If You’re Going Through Hell, Keep Going: Nonlinear Effects of Financial Liberalization in Transition Economies

ABSTRACT

Did increasing the level and pace of financial liberalization during transition expose countries to crises? And if a crisis did strike, did liberalization do more harm or good? Using a database of 28 transition economies over 22 years, this article examines these questions across a host of economic outcomes, including savings and the size of the private sector. The results provide evidence that, while liberalization may initially increase the probability of a crisis, the prospect of a crisis drops dramatically at higher levels of financial openness. Moreover, the benefits of liberalization across several metrics outweigh the risks of these intermediate stages.

KEY WORDS:

• Banking crises
• financial liberalization
• institutions
• nonlinearity
• property rights JEL Classification: P21
• O17
• G20

Acknowledgments

The author wishes to thank Sergei Guriev, Nikolay Ushakov, Pok-Sang Lam, Shaomin Li, Taufiq Choudhry, and two anonymous referees for their extensive comments.

Notes

^1. The countries included in this dataset are Albania, Armenia, Azerbaijan, Belarus, Bosnia, Bulgaria, Croatia, Czech Republic, Estonia, the Former Yugoslav Republic of Macedonia, Georgia, Hungary, Kazakhstan, Kyrgyzstan, Latvia, Lithuania, Moldova, Mongolia, Poland, Romania, Russia, Serbia, Slovakia, Slovenia, Tajikistan, Turkmenistan, Ukraine, and Uzbekistan.

^2. This was in direct contrast to the experience under communism, where there were no “financial crises;” given the administrative role of banks in the socialist economy, financial crises were really governmental or budget crises and not bank runs, losses in the banking system, and/or bank liquidations. Thanks to an anonymous referee on an earlier version of this draft who noted the somewhat obvious lack of financial crises in the Soviet Union, but not the also obvious cause of this fact.

^3. Additionally, random-effects was chosen over a population-averaged model so that the random intercept can capture the combined effect of all omitted country-specific covariates that may cause these countries to be more prone to crisis.

^4. As in Hartwell (2013), I utilize the logarithmic transformation of log (100 + inflation rate) in order to smooth out the rather large changes in inflation that occurred during transition.

^5. Given that the commonly used Demirgüç-Kunt et al. (2008) indicator only runs through 2003, I have painstakingly updated and extended the time series of deposit insurance coverage through 2011 for all transition economies.

^6. Sensitivity tests on other macroeconomic variables featuring in the literature will be used in the next section.

^7. The use of this index will provide a more nuanced approach to institutional quality than in Angkinand, Sawangngoenyuang, and Wihlborg (2010), who use log of real GDP per capita to proxy for institutional effects.

^8. The further advantage of the clustering approach, as De Melo et al. (2001, 17) note, is that this use of principal components “reduce[s] the dimensionality of the initial conditions variables” and deals with multicollinearity of the constituent variables in a much more effective manner. Additionally, they also correctly note that initial conditions often “exert their effect jointly, so that the individual approach suffers from the omitted variables problem and results estimated coefficients that are biased.”

^9. While contract-intensive money is an objective indicator for property rights, this indicator might also catch some effects of financial sector efficiency (i.e., the amount of money held outside of formal financial institutions may be due to the failings of a country’s financial sector instead of the entire institutional system). Even if this is the case, however, as an instrument it is still theoretically sound: The failures of a financial sector could be expected to determine the government’s stance on financial sector liberalization in the following period. I also believe that any financial sector effects it could capture would be more than overwhelmed by the systemic issues (and approach to property rights) that it is meant to represent.

^10. This distinction means that, although similar, Equations (1) and (2) differ slightly and are not simultaneous. The use of methods for solving simultaneous equations, such as maximum likelihood, would thus be inappropriate.

^11. The exigencies of our panel dataset also makes GMM a good choice; in the first instance, the dataset has a shorter time dimension (t = 21 or 22) than country panels (n = 28), a problem that the GMM estimator was designed for.

^12. Additional regressions, not reported, show similar results no matter which coding of a crisis is utilized. The only discrepancy occurs when utilizing the Reinhart and Rogoff (2009) coding for Poland and Russia, which lessens the significance of the EBRD indicator from 1% to 5% and halves the economic significance. Utilizing the Babecký et al. (2014) coding yields only slight changes in magnitude but no changes in statistical significance.

^13. For each specification, a robustness test was attempted that included a full complement of time dummies; however, in every specification, the time dummies were insignificant and did not alter the significance of the main effects. Given the loss of degrees of freedom in the model from including 20 years of dummies, the results reported below do not have time dummies included.

^14. In a simple bivariate regression (not reported), the EBRD bank indicator is continually positively correlated with FDI (for all models, including FE, GLS, and PCSE), showing that this relationship isn’t conditioned on choice of controls.

^15. This specification gives us almost 100 more observations due to the low coverage of the ICRG index.

^16. The objective indicator for external liberalization showed as a poor fit in diagnostics. For this reason, I place more emphasis on the subjective models.

^17. Robustness checks performed in earlier versions of this article, but not reported here, used lags of the crisis dummy out to a horizon of 5 years in order to capture long-term effects of banking crises. Given that the nonlinearity of liberalization means that immediate effects of liberalization should decay rather quickly, there should be less of an effect as time passes. The additional work done in earlier versions of this article, not reported here, uniformly showed that the crisis dummy had little effect beyond a 1-year window, while the effect of the liberalization indicators were unchanged.

^18. Using ivreg2 in Stata 13 with the command “orthog,” the crisis dummy showed as exogenous in each of the other IV-GMM specifications. It was only with private sector size as a dependent variable that it was endogenously determined.