4.4.1 Data
This study will use a panel dataset of 32 Vietnamese banks, including 4 State- owned banks and 28 joint stock commercial banks over the period from 2004 to 2014.
The list and description of variables used are indicated in Table 4.3 and Table 4.4.
Table 4.3: List of sampled banks
No Name of Bank Bank type Data period
1 VietinBank SOCBs 2004-2014
2 VietcomBank SOCBs 2004-2014
3
Bank of Investment and Development of
Vietnam SOCBs 2004-2014
4 AgriBank SOCBs 2004-2014
5 Military Bank JSCBs 2004-2014
6 SacomBank JSCBs 2004-2014
7 Saigon-Hanoi Bank JSCBs 2006-2014
8 Asia Commercial Bank JSCBs 2004-2014
9 TechcomBank JSCBs 2004-2014
10 EximBank JSCBs 2004-2014
11 Vietnam Prosperity Bank JSCBs 2004-2014
12 Saigon Commercial Bank JSCBs 2006-2010,2012-2014
13 Housing Development Bank JSCBs 2006-2014
14 DongA Bank JSCBs 2004-2014
15 Vietnam International Bank JSCBs 2004-2014
16 KienlongBank JSCBs 2004-2014
17 MaritimeBank JSCBs 2006-2014
18 SouthernBank JSCBs 2005-2013
19 SeaBank JSCBs 2004-2014
20 LienVietPost Bank JSCBs 2008-2014
21 AnBinh Bank JSCBs 2009-2014
22 OceanBank JSCBs 2008-2014
23 Mekong Development Bank JSCBs 2008-2014
24 Saigon Bank for Industry and Trade JSCBs 2008-2014
25 Orient Commercial Bank JSCBs 2006-2014
26 National Citizen Bank JSCBs 2005-2014
27 NAM A Bank JSCBs 2004-2014
28 Viet A Bank JSCBs 2004-2014
29 PG Bank JSCBs 2006-2014
30 WesternBank JSCBs 2004-2012
31 Dai A Bank JSCBs 2007-2012
32 Viet Capital Bank JSCBs 2006-2014
Note: SOCBs refers to State-owned Commercial Banks, while JSCBs refers to Joint Stock Commercial Banks
91 Table 4.4: Definition of Variables
Variables Definition Source Expected
sign NPL(-1)
Lagged of the logit transformation of the non-performing loans to total loans ratio
Bank's annual
report (+)
DEBT Net public debt to GDP ratio Ministry of
Finance (+)/(-)
INFSOE Inefficiency of State-owned Enterprises
Vietnam General Statistics Office (+)
GDP The annual growth rate WDI (-)
CRESOE Credit to government and State- owned Enterprises to GDP
Global Financial
Development (+)/(-) SIZE Bank's asset to total banking asset
ratio
Bank's annual
report (+)
ROA Pre-tax profit to total asset ratio Bank's annual
report (-)
COST Cost to income ratio Bank's annual
report (+)/(-)
Dummy Variables
YR2008 Dummy variable for year 2008
(Financial crisis) (+)
YR2011 Dummy variable for year
2011(Housing bubble burst) (+)
YR2013 Dummy variable for year 2013
(Creation of VAMC) (-)
4.4.2 Methodology
Based on the existing literature, the empirical estimation of non-performing loans’
determinants is as followed:
y"# = αy",#()+ βX",#()+ θM#()+ δX",#+ γM#+ 12+ ρ"+ ε"# , α < 1,
Where:
782 is the logit transformation of the NPLs ratio of bank i at time t, (i.e. log()(9:;<9:;< ))
78,2() is the lagged logit transformation of the NPLs ratio of bank i,
X"# is a vector of bank-specific factors
X",#() is a vector of lagged bank-specific controls
M# is a vector of macroeconomic variables
M#() is a vector of lagged macroeconomic factors
92 12 is sets of time dummies
ε"# is disturbance term
ρ" is a bank effect
When analysing the relationship between NPLs and macro- and micro-factors, several issues may arises such as:
- Causality may run in both directions;
- Possibility of autocorrelation between lagged dependent variable and the fixed effect which may result in dynamic panel bias;
- Difference between T (time) and N (number of banks) dimensions in the dataset. There are more banks (N) than years (T).
The OLS technique is not suitable for the estimation of a dynamic panel model because of strict exogeneity of the regressors assumption and the correlation between lagged dependent variable (78,2() ) and the disturbance term (ε"# ) which may causes the OLS estimates to be biased and inconsistent.
The system-generalized method of moments (GMM) estimators as extent developed by Blundell and Bond (1998) are the most appropriate amongst its counterparts to overcome the above problems by instrumenting the predetermined and endogenous variables with their own lags (Bond, 2002; Roodman, 2006; Sarafidis et al., 2006 and Baltagi, 2008). The system-GMM estimators have one-step and two- step variants. Although the two-step system GMM theoretically is considered more efficient than one-step estimates, there is some evidence that two step standard errors tend to be biased downwards in small samples while those for one-step counterpart are effectively unbiased (Arellano and Bond, 1991). In addition, one of the main weaknesses of the system-GMM is the presence of too many instrument sets whose number increases quadratically with T. This potentially causes the two-step system GMM models less reliable for making inference. A large instrument count in system
93 GMM estimates is likely to overfit endogenous variables and weaken the correctness of estimation results (Roodman, 2008).
Therefore, in what follows the study will employ one-step system GMM estimator with robust standard error and the collapsing method as suggested in Roodman (2009) to decrease the instrument count. The lag range used in the instrument matrix is also restricted between lag 2 and lag 4 and in order to ensure that time specific effects do not drive the results; year dummies are included in the study as exogenous instruments.
The validity of instruments is tested using the Hansen test of over-identifying restrictions (Arrelano and Bond, 1991). First-order and second-order serial correlation related to the estimated residuals in first differences is tested using AR(1) and AR(2) statistics. The system GMM estimator is consistent and efficient if there is no second- order serial correlation in the residuals (AR(2) test) and the instrumental variables are valid (Hansen test).