2003) has argued that the empirical power of the instrument derives from its correlation with contemporary health conditions.

25

Their variable is equal to zero except in years of the nonviolent death of a country’s leader, in which case it equals the leader’s number of years in office. This variable is designed to capture exogenous political transitions, with length of term in office acting as a proxy for the scope for policy changes under the new leader. Because the variable is quite sparse, its lack of significance in a probit regression does not mean that it is not economically important. Probit regressions require variation across outcomes, which can be lacking in variables measuring rare events.

26

This measure captures recent civil conflict and is taken from the World Bank’s Peacebuilding data project available at http://www.worldbank.org/research/conflict/papers/peacebuilding.

    Other Resources Citing This Publication