Exogenous Shocks and Growth Crises in Low-Income Countries
A Vulnerability Index*
Author: Ms. Era Dabla-Norris and Yasemin Bal Gunduz

Contributor Notes

Authors’s E-Mail Address: edablanorris@imf.org; ybalgunduz@imf.org
Publication Date:
02 Nov 2012
eISBN:
9781475528510
Language:
English
Keywords:
WP; exchange market pressure index; correlation matrix; policy buffer; growth rate; growth crises risk
Download PDF (3.9 MB)

This paper develops a new index which provides early warning signals of a growth crisis in the event of large external shocks in low-income countries. Multivariate regression analysis and a univariate signaling approach are used to map information from a parsimonious set of underlying policy, structural, and institutional indicators into a composite vulnerability index. The results show that vulnerabilities to a growth crisis in low-income countries declined significantly from their peaks in the early 1990s, but have risen in recent years as fiscal policy buffers were expended in the wake of the global financial crisis.

Abstract

This paper develops a new index which provides early warning signals of a growth crisis in the event of large external shocks in low-income countries. Multivariate regression analysis and a univariate signaling approach are used to map information from a parsimonious set of underlying policy, structural, and institutional indicators into a composite vulnerability index. The results show that vulnerabilities to a growth crisis in low-income countries declined significantly from their peaks in the early 1990s, but have risen in recent years as fiscal policy buffers were expended in the wake of the global financial crisis.

I. Introduction

Low-income countries are subject to a wide variety of exogenous shocks—sharp swings in the terms-of trade, export demand, natural disasters, and volatile financial flows. The amplitude and frequency of such shocks tend to be higher than in advanced and emerging market countries (IMF, 2011). In principle, negative shocks in a standard neoclassical growth model would imply a quick reversion to the steady state level of income, implying a growth “bounce-back” and benign transitory effects. However, the resources, instruments and policy buffers needed to absorb or mitigate shocks are often unavailable in low-income countries or difficult to implement in weak institutional and policy environments. Consequently, large adverse external shocks tend to induce breaks in trend growth rather than cycle fluctuations around a trend, imposing steep output and welfare losses, both directly and often through ensuing prolonged growth slowdowns.1

This paper develops a vulnerability index that can help quantify low-income country risks to growth crises arising from external shocks. While predicting the timing of such events is likely to be an elusive endeavor, flagging the underlying vulnerabilities that predispose countries to growth declines in the event of external shocks can provide a first indication of a possible problem and signal the potential for pre-emptive policy action.

Growth crises episodes in this paper capture the combined effects of growth declines (negative growth) and level drops in the event of shocks, which can endanger the sustainability of a country’s growth path. It is well documented that macroeconomic vulnerabilities—large fiscal and external imbalances, unsustainable debt ratios, and inadequate reserve buffers, weakly diversified economic structures, narrow and concentrated tax bases, and institutional weaknesses serve to reduce resilience to exogenous shocks in low-income countries.2 In line with this literature, a range of economic, structural, and institutional indicators that capture the flow and stock vulnerabilities in the external and fiscal sectors and the real economy are used to gauge whether a country hit by a large external shock is also likely to experience a growth crisis.

Two complementary approaches are used to map information from the underlying indicators into a composite vulnerability index: multivariate regression analysis, and a univariate ‘signaling” approach. The multivariate regression approach uses a correlated panel probit model to estimate the probability of a growth crisis. This approach allows us to take account of correlations among different variables, to test for the statistical significance of individual variables, and to assess the constancy of coefficients across country groups.

The univariate approach entails using each indicator of crisis events separately, identifying critical thresholds that signal such events with the lowest prediction error, and then averaging the indicators into a summary index. In particular, the composite vulnerability index measures the number of indicators exceeding these thresholds, weighted by their relative signaling power. This approach can accommodate differences in data availability across countries and allows for the inclusion of a potentially larger number of vulnerability indicators than the multivariate regression method. At the same time, the results from the probit analysis provide guidance on the conditional statistical significance of the variables used in the univariate approach.

The index can be used to assess vulnerabilities to growth declines in low-income countries over time. The analysis shows that the overall vulnerability index has declined significantly from its peak in the early 1990s. Better policy and economic management, coupled with a favorable external environment, especially terms of trade improvements, and official debt relief contributed to stronger macroeconomic positions and lower vulnerabilities till the onset of the global crisis. More recently, growth crises risks in low-income countries remain elevated and well above pre-crisis years as fiscal buffers have increasingly been used up.

The rest of the paper is organized as follows. Section II provides a brief review of the literature. Section III describes the methodology for identifying crisis episodes. Section IV presents the empirical analyses and main results. Section V describes the univariate approach, and Section VI concludes.

II. Literature Review

This paper is related to several strands of research. A large body of evidence finds that adverse external shocks have a significant negative impact on short- and medium-run growth through their effect on aggregate demand, external balances, and the government’s fiscal position (Collier and Goderis, 2009; Berg et al., 2010).3 Moreover, these effects are asymmetric: while negative shocks impede growth, positive shocks do not necessarily contribute to long-run growth, particularly in resource-rich countries with weak institutions (Collier and Goderis, 2007). Our paper builds on these studies by analyzing a broad range of macroeconomic, institutional, and structural correlates of growth declines in the event of shocks, although we focus on the negative tail of the distribution.

This paper also complements previous studies that examine determinants of growth down-breaks, broadly defined as extended periods of markedly slow growth (see for e.g., Rodrik, 1999, Pritchett, 2000, Hausmann et al., 2006, and Berg et al., 2011). Our study focuses on sharp declines in growth over a shorter timeframe—events that, in our view, are related to, but distinct from, permanent changes in output and possibly of greater concern to policy makers.4 Further, our approach allows for differentiating local recoveries from severe downturns within a longer growth downbreak.

Finally, our paper is related to a large literature on early warning system models focused on a range of crises facing emerging market and advanced economies (see Kaminsky et al., 1998, for currency crises; IMF, 2007 for sudden stops; Baldacci et al., 2011, for fiscal crises; IMF 2010 for financial crises). In low-income countries, volatility in the real economy stemming from large terms of trade declines, contractions in external demand, and natural disasters can manifest in different sorts of severe economic distress—declines in consumption, balance of payments crises, and fiscal stress. An overarching focus on growth is thus an efficient way to capture most of these manifestations.

This paper is also related to Easterly et al. (2000) who estimate a panel probit regression relating growth downturns (negative growth rates) to a large set of structural variables in developing and advanced economies. Our analysis focuses on growth decline episodes conditional on external shocks and captures relevant macroeconomic and structural vulnerabilities that predispose low-income countries to growth crises. In this respect, our paper is most closely related to Dabla-Norris et al. (2011) who empirically examine the benefits of international reserves in smoothing domestic absorption and consumption in response to exogenous shocks in low-income countries. Our paper draws on their methodology of shock and crisis identification to examine the determinants of growth declines.

III. Methodology

A. Identification of Large Exogenous Shocks

Large negative external shock events in countries are identified if the annual percentage change of the relevant shock variable falls below the 10th percentile in the left-tail of the country-specific distribution. In particular, shock episodes include one or more of the following six shocks: (i) external demand; (ii) terms-of-trade; (iii) FDI; (iv) aid; (v) remittances; (vi) climatic shocks (large natural disasters).5 Defining large negative shocks over country-specific distributions implies that each country experiences the same frequency of shocks, so that the focus is on the reaction to the shock. Moreover, in the context of low-income countries, these shocks can be assumed to be exogenous to country-specific fundamentals or policy. The sample used for the analysis spans the period 1990–2009 for 71 low-income countries. For each shock, only the first year of the shock event is considered in the final set, giving us a total of 698 shock observations (out of 1420 observations).

Although country-specific thresholds for each shock are used for identification purposes, the overall distributions of shock versus non-shock episodes are markedly different. Figure 1 shows the annual changes in external demand (partner country growth), terms-of-trade, aid and FDI for the sub-samples of shock versus non-shock episodes. As can be seen from the figures, the lowest quartile of the non-shock sample is markedly higher than the highest quartile of the shock episodes for each individual shock. This suggests that our shock definition captures reasonably severe events. Moreover, the very low and statistically insignificant correlations among shocks (except for the change in FDI to GDP and the terms of trade growth) suggest that these shocks are largely independent (See Appendix Table 1 for the correlation matrix).

Figure 1.
Figure 1.

Identification of External Shock Episodes

Citation: IMF Working Papers 2012, 264; 10.5089/9781475528510.001.A001

Source: Authors’ calculations.

B. Identification of the Dependent Variable: Growth Crisis Events

Within the sample of identified shock events, a growth crisis is defined as a large drop in real GDP per capita. Specifically, we assume that a crisis occurs when the following two conditions hold: (i) the post-shock two-year average (t and t+1) level of real GDP per capita falls below the pre-shock three-year trend; and (ii) growth of real GDP per capita is negative at time t. All other episodes for which either of these conditions fails are considered as normal episodes. Severe state failure events are excluded from the sample as growth in these episodes is likely to be impacted independently of the impact of covariates.6

Table 1 summarizes the median growth rate of real GDP per capita for the identified shock sample, distinguishing between crisis and non-crisis events. The median growth rate is positive for the entire shock sample, implying that not all shocks incur a drop in real growth. Indeed, the unconditional probability of a crisis within the large shock sample is only about 24 percent. However, there is a substantial difference in real GDP per capita growth of more than 6¾ percentage points between crisis and non-crisis cases, which is also statistically significant.

Table 1.

Median Real GDP per capita Growth

(1990–2009)

article image
Source: Authors’ calculations.

In principle, a crisis event can be defined in various ways. We use the above definition to highlight the extreme nature of the event, which is the main focus of our analysis. Figure 2 presents the distribution of real GDP per capita for crisis versus normal episodes. In a majority of cases, the drop in growth (negative value) is also associated with a persistent decline in the level of output. In other words, as discussed above, in only a few episodes is negative real GDP growth followed by a quick recovery to the pre-shock level of output, akin to normal business cycle movements in advanced and emerging market countries. This observation underscores a key vulnerability in low-income countries: once growth drops into the negative territory, output losses tend to be persistent.

Figure 2.
Figure 2.

Distribution of Real GDP per capita: Crises versus Normal Episodes

(1980–2009)

Citation: IMF Working Papers 2012, 264; 10.5089/9781475528510.001.A001

Source: Authors’ calculations.

IV. Empirical Analysis: Econometric Approach

A. Probit Model

The effect of various policy and structural variables on the likelihood of a growth crisis in the presence of large exogenous shocks is assessed by estimating a binary response model for panel data. The general specification for a panel probit model is given by

yit=1growthcrisiseventyit=0normalepisodesP(yit=1|xit,ci)=Φ(xit′β+ci)i=1,...,nandt=1,...,T`

where, y is the observed outcome, Φ is the cumulative normal density function (c.d.f.), xit is the 1xk vector of explanatory variables, and β is kx1 vector of coefficients associated with xit. Different estimators are constructed depending on their assumptions for the panel heterogeneity, (i.e., how ci is treated).7 In this paper, the estimations are carried out step-by-step under different estimators and a correlated pooled probit model is preferred based on the econometric tests for the significance of both the individual specific effect and the sample average for covariates.8 Sixty-one countries are included in the sample over the 1990–2009 period (Appendix Table 2).

Drawing on the literature, a general-to-specific approach was used to reach the preferred specification of the model, starting from a set of twenty-two potential regressors (see Appendix Table 3). The final set of explanatory variables can be grouped into three clusters:

  • Policy variables: These include the ratio of government balance to GDP, reserve coverage (in months of imports of goods and services) a dummy for flexible exchange rate regime, and the exchange market pressure index. The latter is a composite index comprising depreciation of the official exchange rate, change in the stock of international reserves (in months of imports of goods and services), and the black market premium.9

  • Structural and institutional variables: These include the World Bank’s CPIA10, real GDP growth in the previous period, and the country-specific average of real GDP per capita growth over the sample period. The latter is a proxy for cross-country differences in underlying structural and institutional conditions. For instance, long-run historical performance of income per capita can capture shock amplifiers such as the relative diversification of trade and production, internal conflict, and domestic shocks.

  • Shock size: These include growth in trading partners weighted by the ratio of lagged exports to GDP, and the change in export prices weighted by the ratio of lagged exports to GDP. These variables capture exposure to trade-related shocks as countries experiencing larger shocks are more likely to suffer severe growth declines when shocks materialize. Conversely, a very favorable external environment may shield a country with weaker policy fundamentals from growth crises.

All explanatory variables are lagged by one year, except for the variables capturing exogenous shock size, and are thus predetermined with respect to the crisis event.

B. Estimation Results: Benchmark Probit Specifications

Table 2 reports the estimation results for the baseline probit regressions for all countries (Column 1) and for three sub-groups (excluding commodity exporters, oil exporters, and small islands11; Columns 2–4). We find that the probability of a growth crisis increases sharply for countries with weaker institutions (proxied by the CPIA), lower pre-shock GDP growth, and a track record of anemic past real per capita GDP growth. Importantly, sound policy fundamentals, such as higher reserve coverage and fiscal balance, are associated with a lower likelihood of a growth crisis, reflecting the role of counter-cyclical policies in supporting domestic demand. Similarly, lower pre-shock balance of payments pressures reduce the likelihood of a growth crisis. Finally, consistent with Dabla-Norris et al. (2011), a more flexible exchange rate regime sharply reduces the probability of a growth crisis, suggesting that exchange rate flexibility can help facilitate economic adjustment to real shocks.

Table 2.

Probability of Growth Crisis

(Correlated Pooled Probit Regression, 1990–2009)

article image
Source: Authors’ calculations.Note: Estimated by a correlated pooled probit model with cluster-robust standard errors. Significant at 10 percent:*; 5 percent:**; and 1 percent:***, standard errors are in parantheses.

As expected, shock size is significantly associated with the likelihood of a growth crisis. The shock variables enter with a negative sign, indicating that positive shocks to external demand and commodity prices lower the probability of a growth crisis. The impact of the export price shock, however, is largely driven by commodity exporters as it becomes insignificant once this country-group is excluded from the sample (Table 2, Column 2). Finally, both contemporaneous shocks are significant at the 5 percent level in the sample excluding small islands (Table 2 Column 4).

More informative are the marginal effects of these explanatory variables on the probability of a growth decline.12 Table 3 presents average marginal effects (Column 1)13 and the marginal effects of a specific covariate evaluated at the sample median, and worst quartiles (i.e. 75th/25th percentile if estimated coefficient of a covariate is positive/negative) of all other variables (See Appendix Table 4 for sample distribution of covariates). The coefficient estimates provide a gauge of the relative importance of each variable on the crisis probability. As can be seen from Table 3, institutional quality and the exchange rate regime are the strongest predictors of a growth crisis. Moreover, the marginal effects of the policy variables (e.g., fiscal balance, reserve coverage) and institutional quality on the crisis probability are significantly higher for a country with weaker buffers (Column 3) as compared to the median country (Column 2). This suggests that payoffs from improvements in institutional quality and macroeconomic policy frameworks are likely to be highest in these countries.

Table 3.

Benchmark Regression: Average and Conditional Marginal Effects

article image
Source: Authors’ calculations.Note: Average and conditional marginal effects are estimated from the benchmark correlated pooled probit model for all LICs. Standard errors are in parentheses.

Marginal effects of a specific covariate on the response probability averaged across the distribution of covariates in the sample.

All covariates are set at their median for the full sample.

Covariates are set at their 75th(25th) percentile if their estimated coefficient is positive (negative).

The threshold probability is obtained by the minimization of weighted misclassification errors.

Table 4 reports the marginal impact of a change in each explanatory variables from its median value to its best quartile (75th/25th percentile if estimated coefficient is negative/positive), evaluating all other variables at their sample medians. An improvement in the CPIA from 3.4 to 3.8 reduces the probability of a growth crisis by about 3 percent for the median country. Similarly, an increase in reserve coverage from 2.8 to 4.1 months of imports lowers the crisis probability by 2½ percentage points. Table 4 also suggests that shifting from a fixed to a flexible exchange rate regime has a significant impact in the crisis probability, reducing the likelihood of a growth crisis by 9 percentage points. Finally, the results confirm the importance of path dependence as countries with a stronger past track record of real GDP growth have a significantly lower likelihood of a growth crisis.

Table 4.

Marginal Effects of Explanatory Variables on Crisis Probability

(Percentage point, unless otherwise indicated)

article image
Source: Authors’ calculations.

Predicted probabilities from the benchmark probit regressions can be used to call crisis events based on a threshold probability. Following Demirgüc-Kunt and Detragiache (1999), a loss function minimization approach is used to derive the threshold probability. The expected loss function is the weighted average of missed crises (Type I errors) and false alarms (Type II errors), with the weights reflecting costs attached to each type of error. In this paper, asymmetric weights are used that place higher weights on missing crises, resulting in a threshold probability of 0.19, with associated Type I and Type II errors of 20 and 27 percent, respectively (Table 5). A growth crisis is predicted when a country with weak policy buffers and poor institutions is hit by adverse external shocks (Table 3 Column 3). In this case, the 95 percent confidence interval of predicted probabilities also lies above the threshold probability. On the other hand, the predicted probability for the median low-income country (including the 95 percent confidence interval) lies comfortably below the crisis threshold.

Table 5.

Predicted Probabilities

(Percentiles)

article image

Same model is estimated for 1980–2004. Predicted probabilities are for 2005–2009.

Number of growth crises divided by total observations.

The relationship between each covariate and the threshold crisis probability is further illustrated in Figures 3a (other variables evaluated at sample median) and 3b (variables evaluated at their worst quartiles). As can be seen from Figure 3a, only extreme values of some covariates could push the predicted probability above the threshold for the median country. These include an exceptionally weak institutional quality, high exchange market pressure, a significantly poor track record of past GDP growth, and negative tail shocks to external demand and export prices. On the other hand, barring an exceptionally good track record of past GDP growth or an extremely favorable tail shock to external demand, changes in most explanatory variables are insufficient to lower the predicted probability below the threshold for a weakly-positioned country.14

Goodness of Fit

Measuring the “goodness” of fit of the model is not straightforward as the empirical model predicts conditional probabilities that must be compared to actual events. For instance, a predicted probability of less than one still assigns a non-zero probability to the alternative event. Nonetheless, an examination of the distribution of in-sample and out-of-sample predicted probabilities for growth crises versus normal episodes can provide a gauge of the goodness of fit of the model. As can be seen from Table 5, the distributions of predicted probabilities for crisis and non-crisis events are distinct. Moreover, the predicted probabilities are well dispersed in the [0,1] interval, indicating the ability of the empirical model to differentiate between alternative outcomes.15 The median predicted probability for a growth crisis is 0.38 versus 0.10 for normal episodes. Moreover, in contrast to non-crisis episodes, more than seventy-five percent of growth crisis episodes have probabilities above the crisis threshold.

Table 5 also presents the out-of-sample predictions for the benchmark model. These are obtained by estimating the sample for 1990–2004 and calculating predicted probabilities for 2004–2009. Using the threshold probability of 19 percent, the model correctly calls 14 out of 15 crises, with only 7 percent of crises missed (even lower than the in-sample prediction) and false alarms occurring in 34 percent of the non-crises episodes. On account of the sharply lower Type I error, the overall misclassification errors (sum of the two errors as percent of total observations) are even lower than in the in-sample predictions. While out-of-sample predictions tend to be weaker than in-sample performance, these results provide evidence of the index’s goodness-of-fit.

Robustness Check

As an important robustness check we examined whether the benchmark specification holds for all growth crises, and not just those induced by large exogenous shocks, controlling for shock covariates in the regression. As seen in Appendix Table 5 (columns (1) and (2), the results remain qualitatively unchanged. When all growth crises are included in the sample, climatic shocks and shocks to remittances turn significant. A key difference with the benchmark specification reported in Table 2 is that reserves are far more effective in reducing the likelihood of a growth crisis induced by large exogenous shocks.

V. Empirical Analysis: Signaling Approach

A. Methodology: Signaling Approach

The estimation of thresholds for growth crises for each indicator is based on a “signaling” approach (IMF, 2007; IMF, 2011). This consists of defining cut-off values for each individual indicator that discriminate between crisis and non-crisis episodes. If an indicator exceeds the cut-off level, the model issues a signal of an upcoming growth crisis episode. The optimal cutoff point balances Type I and II errors since the lower the threshold, the more signals send (i.e., Type I errors decrease), but at the expense of higher false alarms (i.e., Type II errors increase). Using a higher threshold reduces the number of incorrect signals, but at the expense of increasing missed crises.

Formally, we can define a signaling variable at time t, dt, for the following j time periods as follows:

dt={1for∀j,ifxt-1>C0,otherwise

where

xt refers to an indicator variable, which is a monotonically increasing function of crisis probabilities, and off for xt. The signaling window j is set to one year in the analysis. Two methods are commonly used to determine the optimal value of C: the minimization of the total misclassification errors and the maximization of the signal-to-noise ratio (SNR).16 Under the total misclassification errors (TME) method, for each cut-off point C, the TME value can be expressed as the sum of type I and type II errors,

TMEC(C)=Missedcrises(C)Totalcrises+Falsealarms(C)Totalnon-crises

The optimal threshold C* is the value that minimizes TME(C).

The overall vulnerability index is calculated based on the signaling power of each indicator. This entails two steps. In the first step, an index summarizing a cluster of variables is calculated. Thresholds that yield the best split are used to map indicator values into zero-one scores, with each indicator assigned a weight based on its predictive power.17 In the second step, the predictive power of the cluster indices is evaluated and the indicators are aggregated in the vulnerability index based on their own predictive power and the predictive power of the cluster indices:

OverallIndex=ΣgwgΣiwi,gdi

where wi,g is the weight of each individual indicator i in group g, wg is the weight of the group, and di is a dummy that takes the value of 1 if the indicator is above (below) the threshold, and zero otherwise. As before, the methodology puts a higher weight on not missing crises, motivated by the high costs associated with crises and the benefits from being able to take mitigating steps.

B. Results: Signaling Approach

The analysis uses 13 variables, starting with the set included in the benchmark probit regressions and adding a number of variables identified in the literature as determinants of growth down-breaks. The composition of the growth decline vulnerability index and information on performance of individual indicators are presented in Table 6.

Table 6.

Non-parametric Signaling Approach: Performance of Indicators and Model Fit

article image
Source: Authors’ calculations.

The thresholds are achieved by minimizing type I plus type II errors.

Threshold for the overall index is derived by minimizing the asymmetrically weighted loss function giving more weight to type I error.

Missed crises plus false alarms as percent of total observations.

The variables are grouped into three clusters: overall economy and institutions, external sector, and the fiscal sector. The first cluster includes the CPIA index, the Gini coefficient (measuring the extent of income inequality), real GDP growth, and the sample average of GDP per capita growth. The external sector indicators comprises reserve coverage, growth in the volume of exports of goods and services, exchange market pressure index, and the contemporaneous variables controlling for the country-specific size of exogenous shocks to external demand and export prices. Finally, the fiscal cluster includes government balance, public debt, tax revenue (all in percentage of GDP), and the cumulative growth in real government revenue over the past two years. As before, all indicators except for the contemporaneous shock size variables are lagged by one period.

For each indicator, the first two columns in Table 6 present the crisis versus non-crisis observations. The third column shows the estimated thresholds followed by the corresponding Type I and Type II errors (Columns 4 and 5). The weight of each indicator in the composite vulnerability index is determined by its relative signaling power (defined as one minus the total error; Column 6).18 The top predictor of growth crises is the overall economy and institutions index (accounting for 37 percent of the index weight). Within this cluster, lagged real GDP growth, which provides information on the state of the pre-shock economy, and the Gini coefficient are the main predictors. Aggregate weights for the external and fiscal sector indices within the overall index are broadly similar, with reserve coverage ratio and exchange market pressure index, and government balance and real revenue growth the top predictors, respectively. The overall index threshold of 0.44 is obtained by minimizing an asymmetrically-weighted loss function which penalizes missing crises more.

We estimated univariate and multivariate probit regressions to assess the relative ability of the sub-indices as well as the overall vulnerability index to provide early warning signals of growth crises.19 In the univariate regressions, both the overall index and the three sub-indices are highly significant determinants of growth crises. Figure 4 plots the predicted probabilities against the value of each index, ranging from zero (no signal is issued in any indicator within the cluster) to one (all indicators issue signals).

Figure 3.
Figure 3.
Figure 3.

Effects of Explanatory Variables on Predicted Probability of Growth Crises1/

Citation: IMF Working Papers 2012, 264; 10.5089/9781475528510.001.A001

Source: Authors’ calculations.1/ For a LIC with weak fundamentals, other covariates are fixed at their 75th(25th) percentile if their estimated coefficient is positive (negative), i.e. at their worst quartiles. For the median values LIC other covariates are fixed at their median. 95 percent confidence intervals for predicted probabilities are presented.
Figure 4.
Figure 4.

Predicted Probability of Growth Crises: Performance of the Vulnerability Index and its Sub-Components1/

Citation: IMF Working Papers 2012, 264; 10.5089/9781475528510.001.A001

Source: Authors’ calculations.1/ Predicted probabilities are estimated from univariate probit regressions when the overall index or each of its sub-component is included as the only covariate. 95 percent confidence intervals for predicted probabilities are presented.

As can be seen from Table 7 and Figure 4, the predicted probability for the overall index is well dispersed within a relatively narrow 95 percent confidence band and fares much better than any of its components. The significant impact of overall economy and institutions and external sector indices on the predicted probability provide support to assigning higher weights to these clusters. The impact of the fiscal index, however, is comparatively lower. This could be driven by the limited availability of fiscal indicators during the first half of the sample period. As a result, we assign a somewhat lower weight to the fiscal index.

Table 7.

Distribution of Predicted Probabilities: Vulnerability Index versus its Sub-Components

(Percentiles)

article image
Source: Authors’ calculations.

Multivariate probit regressions, which control for the correlation among sub-components, provide similar results. The economy and institutions and external sector indices are significant at the 1 percent level, while the fiscal index is significant at the 10 percent level. The relative impacts of the sub-components on predicted probabilities are presented in Figure 5 for the median low-income country, and for a country with weak fundamentals.

Figure 5.
Figure 5.

Marginal Effects of Sub-components on Predicted Probability of Growth Crises1/

Citation: IMF Working Papers 2012, 264; 10.5089/9781475528510.001.A001

Source: Authors’ calculations.1/ Predicted probabilities are estimated from a multivariate probit regression when all sub-components of the vulnerability index are included. 95 percent confidence intervals for predicted probabilities are presented.2/ For median LIC, two subcomponents are fixed at their median while the third sub-component is varied from zero (no signals) to one (all indicators within the cluster issue signals). For a LIC with weak fundamentals, two subcomponents are fixed at their 75th percentile while the third one is varied from zero to one.

For a country with weak fundamentals (indicator values in 75th/25th percentile), all indicators issue a signal and the vulnerability index adds up to its highest value of one. When all indicators are fixed at the sample median, the overall vulnerability index is 0.11, safely below the index threshold of 0.44. On the other hand, only strong policies and a benign global environment can protect a country with weak institutions and anemic pre-shock growth from a crisis. Although a growth crisis would not be predicted, the index would still be rather high at 0.37, and the country could be considered at medium vulnerability.

Goodness of Fit

The vulnerability index performs fairly well in identifying crisis-prone countries. Table 8 presents the distribution of the vulnerability index for crisis versus non-crisis episodes. The median predicted vulnerability index for growth crises is 0.66 versus 0.33 for normal episodes. Seventy-five percent of growth crises have vulnerability indices above 0.50, and only ten percent of crisis events have the vulnerability index below 0.31. With respect to in-sample performance, the index correctly calls 83 percent of growth crises with an overall model misclassification error of 28 percent. A secondary threshold of 0.3 is selected to differentiate low versus medium vulnerability cases. At this threshold, less than 9 percent of crisis events are missed while false alarms increase to about 56 percent.

Table 8.

Vulnerability Index

(Percentiles)

article image
Source: Authors’ calculations.

Number of growth crises divided by total observations.

The out-of-sample model performance is obtained by deriving thresholds and weights for each indicator over the 1990–2008 period and evaluating predictions during the 2009 global crisis. The vulnerability index performs well in explaining growth crises in low-income countries during the global crisis. It correctly flags 9 out of 13 countries (70 percent) that experienced a growth crisis in 2009, with false alarms occurring in 28 percent of the non-crisis episodes. Moreover, one out of the four crisis cases missed would have been flagged as being moderately vulnerable by the index.

C. Assessment of Vulnerabilities since 1990s

The vulnerability index exhibited a trend decline since the 1990s until the onset of the twin crises in 2007/08 (food and fuel and the global financial crisis) (Figure 6 and Appendix Figure 1). During this period, most low-income countries made significant strides in achieving macroeconomic stability and undertaking growth-supporting structural reforms. The trend decline in vulnerabilities was underpinned by a build-up of reserve buffers, improvements in fiscal performance and official debt-relief, substantial progress in removing policy-induced exchange market pressures, and strong growth. The benign global environment during the “great moderation” further contributed to an unprecedented drop in the index.

Figure 6.
Figure 6.

Growth Decline Vulnerability Index

1993–2011

(Quartiles of the Index and Distribution of Vulnerability Flags)

Citation: IMF Working Papers 2012, 264; 10.5089/9781475528510.001.A001

Source: Authors’ calculations

Although the vulnerability index rose significantly in the years following the global crisis, it remained lower than levels observed in the early-90s. At the onset of the crisis, while exceptional tail shocks to external demand and export prices contributed to a deterioration in the index, high pre-crisis policy buffers rendered resilience to the risk of a growth crisis for most countries.

Vulnerabilities remained high during 2010–2011 as policy buffers were expended in the wake of the global crisis. The increased vulnerabilities are attributable to the deterioration in fiscal indicators, in particular a worsening government balance and weaker revenue growth, as well as the pronounced decline in growth in 2009, while external sector indicators largely remained sound (Figure 7). Although the aggregate index showed some easing from its peak in 2010, reflecting more buoyant external conditions and stronger than expected growth in low-income countries, fiscal vulnerabilities remain elevated.

Figure 7.
Figure 7.

Vulnerability Index and its Components

(Median, 1993–2011)

Citation: IMF Working Papers 2012, 264; 10.5089/9781475528510.001.A001

Source: Authors’ calculations.

VI. Concluding Remarks

This paper presents an index which provides early warning signals of a growth crisis in low-income countries. The index uses a parsimonious set of macroeconomic and institutional indicators, aggregated using two complementary approaches. The role of these indicators in signaling growth crises is robustly established through regression analysis and a statistical signaling approach.

The vulnerability index provides a useful tool to monitor individual country risks to sharp growth declines arising from external shocks and to inform judgment-based approaches. The results show that country fundamentals, exchange rate regimes, institutional quality, and the size of shocks are important determinants of growth crises in low income countries. The analysis also suggests that the sensitivity of crisis risks to varying policy and institutional fundamentals differs across countries. In particular, a strengthening of policy and institutional frameworks results in a larger reduction in the crisis probability in countries with initially weak buffers.

The vulnerability index and its sub-indices are calculated for a large sample of low-income countries during 1993–2011. The results show that sounder policy fundamentals and a benign global environment contributed to a marked decline in vulnerabilities till the onset of the global financial crisis. While most low-income countries recovered swiftly from the global crisis and have grown strongly since 2010, fiscal risks remain elevated compared to the pre-crisis period.

Appendix Figure 1.
Appendix Figure 1.

Flags Raised by Selected Indicators of the Vulnerability Index

(1993–2011)

Citation: IMF Working Papers 2012, 264; 10.5089/9781475528510.001.A001

Source: Authors’ calculations.
Appendix Table 1.

Correlation Matrix for Exogenous Shocks

(1990–2008)

article image
Source: Authors’ calculations.Note: Correlations among variables are reported for the common sample for which all shock variables are available. Results are very similar for the pair-wise correlation matrix relaxing the common sample restriction and in the sub-sample excluding commodity exporters. Pair-wise correlations are not statistically significant except for the negative correlation between a change in FDI to GDP and a terms of trade shock.
Appendix Table 2.

List of Countries in the Sample

article image
Appendix Table 3.

Variables Used in the Probit Regressions and the Signalling Approach

article image
Appendix Table 4.

Distribution of Explanatory Variables

(1990–2009)

article image
Source: Authors’ calculations.Note: The distribution of explanatory variables is provided for the estimation sample.
Appendix Table 5.

Robustness Check with Alternative Specifications: All Growth Crisis Events

article image
Source: Authors’ calculations.Note: (1) is estimated by correlated random effects probit model based on the significance of country-specific effect and (2) is estimated by a correlated pooled probit model with cluster-robust standard errors. Significant at 10 percent:*; 5 percent:**; and 1 percent:***, standard errors are in parantheses.

Marginal effects of a specific covariate on the response probability averaged across the distribution of covariates in the sample.

References

  • Acemoglu, D., S. Johnson, J. Robinson, and Y. Thaicharoen, 2003, “Institutional Causes, Macroeconomic Symptoms: Volatility, Crises, and Growth,” Journal of Monetary Economics, Vol. 50, pp. 49–123.

    • Search Google Scholar
    • Export Citation

  • Aghion, P., G. Angeletos, A. Banerjee, and K. Manova, 2005, “Volatility and Growth: Credit Constraints and Productivity-Enhancing Investment,” NBER Working Paper No. 11349 (Cambridge, Massachusetts: National Bureau of Economic Research).

    • Search Google Scholar
    • Export Citation

  • Baldacci, E., I. Petrova, N. Belhocine, G. Dobrescu, and S. Mazraani, 2011, “Assessing Fiscal Stress,” IMF Working Paper 11/100 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation

  • Beck, T., A. Demirgüç-Kunt, and R. Levine, 2000, “A New Database on Financial Development and Structure,” World Bank Economic Review, Vol. 14, pp. 597–605.

    • Search Google Scholar
    • Export Citation

  • Becker T., and P. Mauro, 2007, “Output Drops and the Shocks That Matter,” IMF Working Paper 07/172 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation

  • Berg, A., D., O. Jonathan, and J. Zettelmeyer, 2011, “What Makes Growth Sustained” (forthcoming; Journal of Development Economics).

  • Broda, C., 2004, “Terms of Trade and Exchange Rate Regimes in Developing Countries,” Journal of International Economics, Vol. 63, pp. 31–58.

    • Search Google Scholar
    • Export Citation

  • Bruckner, Markus and Antonio Ciccone, 2010a, “International Commodity Prices, Growth, and the Outbreak of Civil War in Sub-Saharan Africa,” The Economic Journal, Vol. 120, Issue 544, pp. 535–550.

    • Search Google Scholar
    • Export Citation

  • Chamberlain, Gary, 1982, “Multivariate Regression Models for Panel Data,” Journal of Econometrics, Vol. 1, pp. 5–46.

  • Collier, P., and B. Goderis, 2009, “Structural Policies for Shock-Prone Developing Countries,” Oxford Economic Papers, Vol. 61, Issue 4, pp. 703–726.

    • Search Google Scholar
    • Export Citation

  • Collier, P., B. Goderis, and A. Hoeffler, 2007, “Commodity Prices, Growth, and the Natural Resource Curse: Reconciling a Conundrum,” CSAE Working Paper Series 2007–15 (Oxford: Center for the Study of African Economy).

    • Search Google Scholar
    • Export Citation

  • Collier, P., B. Goderis, and A. Hoeffler, 2006, “Shocks and Growth: Adaptation, Precaution and Compensation,” Department of Economics (Oxford: Oxford University).

    • Search Google Scholar
    • Export Citation

  • Dabla-Norris, E., J. I. Kim, and K. Shirono, 2011, “Optimal Precautionary Reserves for Low-Income Countries: A Cost-Benefit Analysis,” IMF Working Paper 11/249 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation

  • Demirgüc-Kunt, A. and E. Detragiache, 1999, “Monitoring Banking Sector Fragility: A Multivariate Logit Approach,” IMF Working Paper 99/147 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation

  • Eichengreen, B., A. Rose, C. Wyplosk, B. Dumas, and A. Weber, 1995, “Exchange Market Mayhem: The Antecedents and Aftermath of Speculative Attacks,” Economic Policy, Vol. 10, No. 21, pp. 249–312.

    • Search Google Scholar
    • Export Citation

  • Hausman, R., L. Pritchett, and D. Rodrik, 2006, “Growth Accelerations,” Journal of Economic Growth, Vol. 10, pp. 303–329.

  • Herrera, S. and C. Garcia, 1999, “User’s Guide to an Early Warning System for Macroeconomic Vulnerability in Latin American Countries,” Policy Research Working Paper No. 2233. (Washington: World Bank)

    • Search Google Scholar
    • Export Citation

  • Hnatkovska, V. and N. Loayza, 2005, “Volatility and Growth” in Managing Economic Volatility and Crises: A Practitioner’s Guide, ed. by J. Aizenman and B. Pinto (Cambridge, Massachusetts: Cambridge University Press).

    • Search Google Scholar
    • Export Citation

  • International Monetary Fund, 2010, “The IMF-FSB Early Warning Exercise—Design and Methodological Toolkit.” Available at www.imf.org/external/np/pp/eng/2010/090110.pdf

    • Search Google Scholar
    • Export Citation

  • International Monetary Fund, 2011, “Managing Volatility: A Vulnerability Exercise for Low-Income Countries.” Available at www.imf.org/external/np/pp/eng/2011/030911.pdf

    • Search Google Scholar
    • Export Citation

  • Kaminsky, G., S. Lizondo, and C. M. Reinhart, 1998, “Leading Indicators of Currency Crises,” Staff Papers, International Monetary Fund, Vol. 45(1), pp.1–48.

    • Search Google Scholar
    • Export Citation

  • Kaminsky, G. and C. Reinhart, 1999, “The Twin Crises: The Causes of Banking and Balance of Payments Problem,” American Economic Review, Vol. 89, No. 3, pp. 473–500.

    • Search Google Scholar
    • Export Citation

  • Loayza, N., and C. Raddatz, 2007, “The Structural Determinants of External Vulnerability,” World Bank Economic Review, Vol. 21, No.3, pp. 359–387.

    • Search Google Scholar
    • Export Citation

  • Mundlak, Y., 1978, “On the Pooling of Time Series and Cross Section Data,” Econometrica, Vol. 46, pp. 69–85.

  • Papageorgiou C., C. Pattillo, N. Spatafora, and A. Berg, 2010. “The End of an Era? The Medium- and Long-term Effects of the Global Crisis on Growth in Low-Income Countries,” IMF Working Papers 10/205.

    • Search Google Scholar
    • Export Citation

  • Perry, G., 2009, “Beyond Lending: How Multilateral Banks Can Help Developing Countries Manage Volatility” (Washington: Center for Global Development).

    • Search Google Scholar
    • Export Citation

  • Pritchett, L., 2000. Understanding Patterns of Economic Growth: Searching for Hills among Plateaus, Mountains, and Plains,” World Bank Economic Review, Vol. 14, No.2, pp. 221–250.

    • Search Google Scholar
    • Export Citation

  • Reinhart, C. M. and K. Rogoff, 2004, “The Modern History of Exchange Rate Arrangements: A Reinterpretation,” The Quarterly Journal of Economics, Vol. CXIX, Issue 1, pp. 1–48.

    • Search Google Scholar
    • Export Citation

  • Rodrik, D., 1999, “Where Did All the Growth Go? External Shocks, Social Conflict, and Growth Collapses,” Journal of Economic Growth, Vol. 4, pp. 85–412.

    • Search Google Scholar
    • Export Citation

  • World Bank, 2006, “IDA Countries and Exogenous Shocks” (Washington).

*

This paper was developed as input to the IMF Policy Paper “Managing Volatility-A Vulnerability Exercise for Low-Income Countries” (IMF, 2011). The authors would like to thank Hugh Bredenkamp, Saul Lizondo, Sharmini Coorey, Andy Berg, Catherine Pattillo, Christian Mumssen, Raphael Espinoza, Marco Arena, and Enriko Berkes for helpful comments and suggestions.

1

Studies have found that the negative impact of exogenous shocks on growth and consumption volatility is especially pronounced in low-income countries (Becker and Mauro, 2007; Perry, 2009), and that such impact results mostly from crises or severe recessions rather than normal cyclical fluctuations (Hnatkovska and Loayza, 2005).

2

The importance of these factors in reducing resilience to shocks is well established in the empirical literature (see Collier et al., 2006; Loayza and Raddatz, 2007; Acemoglu et al., 2003; and Rodrik, 1999).

3

Studies find that shocks impact long-run growth through reductions in investment (Aghion et al., 2005), a worsening of economic policy (World Bank, 2006) and, in extreme cases, by increasing the risk of conflict (Bruckner and Ciccone, 2010).

4

If economic agents were liquidity-constrained or short-sighted, they may be more concerned about an immediately apparent sharp decline in growth than a slowdown in the long-run economic growth that has a similar impact in net present value terms (Becker and Mauro, 2007).

5

FDI, aid, and remittances are measured as ratios to GDP. Large natural disasters are identified if the number of people affected and the economic damage was considered to be among the top 25th percentile of the distribution. Data on natural disasters are drawn from the Emergency Events Database. See Dabla-Norris et al. (2011) for details.

6

The severe state failure events are taken from Political Instability Task Force (PITF) dataset. Four types of political crises are included in this dataset: revolutionary wars, ethnic wars, adverse regime changes, and genocides. From this dataset the variable SFTPMMAX, which presents the maximum magnitude of all events in a year, exceeding 3.9 is taken as a severe state failure event.

7

Pooled probit models assume independence of observations over both t and i. A random effects (RE) probit model treats the individual specific effect, ci, as an unobserved random variable with ci|xit∼IN(μc, σ2c) if an overall intercept is excluded, and imposes independence of ci and xi. A fixed effects (FE) probit model treats ci as parameters to be estimated along with β, and does not make any assumptions about the distribution of ci given x¿. This can be problematic in short panels as both β and ci are inconsistently estimated owing to an incidental parameters problem. Finally, a correlated model relaxes independence between ci and xi using the Chamberlain (1982)-Mundlak (1978) device under conditional normality. In this specification, the time average is often used to save on degrees of freedom.

8

Results are available from authors upon request.

9

Drawing from indicators of speculative pressure in the crisis literature (Eichengreen et al., 1995, Kaminsky and Reinhart, 1999, and Herrera and Garcia, 1999), this study uses a composite indicator given by:

EMPIit=ln(xritxrit-1)σΔIn(xr)-resit-resit-1msgt-1σΔres/msgt-1+ln(1+blackprit)σΔln(xr)

where EMPIit is the exchange market pressure index for country i at time t, xr is the exchange rate of national currency to U.S. dollar (an increase indicates a nominal depreciation), res is the stock of international reserves, mgs is the imports of goods and services, blackpr is the black market premium, and σ is the standard deviation of each variable. Weights are inverses of the standard deviation of each component for all countries over the full sample after removing the outliers. Higher levels of EMPI indicate increased pressures on the exchange rate. This version of the index was first used in Bal Gündüz (2009).

10

The World Bank’s CPIA (the Country Policy and Institutional Assessment) is a broad indicator of the quality of a country’s present policy and institutional framework. It is based on 16 criteria which are grouped into four clusters: economic management, structural policies, policy for social inclusion and equity, and public sector management and institutions.

11

Small islands are defined as islands with a population less than 1 million.

12

Owing to the non-linearity of the model, estimated coefficients have no direct interpretation. Marginal effects of variables calculated at preset values of other explanatory variables (at their means or medians) are reported to present the relative impact of each variable.

13

Marginal effects of a specific covariate are averaged across the sample distribution of other covariates.

14

The ranges for individual covariates in Figure 3 are determined after removing outliers on both tails of their distributions in the estimation sample.

15

The less informative the model, the less dispersed the predicted probabilities, the limiting case being the flat sample probability predicted for both types of events.

16

Both approaches were implemented, but as in previous studies, the maximization of the signal-to-noise ratio led to corner solutions associated with very high type I or type II errors. Therefore, we reported thresholds based on the minimization of misclassification errors.

17

The weights of individual indicators are determined on the basis of their goodness of fit.

18

We rebalanced unconstrained weights predicted from the signaling power of indicators to account for potential correlation among variables both within a cluster and across clusters. Unconstrained weights obtained were 51 percent for overall economy and institutions; 28 percent for external index; and 21 percent for fiscal index. Adjustments to the weights were guided by model performance.

19

In the univariate probit regressions, the overall index and its sub-components are entered one-by-one as the only covariate in addition to the constant. In the multivariate probit regressions, all subcomponents of the index are included simultaneously.

Same Series

Other IMF Content

Other Publishers