Determinants of non-performing loans in North Macedonia

https://doi.org/10.1080/23311975.2022.2140488
Page 8 of 40

---

Table 
1. 
Descriptive 
Statistics 
NPL, 
UN, 
LGL, 
LGDP, 
INT 
(in 
level)
NPL
UNEMPLOYMENT
LGROSS_LOANS
LGDP
INTEREST
Mean
8.614000
27.15429
12.24755
11.72997
7.546714
Median
9.300000
28.65000
12.33131
11.71979
7.460,000
Maximum
16.10000
38.60000
12.93018
12.24417
11.04000
Minimum
2.980000
14.70000
11.00777
11.04219
4.600000
Std. 
Dev.
3.367931
7.126063
0.506922
0.282783
1.912628
Skewness
0.027600
−0.355073
−0.899559
−0.306561
0.144364
Kurtosis
2.066099
1.820548
3.031769
2.340022
1.775840
Jarque-Bera
2.552717
5.528290
9.443677
2.366847
4.613966
Probability
0.279052
0.063030
0.008899
0.306229
0.099561
Sum
602.9800
1900.800
857.3287
821.0980
528.2700
Sum 
Sq. 
Dev.
782.6643
3503.874
17.73093
5.517661
252.4119
Observations
70
70
70
70
70
Descriptive 
Statistics 
of 
dNPL, 
dUN, 
dLGL, 
dLGDP, 
dINT 
(in 
first 
differences)
DNPL
DUNEMPLOYMENT
DLGROSS 
LOANS
DLGDP
DINTEREST
Mean
−0.190145
−0.346377
0.027861
0.016882
−0.092029
Median
−0.100000
−0.300000
0.019571
0.039183
−0.100000
Maximum
1.400000
1.100000
0.096301
0.209005
0.470,000
Minimum
−3.400000
−1.600000
−0.016765
−0.195558
−0.400000
Std. 
Dev.
0.735016
0.497813
0.025834
0.102661
0.119946
Skewness
−1.225025
0.008340
1.184244
−0.544334
1.538134
Kurtosis
6.992691
3.770606
3.844841
2.308133
10.34149
Jarque-Bera
63.08993
1.708072
18.18003
4.783645
182.1625
Probability
0.000000
0.425693
0.000113
0.091463
0.000000
Sum
−13.12000
−23.90000
1.922411
1.164847
−6.350000
Sum 
Sq. 
Dev.
36.73690
16.85159
0.045383
0.716668
0.978316
Observations
69
69
69
69
69
(Continued
)
Golitsis et al., Cogent Business & Management (2022), 9: 2140488
https://doi.org/10.1080/23311975.2022.2140488
Page 9 of 40

---

Table 
1. 
(Continued
) 
Correlation 
Matrix 
of 
NPL, 
UN, 
LGL, 
LGDP, 
INT
NPL
UNEMPLOYMENT
LGROSS_LOANS
LGDP
INTEREST
NPL
1
0.8039
−0.7432
−0.7713
0.7514
UNEMPLOYMENT
0.8039
1
−0.9036
−0.9391
0.9692
LGROSS_LOANS
−0.7432
−0.9036
1
0.9489
−0.9518
LGDP
−0.7713
−0.9391
0.9489
1
−0.9608
INTEREST
0.7514
0.9692
−0.9518
−0.9608
1
Golitsis et al., Cogent Business & Management (2022), 9: 2140488
https://doi.org/10.1080/23311975.2022.2140488
Page 10 of 40

---

employed also dropped by 102 thousand to 691.4 thousand. Meanwhile, the labour force partici-
pation rate edged down to 55.3% from 56% a year ago (State Statistical Office of the Republic of 
North Macedonia, 2022). Also, there is value to add that the minimum value of NPLs is the most 
recent one, and it is equal to 2.98%, while all variables, excluding gross loans, according to the 
Jarque–Bera statistics, are normally distributed at a 0.05 level of significance.
Finally, the dummy variable was constructed in an attempt to mainly capture the influences of 
increasing migration to EU countries. According to Gujarati (
2009
), and as known, dummy variables 
are generally used for quantifying certain qualitative factors. In terms of our research, the dummy 
variable essentially divides the data sample into migration, not present (0) and migration present 
(1). Following different time periods and various tests, the dummy, assigned with value 1, even-
tually starts from 2007(Q1) and onwards denoting, for example, the entrance of Bulgaria in the EU, 
which, as stated, provided to RNM citizens the ability to use Bulgarian passports. Subsequently, the 
Bulgaria EU entry led to a relatively massive outward migration wave which was further boosted by 
or could be attributed to, the crisis and post-crisis effects within the country. Additionally, the 
dummy variable could be responsible for capturing the overall political changes and the broader 
dissatisfaction of the citizens of the country which further impact factors related to the migration 
decision-making processes.
4
4. Methodology: The bounds testing co-integration approach. An ARDL
Provided that all variables are integrated either of order zero or one (i.e. two and not three levels of 
integration are present, and no variable is I 2
ð Þ
), and once the co-integrating relations are detected 
within the bounds testing (ARDL) approach, we can examine the long-run impact of the selected 
macroeconomic and financial variables to NPLs. This approach is perceived as a better fit to the 
data set when the sample size is rather small, which is in our case (our data span is from 2005 to 
2022). The ARDL on the bounds testing co-integration form, introduced, as stated, by Pesaran et al. 
(
2001
), can be written as such:
Δ
y
t
¼
β
0
þ
∑ β
i
Δ
y
t i
þ
∑ γ
j
Δ
x
1t j
þ
∑ δ
k
Δ
x
2t k
þ
θ
0
y
t 1
þ
θ
1
x
1t 1
þ
θ
2
x
2t 1
þ
e
t
(1) 
where, θ
i 
are the long-run multipliers, β
0 
is the drift, and e
t 
is the white noise error. This form is 
called by Pesaran et al. (
2001
) as a “conditional Error Correction Model (ECM)” and in practice is an 
“unrestricted ECM”, or an “unconstrained ECM”. An F-test will be used to test the following 
hypothesis, i.e., H
0
: θ
0
¼
θ
1
¼
θ
2
¼
0; against the alternative that H
0 
is not true. A rejection of H
0 
implies that a long-run relationship.
On top of that, and apart from the fact that the primary focus of our investigation is to detect 
the various country-specific key factors that NPLs are dependent on, thus allowing us to focus on 
one equation only, the Autoregressive Distributed Lag (ARDL) co-integration approach is chosen for 
the following additional reasons. First, it is a rather simple technique as opposed, for example, to 
multivariate co-integration procedures including the Johansen and Juselius (
1990
) technique, 
given that co-integrating equations can be estimated by OLS once the lag order is identified. 
Second, no pretesting is needed, including the stationarity tests, given that the level of integration 
of the variables could be of any order provided, for the given though bound testing nature of the 
test (i.e., the procedure cannot be performed in the presence of I 2
ð Þ
series; see Ouattara, 
2004
).