QUANTITATIVE MACROECONOMICS

ARDL MODEL

Manuel Fernández de la Concha Rebollo ([email protected])

1.- Introduction

ARDL is abbreviated from Autoregressive Distributed-lagged model. It is a common model

used when working with time series data.
ARDL is generally given by:
(1) yt = B0xt+ B1xt-1 + B2xt-2 + ... + Bkxt-k + et
For one-off unit change in x there is an impact on y; this impact is capture by B0; B1 is
the impact on y after one period, B2 is the impact after 2 period, and so on. The final impact
on y is Bk. If all the coefficients are collected: {B0, B1, B2, ..., Bk}, they are called impulse
response function of the mapping on xt to yt. The above model is the lagged model
accounting for the changes in {B0, B1, B2, ..., Bk} on x for lagged period t. On the y-axis, the y-
dependent may response to exogenous factor as well; thus, the ARDL may accommodate for
both x and y as: yt = a0 + a1yt-1 + ...+ apyt-p + B0xt + B1xt-1 + ... + Bkxt-k + et
This is called autoregressive distributed-lag model (ARDL).
SHORT-RUN: ARDL(1,1): The first order dynamic linear regressive model or ARDL(1,1) may
be used for short-run analysis. The ARDL(1,1) is given by:
1.yt = a0 + a1yt-1 + B0xt + B2xt-1 + et where t = 1,2,... T
2. In this short-run analysis, yt is stable and would converge to equilibrium with the
condition: -1<a<1. Once stability is established, then the long-run analysis may be
undertaken.

LONG-RUN: The long-run condition is also known as steady state. This steady state is
given as:
(4) yt = (a0 / 1 - a1) + [(B0 + B1) / (1 - a1)]xt + et /(1 - a1)
which may be simplified as: yt = c0 + c1xt + et / (1 - a1). In this steady state, the
equilibrium is yt = yt-1 and xt = xt-1. From this condition, one can now predict the long-
run equilibrium (which is the same as target value) as:
(5) y*t = c0 + c1xt
REPARAMETIZATION: Recall the short-run ARDL as: yt = B0xt+ B1xt-1 + B2xt-2 + ... + Bkxt-
k + et , NOW it can be parametized as:
(6) DYt = d0 +d1yt-1 + d2Dxt +d3xt-1 + et
where d0 = a0, d1 = a1 - 1, d2 = B0, d3 = B0 + B1. NOW let's ad xt to the right-hand side
of the equation:
(7) Dyt = d0 + d1yt-1 + d2Dxt + d3xt-1 + d4xt + et
If one tries to estimate the regression model at this point, there would be a problem
due to perfect multicollinearity. Note that xt and xt-1 = xt - xt-1.
TESTING: Use D-test and LR test for model testing. We started by building
unrestricted model. restrict the model by the following hypothesis formulation: H0:
a1 = 0 and B1 = 0. The alternative hypothesis is: H1:a1 not = 0 and B1 not = 0. The
static model that we want to work with is:
 (8) yt = a + bxt + ut
The F-test applies:
(9) F-test = [(Sut2 - Set2) / 2] / (et2) / (T - 4)
This approximates F-distribution. In the alternative, the log-likelihood ratio test may
also be used:
(10) LR-test = 2(MLLUR - MLLR)
... where MLLUR is the maximum log-likelihood of the unrestricted model and MLLR is
the maximum log-likelihood of the restricted model. This approximates chi-square
distribution.

3. ARDL model

After running the model, using as dependent variable D(GDP) and D(INV) as independent
variable, we run the model for USA:
Dependent Variable: D(GDP)
Method: ARDL
Date: 04/08/19 Time: 20:13
Sample (adjusted): 1947Q4 2018Q4
Included observations: 285 after adjustments
Maximum dependent lags: 4 (Automatic selection)
Model selection method: Akaike info criterion (AIC)
Dynamic regressors (4 lags, automatic): D(INV)
Fixed regressors: GDP(-1) INV(-1) C
Number of models evalulated: 20
Selected Model: ARDL(2, 0)
Note: final equation sample is larger than selection sample

Variable Coefficient Std. Error t-Statistic Prob.*

D(GDP(-1)) 0.148091 0.051989 2.848489 0.0047

D(GDP(-2)) 0.100778 0.052612 1.915479 0.0565
D(INV) 206.9216 18.67757 11.07862 0.0000
GDP(-1) 0.001322 0.000928 1.424291 0.1555
INV(-1) 11.50829 8.804905 1.307031 0.1923
C -131.6811 146.9147 -0.896310 0.3709

R-squared 0.417163 Mean dependent var 151.6526

Adjusted R-squared 0.406718 S.D. dependent var 260.6323
S.E. of regression 200.7515 Akaike info criterion 13.46284
Sum squared resid 11244027 Schwarz criterion 13.53974
Log likelihood -1912.455 Hannan-Quinn criter. 13.49367
F-statistic 39.93858 Durbin-Watson stat 1.988110
Prob(F-statistic) 0.000000

Now we check if there is serial correlation or not, and if the model is stable or not.
To check serial correlation we use the TM test:
Breusch-Godfrey Serial Correlation LM Test:
Null hypothesis: No serial correlation at up to 2 lags

F-statistic 0.494087 Prob. F(2,277) 0.6107

Obs*R-squared 1.013099 Prob. Chi-Square(2) 0.6026

Test Equation:
Dependent Variable: RESID
Method: ARDL
Date: 04/08/19 Time: 20:19
Sample: 1947Q4 2018Q4
Included observations: 285
Presample missing value lagged residuals set to zero.

Variable Coefficient Std. Error t-Statistic Prob.

D(GDP(-1)) -0.001746 0.090458 -0.019301 0.9846

D(GDP(-2)) -0.066338 0.086458 -0.767280 0.4436
D(INV) 1.556908 18.77697 0.082916 0.9340
GDP(-1) 5.56E-05 0.000933 0.059641 0.9525
INV(-1) 3.293619 9.717145 0.338949 0.7349
C -48.73760 158.7686 -0.306972 0.7591
RESID(-1) -0.001642 0.107399 -0.015289 0.9878
RESID(-2) 0.099202 0.100025 0.991778 0.3222

R-squared 0.003555 Mean dependent var -2.55E-14

Adjusted R-squared -0.021626 S.D. dependent var 198.9765
S.E. of regression 201.1165 Akaike info criterion 13.47331

With this value of Chi-Square we cannot reject the null hypothesis, and so we cannot reject
the serial correlation.

For China:

Dependent Variable: D(GDP)

Method: ARDL
Date: 04/09/19 Time: 08:53
Sample (adjusted): 1963 2010
Included observations: 48 after adjustments
Maximum dependent lags: 4 (Automatic selection)
Model selection method: Akaike info criterion (AIC)
Dynamic regressors (4 lags, automatic): D(INV)
Fixed regressors: GDP(-1) INV(-1) C
Number of models evalulated: 20
Selected Model: ARDL(2, 0)
Note: final equation sample is larger than selection sample

Variable Coefficient Std. Error t-Statistic Prob.*

D(GDP(-1)) 0.637977 0.153918 4.144912 0.0002

D(GDP(-2)) -0.384266 0.159028 -2.416346 0.0201
D(INV) -6.978674 6.058417 -1.151897 0.2559
GDP(-1) 0.076148 0.016732 4.551170 0.0000
INV(-1) -1.810462 4.632244 -0.390839 0.6979
C 57.94508 176.2428 0.328780 0.7440

R-squared 0.956163 Mean dependent var 92.26144

Adjusted R-squared 0.950944 S.D. dependent var 112.7684
S.E. of regression 24.97659 Akaike info criterion 9.390224
Sum squared resid 26200.86 Schwarz criterion 9.624124
Log likelihood -219.3654 Hannan-Quinn criter. 9.478615
F-statistic 183.2182 Durbin-Watson stat 1.984489
Now we check if there is serial correlation or not, and if the model is stable or not.
To check serial correlation we use the TM test (viewresidual diagnosticsTM test).
Breusch-Godfrey Serial Correlation LM Test:
Null hypothesis: No serial correlation at up to 2 lags

F-statistic 0.342247 Prob. F(2,40) 0.7122

Obs*R-squared 0.807574 Prob. Chi-Square(2) 0.6678

Test Equation:
Dependent Variable: RESID
Method: ARDL
Date: 04/09/19 Time: 08:56
Sample: 1963 2010
Included observations: 48
Presample missing value lagged residuals set to zero.

Variable Coefficient Std. Error t-Statistic Prob.

D(GDP(-1)) 0.496164 0.747789 0.663507 0.5108

D(GDP(-2)) 0.144736 0.295584 0.489661 0.6270
D(INV) 0.130730 6.197719 0.021093 0.9833
GDP(-1) -0.059119 0.076794 -0.769836 0.4459
INV(-1) 2.260117 5.869411 0.385067 0.7022
C -76.90885 217.3938 -0.353777 0.7254
RESID(-1) -0.451851 0.708965 -0.637340 0.5275
RESID(-2) -0.454198 0.549306 -0.826859 0.4132

R-squared 0.016824 Mean dependent var -6.66E-15

Adjusted R-squared -0.155231 S.D. dependent var 23.61070

To check stability: view stability diagnostics recursive estimate Cusum test and we
obtain:

As the regression is among the red lines of significance we can say that the model is stable.

Bound test.
We are now going to investigate if there is long range associationship among the variables.
To do it we are going to do a Wald test (viewcoefficient -Wald test), using as nul
hypothesis C4=C5=0
Wald Test:
Equation: Untitled

Test Statistic Value df Probability

F-statistic 11.32220 (2, 42) 0.0001

Chi-square 22.64439 2 0.0000

Null Hypothesis: C(4)=C(5)=0

Null Hypothesis Summary:

Normalized Restriction (= 0) Value Std. Err.

C(4) 0.076148 0.016732

C(5) -1.810462 4.632244

Restrictions are linear in coefficients.

Now we compare the F-statistic value with the critical value at 5% taking from Narayan
(2005) (Table Case III: Unrestricted intercept and no trend; page 1988) which is 5.92.
The guideline is that if F-statistics is bigger than the upper bound value, we can reject the
null hypothesis, that it means that there will be long range associationship among the
variables.
The long range equation is:
GDP = b1+ b2 *INV
We estimate now the coefficients and get the residue series, which we will call ECT (error
correction term).
Dependent Variable: GDP
Method: Least Squares
Date: 04/09/19 Time: 11:09
Sample: 1960 2010
Included observations: 51

Variable Coefficient Std. Error t-Statistic Prob.

C 26285.74 3587.192 7.327663 0.0000

INV -678.4509 96.38998 -7.038604 0.0000

R-squared 0.502750 Mean dependent var 1049.809

Adjusted R-squared 0.492602 S.D. dependent var 1149.564
S.E. of regression 818.8560 Akaike info criterion 16.29212
Sum squared resid 32855736 Schwarz criterion 16.36788
Log likelihood -413.4491 Hannan-Quinn criter. 16.32107
F-statistic 49.54195 Durbin-Watson stat 0.303888
Prob(F-statistic) 0.000000
With the residue series we introduce in the regression model (now the variables will be
D(GDP) (-1) D(GDP) (-2) D(INV) and ECT(-1).
Trying to run the OLS model appears the message: Near singular matrix error. Regressors
may be perfectly collinear

4. ARDL MODEL using as variables LOG(GDP) and LOG(INV).

We are going to try to run the model, but instead of using GDP and INV as variabLes
we will use LOG(GDP) and LOG(INV).
 Dependent Variable: LOG(GDP)
Method: ARDL
Date: 04/09/19 Time: 11:52
Sample (adjusted): 1964 2010
Included observations: 47 after adjustments
Maximum dependent lags: 4 (Automatic selection)
Model selection method: Akaike info criterion (AIC)
Dynamic regressors (4 lags, automatic): LOG(INV)
Fixed regressors: GDP(-1) INV(-1) C
Number of models evalulated: 20
Selected Model: ARDL(4, 0)

Variable Coefficient Std. Error t-Statistic Prob.*

LOG(GDP(-1)) 1.399118 0.153228 9.130971 0.0000

LOG(GDP(-2)) -0.791344 0.268166 -2.950952 0.0053
LOG(GDP(-3)) 0.256536 0.232779 1.102061 0.2772
LOG(GDP(-4)) 0.159467 0.109958 1.450247 0.1550
LOG(INV) -0.081153 0.357903 -0.226744 0.8218
GDP(-1) -4.98E-07 2.10E-05 -0.023677 0.9812
INV(-1) 0.005304 0.010637 0.498672 0.6208
C 0.028368 1.073045 0.026437 0.9790

R-squared 0.998786 Mean dependent var 6.528573

Adjusted R-squared 0.998568 S.D. dependent var 1.013537
S.E. of regression 0.038360 Akaike info criterion -3.529770
Sum squared resid 0.057388 Schwarz criterion -3.214851
Log likelihood 90.94958 Hannan-Quinn criter. -3.411263
F-statistic 4582.014 Durbin-Watson stat 2.164613
Prob(F-statistic) 0.000000

*Note: p-values and any subsequent tests do not account for model

Now we check if there is serial correlation or not, and if the model is stable or not.
To check serial correlation we use the LM test (viewresidual diagnosticsLM test).
Breusch-Godfrey Serial Correlation LM Test:
Null hypothesis: No serial correlation at up to 2 lags

F-statistic 1.393543 Prob. F(2,37) 0.2609

Obs*R-squared 3.292350 Prob. Chi-Square(2) 0.1928

Test Equation:
Dependent Variable: RESID
Method: ARDL
Date: 04/09/19 Time: 11:54
Sample: 1964 2010
Included observations: 47
Presample missing value lagged residuals set to zero.

Variable Coefficient Std. Error t-Statistic Prob.

LOG(GDP(-1)) 0.347890 0.493986 0.704250 0.4857

LOG(GDP(-2)) -0.382655 0.891524 -0.429214 0.6703
LOG(GDP(-3)) 0.109770 0.654273 0.167775 0.8677
LOG(GDP(-4)) -0.085232 0.249659 -0.341395 0.7347
LOG(INV) 0.101068 0.361501 0.279578 0.7814
GDP(-1) 5.72E-06 2.13E-05 0.268623 0.7897
INV(-1) 0.000290 0.010635 0.027220 0.9784
C -0.345819 1.083562 -0.319150 0.7514
RESID(-1) -0.467545 0.512971 -0.911446 0.3680
RESID(-2) -0.248535 0.322405 -0.770876 0.4457

R-squared 0.070050 Mean dependent var 5.98E-16

Adjusted R-squared -0.156154 S.D. dependent var 0.035321
S.E. of regression 0.037979 Akaike info criterion -3.517288
Sum squared resid 0.053368 Schwarz criterion -3.123639
Log likelihood 92.65626 Hannan-Quinn criter. -3.369155
F-statistic 0.309676 Durbin-Watson stat 1.957578
Prob(F-statistic) 0.966855

To check stability: view stability diagnostics recursive estimate Cusum test and we
obtain:
As the regression is among the red lines of significance we can say that the model is stable.

Bound test.
We are now going to investigate if there is long range associationship among the variables.
To do it we are going to do a Wald test (viewcoefficient -Wald test), using as nul
hypothesis C6=C7=0.

Wald Test:
Equation: Untitled

Test Statistic Value df Probability

F-statistic 0.167343 (2, 39) 0.8465

Chi-square 0.334685 2 0.8459

Null Hypothesis: C(6)=C(7)=0

Null Hypothesis Summary:

Normalized Restriction (= 0) Value Std. Err.

C(6) -4.98E-07 2.10E-05

C(7) 0.005304 0.010637

Restrictions are linear in coefficients.

Now we compare the F-statistic value with the critical value at 5% taking from Narayan
(2005) (Table Case III: Unrestricted intercept and no trend; page 1988) which is 5.92.
The guideline is that if F-statistics is bigger than the upper bound value, we can reject the
null hypothesis, that it means that there will be long range associationship among the
variables. As the F-statistics isn’t bigger we cannot reject the nul hypothesis.

If the computed F-statistic falls below the lower bound we would conclude that the
variables are I(0), so no cointegration is possible, by definition. If the F-statistic exceeds the
upper bound, we conclude that we have cointegration. Finally, if the F-statistic falls between
the bounds, the test is inconclusive
The long range equation is:
log(GDP) = b1+ b2 *log(INV)
We estimate it now the coefficients and get the residue series, which we will call ECT (error
correction term).
Dependent Variable: LOG(GDP)
Method: Least Squares
Date: 04/09/19 Time: 13:07
Sample (adjusted): 1960 2010
Included observations: 51 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

C 68.47006 14.21018 4.818382 0.0000

LOG(INV) -17.16421 3.929981 -4.367504 0.0001

R-squared 0.280207 Mean dependent var 6.409529

Adjusted R-squared 0.265517 S.D. dependent var 1.056690
S.E. of regression 0.905605 Akaike info criterion 2.677998
Sum squared resid 40.18587 Schwarz criterion 2.753756
Log likelihood -66.28895 Hannan-Quinn criter. 2.706948
F-statistic 19.07509 Durbin-Watson stat 0.120120
Prob(F-statistic) 0.000065

Then the regression is now:

Log(gdp)= a1+b1log(gdp)(-1)+b2log(gdp)(-2)+b3log(gdp)(-3)+b4log(gdp)(-
4)+b5log(inv)+c*ECT(-1)
Being ECT Error Correction term.
The c term must be negative and significant.
Trying to run the model, it wasn’t possible due to the correlation of the regressors with the
dependent variable.
 5. ARDL model being INV dependent variable

We run the model.

Dependent Variable: D(INV)

Method: ARDL
Date: 04/09/19 Time: 19:08
Sample (adjusted): 1963 2010
Included observations: 48 after adjustments
Maximum dependent lags: 4 (Automatic selection)
Model selection method: Akaike info criterion (AIC)
Dynamic regressors (4 lags, automatic): D(GDP)
Fixed regressors: GDP(-1) INV(-1) C
Number of models evalulated: 20
Selected Model: ARDL(2, 0)
Note: final equation sample is larger than selection sample

Variable Coefficient Std. Error t-Statistic Prob.*

D(INV(-1)) 0.051630 0.155720 0.331559 0.7419

D(INV(-2)) -0.216771 0.154858 -1.399803 0.1689
D(GDP) -0.002904 0.003165 -0.917662 0.3640
GDP(-1) 7.50E-05 0.000347 0.216138 0.8299
INV(-1) -0.240159 0.123622 -1.942689 0.0588
C 9.077043 4.701278 1.930761 0.0603

R-squared 0.186736 Mean dependent var -0.034711

Adjusted R-squared 0.089919 S.D. dependent var 0.645082
S.E. of regression 0.615397 Akaike info criterion 1.983369
Sum squared resid 15.90595 Schwarz criterion 2.217269
Log likelihood -41.60086 Hannan-Quinn criter. 2.071760
F-statistic 1.928751 Durbin-Watson stat 1.917633
Prob(F-statistic) 0.109787

*Note: p-values and any subsequent tests do not account for model

Now we check if there is serial correlation or not, and if the model is stable or not.
To check serial correlation we use the LM test (viewresidual diagnosticsLM test).
Breusch-Godfrey Serial Correlation LM Test:
Null hypothesis: No serial correlation at up to 2 lags

F-statistic 0.397787 Prob. F(2,40) 0.6744

Obs*R-squared 0.936071 Prob. Chi-Square(2) 0.6262

Test Equation:
Dependent Variable: RESID
Method: ARDL
Date: 04/09/19 Time: 19:11
Sample: 1963 2010
Included observations: 48
Presample missing value lagged residuals set to zero.

Variable Coefficient Std. Error t-Statistic Prob.

D(INV(-1)) -0.380276 0.498841 -0.762319 0.4503

D(INV(-2)) -0.017166 0.382411 -0.044889 0.9644
D(GDP) -0.000931 0.003398 -0.273936 0.7855
GDP(-1) -5.01E-05 0.000369 -0.135841 0.8926
INV(-1) -0.209720 0.280242 -0.748356 0.4586
C 7.932976 10.61186 0.747557 0.4591
RESID(-1) 0.619920 0.710241 0.872830 0.3880
RESID(-2) 0.137469 0.525077 0.261808 0.7948

R-squared 0.019501 Mean dependent var -1.23E-15

Adjusted R-squared -0.152086 S.D. dependent var 0.581743
S.E. of regression 0.624415 Akaike info criterion 2.047008
To check stability: view stability diagnostics recursive estimate Cusum test and we
obtain:

As the regression is among the red lines of significance we can say that the model is stable.

Bound test.
We are now going to investigate if there is long range associationship among the variables.
To do it we are going to do a Wald test (viewcoefficient -Wald test), using as nul
hypothesis C4=C5=0.
Wald Test:
Equation: Untitled

Test Statistic Value df Probability

F-statistic 2.123038 (2, 42) 0.1323

Chi-square 4.246077 2 0.1197

Null Hypothesis: C(4)=C(5)=0

Null Hypothesis Summary:

Normalized Restriction (= 0) Value Std. Err.

C(4) 7.50E-05 0.000347

C(5) -0.240159 0.123622

Restrictions are linear in coefficients.

Now we compare the F-statistic value with the critical value at 5% taking from Narayan
(2005) (Table Case III: Unrestricted intercept and no trend; page 1988) which is 5.92.
The guideline is that if F-statistics is bigger than the upper bound value, we can reject the
null hypothesis, that it means that there will be long range associationship among the
variables. As the F-statistics isn’t bigger we cannot reject the nul hypothesis.

If the computed F-statistic falls below the lower bound we would conclude that the
variables are I(0), so no cointegration is possible, by definition. If the F-statistic exceeds the
upper bound, we conclude that we have cointegration. Finally, if the F-statistic falls between
the bounds, the test is inconclusive