Chapter

Chapter 3. Growth Dynamics: The Myth of Economic Recovery

Author(s):
Christopher Crowe, Simon Johnson, Jonathan Ostry, and Jeronimo Zettelmeyer
Published Date:
August 2010
Share
  • ShareShare
Show Summary Details

Valerie Cerra • Sweta Chaman Saxena1

3.1. Introduction

Although researchers have documented that many financial crises are associated with severe recessions (Kaminsky and Reinhart, 1999)), very little attention has been paid to whether countries recover from such large negative shocks in the sense that output losses are reversed. A few recent studies show persistent output loss from financial crises in a small set of countries. For instance, Cerra and Saxena (2005a) demonstrate that six Asian crisis countries suffered permanent output loss from the Asian crisis, and Cerra and Saxena (2005b) show that only a tiny fraction of the output loss from Sweden’s banking crisis in the early 1990s was recuperated. The graphs in Figure 3.1 illustrate persistent output loss for selected countries following the 1997–98 Asian financial crisis and the debt crisis of the early 1980s.

In addition to financial crises, many countries experience large negative political shocks, which could include violent conflicts such as civil wars as well as a deterioration in the country’s governance. Such political shocks have the potential for significant disruption to economic activity, as illustrated for a few episodes of civil war (Figure 3.2).

This chapter systematically documents the behavior of output following financial and political crises in a large set of 190 countries. Whereas the graphs in Figure 3.1 and Figure 3.2 are suggestive, our aim is to formally analyze the impact of financial and political shocks on output in a broad set of countries, particularly whether output losses are recovered. Financial shocks are comprised of currency, banking, and twin financial crises. For political shocks, we examine civil wars, a deterioration in the quality of political governance, and twin political crises comprised of both shocks. We choose civil wars rather than interstate conflicts to ensure that the war occurs on the country’s own soil. The military theater for some interstate conflicts may not directly encompass all parties to the conflict. In addition, the increase in wartime

Figure 3.1.(A) Asian crisis. (B) Debt crisis.

Figure 3.2.Protracted civil wars.

spending for an international conflict may boost economic activity in some countries. We also wish to examine the economic impact of a deterioration in a country’s political governance or institutional quality. Acemoglu, Johnson, and Robinson (2001)) and Acemoglu and others (2003)) use constraints on the power of the political executive as a measure of institutional quality and find that it is linked to growth and volatility. Thus, we use this measure to study the shock to political governance.

Potential endogeneity of the financial or political crisis is an important issue in estimating the output impact of the crisis. That is, the crisis itself may be a function of a slowdown of economic growth or changes in expectations of future growth. We attempt to address this issue using a few methods, which are far from definite, but nonetheless uncover some interesting facts. In particular, we find that the forecasts of growth from an autoregressive model and from consensus surveys are optimistic relative to actual growth occurring during and after a crisis.

3.2. Methodology

We are interested in examining the impact of financial and political crises on output. We follow the methodology used by Romer and Romer (1989)) to identify the impact of monetary policy shocks on output. We construct qualitative indicators of financial and political crises and estimate impulse response functions to the shock. Given that our data consists of a large set of countries, we estimate the models using panel data analysis, and we provide group averages of the impulse responses of output to each type of shock. We are also able to partition the country samples to examine any differential impact of a shock on countries according to their income level or region.

We formally test the statistical relationship between growth and the shock by estimating impulse response functions for each different type of shock. In particular, we estimate a univariate autoregressive model in growth rates, which accounts for the nonstationarity of output (Nelson and Plosser, 1982)) and for serial correlation in growth rates. 2 We control for country fixed effects, which F-tests indicate are present. 3 We estimate an AR(4), as we find insignificant coefficients beyond the fourth lag. We estimate the model on all of the available data from 190 countries over the period 1960 through 2001. We then extend the estimation equation to include the current and lagged impacts of the shock. Thus, we estimate the following model:

where g is the percentage change in real GDP and D is a dummy variable indicating a financial or political crisis. The impulse response functions to each crisis type are shown with one standard error bands drawn from a thousand Monte Carlo simulations.

3.3. Data

We use GDP growth rates from the World Bank’s World Development Indicators (WDI). This dataset contains the largest sample of countries. Our data consists of unbalanced panels of annual observations spanning 190 countries from 1960–2001. We also disaggregate the results based on World Bank classifications. The countries are split into seven regional groups—Africa, Asia, Industrial Countries, Latin America, Middle East, Transition Countries, and Western Hemisphere Islands—and four income groups—low income (per capita real GDP less than or equal to $735), lower middle income (per capita real GDP between $736 and $2935), upper middle income (per capita real GDP between $2936 and $9075), and high income (per capita real GDP over $9075). For robustness, we also use GDP growth rates computed using the Penn World Tables data (Heston, Summers, and Aten, 2006)), converting per capita growth to aggregate GDP growth rates using population growth rates.

We form a panel data set for currency crises by constructing an exchange market pressure index (EMPI) for each country. The EMPI is defined as the percentage depreciation in the exchange rate plus the percentage loss in foreign exchange reserves. This formulation makes indices comparable across countries. 4 A dummy variable for a crisis is formed for a specific year and country if the EMPI is in the upper quartile of all observations across the panel.

We obtain banking crisis dates on a large set of countries from Caprio and Klingebiel (2003). We confine the analysis to systemic banking crises. Moreover, because the end of a banking crisis is often highly uncertain, we restrict our analysis to the initial shock stemming from the first year of a banking crisis.

To check the robustness of our results, we also use banking and currency crisis dates from Kaminsky and Reinhart’s (1999) influential study on twin crises. However, the drawback of this source is that there are only 23 countries included in the study, which prevents us from examining the regional and income group disaggregations.

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

The data on the quality of the government comes from Polity International IV data set. The constraint on the executive variable is constructed by the Polity IV project by coding the authority characteristics of states in the world. The variable measures the extent of regular institutional constraints on executive power. These constraints arise from accountability groups, such as legislatures and judiciaries that have equivalent or greater effective authority, or can impose constraints on executive behavior in most activities. The scale ranges from 1 (weak constraints on executive power) to 7 (strong constraints on executive power).

Data on consensus forecasts of economic growth in the current year and next year is obtained from the commercial database produced by Consensus Economics Inc. The data is gathered monthly by surveying the expectations of analysts typically from large banks and financial firms within the country. The data cover 35 high-income, emerging market, and transition countries. Coverage begins in late 1989 for high-income countries, with additional countries added from 1990 to 1995. The frequency is monthly for high-income and Asian emerging market countries and bimonthly for Latin American and transition countries.

3.4. Results

3.4.1. Impulse Responses

The impact of a currency crisis on output is negative and highly persistent (Figure 3.3). The loss in output averages about 4 percent for the entire panel of countries. The depth of the loss varies by income group, with output loss averaging only 1 percent for high-income countries, but close to 5 percent for all other income groups. All groups except countries in the Middle Eastern region experience output loss, which persists even at a 10-year horizon. Indeed, no income group or regional group experiences a rebound in output by more than ½ of a percent relative to the point of deepest loss.

The output impact of a banking crisis is nearly twice as large (7½ percent loss) as a currency crisis and just as persistent (Figure 3.4). Output loss at a 10-year horizon exceeds 6 percent for all groups except the Latin American region and lower-middle-income countries. Whereas currency crises have a modest impact on high-income countries, banking crises lead to severe output loss in this group.

Output loss from twin financial crises is deeper than either of the individual crises (Figure 3.5). By three years after the crisis, output loss reaches and remains at 10 percent. The persistence of the loss is robust to all regional and income groups except for the Latin American subset.

In contrast to the extreme persistence of output loss following financial crises, output partially rebounds from a civil war (Figure 3.6). On average for the panel of countries, output declines by 6 percent initially. Half of the loss is recuperated after four years, but 3 percentage points of cumulative loss remain even after a decade. These results likely reflect the combination of both permanent and temporary effects. Physical infrastructure is damaged in most war situations and constrains output, but infrastructure may be repaired within a short time after the conflict ceases. Also, parameter uncertainty is large, and the standard error bands encompass zero for several groups. This result reflects the wide range of experiences of different countries to postconflict situations. The positive impact of civil war on output for industrial countries should be treated with caution, as the result is driven by limited episodes for this set of countries. 5

Figure 3.3.Impulse responses: Currency crises.

Figure 3.4.Impulse responses: Banking crises.

Figure 3.5.Impulse responses: Twin financial crises

(currency and banking).

Figure 3.6.Impulse responses: Civil wars.

The average output loss following a deterioration in constraints on executive power is as large and persistent as that for currency crises, but the impact varies markedly by the different regional and income groups (Figure 3.7). Weaker constraints on executive power (fewer institutional checks and balances) are associated with output gains for Asia and the Middle East, but persistent losses for Africa, Latin America, transition countries, and islands in the Western Hemisphere. The difference by income group is monotonic. High-income countries have significant output gains when executive discretion strengthens. Upper-middle-income countries experience output losses of several percent, but the standard error bands are large. Lower-middle-income and low-income countries experience output losses of about 5 percent. In contrast to the partial rebound in output observed for civil wars, the output loss associated with this measure of a deterioration in governance leads to sustained output loss, averaging 4 percent for the full set of countries. The different impact of changes in executive power on income groups may reflect large differences in initial starting levels. In high-income countries, power is widely distributed among institutions with strong constraining mechanisms. The average level is 6.1 on the scale from 1 to 7. The average level declines monotonically with income groups to 2.8 in low-income countries. The results suggest that an optimal sharing of power may lie between the bounds. At low levels of power sharing, executive power may be too discretionary and contribute to cronyism. Too much power sharing, on the other hand, may cripple decision making.

Figure 3.7.Impulse responses: Stronger executive power.

Twin political crises (civil wars combined with fewer controls on executive discretion) have the most severe overall impact on output of any large negative shock that we study (Figure 3.8). Output declines by about 16 percent on average for our broad set of countries. Moreover, the loss is persistent, with no discernible rebound. The output loss is particularly devastating for low-income countries, reaching 20 percent.

Figure 3.8.Impulse responses: Twin political crises

(war and stronger executive power).

3.4.2. Distribution of Shocks

The results above show that the impact of a crisis varies among different country groups. For several types of shocks, output loss is more severe for lower-income countries than high-income countries. In this section, we calculate the frequency of each type of shock for each country subsample. The analysis consists of all country-year observations in which data on the growth rate and data on the shock indicator are available.

The frequency of shocks varies considerably across different country subgroups (Table 3.1). In particular, the frequency increases sharply and nearly monotonically as the income level of the country group falls. Financial crises occur almost twice as often in low-income countries than in high-income countries. 6 Political crises are even more unequally distributed. Civil wars occur in 18 percent of all years in low-income countries, but are not observed in high-income countries. A deterioration of constraints on executive power is not observed in high-income countries in the most recent decade of available data, but occurs in nearly a quarter of the years for low-income countries.

Table 3.1Probability of Shocks
All Available Years1992–2001
Financial CrisesPolitical CrisesFinancial CrisesPolitical Crises
CurrencyBankWarConstraintsCurrencyBankWarConstraints
Africa3741263441623
Asia233167194109
Industrial countries2221128141
Latin America315105275108
Middle East264104162512
Transition countries276642061113
Western2120311503
Hem. islands
High income2020220100
Upper middle2352316408
income
Low middle income3051062661312
Low income3641863241923

These computations indicate that the higher frequency of crises in lower-income countries relative to high-income countries, especially in the recent decade, compounds the generally larger output loss associated with the crises. 7 Indeed, multiplying full sample estimates for the long-term output loss of a crisis (the loss at the 10-year horizon in the impulse response) by the probability of a crisis in each year, we find that a 10-year accumulation of financial and political shocks could reduce the long-term level of output in low-income countries by as much as 25 percentage points more than in high-income countries. Currency crises produce the greatest differential effects between high-income countries and the other three groups because of significant differences in both frequency and magnitude. In addition, the considerably higher frequency of governance shocks (less constraints on executive discretion) in low- and low-middle-income countries in the recent period would intensify the differences in output loss between the income groups.

3.4.3. Robustness Checks

We also check the robustness of our results using alternative sources of crisis dates and growth rates, and controls for some common shocks.

The results are robust to the source of data on crisis dates and growth rates. As a check of the robustness of our results on the persistent output loss from financial crises, we examine the impulse response functions using the crises dates from Kaminsky and Reinhart (1999). As shown in Figure 3.9, we continue to find deep and persistent output loss from currency, banking, and twin financial crises using this alternative set of crisis dates. The size of the loss is, in fact, larger by 1–4 percentage points. We also substitute growth rates from the latest release of the Penn World Tables, and find similar impulse response functions (Figure 3.10).

The results are also robust to controls for common shocks. In Figure 3.11, we add oil price changes into the regression. In Figure 3.12, we allow for arbitrary common shocks by including period effects. The impulse response functions are not affected much by either of these controls.

In addition to examining robustness, we also note that our estimates of the extent of output loss may be conservative because of the use of fixed effects estimation. As mentioned, fixed effects estimation provides a downward bias to the coefficient estimate on a lagged dependent variable in a panel regression (Nickell, 1981), although we argue it should be very small in this dataset given the fairly long time series. To the extent that true growth rates are more serially correlated than our estimates, the impact of a shock on output would be magnified compared to our estimates.

Figure 3.9.Impulse responses: Financial crises using Kaminsky and Reinhart (1999) dates.

Figure 3.10.Impulse responses: Penn World Tables dataset.

Figure 3.11.Controlling for oil price changes.

Figure 3.12.Impulse responses: Controlling for common period shocks.

3.5. Exogeneity

Our estimating equation from section 3.2 assumes that we can treat the occurrence of a crisis as a contemporaneously exogenous event with respect to output growth. However, the other polar case—in which output growth is contemporaneously exogenous with respect to the crisis and the crisis has only a lagged effect on output—is also plausible. Second, we estimate equation (1) as a single equation, ignoring any feedback in which the probability of a crisis is affected by output growth. In this section, we discuss the impact of alternative assumptions and provide some evidence to address these issues of exogeneity.

Formally, suppose that we can model the relationship between growth and crisis as a bivariate system of equations, conceptually analogous to a bivariate vector autoregression, although not linear. That is, we augment the growth equation with another equation specifying that the probability of a crisis depends on contemporaneous and lagged growth and lags of the crisis dummy.

In section 3.2, we assume zeros for all coefficients in equation (2). This assumption implies that the crisis is strictly exogenous for growth, and serially uncorrelated.

Using this framework, we can modify our assumption in two key ways. First, if δ0 = 0, then output growth is contemporaneously exogenous, and the crisis has only a lagged effect on growth. The original assumption and this modified assumption are similar to the two alternative triangular factorizations that can be imposed for a Cholesky decomposition of a bivariate VAR. The second important modification would be to allow nonzero coefficients in equation (2)., so that output growth can affect the probability of a crisis and the crisis dummy can be serially correlated.

The data indicates that lower growth is associated with a higher probability of a crisis within the same year. By assuming that the crisis is exogenous, we attribute the low growth to the impact of the crisis. But if, instead, output growth is exogenous (δ0 = 0), then the crisis may be a result, rather than a cause, of the low growth. In this case, the impact of the crisis would only be because of the lagged effects in equation (1). The output loss shown in Figure 3.3Figure 3.12 may thus be exaggerated.

The next two sections below provide some evidence to address this question of the contemporaneous relationship between growth and a crisis. In section 3.5.1, we generate dynamic forecasts from a univariate autoregressive model of growth. We compute errors between forecasts from the model and actual growth following a crisis under each polar case for contemporaneous exogeneity of output growth and crisis. In section 3.5.2, we analyze consensus forecasts of expected growth in an attempt to disentangle whether a crisis leads to output loss, or whether actual or expected output loss leads to the crisis. The results provide some evidence pointing to growth optimism at the time of a crisis.

In Section 3.5.3, we explicitly impose the alternative assumption that output growth is exogenous by setting δ0 = 0 in the estimation equation. In this case, the crisis has only lagged effects on output growth, and we thus generate impulse response functions that correspond to this alternative assumption.

In Section 3.5.4, we discuss the impact of allowing nonzero coefficients in equation (2). We provide some evidence from probit models showing that low growth would increase the probability of a future crisis. Moreover, crises are serially correlated. The implication of this evidence is that the results for output loss shown in Section 3.4 may be underestimated because we have ignored these feedback effects to the probability of a crisis.

In Section 3.5.5, we consider some additional potential specification errors, such as the impact of expectations and the possibility of third variables driving both growth and crises.

3.5.1. Forecast Errors

We compare the actual level of output following a financial or political crisis with the level of output predicted from a univariate AR model that controls for normal business cycle dynamics. The forecast error provides a measure of the impact of the crisis. We show how the results differ under alternative extreme assumptions that all contemporaneous correlation between output and crisis can be attributed to (1) crisis innovations or (2) output innovations.

For the panel of countries, we estimate a univariate autoregressive model in growth rates to account for the business cycle and any ex-ante slowdown in growth:

For each crisis date, t, in each country, we then compare current and subsequent actual growth rates to those of dynamic forecasts constructed using coefficient estimates from the AR(4) model. However, contemporaneous correlation between current growth and the crisis must be distributed between the two variables, as current growth may be unexpectedly low because of the crisis or the crisis may occur because of the negative innovation in growth. To account for these possibilities, we construct two sets of forecast errors that correspond to each extreme assumption.

Assuming that low growth in time t occurs because of the crisis innovation in time t, we form 1-, 2-, 3-, and 4-period ahead dynamic growth forecasts using only growth data through time t −1. The forecast errors are given by:

Alternatively, if we fully attribute any slowdown in growth in the year of a crisis (contemporaneous correlation between growth and crisis) to growth innovations, then we can construct forecast errors from our AR(4) model using growth information through time t. Under this assumption, the growth slowdown is responsible for the crisis in time tand the forecast errors will pick up only lagged effects from the crisis to future growth. Thus, the forecast errors are constructed as:

We compute such sets of forecast errors for each crisis in our panel data, and compute the average forecast error across the sample at each horizon. Figure 3.13 presents forecast errors of the level of output by accumulating the 1-, 2-, 3-, and 4-year ahead forecast errors of output growth as described in equation (4) and equation (5). Forecast errors shown by solid lines assume that any correlation between growth and a financial or political crisis is attributed to the crisis in the year it occurs, whereas those shown by dashed lines attribute the correlation entirely to the innovation in the growth rate.

The results show that output loss occurs irrespective of which polar assumption is used to attribute contemporaneous correlation between output and crisis innovations. The magnitude of output loss is smaller if crises have only lagged effects on growth, corresponding to δ0 = 0 in equation (1). The attribution of contemporaneous correlation to growth versus crisis innovations impacts the magnitude of output loss more for political crises than for financial crises. However, on average across the panel of countries, actual growth rates fall short of those that are projected to account for normal business cycle fluctuations for all four types of shocks, regardless of alternative assumptions on contemporaneous exogeneity.

Figure 3.13.Forecast errors.

3.5.2. Consensus Growth Forecasts

We consider the possibility that a crisis may occur not only because of an ex-ante decline in output growth, but also because economic agents may expect a future slowdown. For instance, suppose that economic agents revise downward their forecast of growth, and as a result, they take actions that induce a financial crisis, start a civil war, or weaken the constraints on executive power. In this situation, the contemporaneous correlation between the crisis and growth should be attributed to growth, and the crisis would have only a lagged impact on growth.

To account for changes in growth forecasts, we collect consensus forecasts of economic growth in the crisis year and subsequent year for a set of industrial and emerging market countries. For each type of financial and political crisis, we regress the crisis dummy indicator on the difference between the midyear consensus forecast of growth in the current crisis year and the actual growth outturn, as well as the difference between the midyear forecast of growth in the following year and its actual outturn. 8 We also compare the timing of revisions to expected growth using a difference of differences specification. We regress the crisis dummy on the changes in consensus forecasts [([(FitgitFi,t1git)(Fi,t+1gi,t+1Fitgi,t+1)], where Fit denotes the consensus forecast of country i in the middle of year t. If the revision to growth forecasts occurs before the crisis, the first term would be negative and the second term zero. In contrast, the first term would be zero and the second part of the expression would be positive if growth forecasts are revised downward after the crisis (that is, if Fi,t+1gi,t+1<Fitgi,t+1.

For financial crises, we find robust evidence of growth optimism. Table 3.2 shows that the midyear consensus forecast of growth in the year of a crisis is 0.8 percentage points higher than the actual outturn for currency crises and 2.4 percentage points higher for banking crises. The expectational error is even larger for growth in the subsequent year: 1.9 percentage points too optimistic for currency crises and 4.7 percentage points too optimistic for banking crises. Moreover, growth revisions lag, rather than lead, a financial crisis, especially for banking crises. Growth is revised downward by 6 percentage points more after the banking crisis starts compared with any ex-ante revision.

The evidence on consensus forecasts for political crises is insignificant. The weak results may reflect that the sample of countries and time period available for consensus forecast data is quite restricted compared with the broad sample of countries available for the impulse response analysis. The consensus forecast sample includes only one low-income country. Thus the results tend to be biased toward higher-income countries, which do not suffer as large output losses from political crises, especially from the decline in constraints on executive power.

Table 3.2Consensus Forecast Errors for Growth
Financial CrisisPolitical Crisis
CurrencyBankWarConstraints
Current year expectation error0.83***2.43***-0.100.97
Standard errors0.310.521.031.13
Number of countries34333434
Number of observations408352229386
Next year expectation error1.88***4.47***-0.59-0.49
Standard errors0.510.840.871.67
Number of countries34303433
Number of observations376328205362
Revision of expectations0.385.84***0.70-0.23
Standard errors0.5610.71.281.76
Number of countries34343434
Number of observations378326199358
Notes: Sample is 1990–2004, as available. July is used for consensus forecasts, except for Latin American countries, where June is used. Regressions include fixed effects. Asterisks

denote significance at the 1 percent level.

Notes: Sample is 1990–2004, as available. July is used for consensus forecasts, except for Latin American countries, where June is used. Regressions include fixed effects. Asterisks

denote significance at the 1 percent level.

3.5.3. Output Growth Contemporaneously Exogenous

In this section, we consider the case that a crisis has only a lagged effect on output growth. That is, we impose a zero restriction on the coefficient of the contemporaneous value of the dummy variable in equation (1), δ0 = 0. This assumption implies that output is contemporaneously exogenous with respect to a crisis. We examine the output loss associated with this alternative specification. In particular, we estimate a revised equation in which crises impact GDP growth only through lags, and we control for the dynamics of growth rates:

Figure 3.14 shows the impulse responses from this regression. The change in the assumption of contemporaneous exogeneity has a smaller impact on the results for financial crises than political crises. The finding of persistent output loss remains robust for financial crises. On the other hand, the magnitude of loss is dampened when the contemporaneous decline in output is attributed to the output innovation. But even under this assumption, the lagged effects of currency, banking, and twin financial crises still result in 2½ percent, 4 percent, and 5 percent of output loss, respectively, by the end of 10 years. For wars and the weakening of executive constraints, output falls initially, but at the end of 10 years it is only one percentage point lower than its initial level. Output loss from twin political crises remains at 4 percent at the end of 10 years, but the uncertainty bands are large.

Figure 3.14.Impulse responses: Lagged response of output to crises.

3.5.4. Feedback to the Probability of Crisis

The results presented in Section section 3.4 ignore the possibility that growth affects the probability of a future crisis or that crises are serially correlated. We relax this assumption by estimating a probit model for equation (2), using each type of crisis indicator as a dependent variable. We impose only the restriction that γ0 = 0, so that the crisis is contemporaneously exogenous. The results in Table 3.3 show that even when controlling for lags of the crisis itself, 9 the first lag of growth has a significant inverse relationship with the probability of each type of crisis. That is, lower (lagged) growth leads to a higher probability of crisis. Moreover, currency crises and civil wars are positively serially correlated. Under the assumption that γ0 = 0, the omission of such feedback effects—higher probability of crisis resulting from both lower lagged growth and positive serial correlation—from the results shown in Section 3.4 imply that the impulse responses underestimate the overall impact of a crisis on output.

Table 3.3Probit Results for Financial and Political Crises
Financial CrisisPolitical Crisis
Dep var/shock typeCurrencyBankingCivil WarConstraints
Constant-0.783***-0.906***-1.767***-1.815***-2.253***-1.761***
0.0320.2200.0450.3180.0550.041
Shock (-1)0.326***0.525***2.758***-0.140
0.0440.2090.1630.200
Shock (-2)0.160***0.0470.073-0.112
0.0450.2230.2130.197
Shock (-3)0.169***0.0740.671***0.292*
0.0450.2120.1760.153
Growth-0.093-0.169***
0.0310.042
Growth (-1)-0.010***-0.016-0.028***-0.030-0.012**-0.017***
0.0030.0290.0060.0370.0060.005
Next year growth0.0540.192*
expectations0.0710.102
Number of4,6913123,0432865,0784671
observations
Sample1964–20021990–20011975–20021990–20011962–19971962–1997
Notes: Standard errors are shown below coefficient values:

significant at, or below, 1 percent;

significant at, or below, 5 percent;

significant at, or below, 10 percent.

Notes: Standard errors are shown below coefficient values:

significant at, or below, 1 percent;

significant at, or below, 5 percent;

significant at, or below, 10 percent.

3.5.5. Expectations and Omitted Variables

To summarize, we provide some evidence of growth optimism at the time of a crisis, suggesting that the crisis is contemporaneously exogenous with respect to output growth. If this finding is invalid, and instead growth is contemporaneously exogenous, we then show that even the lagged effects of the crisis would reduce output. Lower output and serial correlation of crises would act to further reduce output by feedback effects that increase the probability of a future crisis.

Nevertheless, other potential specification errors need to be considered. Even the lagged impact of a crisis on growth could overstate the output loss from crisis innovations if we allow for expectational factors and omitted variables. For instance, suppose that the crisis impacts output growth with a lag, as in Section 3.5.3 previously, where δ0 = 0. But suppose also that the probability of a crisis depends on expectations of lower future growth and growth depends on some omitted variable, Z, that relevant economic agents can observe. The system of equations becomes

Even though the crisis dummy is predetermined with respect to output growth in equation (7), it can still be correlated with the error term that contains the omitted variable. We make use of our consensus forecast data to test the specification of equation (8), although the consensus forecast data limits the countries and time periods and we can only show results for financial crises. This information suggests that the expectation of future growth is not a significant determinant of a currency crisis once we include actual growth in the regression equation (shown in the second column of results in Table 3.3). The expectation of future growth is significant at the 10 percent level for the probability of a banking crisis (column 4 of Table 3.3), but the sign is perversely positive, again suggesting excessive growth optimism.

Finally, we cannot rule out the possibility of an omitted variable in both equations that causes a crisis and reduces growth:

In this situation, omitting variable Z implies that the coefficient on the crisis dummy variable may be overestimated as it captures the correlation with the error term rather than the pure effect of the (lagged) crisis on growth. This possibility is quite plausible, given that crises and growth are likely to be related to or driven by other macroeconomic variables.

3.6. Conclusion

Using panel data for a large set of high-income, emerging market, developing, and transition countries, this chapter documents that the large output loss associated with financial crises and some types of political crises is highly persistent. Impulse response functions show that less than 1 percentage point of the deepest output loss is regained by the end of 10 years following a currency crisis, banking crisis, deterioration in political governance, twin financial, or twin political crisis. Of the large negative shocks examined, a partial rebound in output is observed only for civil wars. Moreover, the magnitude of persistent output loss ranges from around 4 percent to 16 percent for the various shocks.

The chapter provides some suggestive, although not definitive, evidence of causality. Financial crises are associated with growth optimism. Forecasts of economic growth, whether measured by projections from a univariate autoregressive model or by consensus forecasts of financial experts, tend to be higher than actual growth outturns. However, this evidence cannot rule out the possibility of a third factor that precipitates a crisis and leads to a reversal of growth optimism.

The results pose a challenge to explain the observed behavior of output following the various negative shocks. Temporary output losses could be explained by allowing for variable capacity utilization or other elements of business cycle models, but the puzzle is to explain the permanent effects. It would be useful, therefore, to develop theoretical models with propagation mechanisms that are persistent, especially for low-income and emerging market economies.

References

    AcemogluDaronSimonJohnson and JamesRobinson2001“Colonial Origins of Comparative Development: An Empirical Investigation,”American Economic ReviewVol. 91No. 5pp. 1369–1401.

    • Search Google Scholar
    • Export Citation

    AcemogluDaronSimonJohnson and JamesRobinsonandYunyongThaicharoen2003“Institutional Causes, Macroeconomic Symptoms: Volatility, Crises and Growth,”Journal of Monetary EconomicsVol. 50No. 1pp. 49–123.

    • Search Google Scholar
    • Export Citation

    AguiarMark and GitaGopinath2007“Emerging Market Business Cycles: The Cycle is the Trend,”Journal of Political EconomyVol. 115No. 1pp. 69–102.

    • Search Google Scholar
    • Export Citation

    CaprioGerard and DanielaKlingebiel2003“Episodes of Systemic and Borderline Financial Crises.”Available via Internet: http://go.worldbank.org/5DYGICS7B0.

    • Search Google Scholar
    • Export Citation

    CerraValerie and Sweta C.Saxena2005a“Did Output Recover from the Asian Crisis?”IMF Staff PapersVol. 54No. 1pp. 1–23.

    CerraValerie and Sweta C.Saxena2005b“Eurosclerosis or Financial Collapse: Why Did Swedish Incomes Fall Behind?”IMF Working Paper 05/29 (Washington: International Monetary Fund).

    • Search Google Scholar
    • Export Citation

    HestonAlanRobertSummers and BettinaAten2006“Penn World Table Version 6.2,”(Philadelphia: Center for International Comparisons of Production Income and Prices at the University of Pennsylvania).

    • Search Google Scholar
    • Export Citation

    JudsonRuth and AnnOwen1999“Estimating Dynamic Panel Data Models: A Guide for Macroeconomists,”Economics LettersVol. 65No. 1pp. 9–15.

    • Search Google Scholar
    • Export Citation

    KaminskyGraciela and CarmenReinhart1999“The Twin Crises: The Causes of Banking and Balance of Payments Problems,”American Economic ReviewVol. 89No. 3pp. 473–500.

    • Search Google Scholar
    • Export Citation

    NelsonCharles R. and Charles I.Plosser1982“Trends and Random Walks in Macroeconomic Time Series: Some Evidence and Implications,”Journal of Monetary EconomicsVol. 10No. 2pp. 139–62.

    • Search Google Scholar
    • Export Citation

    NickellStephen J.1981“Biases in Dynamic Models with Fixed Effects,”EconometricaVol. 49No. 6pp. 1417–26.

    Polity IV Project“Integrated Network for Societal Conflict Research (INSCR) Program,”Center for International Development and Conflict Management (CIDCM) University of Maryland College Park 20742.Available via Internet: http://www.cidcm.umd.edu/inscr/polity.

    • Search Google Scholar
    • Export Citation

    RomerChristina D. and David H.Romer1989“Does Monetary Policy Matter? A New Test in the Spirit of Friedman and Schwartz,”NBER Macroeconomics AnnualVol. 4pp. 121–70.

    • Search Google Scholar
    • Export Citation

    SarkeesMeredith R.2000“The Correlates of War Data on War: An Update to 1997,”Conflict Management and Peace ScienceVol. 18No. 1pp. 12344.Available via Internet: http://www.correlatesofwar.org

    • Search Google Scholar
    • Export Citation

    SingerJ. David and MelvinSmall1994“Correlates of War Project: International and Civil War Data, 1816?1992,#x201D;Inter-University Consortium for Political and Social Research [distributor] Ann Arbor MI.Available via Internet: http://ssdc.ucsd.edu/ssdc/icp09905.html.

    • Search Google Scholar
    • Export Citation
Note: This chapter is a slightly revised version of an article that appeared in American Economic Review Vol. 98, No 1 (March 2008), pp. 439–57. Reprinted with permission.
1We are grateful to Mark Gertler, David Hendry, Adrian Pagan, Roberto Rigobon, anonymous referees, and seminar participants at the Reserve Bank of India for very useful comments and suggestions. We owe Charles Nelson for his continued encouragement in our work. The chapter is dedicated to the memory of Paul A. Cerra.
2Panel unit roots tests using the data in this study strongly suggest the presence of unit roots in output. Results are available upon request.
3The country-fixed effects are correlated with the lagged dependent variables in the autoregressive equation. However, Nickell (1981) shows that the order of bias is 1/T, which is small for this dataset. Indeed, Judson and Owen (1999) calculate that the bias of the least squares dummy variable (LSDV) estimator is approximately 2 to 3 percent on the lagged dependent variable and less than 1 percent on other regressors for a panel of size N = 100, T = 30, and low persistence.
4The crisis literature often normalizes reserves and exchange rate movements by their within country standard deviations, but then the magnitudes of the EMPI are only comparable within countries. We dropped interest rates because of the scarcity of data.
5No civil war episodes were observed for Western Hemisphere islands or high-income countries. The three industrial countries not included in the high-income group are Malta, South Africa, and Turkey. Other civil war results should also be treated with caution, as some episodes of civil war reflect events confined to a small region of a country and therefore may not have a systemic economic impact. For instance, the inclusion of smaller regional conflicts within the definition of civil war could explain the relatively small permanent output effects.
6The higher average frequency of currency crises relative to banking crises is partly a matter of data construction. Dummies for currency crises are formed from the highest quartile of EMPI, but banking crisis dummies reflect only the first year of banking crises. This construction, however, does not impact the relative frequency of crises across country groups.
7The higher frequency and deeper relative output loss for low-income and emerging market countries, combined with the persistence of output loss, implies that such crises contribute to greater volatility of permanent shocks for these countries than for high-income countries. Similarly, Aguiar and Gopinath (2007) find that business cycle fluctuations are driven primarily by permanent shocks in emerging markets. Using parameter estimates from Mexico and Canada as benchmark emerging market and developed countries, respectively, they exploit consumption and net export shares to GDP to identify the underlying productivity processes.
8The regressions in this section include fixed effects.
9Lags of banking crises are not included because by construction the variable measures only the first year in a string of possible banking crisis years.

    Other Resources Citing This Publication