The ARDL Method

Jamee K. Moudud
Spring 2023
 Reference: Pesaran, M. Hashem, Yongcheol Shin, and Richard
J. Smith. 2001. "Bounds Testing Approaches to the Analysis of
Level Relationships." 16 (3): 289-326. [PSS]
 Unlike traditional cointegration methods the ARDL method
derived by Pesaran, Shin, and Smith (PSS) does not require
prior unit root testing. As we know from prior work on
cointegration unit root tests can give ambiguous results at
times, an issue that is exacerbated when the sample size is
small.
 The ARDL model of PSS is basically based on the F test.
recap

 To every ARDL model there is a correspond ECM and

cointegrating equation.
 12.4
Cointegration

12.4.2
The Error Correction
Model

Consider a general model that contains lags of y

and x
 Namely, the autoregressive distributed lag
(ARDL) model, except the variables are
nonstationary:
yt  δ  θ1 yt 1  δ 0 xt  δ1 xt 1  vt

Chapter 12: Regression with Time-Series Data:

Principles of Econometrics, 4th Edition Page 4
Nonstationary Variables
 12.4
Cointegration

12.4.2
The Error Correction
Model

If y and x are cointegrated, it means that there is a

long-run relationship between them
 To derive this exact relationship, we set
yt = yt-1 = y, xt = xt-1 = x and vt = 0
 Imposing this concept in the ARDL, we obtain:
y 1  θ1   δ   δ0  δ1  x

This can be rewritten

y  β  βinx the form:
1 2

Chapter 12: Regression with Time-Series Data:

Principles of Econometrics, 4th Edition Page 5
Nonstationary Variables
 12.4
Cointegration

12.4.2
The Error Correction
Model
Add the term -yt-1 to both sides of the ARDL
equation:
yt  yt 1  δ  θ1  1 yt 1  δ0 xt  δ1 xt 1  vt

 Add the term – δ0xt-1+ δ0xt-1:

yt  δ  θ1  1 yt 1  δ0  xt  xt 1   δ0  δ1  xt 1  vt

 δ
 Manipulating
yt  θ1  1  this we
 yt 1 
 δ0  δ1 
get: 
xt 1   δ0 xt  vt
 θ  1  θ  1 
 1 1 
Chapter 12: Regression with Time-Series Data:
Principles of Econometrics, 4th Edition Page 6
Nonstationary Variables
 12.4
Cointegration

12.4.2
The Error Correction
Model
Or:
Eq. 12.10
yt  α  yt 1  β1  β 2 xt 1   δ0 xt  vt

 This is called an error correction equation

 This is a very popular model because:
It allows for an underlying or fundamental link
between variables (the long-run relationship)
It allows for short-run adjustments (i.e.
changes) between variables, including
adjustments to achieve the cointegrating
relationship
Chapter 12: Regression with Time-Series Data:
Principles of Econometrics, 4th Edition Page 7
Nonstationary Variables
 What is the interpretation of the α?
α = error correction coefficient
If α > 0 then in the ECM (error correction model)
yt  α  yt 1  β1  β 2 xt 1   δ0 xt  vt

Whenever yt-1 – β1 – β2xt-1 > 0 then Δyt < 0

yt-1 – β1 – β2xt-1 < 0 then Δyt > 0
This condition ensures that the dynamical system
represented by the ECM is stable.

Chapter 12: Regression with Time-Series Data:

Principles of Econometrics, 4th Edition Page 8
Nonstationary Variables
 Consider the ECM:
yt  α  yt 1  β1  β 2 xt 1   δ0 xt  vt

If cointegration is found in the regression

yt = β1 + β2xt + ut  ut is stationary then
yt-1 = β1 + β2xt-1 + ut-1  ut-1 is also stationary

ECM is thus:
∆yt = -αut-1 + δ0∆xt + vt
Chapter 12: Regression with Time-Series Data:
Principles of Econometrics, 4th Edition Page 9
Nonstationary Variables
 Procedure of the Engels-Granger Method:
1) Carry out unit root test on each variable
2) Provided they are both I(1) then run the
regression: yt = β1 + β2xt + ut
3) Save regression residual ut and test it for unit
root: If ut is stationary then yt and xt are said to
cointegrated.
4) Estimate the ECM: ∆yt = -αut-1 + δ0∆xt + vt
If α > 0 and significant then xt is “pulling” at
yt (xt Granger causes yt)
Chapter 12: Regression with Time-Series Data:
Principles of Econometrics, 4th Edition Page 10
Nonstationary Variables
ARDL cointegration method

 The ARDL model of PSS is basically based on the F test.

 To understand the ARDL method, we need to recall that an F
test (involving the testing of the joint significance of a group of
explanatory variables) can be used to investigate whether
adding the additional explanatory variables increases the
explained sum of squares (ESS) and thus the R2 of the
regression in question. That is, if R2new corresponds to the
regression with the additional variables and R2old corresponds
to the original regression then the F statistic is defined as:
  
 F = [(R2new - R2old)/N]/[(1 - R2new)/df]
  
 So if F > Fc at some level of significance then we can conclude
that R2new is significantly different from R2old, i.e. the additional
group of variables is contributing to additional variations of the
dependent variable around its mean.
To Add or Not to Add New
Variables?
 Consider the infant mortality rate model.
 We want to see how per capita GNP (PGNP) and female
literacy rate (FLR) affect the child mortality rate (CM).
Dependent Variable: CM    
Method: Least Squares    
Date: 03/08/11 Time: 10:31    
Sample: 1 64      
Included observations: 64    
         
         
Variable Coefficient Std. Error t-Statistic Prob.  
         
         
C 157.4244 9.845583 15.98935 0.0000
PGNP -0.011364 0.003233 -3.515661 0.0008
         
         
R-squared 0.166217    Mean dependent var 141.5000
Adjusted R-squared 0.152769    S.D. dependent var 75.97807
S.E. of regression 69.93413    Akaike info criterion 11.36374
Sum squared resid 303228.5    Schwarz criterion 11.43120
Log likelihood -361.6396    Hannan-Quinn criter. 11.39031
F-statistic 12.35987    Durbin-Watson stat 1.931458
Prob(F-statistic) 0.000826      
         
         
Dependent Variable: CM    
Method: Least Squares    
Date: 03/08/11 Time: 10:42    
Sample: 1 64      
Included observations: 64    
         
         
Variable Coefficient Std. Error t-Statistic Prob.  
         
         
C 263.6416 11.59318 22.74109 0.0000
PGNP -0.005647 0.002003 -2.818703 0.0065
FLR -2.231586 0.209947 -10.62927 0.0000
         
         
R-squared 0.707665    Mean dependent var 141.5000
Adjusted R-squared 0.698081    S.D. dependent var 75.97807
S.E. of regression 41.74780    Akaike info criterion 10.34691
Sum squared resid 106315.6    Schwarz criterion 10.44811
Log likelihood -328.1012    Hannan-Quinn criter. 10.38678
F-statistic 73.83254    Durbin-Watson stat 2.186159
Prob(F-statistic) 0.000000      
         
         
 This is a process of incremental addition of variables and the
question is: is the increase in the adjusted R2 “significant”?

 In short, is it worthwhile adding additional variables or by

making a model parsimonious are we risking the problem of
underspecification?
 We use this test:
 F = [(R2new – R2old)/N]/[(1 - R2new)/df]
 = [(R2new – R2old)/# of new regressors]/[(1 - R2new)/(n - # of
parameters in the new model)]
 R2new = 0.7077, R2old = 0.1662
 F = [(0.7077-0.1662)/1]/[(1 – 0.7077)/(64 – 3)] = 113.05 
highly significant.

 So adding the FLR provides more information regarding what

makes CM fluctuate around its mean value.
ARDL Model

 Recall the basic form of an ARDL (equation 1):

  
 yt = β0 + β1yt-1 + .......+ βkyt-p + α0xt + α1xt-1 + α2xt-2 + ...+ αqxt-q +
εt       
 Consider a conventional ECM from an Engels-Granger test
(equation 2):

          Δyt = β0 + Σ βiΔyt-i + ΣγjΔx1t-j + ΣδkΔx2t-k + φzt-1 + et          
 Where it may be recalled that the sum of the first differences
arise from “packaging” the ARDL equation: they are a
reflection of the dynamic nature of relationship linking y to x 1
and x2. As we saw these sums of the first differences have the
added advantage of eliminating serial correlation.
 As we also saw, the z in the above equation is the residual of
the cointegrating equation (equation 3):
  
 yt = α0 + α1x1t + α2x2t + zt            
  
 In equation 2, zt-1 = yt - α0 - α1x1t-1 - α2x2t-1. In short, the ECM in
2 has been restricted to include the regression estimators α0, α1,
and α2 obtained from equation 3.
  
 Consider next the following ECM (equation 4)
  
 Δyt = β0 + ΣβiΔyt-i + ΣγjΔx1t-j + ΣδkΔx2t-k + θ0yt-1 + θ1x1t-1 + θ2 x2t-
1 + et       

  
 This ECM in equation 4 is similar to that in equation except
that the level variables yt-1, x1t-1, and x2t-1 are not restricted to
have the estimated coefficients from (3). Instead they take on
the coefficient values (θ0, θ1, and θ2) obtained by running
equation 4 as an OLS. Equation 4 is therefore an unrestricted
ECM or what PSS call a conditional ECM.
 The lag lengths on the lags of the first differences have to be chosen
on the basis of the SC or AIC (which in Eviews have to minimized),
care being taken that all serial correlation is eliminated, i.e.
eliminating serial correlation takes precedence over SC or AIC
minimization.
  Assuming no serial correlation, a Wald coefficient restrictions test is
carried on the levels variables in equation 4 for their joint
significance:
 H0:  θ0 = θ1 = θ2 = 0
 H1:  θ0 ≠ θ1 ≠ θ2 ≠ 0
  In the event that F is significant (in terms of the special critical values
tabulated by PSS—see below) one can reject the null in favor of the
alternative.
 Δyt = β0 + ΣβiΔyt-i + ΣγjΔx1t-j + ΣδkΔx2t-k + θ0yt-1 + θ1x1t-1 + θ2 x2t-
1

 + et       
 Rejection of the null suggests that the levels variables (yt-1, x1t-1,
& x2t-1) jointly explain variations of the Δyt.
 This can only be true if the linear combination
 (θ0yt-1 + θ1x1t-1 + θ2 x2t-1)
 is I(0) since all the other RHS terms have to be I(0) and Δy t is
I(0), i.e. that yt, x1t, and x2t are cointegrated.
Bounds test

 Significance of the F statistic is determined by the critical

values provided by Pesaran et al (2001). There are two sets of
critical values.
 One sets assumes that all the variables are I(1) (the upper
bound) and the other that they are I(0) (the lower bound).
 If at a given level of significance the computed F is greater than
the upper bound then the null of the non-existence of a long-run
relationship between the variables can be rejected.
 If the computed F falls below the lower bound the null cannot
be rejected.
 Finally if the computed F is between the two bounds the result
is inconclusive.
 Non-existence of a long-run relationship = no cointegration
 Rejecting the null of the non-existence of a long-run
relationship
 = cointegration
 In short if F is significant (higher than the upper bound) one
can reject the null and the variables are cointegrated with the
underlying variables being I(1)
 If F is insignificant (below the lower bound) one cannot reject
the null and the variables are not cointegrated.
 Note that if the underlying levels variables were I(0) an ECM
makes no sense and therefore there cannot be a cointegrating
relationship between them (remember that the Granger
Representation theorem states that a cointegrating relationship
has to involve an ECM so as to ensure that the long-term
movements of the variables are “tied” together). This is the
“insignificant F” scenario.
 On the other hand, if the underlying variables are indeed I(1)
(but you don’t know this ex ante because you did not do a unit
root test) then a significant F statistic implies the existence of
an ECM.
 Since Δyt has to be I(0) the only way that potentially I(1)
variables can explain the variations of Δyt is if they are
cointegrated.
 Caution: Of course Δyt could be I(1) so that yt is I(2) but then
the ARDL method cannot be applied in such a situation.
 “US & European Gas Prices.wf1”
 Estimation in Eviews: “Quick”  EUR C US  “Method”:
ARDL  Choose “Rest. Constant” or “Rest. Trend”.
 Let’s say we (arbitrarily) accept the default maximum lag
length at 4.
 So: Dependent = 4 & Regressors = 4.
 preliminary ARDL
Dependent Variable: EUR    
Method: ARDL      
Date: 03/26/23 Time: 10:56    
Sample: 1995M05 2011M03    
Included observations: 191    
Dependent lags: 4 (Automatic)    
Automatic-lag linear regressors (4 max. lags): US
Deterministics: Restricted constant and no trend (Case 2)
Model selection method: Akaike info criterion (AIC)
Number of models evaluated: 20  
Selected model: ARDL(4,0)    
         
         
Variable Coefficient Std. Error t-Statistic Prob.*
         
         
EUR(-1) 1.055477 0.069878 15.10464 0.0000
EUR(-2) 0.073369 0.104255 0.703740 0.4825
EUR(-3) 0.113827 0.105746 1.076426 0.2831
EUR(-4) -0.283568 0.069044 -4.107066 0.0001
US 0.052619 0.014258 3.690571 0.0003
C -0.007739 0.060952 -0.126973 0.8991
         
         
R-squared 0.987834    Mean dependent var 5.480838
Adjusted R-squared 0.987506    S.D. dependent var 3.261771
S.E. of regression 0.364595    Akaike info criterion 0.850848
Sum squared resid 24.59191    Schwarz criterion 0.953014
Log likelihood -75.25603    Hannan-Quinn criter. 0.892230
F-statistic 3004.375    Durbin-Watson stat 1.814474
Prob(F-statistic) 0.000000      
         
         
*Note: p-values and any subsequent test results do not account for model
        selection.    
Preliminary Cointegrating
equation
 The above ARDL equation is not of much interest except that it has
“buried” inside it the cointegrating relationship and the ECM.
 “View”: ARDL Diagnostics  Cointegrating Relation

 Deterministics: Rest. constant (Case 2)

 CE = EUR(-1) - (1.286711*US - 0.189249)



Preliminary Cointegrating
equation
 Dependent variable: EUR

Variable * Coefficient Std. Error t-Statistic Prob.

US 1.286711 0.302119 4.258957 0.0000

C -0.189249 1.503984 -0.125832 0.9000

Note: * Coefficients derived from the CEC regression.

Preliminary ECM

 Click on ARDL Menu and then check for serial correlation.

Breusch-Godfrey Serial Correlation LM Test:  

Null hypothesis: No serial correlation at up to 2 lags
         
         
F-statistic 8.623045    Prob. F(2,183) 0.0003
Obs*R-squared 16.44978    Prob. Chi-Square(2) 0.0003
         
         
         
 Oops! There is serial correlation (p-value of chi2 < 0.05).
 Increase lag length on the ARDL, say “Dependent” = 8 and
“Regressors” = 8. Try again. Test ECM for serial correlation
Bingo!

Breusch-Godfrey Serial Correlation LM Test:  

Null hypothesis: No serial correlation at up to 2 lags
         
         
F-statistic 1.853055    Prob. F(2,175) 0.1598
Obs*R-squared 3.878114    Prob. Chi-Square(2) 0.1438
         
         
 For the remainder of the lecture go to “Notes on the ARDL
Method.docx”.