Independent central banks and banking crisis liquidity

See Fernández-Albertos (2015) for a recent literature review on the politics of central bank independence.

Anecdotal evidence found in Drechsler et al. (2016) and Crosignani et al. (2017) show that restrictions on lending to Eurozone sovereigns may have indirectly and unwittingly led the European Central Bank to expand its liquidity support to Eurozone banks.

These three implications are found in the Online Supplementary Material.

Assuming the commercial bank holds no capital is without loss of generality. As shown in Repullo (2005), adding a capital requirement to the model reduces the central bank’s liquidity provision threshold, but such an effect is independent of the variables studied here.

In a seminal paper, Gorton (1988) shows that banking panics occur because households respond to exogenous events such as unexpected macroeconomic shocks rather than bank specific fundamentals. Thus, while banking panics are explainable non-random events, they may be considered random from the perspective of individual banks experiencing a liquidity withdrawal.

This assumption is made without loss of generality and serves two functions in the model. First, it is necessary for a liquidity shortfall in period 1 to have a possibility of producing a banking crisis in period 1. Second, it simplifies the household’s payoff when the central bank does not lend to the bank (i.e., the right hand side of the inequality that produces (3)).

See Kono (2006) for an example of this approach to modelling democracy.

For example, even in Rogoff (1985) the optimal central bank strategy places a non-zero weight on employment stabilization.

The model assumes zero opportunity cost to the central bank from its liquidity operations. This assumption is without loss of generality. Indeed, it is trivial to add such costs to the model. See Repullo (2005).

Since financial safety nets cannot be negative, it is assumed that \(\bar {v} \geq 0\).

Note that since the decision over central bank independence is made in period 0, the government’s decision is based on the equilibrium, and not observed, values of \(\bar {v}^{*}_{n}\), \(\bar {v}^{*}_{i}\), and π^∗.

I thank an anonymous reviewer for suggesting this hypothesis.

This step implies that the dependent variable will be measured with error, imparting a well-known downward bias in the estimated regression coefficients (Green 2003, p. 84).

See Laeven and Valencia (2013) for detailed criteria defining a distressed banking system and the list of remedial state interventions.

What this article refers to as government borrowing restrictions, Garriga (2016) labels as central bank financial independence.

The large N, small T setting also precludes popular estimators that rely more on within-unit variation such as fixed effects, and the GMM estimators of Arellano and Bond (1991) and Blundell and Bond (1998). Moreover, simulations in Clark and Linzer (2015) show that in a panel with sluggish independent variables, many units, and few observations per unit, the preference for a fixed or random effects estimator hinges on the correlation between the unit effects and the within-unit mean of the independent variable. Given that across a wide range of specifications using the variables in Table 1 the highest correlation between the unit effects and the within-unit mean of government borrowing restrictions is 0.25, the random effects estimator outperforms the fixed effects estimator in terms of total root mean squared error (Clark and Linzer 2015, p. 406). Likewise, the highest correlation between the unit effects and the within-unit mean of the credit to deposit ratio is 0.10.

For example, if the measure of capital account openness of Chinn and Ito (2008) is substituted for net capital inflows and the index of creditor rights from Djankov et al. (2007) is substituted for the democracy measure of Pemstein et al. (2010), results are very similar to the main event study.

0.235 units on the government borrowing restrictions index roughly corresponds to the difference in scores between Indonesia (0.212) and Malaysia (0.514) during their respective crises in 1997.

These results were also found to be stable and statistically significant across a wide range of event window lengths. Details on this robustness check are available in the Online Supplementary Material.

Lagged dependent variables not reported in Table 1, but are available upon request. Tests using the actest command in Stata 14 indicate that autocorrelation is purged from the model with the inclusion of three lagged dependent variables. Furthermore, support for the hypotheses hold up to and including the addition of five lagged dependent variables.

The advice of Wilkins (2018) notwithstanding, in panel data settings the inclusion of a lagged dependent variable in a small T setting opens the possibility of Nickell bias (Nickell 1981). While this possibility led to the exclusion of lagged dependent variables in the random effects estimations in Table 1, the results of columns (6) and (7) continue to hold up to and including three lagged dependent variables. These results are available upon request. Estimations using Newey-West standard errors, which account for serial correlation of the error term, are also very similar to the main results and are found in the Online Supplementary Material.

States with a Unified Democracy Score above the average in the sample, equal to 0.35, are classified as democracies. Non-democracies are states with scores below 0.35.

Results using the binning procedure are found in the Online Supplementary Material.

The absence of a response by central banks to increased leverage in autocracies may reflect more than the absence of credible central bank independence. In particular, autocracies also tend to have weak institutions that are incapable of a quick, decisive response to a crisis irrespective of the degree of leverage in the banking system. I thank an anonymous reviewer for this suggestion.