See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/328854930

An augmented Autoregressive Distributed Lag bounds test for cointegration

Article  in  Economic Modelling · November 2018

DOI: 10.1016/j.econmod.2018.11.001

CITATIONS READS

9 1,523

3 authors, including:

Chung Yan Sam Soo khoon Goh

Universiti Sains Malaysia Universiti Sains Malaysia
8 PUBLICATIONS   155 CITATIONS    44 PUBLICATIONS   453 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Outward FDI & ASEAN View project

Developed a Bootstrap ARDL for Cointegration View project

All content following this page was uploaded by Soo khoon Goh on 03 June 2019.

The user has requested enhancement of the downloaded file.

 Economic Modelling 80 (2019) 130–141

Contents lists available at ScienceDirect

Economic Modelling
journal homepage: www.journals.elsevier.com/economic-modelling

An augmented autoregressive distributed lag bounds test for cointegration

Chung Yan Sam a, Robert McNown b, Soo Khoon Goh a, *
a
Centre for Policy Research & International Studies, Universiti Sains Malaysia, 11800, Penang, Malaysia
b
University of Colorado, Boulder, CO, 256 UCB, 80309, United States

A R T I C L E I N F O A B S T R A C T

JEL classification numbers: An augmented autoregressive distributed lag (ARDL) bounds test for cointegration involves an extra F-test on the
C12 lagged levels of the independent variable(s) in the ARDL equation. Originally, this testing strategy was introduced
C32 using the bootstrap procedure. This paper provides both the small sample and asymptotic critical values for easier
C63
implementation of the test, making it applicable for a broader range of researchers. Two advantages of this
Keywords: augmented ARDL bounds test are that the assumption of an I(1) dependent variable is not necessary, and a clear
ARDL bounds test
conclusion on the cointegration status is provided by the three tests. The augmented ARDL bounds test is
Cointegration
Degenerate case
demonstrated using an empirical study on government taxation and expenditures. The tests support the tax-and-
Lagged independent variable(s) test spend hypothesis of the budgetary policy for the US, the UK, and France.

1. Introduction make a valid conclusion. Degenerate cases as pointed out by PSS imply
non-cointegration. The degenerate cases arise when either the lagged
In the early 2000s, Pesaran et al. (2001), henceforth PSS, introduced a level of the dependent variable or lagged level(s) of the independent
cointegration testing approach called the autoregressive distributed lag variable(s) in the error correction term are found to be insignificant. The
(ARDL) bounds test. This approach became popular as it breaks the case with insignificance of the lagged level of the dependent variable is
traditional restriction of cointegration tests in that the tested variables known as the degenerate lagged dependent variable case (named degenerate
must be non-stationary and all the variables are integrated of the same case #2 in MSG) while the case of insignificant lagged levels of the in-
order. Some researchers favour this approach as many of the applications dependent variables is known as the degenerate lagged independent vari-
involve economic variables of mixed or unknown order of integration. able(s) case (called degenerate case #1 in MSG). This incomplete error
The conventional cointegration testing restriction, as in the correction term does not close the residual gap between the dependent
Engle-Granger test (1987) or the Johansen test (1991, 1995), raises and independent variable(s) and thus, cointegration does not hold.
problems in conducting cointegration analysis involving mixed orders of In conducting the bounds test, the significance of the overall F-test
variables. In these cases, researchers may either transform the variables suggests the lagged level of the variables are jointly significant. However,
into stationary form, precluding a finding of cointegration, or inappro- this significance of the F-test may arise solely from either the lagged level
priately drop some variables. of the dependent variable or the lagged level of the independent vari-
Nevertheless, PSS made some assumptions in developing the bounds able(s) alone. Therefore, the t-test for the lagged level of the dependent
testing approach. These include the exogeneity of the independent var- variable is needed to rule out the degenerate lagged dependent variable case.
iables, the dependent variable must be I(1), and the absence of degen- The additional assumption that the dependent variable is I(1), rules out
erate cases. However, as pointed out by McNown et al. (2018), MSG the occurrence of the degenerate lagged independent variable(s) case. The
henceforth, these assumptions were sometimes ignored by researchers, reason is that if only the lagged level dependent variable is significant,
possibly leading to misleading conclusions (see Goh and McNown, 2015; then the ARDL equation reduces to a (generalized) Dickey-Fuller equa-
Goh et al., 2017a; Goh et al., 2017b). tion. The significance of this lagged dependent variable term implies that
PSS introduced two tests for cointegration: the overall F-test on all the the dependent variable is I(0).
lagged level variables and the t-test on the lagged level of the dependent Rather than assuming the dependent variable to be I(1), MSG intro-
variable. But these tests need to work with the assumption of I(1) duced an additional test to test on the significance of the lagged levels of
dependent variable to rule out the possibilities of degenerate cases to independent variable(s). The advantage of the additional test is to

* Corresponding author.
E-mail addresses: skgoh@usm.my, sookhoongoh@gmail.com (S.K. Goh).

https://doi.org/10.1016/j.econmod.2018.11.001
Received 2 June 2018; Received in revised form 27 October 2018; Accepted 1 November 2018
Available online 10 November 2018
0264-9993/© 2018 Elsevier B.V. All rights reserved.
C.Y. Sam et al. Economic Modelling 80 (2019) 130–141

overcome the reliance on the assumption of an I(1) dependent variable to 2. The asymptotic distribution
rule out the degenerate case. This reduces the risk of false conclusions
based on standard unit root tests that have low power. This is also PSS introduced 5 Cases in describing different specifications of the
consistent with the spirit of ARDL bounds testing that makes minimal ARDL model, and these cases are restated in the Appendices. Considering
assumptions about the orders of integration of the variables in the the most general PSS Case V (unrestricted intercepts and trends, the
analysis. single ECM (error correction model) equation is written as
By combining this new test with the two tests presented by PSS, we
gain a complete picture of the cointegration status of the system. If all the X
p1
0 0
Δyt ¼ c0 þ c1 t þ π yy yt1 þ πyx:x xt1 þ ψi Δzti þ ω Δxt þ ut ; (1)
three tests (overall F-test on lagged level variables, t-test on the lagged i¼1
level of the dependent variable, and F-test on the lagged levels of the
0
independent variable(s)) are found to be significant, we can conclude 0
where zt ¼ ðyt ; xt Þ , π yy and πyx:x are the coefficients defining the coin-
that there is cointegration. If the overall F-test and the t-test on the lagged tegrating relation between yt and xt . The π yy and πyx:x may be expressed
dependent variable are found to be significant but not the test on the 0 0
as αyy βyy and αyx:x βyx:x respectively, where αyy , βyy and αyx:x , βyx:x are (1,
lagged independent variable(s), this indicates that it is a degenerate lagged
independent variable(s) case. Another possibility is when the overall F-test 1) and (k, r) matrices of full column rank, respectively. The k is the
and the test on the lagged level of the independent variable(s) are found dimension of the forcing variables and r is the forcing variable cointe-
to be significant but not for the t-test on the lagged level of the dependent gration rank. PSS presented two theorems establishing the limiting null
variable. This falls into the case of a degenerate lagged dependent variable. distributions of the Wald statistic for the significance of all lagged vari-
Either of the degenerate cases will imply a case of no cointegration; ables and the t-statistic on the lagged level of the dependent variable.
therefore all three tests must be applied to reach a valid conclusion. Theorem 1. states, if Assumptions 1–4 and 5a hold (these assumptions are
To conduct the test on the lagged levels of the independent vari- 0
restated in the Appendices), then under H0 : π yy ¼ 0 and πyx:x ¼ 0 , as
able(s), MSG used a bootstrap procedure to generate its critical values.
T → ∞, the asymptotic distribution of the Wald statistic W has the
However, this involves programming and computation that are not
representation
convenient or user friendly. Based on theorems that establish the limiting
distributions of this test statistic, this paper provides tables of critical 0 11
Z 1 Z 1
0 0
values for the test on the lagged levels of the independent variables. By dWu ðaÞFkrþ1 ðaÞ @ Fkrþ1 ðaÞFkrþ1 ðaÞ daA
0
W ⇒ zr zr þ
providing the tables of critical values, this eases the implementation of 0 0
(2)
the test, so that it becomes accessible to a broader range of researchers. Z 1
This is the primary contribution of the paper. This contribution will be of Fkrþ1 ðaÞdWu ðaÞ
0
use to those wishing to apply the widely employed ARDL procedure with
the three tests combined to arrive at a clear conclusion on the status of where zr  Nð0; Ir Þ is distributed independently of the second term (2), W is a
cointegration. Wiener process and
Many researchers used the ARDL bounds test as a cointegration test,
8 9
for example, Muscatelli and Spinelli (2000); Rushdi et al. (2012); Jalil >
> Wkrþ1 ðaÞ Case I >>
>
>   >
>
et al. (2013); Kim and Baek (2013); and Asteriou et al. (2016). Muscatelli >
> >
Case II >
0
>
> Wkrþ1 ðaÞ ; 1 >
>
and Spinelli adopted the ARDL bounds test to examine the stability of the >
> >
>
>
> >
>
demand for money in Italy; Rushdi et al. investigate the long run rela- < ~ krþ1 ðaÞ =
Fkrþ1 ðaÞ ¼ W Case III
tionship between real stock returns and inflation in Australia; Jalil et al. >  0 >
>
> >
>
>
> ~ krþ1 ðaÞ ; a  1 Case IV >
>
0
study the long run relationship between tourism and economic growth in > W >
>
> 2 >
>
Pakistan; Kim and Baek re-examined the relationship among volume of >
> >
>
>
> >
>
crude oil imports, the price of imported crude oil and economic growth in : b krþ1 ðaÞ ;
W Case V
Korea; and Asteriou et al. examined the effect of exchange rate volatility
on international trade volumes in Mexico, Indonesia, Nigeria and Turkey. r ¼ 0; :::; k, a 2 ½0; 1. Note that the limiting distribution of the Wald statistic
Many of these studies carry important economic policy implications. W depends on the dimension and cointegration rank of the forcing variables
Hence, it is imperative that researchers adopt a more robust version of 0
fxt g, k and r respectively. Referring to (2), the first component zr zr  χ 2 ðrÞ
the augmented ARDL test in order to avoid flawed results or give corresponds to testing for the exclusion of the r-dimensional stationary vector
incorrect conclusions. The accuracy and robustness of the results are 0
βxx xt1 whereas the second term is a non-standard Dickey-Fuller unit-root
important to provide sound recommendations to policy makers and in-
distribution, corresponding to the test for the exclusion of the ðk  r þ
stitutions for efficient and effective decisions on government policies and 0

for further analysis of policies. 1Þ-dimensional I(1) vector ðβ? ?

y ; β Þ zt1 . The theorem above shows the gen-
In this paper, besides providing the tables of critical values of the F- eral limiting distribution of W. PSS derived two polar cases when fxt g are I(0)
test on the lagged levels of the independent variable(s), the limiting and I(1):
distribution of the test statistic is discussed as well. Theorems are given in
Corollary 1. (Limiting distribution of W if fxt g  Ið0Þ.) If Assumptions
deriving the limiting distribution. Lastly, we illustrate the use of the 0
1–4 and 5a hold, then under H0 : π yy ¼ 0 and πyx:x ¼ 0 , as T → ∞, the
generated critical values in the augmented ARDL bounds testing frame-
limiting distribution of the Wald statistic W has the representation
work by revisiting the spend-tax relationship of the US, the UK and
France.  2
R1
This article is constructed as follows. Section 2 discusses the limiting 0
FðaÞdW u ðaÞ
0
distribution of the augmented test on the lagged level of the independent W ⇒ zk zk þ R1 (3)
variable(s). Section 3 explains the simulation setup and presents the ta- 0
FðaÞ2 da
bles of the critical values. Section 4 demonstrates the use of the
augmented ARDL testing framework by revisiting the spend-tax rela- where zk  Nð0; Ik Þ is distributed independently of the second term (3).
tionship of the US, the UK and France. Finally, Section 5 concludes the Corollary 2. (Limiting distribution of W if fxt g  Ið1Þ.) If Assumptions
study. 0
1–4 and 5a hold, then under H0 : π yy ¼ 0 and πyx:x ¼ 0 , as T → ∞, the
limiting distribution of the Wald statistic W has the representation

131
C.Y. Sam et al. Economic Modelling 80 (2019) 130–141

0 11
Z 1 Z 1 Z 1
3. Simulation setup and tables of critical values
0 0
W⇒ dWu ðaÞFkþ1 ðaÞ @ Fkþ1 ðaÞFkþ1 ðaÞ daA Fkþ1 ðaÞdWu ðaÞ:
0 0 0 A simulation is used to generate critical values and its setup is similar
(4) to PSS' setting. An I(1) dependent variable, yt , is generated from yt ¼
yt1 þ ε1t and I(1) or I(0) independent variables, xt , are generated from
0
Notice from Corollary 1, when fxt g  Ið0Þ, r ¼ k, the first term in (3) 0
xt ¼ Pxt1 þ ε2t , t ¼ 1;:::;T, where y0 ¼ 0, x0 ¼ 0 and εt ¼ ðε1t ; ε2t Þ are
becomes k-dimension chi-squared and this comes from the k-dimensional drawn as ð1 þ kÞ independent standard normal variables. The P matrix
stationary fxt g, whereas the second term in (3) corresponds to fyt g as determines the integration order of the generated xt . If xt is purely I(1),
I(1). For Corollary 2, when fxt g  Ið1Þ, r ¼ 0 and all the variables fyt g P ¼ Ik whereas P ¼ 0 if xt is purely I(0). Following the PSS setup, the
and fxt g are I(1), thus, W is distributed as a k þ 1-dimensional non- stochastic simulation uses T ¼ 1000 (sample size for the asymptotic case)
standard Dickey-Fuller unit-root distribution. 0
and N ¼ 40,000 replications for the F-statistic for testing δ ¼ 0 in the
0 0
Based on the above theorem and corollaries, we derive the limiting regression Δyt ¼ ϕyt1 þ δ xt1 þ a wt þ ξt , t ¼ 1; :::; T where xt ¼
0
distribution of the Wald test on the lagged levels of the independent ðx1t ; …; xkt Þ . The choice of wt depends on the considered cases and it is
variables. To test for the exclusion of xt1 from (1), define the null hy- 0
treated as wt ¼ 0, wt ¼ 1, wt ¼ ð1; tÞ for Cases I, III, and V respectively.
pothesis, H0 : πyx:x ¼ 0, and the alternative hypothesis, H1 : πyx:x 6¼ 0.
Two sets of critical value bounds are generated covering all the
Theorem 2. (Limiting distribution of W on the lagged levels of the inde- possible classifications of fxt g  Ið0Þ, fxt g  Ið1Þ and mutually cointe-
pendent variables). If Assumptions 1–4 and 5a hold, then under H0 : πyx:x ¼ grated processes for sizes 0.100, 0.025, 0.050 and 0.010 as summarised
0, as T → ∞, the limiting distribution of the Wald statistic W has the in Tables 1–3 below. The lower bound values in column I(0) require that
representation the forcing variables be purely I(0); whereas the upper bound values in
column I(1) assume that the forcing variables are purely I(1). Besides the
0
W ⇒ zr zr asymptotic case, small sample critical values ranging from sizes 30 to 80
0 11 are considered as well, following Narayan's (2005) strategy in generating
Z 1 Z 1 Z 1
0 0 the small sample critical values for the overall F-statistic on lagged level
þ dWu ðaÞFkr ðaÞ @ Fkr ðaÞFkr ðaÞ daA Fkr ðaÞdWu ðaÞ
0 0 0 variables.
The ‘I(0)’ columns refer to the lower bounds of the critical values
(5)
obtained when xt is purely I(0), while the ‘I(1)’ columns refer to the
where zr  Nð0; Ir Þ is distributed independently of the second term (5) and upper bounds obtained when xt is purely I(1). ∞ indicates the asymptotic
cases, approximated with sample sizes of 1000. If a test statistic exceeds
8 9
< Wkr ðaÞ Case I = the upper bound, the test is significant and the null hypothesis is rejected;
Fkr ðaÞ ¼ W~ ðaÞ Case III the null hypothesis is not rejected if the test statistic is less than the lower
: b kr ;
W kr ðaÞ Case V bound. The test is inconclusive if it falls between the bounds.

r ¼ 0; :::; k, a 2 ½0; 1. Similar to the test on the lagged level of the dependent 4. Empirical application: revisiting the US, the UK, and France
variable, the restrictions on the deterministic coefficients are ignored in Cases II spend-tax correlations
and IV. Thus, these two cases are subsumed under Case III and Case V
respectively (see PSS p.299). Two polar cases are considered i.e. when fxt g is Following the discussion of the test on the lagged levels of the inde-
I(0) and I(1) in the following corollaries: pendent variable(s), this section provides a demonstration of the three-
Corollary 3. (Limiting distribution of W if fxt g  Ið0Þ.) If Assumptions part augmented ARDL bounds test by revisiting the government spend-
0 tax relations of the US, the UK and France. The relationship between
1–4 and 5a hold, then under πyx:x ¼ 0 , as T → ∞, the limiting distribution of
government revenue and government expenditures has been studied over
the Wald statistic W has the representation
the years. It is important to understand this relationship as it is related to
0
W ⇒ zk zk (6) the generation of government fiscal deficits and sustainability, and thus,
strategies for government policies and their implications. There is sub-
where zk  Nð0; Ik Þ. stantial literature that discusses the relationship. However, the findings
are mixed, from country to country or within countries over different
Corollary 4. (Limiting distribution of W if fxt g  Ið1Þ.) If Assumptions time periods.
0
1–4 and 5a hold, then under πyx:x ¼ 0 , as T → ∞, the limiting distribution of There are four main hypotheses concerning spend-tax behavior. The
the Wald statistic W has the representation tax-and-spend hypothesis was promoted by Friedman (1978). This theory
0 11 hypothesizes unidirectional causality running from tax revenues to
Z 1 Z 1 Z 1 government expenditures. Friedman explained that increasing taxes will
0 0
W⇒ dWu ðaÞFk ðaÞ @ Fk ðaÞFk ðaÞ daA Fk ðaÞdWu ðaÞ: (7) not reduce fiscal deficits, as it will result in more spending. However,
0 0 0
cutting taxes will eventually lead to spending cuts. From the reverse
perspective, Peacock and Wiseman (1961, 1979) argued for the
spend-and-tax hypothesis, i.e. unilateral causality running from govern-
Remark 1. This test involves only the lagged level of the independent ment spending to revenues. They explained that this phenomenon is due
variables. Referring to Corollary 3 above, when fxt g  Ið0Þ, W is based to the occurrence of a crisis that initially forces government spending to
only on the k-dimensional stationary vector βxx xt1 . In this case, as
0 increase. Taxes are then imposed for the adjustment to the desired level
0 of spending. The third hypothesis, fiscal synchronization, explains the
T → ∞, W approaches to zk zk which has a limiting chi-squared distri-
situation of bilateral causality between spending and taxation. Musgrave
bution with k-degrees of freedom.
(1966) and Meltzer and Richard (1981) argued that the governments
Remark 2. Following Corollary 4, when fxt g  Ið1Þ and as T → ∞, W change their spending and taxation levels concurrently based on an
is distributed as a k-dimensional non-standard Dickey-Fuller unit-root optimal spending program. Finally, the institutional separation hypoth-
distribution. esis, raised by Baghestani and McNown (1994), describes a situation with
no long run causal relation between spending and taxation. They argued
that this happens because the allocation of the spending is taken

132
C.Y. Sam et al. Economic Modelling 80 (2019) 130–141

Table 1
Case I (No intercepts; no trends).
N k¼1 k¼2 k¼3 k¼4 k¼5 k¼6 k¼7

I(0) I(1) I(0) I(1) I(0) I(1) I(0) I(1) I(0) I(1) I(0) I(1) I(0) I(1)

p ¼ 0.010

30 7.79 11.61 5.52 8.45 4.59 7.21 4.13 6.67 3.92 6.24 3.69 6.12 3.62 5.83
35 7.48 11.31 5.38 8.04 4.53 7.02 4.04 6.33 3.79 5.94 3.52 5.71 3.39 5.59
40 7.52 11.18 5.32 8.08 4.38 6.78 3.94 6.26 3.63 5.89 3.39 5.57 3.24 5.37
45 7.21 10.98 5.19 7.86 4.38 6.61 3.84 6.12 3.57 5.60 3.35 5.34 3.19 5.22
50 7.03 11.01 5.06 7.96 4.26 6.72 3.80 6.00 3.47 5.52 3.22 5.24 3.10 4.95
55 7.05 11.00 4.94 7.73 4.23 6.50 3.70 5.87 3.43 5.50 3.22 5.15 3.04 4.96
60 7.17 10.72 4.95 7.64 4.13 6.47 3.71 5.78 3.35 5.43 3.13 5.13 2.96 4.93
65 7.12 10.52 4.99 7.64 4.07 6.40 3.67 5.78 3.34 5.33 3.13 5.13 2.95 4.85
70 6.82 10.55 4.96 7.49 4.11 6.45 3.60 5.74 3.32 5.35 3.16 5.02 2.97 4.81
75 7.05 10.74 4.90 7.53 4.06 6.42 3.63 5.68 3.31 5.33 3.10 4.98 2.92 4.78
80 6.87 10.79 4.92 7.54 3.98 6.23 3.56 5.62 3.26 5.24 3.02 5.06 2.91 4.75
∞ 6.65 10.00 4.65 7.13 3.75 5.93 3.37 5.35 3.05 4.85 2.84 4.55 2.65 4.31

p ¼ 0.025

30 5.65 8.90 4.31 6.77 3.67 5.93 3.34 5.46 3.19 5.17 3.01 5.07 2.94 4.96
35 5.45 8.82 4.17 6.58 3.57 5.75 3.27 5.31 3.09 5.03 2.92 4.82 2.82 4.78
40 5.46 8.67 4.12 6.52 3.50 5.70 3.20 5.19 2.99 4.93 2.79 4.74 2.73 4.57
45 5.34 8.58 4.04 6.43 3.49 5.55 3.13 5.10 2.92 4.75 2.78 4.59 2.64 4.49
50 5.28 8.62 3.99 6.40 3.40 5.53 3.10 5.08 2.91 4.73 2.71 4.51 2.62 4.35
55 5.23 8.62 3.92 6.25 3.38 5.47 3.02 4.97 2.83 4.72 2.71 4.45 2.56 4.30
60 5.37 8.44 3.90 6.27 3.32 5.40 3.03 4.94 2.81 4.64 2.64 4.44 2.52 4.27
65 5.18 8.30 3.93 6.27 3.32 5.42 3.02 4.95 2.78 4.60 2.65 4.39 2.49 4.24
70 5.20 8.31 3.88 6.19 3.33 5.38 2.99 4.90 2.77 4.61 2.63 4.38 2.50 4.15
75 5.26 8.48 3.90 6.21 3.31 5.41 2.98 4.88 2.78 4.54 2.61 4.30 2.49 4.19
80 5.19 8.40 3.89 6.28 3.24 5.27 2.94 4.81 2.73 4.51 2.58 4.44 2.48 4.14
∞ 5.10 8.05 3.73 5.99 3.12 5.05 2.82 4.57 2.58 4.27 2.44 4.00 2.30 3.86

p ¼ 0.050

30 4.19 7.07 3.40 5.59 3.00 4.96 2.75 4.61 2.62 4.41 2.53 4.34 2.48 4.22
35 4.08 6.93 3.32 5.47 2.92 4.84 2.71 4.54 2.57 4.30 2.48 4.18 2.37 4.09
40 4.09 6.87 3.29 5.38 2.86 4.84 2.65 4.45 2.51 4.26 2.39 4.09 2.32 3.95
45 3.99 6.81 3.23 5.35 2.85 4.70 2.61 4.36 2.45 4.13 2.37 4.02 2.27 3.89
50 3.98 6.78 3.21 5.31 2.81 4.68 2.57 4.35 2.45 4.12 2.31 3.97 2.25 3.84
55 4.01 6.79 3.15 5.26 2.76 4.68 2.54 4.30 2.40 4.09 2.30 3.91 2.20 3.78
60 4.06 6.70 3.11 5.26 2.73 4.62 2.53 4.25 2.40 4.06 2.27 3.88 2.18 3.77
65 3.99 6.71 3.15 5.25 2.74 4.63 2.51 4.27 2.36 4.02 2.27 3.86 2.16 3.76
70 3.93 6.66 3.14 5.24 2.73 4.61 2.51 4.25 2.35 4.02 2.25 3.85 2.16 3.68
75 4.00 6.79 3.12 5.19 2.73 4.61 2.50 4.23 2.36 3.98 2.25 3.81 2.15 3.71
80 3.93 6.75 3.13 5.23 2.71 4.55 2.49 4.18 2.34 3.97 2.23 3.91 2.15 3.67
∞ 3.87 6.57 3.02 5.05 2.61 4.41 2.39 4.01 2.22 3.81 2.12 3.60 2.02 3.47

p ¼ 0.100

30 2.88 5.13 2.53 4.42 2.32 3.99 2.18 3.78 2.12 3.67 2.04 3.60 2.01 3.53
35 2.88 5.07 2.49 4.32 2.28 3.96 2.15 3.76 2.08 3.59 2.01 3.51 1.95 3.46
40 2.83 5.05 2.48 4.28 2.24 3.92 2.12 3.71 2.05 3.57 1.96 3.47 1.92 3.38
45 2.79 5.07 2.45 4.28 2.21 3.87 2.09 3.65 2.00 3.49 1.94 3.42 1.89 3.32
50 2.78 5.07 2.42 4.22 2.21 3.83 2.07 3.63 1.99 3.47 1.91 3.38 1.86 3.30
55 2.77 5.08 2.38 4.22 2.19 3.88 2.03 3.60 1.97 3.48 1.90 3.37 1.84 3.28
60 2.84 5.03 2.37 4.21 2.17 3.82 2.04 3.59 1.97 3.43 1.89 3.34 1.82 3.25
65 2.77 5.05 2.41 4.22 2.17 3.82 2.03 3.58 1.94 3.42 1.88 3.32 1.82 3.24
70 2.77 5.00 2.39 4.21 2.15 3.82 2.04 3.57 1.95 3.42 1.87 3.31 1.81 3.22
75 2.79 5.07 2.40 4.16 2.16 3.80 2.03 3.56 1.94 3.39 1.88 3.29 1.81 3.21
80 2.80 5.00 2.37 4.18 2.16 3.77 2.02 3.53 1.94 3.40 1.85 3.36 1.81 3.21
∞ 2.72 5.00 2.32 4.09 2.09 3.69 1.95 3.46 1.85 3.30 1.77 3.17 1.72 3.06

independently from the decisions of taxation. This theory implies that establish the existence and directions of Granger-causality, the cointe-
large, uncontrolled budget deficits can arise as the taxation and expen- gration test serves as an important indicator of budget sustainability
diture decisions may be inconsistent and unsustainable. through the existence of a long-run relationship between expenditures
and revenues. Both the short-run causality and long-run cointegration
relationships should be studied to reveal the government budget policy
4.1. Analytical framework
status. These perspectives can be presented within the ARDL framework.
In this empirical study, the augmented ARDL bounds test, is adopted to
Various econometric techniques have been used to study the causality
analyze the spending and taxation relationship. Note that the ARDL
relation between government taxation and government spending. These
bounds test is suitable for cases when only one variable is endogenous.
include panel estimation, cross-sectional analysis, and cointegration
According to PSS, the test is valid only under the assumption of exoge-
tests. As noted, different methods give different outcomes. Nevertheless,
neity of the independent variables. However, based on MSG's investiga-
recent studies commonly favour the cointegration approach. As claimed
tion, violation of this exogeneity assumption does not adversely affect the
by Fasano and Wang (2002), cointegration or error-correction frame-
ARDL test performance. Therefore, we proceed with this
work provides a more comprehensive test of causality. In addition, to

133
C.Y. Sam et al. Economic Modelling 80 (2019) 130–141

Table 2
Case III (Unrestricted intercepts; no trends).
N k¼1 k¼2 k¼3 k¼4 k¼5 k¼6 k¼7

I(0) I(1) I(0) I(1) I(0) I(1) I(0) I(1) I(0) I(1) I(0) I(1) I(0) I(1)

p ¼ 0.010

30 7.70 12.34 5.41 9.16 4.60 7.72 4.15 6.83 3.94 6.48 3.72 6.16 3.62 6.12
35 7.48 12.31 5.35 8.75 4.51 7.23 4.03 6.63 3.70 6.15 3.58 5.82 3.42 5.77
40 7.43 12.40 5.14 8.58 4.35 7.24 3.90 6.39 3.71 6.02 3.37 5.74 3.23 5.44
45 7.41 12.13 5.06 8.38 4.34 7.06 3.80 6.33 3.53 5.83 3.33 5.52 3.19 5.36
50 7.36 12.05 5.14 8.24 4.29 7.07 3.69 6.17 3.46 5.81 3.31 5.46 3.08 5.20
55 7.04 12.09 5.05 8.31 4.18 6.89 3.75 6.19 3.42 5.68 3.22 5.36 3.06 5.12
60 7.03 11.84 5.03 8.10 4.24 6.83 3.68 6.09 3.38 5.63 3.20 5.26 3.05 5.03
65 7.22 11.85 4.99 8.24 4.12 6.76 3.68 5.98 3.38 5.58 3.07 5.19 2.98 4.93
70 6.89 11.69 4.97 8.09 4.13 6.79 3.60 5.86 3.32 5.50 3.11 5.15 2.98 4.92
75 6.96 11.70 4.90 8.08 4.13 6.79 3.66 5.97 3.29 5.47 3.09 5.10 2.93 4.86
80 6.98 11.49 4.98 8.02 4.09 6.71 3.61 5.87 3.28 5.33 3.02 5.07 2.94 4.86
∞ 6.53 11.05 4.71 7.68 3.83 6.33 3.33 5.47 3.05 5.02 2.84 4.69 2.62 4.42

p ¼ 0.025

30 5.61 9.56 4.25 7.24 3.64 6.22 3.36 5.60 3.19 5.33 3.04 5.18 2.96 5.06
35 5.47 9.52 4.12 6.96 3.59 6.04 3.26 5.50 3.06 5.15 2.95 4.89 2.81 4.81
40 5.52 9.48 4.04 6.94 3.49 5.92 3.18 5.35 3.00 5.08 2.79 4.81 2.73 4.62
45 5.46 9.30 3.99 6.78 3.46 4.34 3.11 5.34 2.92 4.94 2.77 4.70 2.67 4.54
50 5.36 9.28 4.04 6.74 3.45 5.87 3.04 5.22 2.90 4.92 2.75 4.64 2.60 4.48
55 5.24 9.30 3.95 6.76 3.37 5.74 3.05 5.21 2.87 4.85 2.68 4.57 2.59 4.41
60 5.25 9.19 3.94 6.67 3.38 5.75 3.01 5.13 2.79 4.79 2.66 4.56 2.57 4.33
65 5.35 9.29 3.90 6.72 3.31 5.63 3.05 5.08 2.80 4.74 2.63 4.50 2.51 4.31
70 5.12 9.16 3.96 6.65 3.35 5.64 2.98 5.03 2.79 4.72 2.61 4.47 2.51 4.28
75 5.24 9.04 3.88 6.67 3.30 5.60 3.02 5.10 2.76 4.70 2.63 4.41 2.50 4.26
80 5.23 9.03 3.87 6.65 3.31 5.60 2.98 5.00 2.75 4.60 2.59 4.43 2.48 4.25
∞ 4.95 8.84 3.74 6.42 3.15 5.40 2.79 4.73 2.59 4.41 2.43 4.14 2.29 3.97

p ¼ 0.050

30 4.22 7.51 3.36 5.90 2.96 5.14 2.80 4.70 2.65 4.54 2.53 4.37 2.47 4.30
35 4.19 7.42 3.31 5.64 2.93 5.05 2.70 4.67 2.59 4.40 2.47 4.22 2.40 4.15
40 4.10 7.41 3.26 5.68 2.85 4.97 2.64 4.54 2.49 4.36 2.37 4.15 2.32 4.02
45 4.11 7.33 3.21 5.62 2.81 4.94 2.60 4.55 2.46 4.24 2.36 4.09 2.26 3.96
50 4.03 7.29 3.22 5.62 2.83 4.94 2.55 4.49 2.45 4.25 2.32 4.03 2.24 3.93
55 4.01 7.36 3.13 5.54 2.79 4.92 2.55 4.51 2.42 4.17 2.29 4.02 2.22 3.86
60 4.00 7.29 3.15 5.56 2.78 4.88 2.54 4.42 2.37 4.18 2.28 3.98 2.19 3.84
65 4.08 7.28 3.14 5.59 2.74 4.85 2.55 4.38 2.36 4.13 2.26 3.95 2.17 3.82
70 3.90 7.22 3.14 5.54 2.75 4.80 2.51 4.37 2.37 4.13 2.25 3.91 2.16 3.79
75 4.01 7.17 3.12 5.50 2.74 4.79 2.54 4.40 2.36 4.12 2.25 3.90 2.15 3.76
80 3.98 7.09 3.12 5.54 2.73 4.77 2.51 4.39 2.34 4.05 2.23 3.90 2.14 3.75
∞ 3.79 7.21 3.01 5.42 2.62 4.65 2.39 4.18 2.24 3.90 2.12 3.72 2.01 3.58

p ¼ 0.100

30 2.88 5.34 2.53 4.54 2.30 4.11 2.22 3.84 2.12 3.76 2.06 3.62 2.02 3.55
35 2.85 5.31 2.48 4.43 2.30 4.04 2.14 3.82 2.07 3.67 2.00 3.52 1.95 3.48
40 2.84 5.37 2.45 4.46 2.23 4.03 2.11 3.76 2.03 3.63 1.95 3.49 1.92 3.39
45 2.87 5.30 2.41 4.43 2.20 3.97 2.08 3.76 2.00 3.58 1.94 3.46 1.88 3.38
50 2.82 5.29 2.40 4.44 2.23 3.99 2.06 3.74 2.00 3.57 1.90 3.42 1.86 3.34
55 2.83 5.36 2.38 4.37 2.20 4.02 2.06 3.72 1.99 3.54 1.90 3.41 1.85 3.31
60 2.80 5.31 2.41 4.40 2.19 3.99 2.05 3.67 1.95 3.55 1.89 3.41 1.83 3.31
65 2.86 5.34 2.42 4.37 2.15 3.94 2.05 3.68 1.96 3.51 1.87 3.39 1.83 3.28
70 2.73 5.32 2.38 4.43 2.17 3.91 2.04 3.67 1.95 3.49 1.86 3.36 1.82 3.25
75 2.78 5.24 2.37 4.34 2.16 3.94 2.04 3.68 1.92 3.51 1.87 3.37 1.80 3.26
80 2.80 5.24 2.36 4.36 2.16 3.93 2.04 3.66 1.92 3.45 1.86 3.35 1.80 3.25
∞ 2.67 5.31 2.31 4.33 2.08 3.86 1.96 3.58 1.86 3.39 1.78 3.25 1.72 3.15

taxation-spending example to illustrate the augmented ARDL bounds revenues excluding grants (current local currency units) and deflatort is
tests, without concerns over the possible endogeneity of both variables. the GDP deflator (ratio of current local currency to base year local cur-
rency). The source of data is the World Bank Database. The time series
4.2. Data and unit root tests plots of LREXP and LRREV for the US, the UK and France as shown in
Fig. 1 show that all these series have rising trends, suggesting that a linear
In investigating the spend and tax relationship, the annual data on the trend should be included in the unit root tests. As the ARDL bounds test
logarithm of real government consumption expenditures (LREXP) and only allows for time series to be either I(0) or I(1), a series of unit root
the logarithm of real tax revenues (LRREV), for the period of 1972–2015 tests were conducted to ensure the testing variables are not of integration
are used. The LREXP is computed as logðgovernment expendituret order higher than one.
=deflatort Þ and the LRREV is given by logðtax revenuet =deflatort Þ where Table 4 summarizes the results for two unit-root tests: the Augmented
Dickey-Fuller (ADF) test and the Phillips and Perron (PP) test. The
expendituret is general government final consumption expenditures but
implementation procedures for these tests included intercept and trend
excluding government military expenditures that are part of government
components in the test equation for testing the level of the series, and
capital formation (local currency units), revenuet is government tax

134
C.Y. Sam et al. Economic Modelling 80 (2019) 130–141

Table 3
Case V (Unrestricted intercepts; unrestricted trends).
n k¼1 k¼2 k¼3 k¼4 k¼5 k¼6 k¼7

I(0) I(1) I(0) I(1) I(0) I(1) I(0) I(1) I(0) I(1) I(0) I(1) I(0) I(1)

p ¼ 0.010

30 7.82 12.57 5.46 9.05 4.67 7.58 4.30 7.01 4.00 6.66 3.77 6.25 3.70 6.00
35 7.65 12.51 5.33 8.92 4.54 7.59 4.02 6.69 3.87 6.26 3.58 5.91 3.43 5.69
40 7.46 12.21 5.25 8.82 4.42 7.32 3.94 6.55 3.69 5.96 3.46 5.68 3.33 5.45
45 7.33 12.39 5.14 8.79 4.34 7.17 3.82 6.37 3.59 5.91 3.37 5.61 3.20 5.33
50 7.19 12.12 5.12 8.62 4.27 6.98 3.76 6.19 3.52 5.72 3.29 5.44 3.10 5.15
55 7.13 12.11 5.00 8.43 4.13 7.04 3.71 6.19 3.42 5.67 3.23 5.34 3.05 5.13
60 7.04 12.22 5.16 8.31 4.13 6.99 3.67 6.08 3.42 5.67 3.16 5.33 3.01 5.07
65 7.08 11.85 5.06 8.28 4.14 6.86 3.64 6.11 3.37 5.62 3.17 5.25 2.97 5.01
70 7.03 11.80 4.89 8.38 4.08 6.70 3.66 5.99 3.28 5.52 3.10 5.20 2.94 5.04
75 7.07 12.22 4.92 8.28 4.06 6.83 3.61 6.00 3.30 5.51 3.04 5.24 2.95 4.91
80 7.10 12.05 4.90 8.30 4.02 6.77 3.56 5.97 3.28 5.50 3.00 5.17 2.91 4.86
∞ 6.65 11.76 4.58 7.94 3.86 6.46 3.40 5.68 3.05 5.12 2.81 4.79 2.68 4.54

p ¼ 0.025

30 5.55 9.51 4.22 7.12 3.17 6.16 3.37 5.67 3.24 5.37 3.05 5.12 3.02 5.01
35 5.56 9.52 4.18 7.05 3.58 6.10 3.30 5.47 3.12 5.16 2.94 4.91 2.83 4.77
40 5.48 9.29 4.05 7.02 3.57 5.98 3.16 5.39 3.01 4.99 2.85 4.79 2.78 4.64
45 5.39 9.40 4.03 6.86 3.48 5.85 3.15 5.26 2.96 4.93 2.80 4.74 2.66 4.57
50 5.36 9.31 4.03 6.84 3.42 5.76 3.09 5.18 2.94 4.89 2.73 4.63 2.62 4.42
55 5.32 9.27 3.96 6.80 3.38 5.82 3.05 5.23 2.83 4.80 2.73 4.57 2.57 4.38
60 5.28 9.45 3.98 6.74 3.36 5.76 3.02 5.11 2.84 4.81 2.67 4.58 2.55 4.39
65 5.30 9.19 4.03 6.74 3.34 5.73 3.01 5.11 2.82 4.77 2.65 4.49 2.51 4.33
70 5.30 9.21 3.84 6.82 3.28 5.72 3.01 5.07 2.77 4.74 2.62 4.50 2.49 4.28
75 5.26 9.53 3.92 6.74 3.31 5.76 3.00 5.09 2.78 4.76 2.58 4.50 2.50 4.25
80 5.20 9.23 3.84 6.76 3.26 5.66 2.96 5.06 2.74 4.76 2.57 4.45 2.48 4.24
∞ 5.02 9.19 3.67 6.53 3.17 5.45 2.81 4.94 2.60 4.51 2.41 4.23 2.31 4.04

p ¼ 0.050

30 4.22 7.33 3.35 5.77 2.97 5.09 2.74 4.71 2.67 4.49 2.56 4.34 2.54 4.24
35 4.10 7.32 3.30 5.72 2.93 5.05 2.71 4.61 2.57 4.40 2.46 4.18 2.41 4.09
40 4.15 7.09 3.22 5.66 2.90 4.96 2.63 4.54 2.51 4.24 2.42 4.11 2.34 4.00
45 4.06 7.29 3.24 5.62 2.84 4.92 2.62 4.51 2.49 4.27 2.37 4.08 2.29 3.98
50 4.05 7.16 3.20 5.54 2.82 4.86 2.58 4.44 2.45 4.20 2.32 4.03 2.25 3.86
55 3.99 7.22 3.20 5.55 2.79 4.85 2.55 4.47 2.40 4.19 2.32 3.98 2.23 3.85
60 4.03 7.32 3.17 5.59 2.77 4.82 2.52 4.42 2.40 4.17 2.28 3.98 2.19 3.84
65 3.98 7.25 3.23 5.55 2.76 4.82 2.53 4.39 2.40 4.12 2.28 3.95 2.16 3.81
70 3.99 7.22 3.10 5.58 2.74 4.82 2.53 4.40 2.37 4.14 2.25 3.94 2.16 3.77
75 4.02 7.32 3.13 5.62 2.73 4.86 2.53 4.38 2.37 4.11 2.24 3.94 2.15 3.76
80 3.98 7.22 3.12 5.57 2.70 4.80 2.50 4.35 2.34 4.13 2.22 3.92 2.15 3.74
∞ 3.87 7.25 3.00 5.45 2.64 4.70 2.40 4.31 2.24 3.98 2.11 3.75 2.02 3.62

p ¼ 0.100

30 2.91 5.20 2.53 4.43 2.30 4.04 2.18 3.79 2.13 3.67 2.07 3.56 2.06 3.48
35 2.87 5.17 2.49 4.41 2.27 3.99 2.16 3.74 2.08 3.62 2.00 3.47 1.97 3.40
40 2.84 5.11 2.44 4.37 2.25 3.96 2.11 3.72 2.03 3.54 1.98 3.45 1.93 3.35
45 2.85 5.24 2.42 4.32 2.23 3.93 2.10 3.69 2.03 3.56 1.95 3.45 1.90 3.37
50 2.82 5.12 2.45 4.33 2.21 3.94 2.07 3.67 2.00 3.52 1.92 3.38 1.86 3.30
55 2.79 5.20 2.41 4.32 2.21 3.94 2.05 3.69 1.98 3.51 1.91 3.37 1.84 3.30
60 2.81 5.22 2.41 4.36 2.19 3.91 2.05 3.66 1.97 3.51 1.89 3.38 1.84 3.30
65 2.75 5.22 2.44 4.35 2.20 3.92 2.03 3.64 1.97 3.46 1.89 3.36 1.82 3.27
70 2.82 5.18 2.39 4.36 2.16 3.94 2.04 3.66 1.94 3.48 1.87 3.36 1.82 3.26
75 2.77 5.26 2.39 4.39 2.16 3.96 2.04 3.66 1.93 3.46 1.86 3.34 1.82 3.25
80 2.76 5.18 2.36 4.35 2.12 3.94 2.03 3.63 1.94 3.48 1.86 3.34 1.81 3.22
∞ 2.74 5.30 2.30 4.34 2.11 3.89 1.97 3.62 1.87 3.44 1.78 3.27 1.72 3.17

only the intercept is included when testing the first differenced series. 4.3. Empirical results
The lag length determination for ADF test is based on the Modified
Akaike Information Criterion, while the PP test uses the Bartlett kernel The various cointegration and non-cointegration cases involving
for its spectral estimation and Newey-West method for its bandwidth LREXP and LRREV are tested within the ARDL equation, first for LREXP,
determination. the general equation is specified as:
From Table 4, the two unit root tests suggest that both LREXP and
LRREV for the three countries are I(1). No tests suggest that these vari- X
p1
0
ΔLREXP ¼ c10 þ c11 t þ π 1yy LREXPt1 þ π 1yx:x LRREVt1 þ ψ y;i1 ΔLREXPti
ables are of an integration order higher than one. If any variable is I(0), i¼1
the ARDL equation using this I(0) variable as the dependent variable X
q1 X
r
0 0
violates PSS' assumption. However, as pointed out by MSG, one should þ ψ x;j1 ΔLRREVtj þ ω1 ΔLRREVt þ δk1 D1t;k þ u1t
proceed to test for cointegration using the full three-test framework in the j¼1 k¼1

ARDL, even if a variable is I(0). The tests would suggest a degenerate (8)
lagged independent variable case if the dependent variable is indeed I(0).

135
C.Y. Sam et al. Economic Modelling 80 (2019) 130–141

Table 4
Unit root tests.
Variables ADF PP
st
Level 1 Differenced Level 1st Differenced

US
LREXP 2.4365 2.7222* 1.7709 2.7988*
LRREV 1.9937 5.1390*** 2.6098 5.1874***
UK
LREXP 2.3242 3.2227** 1.9429 3.6054***
LRREV 2.5854 3.6601*** 2.6504 7.3637***
France
LREXP 2.5369 2.7844* 2.4249 3.6812***
LRREV 1.7650 6.2735*** 1.6477 6.2720***

Note: *, **, *** indicate statistical significance at 10%, 5% and 1% levels.

(i) F-test on the lagged levels of all variables: H0 : π yy ¼ π yx:x ¼ 0

against H1 : any π yy ; π yx:x 6¼ 0
(ii) t-test on the lagged level of the dependent variable: H0 : π yy ¼ 0
against H1 : π yy 6¼ 0
(iii) F-test on lagged level of the independent variable: H0 : π yx:x ¼ 0
against H1 : π yx:x 6¼ 0

If the above null hypotheses (i)H0 : π yy ¼ π yx:x ¼ 0 and (ii) H0 : π yy ¼

0 are rejected but hypothesis (iii) H0 : π yx:x ¼ 0 is not, a case of degenerate
lagged independent variable arises. Note that the absence of the lagged
level of the independent variable reduces the equation to a generalized
Dickey-Fuller equation. The statistical significance of the overall F-sta-
tistic is due solely to the significance of the coefficient on the lagged level
of the dependent variable, indicating that the dependent variable is I(0).
Since this degenerate lagged independent variable case implies the absence
of cointegration, one will not be misled by a stationary dependent vari-
able into falsely concluding that there is cointegration, as long as the
extra test on the lagged level of independent variable provides the correct
result. On the other hand, if the null hypotheses from (i) and (iii) are
rejected but (ii) is not, then this falls into the degenerate lagged dependent
variable case. Cointegration can be concluded only when all three hy-
potheses (i), (ii), (iii) are rejected. Otherwise, there is no cointegration.
After determining all variables are integrated of order not greater
than one, an ARDL(p,q) for equation (8) and an ARDL(a, b) for equation
(9) are specified. For the sake of parsimony and to avoid over-
parameterization, the maximum lag length is set to 4 and the number
of dummies up to a maximum of 5. Since there are visible trends in all the
series, we apply tests that allow unrestricted deterministic linear trends,
namely Cases III and V. Case III describes the ARDL equation that allows
drift in the individual series, while Case V also allows for a quadratic
trend in the individual series. For Case III the deterministic trend term is
omitted from equations (8) and (9). Note that we do not consider the case
of unrestricted intercept and restricted trend (Case IV) since a deter-
ministic trend in the long run relation would be inappropriate for ana-
lysing the spend-tax hypothesis. If the deterministic trend appears in the
Fig. 1. Time series plots of LREXP and LRREV from period 1972 to 2015 for (a) long run relation, this indicates that the two series trend away from each
the US, (b) the UK, (c) France. other, thus violating the budgetary equilibrium.
We adopted the general-to-specific modelling approach presented for
the ARDL model by PSS, with the optimal lag length determined by the
where Dt;k are dummy variables to deal with possible structural breaks in Akaike Information Criterion (AIC). Before proceeding with the ARDL
the equation. The dummy is specified as one in the identified year and cointegration tests, Table 5 summarizes the diagnostic details of these
zero otherwise. An analogous ARDL equation for LRREV is given as models. These include the correlogram for white noise residuals, the
residual normality test (Jarque-Bera), the Lagrange multiplier test for no
X
a1
0
ΔLRREV ¼ c20 þ c21 t þ π 2yy LRREVt1 þ π 2yx:x LREXPt1 þ ψ y;l2 ΔLRREVtl residual serial correlation up to 4 lags, and the heteroscedasticity test
l¼1 (Breusch-Pagan-Godfrey). In addition, the Ramsey RESET test of no
X
b1
02
X
c
0
functional mis-specification is presented. All the equations passed these
2
þ ψ x;j ΔLREXPtm þ ω ΔLREXPt þ δk2 D2t;n þ u2t diagnostics except when France's LRREV was used as the dependent
m¼1 n¼1
variable, both with and without the restricted deterministic trend. These
(9)
models failed to pass the RESET test at 0.10 level suggesting that they
The null hypotheses for all three tests are defined as follows: may not account for some important nonlinearities, thus causing biased

136
C.Y. Sam et al. Economic Modelling 80 (2019) 130–141

Table 5 Table 7
Diagnostic checking for the optimal models. Summary of upper bound critical values and sources.
Equations Q-stat(12) χ 2SC ð4Þ JB χ 2BPG FRESET(1) References p ¼ 0.100 p ¼ 0.050 p ¼ 0.010

Without deterministic trends Narayan (2005) (k ¼ 1, N ¼ 40)

F III
overall
5.000 6.160 8.825
US
F Voverall 6.630 7.980 11.230
LREXP 6.3139 5.6514 1.2524 1.9305 1.6749
Pesaran et al. (2001) (k ¼ 1)
LRREV 7.2398 6.2357 0.5640 7.4768 0.0239
t III 2.91 3.22 3.82
UK DV

LREXP 9.1271 1.7940 1.2238 5.9216 0.2477 t VDV 3.40 3.69 4.26
LRREV 9.3449 3.0024 0.3114 2.9229 0.0036 Critical values in Tables 2 and 3 (k ¼ 1, N ¼ 40)
France F III
IDV
5.37 7.41 12.40
LREXP 8.7416 0.7086 1.2065 3.3676 1.6619 F VIDV 5.11 7.09 12.21
LRREV 7.0027 4.3332 1.0647 3.8157 3.5468*
Note: F III V
overall and F overall refer to overall F statistics for cases III and V respectively;
With deterministic trends III V
t DV and t DV refer to t-statistics for the dependent variable for cases III and V
US respectively; F III V
IDV and F IDV refer to F-statistics for the independent variable for
LREXP 6.3150 5.9700 1.2616 1.9057 1.8651 cases III and V respectively. The number of explanatory variables in all cases is
LRREV 6.3200 4.1000 0.5167 5.1814 0.2141 k ¼ 1, and N ¼ 40 is the sample size.
UK
LREXP 12.444 1.8823 0.0216 12.915 0.4260
LRREV 17.927 4.3210 0.5263 5.3995 0.0950
France values for the lagged dependent variable t-test. Therefore, for the t-test on
LREXP 8.7537 0.6542 2.0668 3.3195 0.1880 the lagged dependent variable, we use PSS critical values reported on pp.
LRREV 7.5806 5.0282 2.2753 6.7115 8.2134*** 303–304 in PSS (2001). For the F-test on the lagged independent vari-
Note: *, **, *** indicate statistical significance at 10%, 5% and 1% levels, able, we refer to Tables 2 and 3 in this paper.1 Table 7 summarizes the
respectively. Q-stat(12) denotes Q statistic at lag 12, χ 2SC ð4Þ indicates Lagrange types of upper bound critical values we use in these tests. Test statistic
Multiplier statistic against order 4 autocorrelation, JB indicates Jarque-Bera that exceeds the upper bound is sufficient to establish statistical signifi-
statistic for normality test, χ 2BPG denotes Breusch-Pagan-Godfrey chi-square sta- cance of each test.
tistic for heteroscedasticity test. FRESET(1) is the F-statistic for Regression Speci- The bounds testing procedure leaves open the possibility of an
fication Error Test against order 1 for no functional form mis-specification. inconclusive test result if the test statistic falls between the lower and
upper bounds. In this case there are two possible courses of action. The
first is to test the orders of integration of each explanatory variable. If
parameter estimators. However, recursive estimation of the cumulative they are all I(0), then use the lower bound as the true critical value; if
sum and cumulative sum of squares plots show that these OLS estimated they are all I(1), then the upper bound is appropriate. One problem with
regression coefficients are stable over the sample period. this approach is that standard unit root tests have low power, leaving the

Table 6
Model specification and cointegration tests.
Without deterministic Optimum lag AIC Trend term Dummies F III
overall t III
DV F III
IDV
Conclusion
trends length

US
LREXP (2, 0) 6.1092 – – 11.4789*** 4.7902*** 20.8821*** Cointegrated
LRREV (4, 3) 3.4181 – D09 3.8705 2.6198 4.9827 No-cointegration
UK
LREXP (1, 4) 5.3153 – D95 7.9541** 3.9181*** 15.8400*** Cointegrated
LRREV (4, 3) 4.4860 – D93, D09 0.1842 0.1874 0.0012 No-cointegration
France
LREXP (1, 3) 6.2422 – – 9.0783*** 2.9578* 6.1522* Cointegrated
LRREV (0, 0) 4.6805 – D95 6.7952 1.6657 1.7466 No-cointegration

With deterministic trends Optimum lag AIC Trend term Dummies F Voverall t VDV F VIDV Conclusion
length

US
LREXP (2, 0) 6.061 Insignificant – 10.6098** 1.5076 7.7421** Degenerate lagged dependent
variable
LRREV (4, 0) 3.4645 ** D09 5.4485 3.1912 2.5300 No-cointegration
UK
LREXP (1, 3) 5.3509 *** D95 17.4321*** 5.8325*** 23.1628*** Cointegrated
LRREV (4, 2) 4.5643 *** D93, D09 4.0576 2.6053 4.03E-07 No-cointegration
France
LREXP (1, 3) 6.2403 Insignificant – 5.0373 3.1740 5.7531* No-cointegration
LRREV (0, 0) 4.6598 Insignificant D95 2.2812 1.9075 1.1875 No-cointegration

Note: *, **, *** indicate statistical significance at 10%, 5% and 1% levels, respectively, with the references from Table 7. AIC denotes the Akaike Information Criterion.
overall and F overall are the statistics for testing π yy ¼ 0 and π yx:x ¼ 0 without and with an unrestricted deterministic trend in equations (8) or (9). t DV and t DV are t-statistics
F III V III V

for π yy ¼ 0 without and with an unrestricted deterministic trend. F III IDV and F V
IDV are F-statistics for π yx:x ¼ 0 without and with an unrestricted deterministic trend. Dummy
variable D## is specified as one at the particular year ## and zero otherwise. For instance, D95 is specified as 1 at year 1995 and 0 for other years.

Finding most diagnostics satisfactory, we move to the tests for coin-

tegration presented in Table 6. For the overall F-statistics, we use the
critical values provided by Narayan (2005), which are appropriate for 1
To obtain F-statistic for a single lagged level independent variable, simply
small samples. However, Narayan does not provide small sample critical square the t-statistic value.

137
C.Y. Sam et al. Economic Modelling 80 (2019) 130–141

researcher with some uncertainty about the correct critical value. three null hypotheses are not rejected, indicating no cointegration when
Another problem is that there could be a mixture of I(0) and I(1) series in LRREV is the dependent variable. However, for the UK LREXP equation
the model. A more general approach is to employ a bootstrap procedure cointegration is supported with strong evidence, regardless of the treat-
to generate critical values that are appropriate for the combination of ment of trends (Cases III and V). It would be safe to conclude that in the
stochastic properties of the several variables in the model. This meth- UK revenues do not adjust to budgetary conditions, but expenditures do.
odology, presented in McNown et al. (2018), eliminates the possibility of Second, all three test statistics from the US LREXP equation without
inconclusive tests. deterministic trend are highly significant, suggesting cointegration when
There are several interesting observations from Table 6. First, in the LREXP is the dependent variable. On the other hand, the same equation
LRREV equation for all three countries, with and without trend, all the with deterministic trend has an insignificant t VDV , even at the 10% level,

Diagram 1. Summary procedures for implementing the augmented ARDL bounds test.

138
C.Y. Sam et al. Economic Modelling 80 (2019) 130–141

implying the US LREXP falls into degenerate lagged dependent variable case. regressors to have mixed or unknown orders of integration, I(0) or I(1),
Without a consensus from these two specifications, the AIC and the sig- which conventional cointegration tests do not permit. However, there is
nificance of the deterministic trend term are informative. From Table 6, the possibility of falling into the degenerate cases of non-cointegration
the deterministic trend term is insignificant in the equation with deter- from the ARDL bounds test. This has often been ignored in empirical
ministic trend, and its AIC value (6.061) is higher than in the equation applications.
without deterministic trend (6.1092). Therefore, the equation without Cointegration in the ARDL framework requires that the coefficients
trend is the preferred specification, leading to the conclusion of cointe- on the lagged levels of all variables in an unrestricted error correction
gration in the US LREXP equation. equation be jointly significant. However, significance of the overall F-
The same reasoning applies to France's LREXP equation. The AIC statistic is insufficient to establish cointegration, as it cannot rule out the
values for the equations with and without deterministic trend are possibility of degenerate cases. In the ARDL bounds test by Pesaran et al.,
6.2403 and 6.2422, respectively, and the trend term is not statisti- the degenerate lagged dependent variable case can be determined by car-
cally significant when it is included. Therefore, the equation without rying out the t-test on the lagged level of the dependent variable. The
trend is preferred. The conclusion is that cointegration is found for degenerate lagged independent variable(s) can be excluded under the
France when LREXP is the dependent variable. assumption of I(1) dependent variable. By not relying on unit root testing
Third, none of the equations fall into the degenerate lagged independent to ensure the integration order of the dependent variable, McNown et al.
variable case, which is consistent with the I(1) findings of the unit root (2018) introduced the test on the lagged level of the independent vari-
tests. This strengthens our earlier statement in the Introduction that the able(s) that relies on bootstrap methodology. As an alternative to the
augmented ARDL, by adding the test on the lagged independent vari- bootstrap method, this paper presents the limiting distributions of the
able(s), in addition to completing the test for cointegration, provides test on the lagged level of the independent variable(s) and tables of
information on the integration order of the dependent variable. There- bounds for the critical values, including both small sample and asymp-
fore, in this example, the augmented ARDL avoids imposing the totic cases. This allows for a straightforward test of the degenerate lagged
assumption that the dependent variable is I(1), as required in the PPS independent variable(s) case that can be used concurrently with the other
test. In summary, this example illustrates the steps in conducting the two tests from Pesaran et al. (2001). Relaxing the assumption of an I(1)
augmented ARDL bounds testing. Diagram 1 below summarizes the dependent variable, the test on the lagged level of independent vari-
procedures for carrying out the augmented ARDL test. able(s) can be combined with the existing tests from Pesaran et al. (2001)
Returning to the cointegration tests for the cases of the US, the UK and to draw a conclusion on the status of cointegration.
France, the finding of cointegration for the expenditures equation but not This augmented ARDL bounds testing framework was applied to the
for the revenues equation indicates that revenues serve as the forcing spend-tax relationship using generated critical values presented in this
variable in this bivariate system. Expenditures respond significantly to paper. This application illustrates that the augmented ARDL bounds test
budgetary conditions, but revenues do not. This can be interpreted as provides greater insight into the cointegration status and the integration
evidence supporting the tax-and-spend hypothesis of budgetary policy. order of the tested variables. If one encounters the case of degenerate
These findings for the tax-and-spend hypothesis for the US and the UK lagged independent variable(s), the ARDL equation is reduced to the
concur with most of the empirical studies, such as Bohn (1991); West- Dickey-Fuller unit root equation and the dependent variable is indicated
erlund et al. (2011); Korena and Stiassny (1998); Hasan and Lincoln as I(0); otherwise, it is I(1). If the testing suggests a degenerate dependent
(1997). This suggests that the budgetary policies in these countries variable or non-cointegration, it indicates that the dependent variable
support Friedman's theory. However, our finding for France differs from does not enter into the cointegrating equation embedded in that ARDL
Owoye's (1995) and Afonso and Rault's (2009). In Owoye's finding, based equation. The movement of the dependent variable does not respond to
on the Engle-Granger cointegration test, France practises a fiscal syn- the movement of the independent variables, again indicating non-
chronization policy while Afonso and Rault found evidence that France cointegration.
practises the spend-and-tax policy based on bootstrap panel causality
tests. The differences in these findings may be due to a possible Acknowledgement
mis-specification of the LRREV equation for France, as suggested by the
RESET test. A preferred specification might involve a non-linear func- We would like to thank the editor and an anonymous referee for their
tional form. Further investigation of this possibility would require a highly constructive and valuable comments which greatly improved the
non-linear ARDL modelling and testing approach with the generation of quality of this paper. We also like to thank Dr. Tang Tuck Cheong from
critical values embedded in this non-linear framework. University of Malaya, Malaysia in suggesting the idea of generating tables
of critical values for the bootstrap ARDL test. This work was supported by
5. Conclusions the Ministry of Education, Malaysia under Fundamental Research Grant
Scheme [FRGS, 203/CDASAR/6711503] and Bridging Grant,
Pesaran et al. (2001) introduced the ARDL bounds approach for 304.CDASAR.6316489.
testing cointegration. One advantage of this approach is that it allows

Appendix A

In deriving the bounds testing methodology, Pesaran et al. (2001) let the data-generating process for VAR(p) of ðk þ 1Þ-vector random process,
fzt g∞
t¼1 to be

ΦðLÞðzt  μ  γtÞ ¼ εt ; t ¼ 1; 2; … (A)

0 0
0 0
where L is the lag operator, μ and γ are unknown ðk þ 1Þ-vectors of intercept and trend coefficients where μ ¼ ðμy ; μx Þ and γ ¼ ðγ y ; γx Þ the ðk þ 1; k þ 1Þ
P
matrix lag polynomial ΦðLÞ ¼ Ikþ1  pi¼1 Φi Li with fΦi gpi¼1 ðk þ 1; k þ 1Þ matrices of unknown coefficients. The equation A can be written in vector
ECM form as

139
C.Y. Sam et al. Economic Modelling 80 (2019) 130–141

X
p1
Δzt ¼ a0 þ a1 t þ Πzt1 þ Γ i Δzti þ εt ;
i¼1

 
0
0 π yy πyx
where zt and Π can be partitioned as zt ¼ ðyt ; xt Þ and Π ¼ respectively. PSS thus imposes 5 assumptions for the bounds testing approach
πxy Πxx
and these assumptions are as follows:
 P 
Assumption 1. The roots of Ikþ1  pi¼1 Φi zi  ¼ 0 are either outside the unit circle jzj ¼ 1 or satisfies z ¼ 1.
Assumption 2. The vector error process fεt g∞
t¼1 is INð0; ΩÞ, where Ω is positive definite.

Assumption 3. The k-vector πxy ¼ 0.

Assumption 4. The matrix Πxx has rank r, 0  r  k.
0
? ?
Assumption 5a. If rank ðΠÞ ¼ r, the matrix ðα? ?
y ; α Þ Γðβy ; β Þ is full rank k  r þ 1; 0  r  k.
0
Assumption 5b. If rank ðΠÞ ¼ r þ 1, the matrix α? Γβ? is full rank k  r; 0  r  k.
Pesaran et al. (2001) followed Johansen's (1991, 1995) methodology of introducing five cases of interest describing the different specifications of
the ARDL model. These cases are categorized according to the types of restrictions on the intercept and deterministic trend. The summary of the ARDL
model specifications are listed below:
Case 1. (No intercepts; no trends):

X
p1
0 0
Δyt ¼ π yy yt1 þ πyx:x xt1 þ ψi Δzti þ ω Δxt þ ut ;
i¼1

Case 2. (Restricted intercepts; no trends)

  X
p1
0 0
Δyt ¼ π yy yt1  μy þ πyx:x ðxt1  μx Þ þ ψi Δzti þ ω Δxt þ ut ;
i¼1

Case 3. (Unrestricted intercepts; no trends)

X
p1
0 0
Δyt ¼ c0 þ π yy yt1 þ πyx:x xt1 þ ψi Δzti þ ω Δxt þ ut ;
i¼1

Case 4. (Unrestricted intercepts; restricted trends)

  X
p1
Δyt ¼ c0 þ π yy yt1  γ y t þ πyx:x ðxt1  γx tÞ þ ψ'i Δzti þ ω' Δxt þ ut ;
i¼1

Case 5. (Unrestricted intercepts; unrestricted trends)

X
p1
0 0
Δyt ¼ c0 þ c1 t þ π yy yt1 þ πyx:x xt1 þ ψi Δzti þ ω Δxt þ ut :
i¼1

For the details of the description and derivation of the bounds testing framework please refer to Pesaran et al. (2001).

Appendix B. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.econmod.2018.11.001.

References Asteriou, D., Masatci, K., Pilbeam, K., 2016. Exchange rate volatility and international
trade: international evidence from the MINT countries. Econ. Modell. 58, 133–140.
https://doi.org/10.1016/j.econmod.2016.05.006.
Afonso, A., Rault, C., 2009. Spend and tax: a panel data investigation for the EU. Econ.
Baghestani, H., McNown, R., 1994. Do revenues or expenditures respond to budgetary
Bull. 29 (4), 2545–2551.
disequilibria? South. Econ. J. 60, 24–31. https://doi.org/10.2307/1059979.

140
C.Y. Sam et al. Economic Modelling 80 (2019) 130–141

Bohn, H., 1991. Budget balance through revenue or spending adjustments. J. Monetary Korena, S., Stiassny, A., 1998. Tax and spend, or spend and tax? An international study.
Econ. 27, 333–359. https://doi.org/10.1016/0304-3932(91)90013-E. J. Pol. Model. 20 (2), 163–191. https://doi.org/10.1016/S0161-8938(96)00073-7.
Engle, R.F., Granger, Clive W.J., 1987. Co-integration and error correction: McNown, R., Sam, C.Y., Goh, S.K., 2018. Bootstrapping the autoregressive distributed lag
representation, estimation and testing. Econometrica 55 (2), 251–276. https:// test for cointegration. Appl. Econ. 50, 1509–1521. https://doi.org/10.1080/
doi.org/10.2307/1913236. 00036846.2017.1366643.
Fasano, U., Wang, Q., 2002. Testing the Relationship between Government Spending and Meltzer, A.H., Richard, S.F., 1981. A rational theory of the size of government. J. Polit.
Revenue: Evidence from GCC Countries. IMF Working Paper, WP/02/201. Econ. 89, 914–927. Retrieved from. http://www.jstor.org/stable/1830813.
Friedman, M., 1978. The limitation of tax limitation. Pol. Rev. 5, 7–14. Muscatelli, V.A., Spinelli, F., 2000. The long-run stability of the demand for money: Italy
Goh, S.K., Sam, C.Y. Sam, McNown, R., 2017a. Re-examining foreign direct investment, 1861-1996. Econ. Lett. 90 (1), 72–76.
exports, and economic growth in Asian economies using a bootstrap ARDL test for Musgrave, R., 1966. Principles of budget determination. In: Cameron, H., Henderson, W.
cointegration. J. Asian Econ. 51, 12–22. https://doi.org/10.1016/ (Eds.), Public Finance Selected Reading. Random House, New York, pp. 15–27.
j.asieco.2017.06.001. Narayan, P.K., 2005. The saving and investment nexus for China: evidence from
Goh, S.K., Yong, J.Y., Lau, C.C., Tang, T.C., 2017b. Bootstrap ARDL on energy-growth cointegration tests. Appl. Econ. 37 (17), 1979–1990. https://doi.org/10.1080/
relationship for 22 OECD countries. Appl. Econ. Lett. https://doi.org/10.1080/ 00036840500278103.
13504851.2017.1284980. Owoye, O., 1995. The causal relationship between taxes and expenditures in the G7
Goh, S.K., McNown, R., 2015. Examining the exchange rate-regime monetary policy countries: cointegration and error correction models. Appl. Econ. Lett. 2 (1), 19–22.
autonomy nexus: evidence from Malaysia. Int. Rev. Econ. Finance (35), 292–303. https://doi.org/10.1080/135048595357744.
https://doi.org/10.1016/j.iref.2014.10.006. Peacock, A.T., Wiseman, J., 1961. The Growth of Public Expenditure in the United
Hasan, M., Lincoln, I., 1997. Tax then spend or spend then tax? Experience in the UK Kingdom. Princeton University Press, Princeton, NJ.
1961–1993. Appl. Econ. Lett. 4, 237–239. https://doi.org/10.1080/758518502. Peacock, A.T., Wiseman, J., 1979. Approaches to the analysis of government expenditure
Jalil, A., Mahmood, T., Idrees, M., 2013. Tourism-growth nexus in Pakistan: evidence growth. Publ. Finance Q. 7, 3–23. https://doi.org/10.1177/109114217900700101.
from ARDL bounds tests. Econ. Modell. 35, 185–191. https://doi.org/10.1016/ Pesaran, M.H., Shin, Y., Smith, R.J., 2001. Bounds testing approaches to the analysis of
j.econmod.2013.06.034. level relationships. J. Appl. Econom. 16, 289–326. https://doi.org/10.1002/jae.616.
Johansen, Søren, 1991. Estimation and hypothesis testing of cointegration vectors in Rushdi, M., Kim, J.H., Silvapulle, P., 2012. ARDL bounds tests and robust inference for
Gaussian vector autoregressive models. Econometrica 59 (6), 1551–1580. https:// the long run relationship between real stock returns and inflation in Australia. Econ.
doi.org/10.2307/2938278. Modell. 29, 535–543. https://doi.org/10.1016/j.econmod.2011.12.017.
Johansen, Søren, 1995. Likelihood-based Inference in Cointegrated Vector Autoregressive Westerlund, J., Mahdavi, S., Firoozi, F., 2011. The tax-spending nexus: evidence from a
Models. Oxford University Press. panel of US state-local governments. Econ. Modell. 28 (3), 885–890. https://doi.org/
Kim, H.S., Baek, J., 2013. Assessing dynamics of crude oil import demand in Korea. Econ. 10.1016/j.econmod.2010.10.016.
Modell. 35, 260–263. https://doi.org/10.1016/j.econmod.2013.07.010.

141

View publication stats