The determinants of Asian banking crises—Application of the panel threshold logit model

• Published date: 01 March 2022
• Author: Chung‐Hua Shen,Hsing‐Hua Hsu
• Date: 01 March 2022
• DOI: http://doi.org/10.1111/irfi.12354

ORIGINAL ARTICLE

The determinants of Asian banking crises—

Application of the panel threshold logit model

Chung-Hua Shen | Hsing-Hua Hsu

Institute of Banking and Money, Nanjing

Audit University, Nanjing, Jiangsu Province,

China

Correspondence

Hsing-Hua Hsu, Institute of Banking and

Money, School of Finance, Nanjing Audit

University, No. 86, West Yushan Road, Pukou

District, Nanjing, Jiangsu Province, China.

Email: helenhsu@nau.edu.cn

Abstract

Considering binary dependent variable, this study extends

the panel threshold model into “panel threshold logit model

(PTLM).”Our PTLM is applied on investigating the effect of

early warning indicators on banking crises in 10 Asian econ-

omies. The ratio of short-term debt to foreign reserves

serves as the threshold variable. Results confirm the exis-

tence of the threshold effect in the determinants of banking

crises, and most of the early warning indicators perform dif-

ferently in the two debt regimes. Our empirical results sug-

gest that important information may be missed in analyzing

crises when conventional logit model is used.

KEYWORDS

banking crisis, early warning signal, panel threshold logit model,

panel threshold model

JEL CLASSIFICATION

C10; C58; E44; G15

1|INTRODUCTION

The application of the panel threshold model alaHansen (1999) has recently attracted considerable attention. The

model uses a threshold variable to separate the data into two panel regimes with consideration for individual fixed

effect. The threshold effect indicates that the response of a concerned variable to an exogenous shock differs in var-

ious economic regimes. The threshold concept provides a more sensible economic interpretation than the conven-

tional linear model. Given the derivations of estimation and testing methods, we demonstrate its applicability to a

This paper is in memory of my dear advisor and hard-working academic partner, Professor Chung-HuaShen, who passed away during which the paper is

being written.

Received: 24 August 2020 Revised: 16 January 2021 Accepted: 24 May 2021

DOI: 10.1111/irfi.12354

© 2021 International Review of Finance Ltd.

248 International Review of Finance. 2022;22:248–277.wileyonlinelibrary.com/journal/irfi

real case by investigating the determinants of banking crisis. An unprecedented interest has recently emerged in

identifying thecauses of banking crisis. This trend occurredbecause banking crises usually have tremendously adverse

effects on the economic and financial conditions of a country.

1

However, most crisis warning models do not provide

satisfactory results in predicting bankingcrises. Rose and Spiegel (2011) said, “…we were essentially unable to find an

empirical model of the crisis that could link any plausible set of potential causes of the crisis to its manifestations.”

2

While visiting the London School of Economics in November 2008, Queen Elizabeth II even asked the academics,

“Why did nobody noticeit?”Hence, developing improvedearly warning indicatorsfor predicting crisis is important.

Previous studies typically use the logit or probit models to predict crisis. Three types of logit models are com-

monly used in building an early warning system for crisis. The first type of the logit model assigns one for a crisis

period and zero for a non-crisis period. For example, Davis and Karim (2008) find that the logit model is the most

appropriate approach to explain crises.

3

Berg and Pattillo (1999a) and Demirgüç-Kunt and Detragiache (2000) also

show that the probabilities of banking crises estimated through their multivariate logit model provide a more accu-

rate basis than the signals approach developed by Kaminsky and Reinhart (1999) (see also Seo and Shin (2016) and

Mensi et al. (2017) who use similar logit models). Using logit regressions, Berger and Bouwman (2017) also examine

if the amount of liquidity created by the banking sector at any point in time can be a predictor of an impending crisis.

The second type of the logit model considers a multinomial logit approach. Caggiano et al. (2014) develop a mul-

tinomial logit approach to predict a banking crisis in which they assume that the crisis involves three states, namely,

the tranquil period, the onset of the crisis, and the duration period (i.e., crisis years other than the first year).

Caggiano et al. (2016) compare the predictability of a crisis over the period 1982–2010 by using a multinomial logit

model and a logit model to estimate the first year of the crisis. When using the full sample, they find that the multi-

nomial logit model outperforms the logit model. Cesa-Bianchi et al. (2018) estimate a new data set for 38 advanced

and emerging economies over 1970–2011 using the traditional ordinary least squares and panel logit methods.

The third type considers a dynamic logit model. Candelon et al. (2014) estimate a dynamic logit model in both

country-by-country and panel frameworks when predicting currency crises for 16 countries. They find that their new

early warning system has a better predictive capability than the static logit model (Candelon et al., 2014).

However, instead of employing traditional logit or probit models to examine crisis, we extend the panel thresh-

old model of Hansen (1999) by using a binary variable as the dependent variable. We call this new model the panel

threshold logit model (PTLM). Our PTLM aims to improve traditional logit models in analyzing crises by developing

an econometric framework that can be applied to panel data with threshold effects as applied by Hansen (1999) and

by combining the estimation method of conditional (fixed effects) logit of Chamberlain (1980). Extending the contin-

uous dependent variable of Hansen (1999) to a binary variable entails a challenge in the estimation of the individual

fixed effect. We adopt the conditional maximum likelihood method of Chamberlain (1980) to eliminate

individual fixed effect.

4

We show the procedure of estimating the parameters in regimes and threshold values. We

also discuss the testing method of equal parameters between the two regimes.

In the empirical design, we use Asian economies because Davis et al. (2011) find substantially different crisis

determinants across regions, indicating the unsuitability of using global samples. Following the literature, we adopt

10 Asian economies to illustrate how to use our PTLM to investigate the determinants of the banking crisis from

1987 to 2016.

5

In terms of threshold effects, the threshold model is applied widely in various fields. For example, in macroeco-

nomics, Bick (2010) uses the inflation rate as a threshold and finds that inflation and economic growth are positively

and negatively correlated in low and high inflation regimes, respectively. In corporate finance, Shen and Wang (2005)

investigate the sensitivity of investment to cash flow in different financial leverage regimes. Daniel et al. (2008) find

that firms are more likely to manage earnings upward when premanaged earnings are below the expected dividend

thresholds than when they are not. In terms of development, Baharumshah et al. (2017) reveal that foreign capital

inflow affects economic growth in low, middle, and high financial development indicators. In the growth

field, Reinhart and Rogoff (2010) use the ratio of gross external debt to GDP as the threshold to examine different

economic growth trends.

6

SHEN AND HSU 249