Chapter 9

Dynamic Models,
Autocorrelation
When is correlation more likely?
In case of time series data one period’s observation is more
likely to be correlated with another period’s observation.
Cross sectional data do not usually exhibit this feature.
Example, Fowzia’s income is not supposed to be related
with Kazori’s income at a certain point of time.
Observations on two persons’ income in this case is an
example of cross sectional observation. But if one
collects data on Fowzia’s income from 2000 to 2011
then, of course, he or she might notice a correlation
between two successive periods’ income. The latter is an
example of time series observations that are correlated.
 What is dynamic relation?
If the action or decision of one period impacts or spreads
over other periods then the relation is dynamic. More
specifically a change in the level of an explanatory
variable may have behavioural implications beyond the
time peiod in which it occurred. The consequences of
economic decisions that result in changes in economic
variables can last a long time.

Example: when income tax increases, consumers reduce

their expenditures on gooods and services, which
reduces profits of suppliers, which reduces the demand
for inputs, which reduces profits of the input suppliers,
and so on. The effect of the tax increase ripples through
the economy. These effects do not occur instantaneously
but are sperad, or distributed, over future time periods.
That is economic actions or decisions taken at one point
in time t have effects on the economy at time t, but also
at times t+1, t+2, and so on.
 How to model dynamic relation?
Three different ways are proposed to model dynamic
relationships.
1. One way is to specify that the dependent variable y t is a
function of current and past values of an explanatory
variable. That is, y f ( x , x , x .....) (9.1)
t t t 1 t 2
2. Second way is to include the lagged dependent variable as
one of the explanatory variables.
yt f ( y , xt ) (9.2)
t 1
Assuming y as unemployment and x interest rate, the above
relation says that current unemployment depends on
previous unemployment and current interest rate. If a
positive relation is supposed, the relation will mean that
higher current unemployment would follow higher
unemployment of past period.
 Modeling dynamic relationship cont.
3. Another way of modeling dynamic relationship is via the error
term. y f ( x ) e
t t t e f (e ) (9.3)
t t 1
e f (et )
(9.3) implies that t 1
The dynamic nature of this relationship is such that the impact of
any unpredictable shock that feeds into the error term will be
felt not just in period t, but also in future periods. The current
error et affects not just the current value of the dependent
variable y but also its future values y , y ....
t t 1 t 2
The dynamic model discussed above with lag in the error term
leads us into other models that involve lags in the dependent
variable y and lags in one or more explanatory variables.

These are the three ways how dynamics can enter a regression
relationship- lagged values of the explanatory variable, lagged
values of the dependent variable, and lagged values of the
error term.
Throughout the analysis we would assume the variables
stationary- mean reverting, not wandering or trending.
 What is autocorrelation?
Autocorrelation is defined as the feature of likely correlation
between error terms.

Think of an intuition about some shocks in a typical economy

that take several periods to work through the system. If we
consider an error term in any one period, it will contain not
only the effect of a crurrent shock, but also the carryover from
previous shocks. This carryover will be related to, or
correlated with, the earlier shocks. When circumstances such
as these lead to error terms that are correlated, it is claimed
that autocorrelation exists.
 Area Response Model
Let A denotes area cultivated, P price then we can design an
area response model as
ln At    ln Pt et ..... ..... (9.4)
1 2
Simplify this as yt   xt et ..... ..... (9.5)
1 2
The question now: how to include lags in the error term?
It is likely that farmers’ decision about area planted dpends on
their perceptions about future price and government policies.
These variables are not included in the model therefore
should be reflected in error term. If the perceptions change
slowly over time, or at least not in a completely random
manner, the current perceptions that form part of e will be
t
related to perceptions held in the previous period that form
part of the lagged error e
t 1
A model that captures these effects is e e v .... .... (9.8)
t t 1 t
 About the random component

The random component e in time period t is composed of two

t
parts:
- et  1 is a carryover from the random error in the previous
period, due to the inertia in economic systems, the magnitude
of the parameter  determines the extent of the carrryover
and
- vt is a new shock to the level of the economic variable. In this
particular example, carryover may be the farmers’ perceptions
about pricing and govt policies. New shock may emerge from
a new policy announced by the government.
- The parameter  determines how quickly the effect of a
shock dissipates. The larger the magnitude of  the greater
the carryover from one period to another and the more slowly
the shock dissipates.
 First order autoregressive model
The model et e vt is known as first order
t 1
autoregressive AR(1) model. The order is determined by how
many periods apart. In this case there is only one period lag
therefore first order. The existence of AR(1) error has
implications on the estimation of our original parameters 
and  . 1
2
 Properties of AR(1) errors
The regression model yt   xt et ..... ..... (9.5)
1 2
Error is assumed to follow AR(1) process.
et e vt .... .... (9.8)
t 1
The random errrors vt are assumed to be uncorrelated random
variables with mean zero and a constant variance  2.
var(vt )  v2 cov(v ,vs) 0; t s v
E(vt ) 0 t
One additional assumption is that absolute value of rho is less
than one,  1  1 which implies the stationarity of error
term et, i.e., do not change from period to period.
 Properties of error term
* E(et ) 0
 v2
* var(et )  e2 
1  2
Since  2 is constant, the error e t is also homoskedastic.
v
Covariance between errors for different time periods would be
nonzero, which can be shown
cov(et ,e )  e2 k k 0 .......... (9.13)
t k
cov(et ,e )  2 k
corr(et ,e )  t k  e  k
t k var(et ) var(e )  e2
t k
corr (et ,e ) 
t 1

This means  defines correlation between errors that are one

peirod apart, which is alternatively called autocorrelation coef.
 Properties of AR(1) Model
Var (et ) E(et2)  (E(et ))2 E(et2)
. 2
E e vt
 
 

 t 1


E( 2e2  2e vt vt2)

t 1 t 1
 2E(e2 )  E(vt2)
t 1
 2 e2  v2
 e2  2 e2  v2
 e2   2 e2  v2
 e2(1  2)  v2
 2
 e2  v
1  2
 Cov(et ,e )  k e2
t k
.
Proof: Given, et e vt
.
t 1
   
Cov(et ,e ) E  et  E(et )e  E(e ) E et e  E  e vt e 
   
      
t 1    t  1 t  1   t  1  t  1  t  1
E(e2 )
t 1
 e2
  
Cov(et ,e ) E  et  E(et )e  E(e ) E(et e ) ..... (2)
 
t 2    t  2 t  2  t 2
et e vt
t 1
e e v
t 1 t 2 t 1
e  2e  v
t 1 t 2 t 1
 Proof
.

et e vt  2e  v vt

t 1 t 2 t 1

 2
Cov(et ,e ) E(et e ) E  ( e  v vt )e 


t 2 t 2  t 2 t 1 t  2
Cov(et ,e ) E   2e2  v e vt e   2E(e2 )  2 e2
 

t 2  t 2 t 1 t 2 t  2 t 2
Cov(et ,e )  e2
t 1
Cov(et ,e )  2 e2
t 2
Cov(et ,e )  k e2
t k
 Some features of AR(1) model
- AR(1) models are widely used to describe autocorrelation
although a variety of such dynamic models can illustrate
autocorrelation.
- The correlation coefficients in one, two or three periods’ apart
are  ,  2,  3 .... respectively.
Since absolute value of rho is less than one, the maximum value
of correlation coefficient would be the one that is one period
apart. This would mean as periods pass autocorrelation
between errors decline and becomes negligible.

In order to investigate the nature of autocorrelation, we use

eviews file chp9_bangla. Next slide shows the estimation
output and the following slide shows the nature of
autocorrelation among errors
 Estimation output

LOG(A) = 3.89325574527 + 0.776118786922*LOG(P)

(se) = (0.061) (0.277)
 residuals
E_CAP
Time e_cap .
• -0.3030292011185984 .8
• 0.2544366298171003
.6
• 0.1815150336078108
• 0.5030530915874198 .4
• 0.275078103431704
• -0.1154826732252646 .2
• -0.4371465657691693
• -0.4234882069386172 .0
• -0.3665888539692065
• -0.2535964664549816 -.2
• -0.145199621163959
• 0.09066624986925984 -.4
• 0.3042675472404802
-.6
• 0.6555207192065731
-.8
5 10 15 20 25 30
 What do you observe?
Both numerics and graph show that there is a tendency of
negative residuals to follow negative residuals and positive
residuals to follow positive residuals. This kind of behaviour is
consistent with an assumption of positive correlation between
successive residuals. If the errors are uncorrelated then any
such particular pattern is not expected. If errors are negatively
correlated then positive residuals would be followed by
negatives.
Recall the definition of sample correlation
cov̂(xt , yt )  (xt  x )( yt  y)
rxy  
var̂(xt ) var̂( yt ) ( x  x ) 2 ( y  y) 2
 t  t
 Sample autocorrelation coefficient
The definition of sample correlation coefficient
cov̂(xt , yt )  (xt  x )( yt  y)
rxy  
var̂(xt ) var̂( yt ) ( x  x ) 2 ( y  y) 2
 t  t
If xt and yt are replaced by ê and eˆ respectively then we
t t  1
obtain sample correlation between ê and eˆ .
T t t 1
 eˆt eˆ
cov̂(et ,e ) t 1
r t  1  t 2
1 var̂(et ) T 2
 eˆ
t 1
2
The summation starts from t=2 to accommodate the lagged
residual
Using the example we obtain r1=0.404, suggesting that the
residuals are moderately correlated. A hypothesis test of if r 1
is significantly different from zero is indeed needed to draw a
firm conclusion.
 Consequences of Autocorrelation

When the errors follow an AR(1) model et e vt the

t 1
least squares assumption of cov(e ,e ) 0 for t s is
t s
violated. The consequences are:
1. The least squares estimator is still a linear unbiased
estimator, but it is no longer best. It is possible to find an
alternative estimator with a lower variance.
2. The formulas for the standard errors usually computed for
the least squares estimator are no longer correct, and
hence confidence intervals and hypothesis tests that use
these standard errors may be misleading.
 Remedy
In presence of autocorrelation, standard errors obtained through
least squares estimation are incorrect.
Heteroskedasticity and autocorrelation consistent (HAC)
standard errors or Newey-West approach can be applied to
obtain the correct standard errors.
Message:
HAC standard errors are rather larger than the incorrect
standard errors. This implies if autocorrelation is not taken
into account, the reliability of usual least squares estimates
would be overstated.
Look at eviews file and check that HAC standard errors are
larger than the usual ones.
 How to estimate an AR(1) model?
Three procedures are followed:
1. Least squares estimation
2. Nonlinear least squares estimation
3. More general least squares estimation.

LEAST SQUARES ESTIMATION

While estimating e e v we can proceed with least
t t 1 t
squares estimation without recognising AR(1) error. The
estimate of above model is ˆ ˆ
et 0.40e
t 1
Estimate of original model is yˆt 3.8930.776xt
 2. NONLINEAR LEAST SQUARES ESTIMATION

As a better estimation process, we can employ nonlinear

estimation approach.
Consider the original model yt    xt  et
1 2
AR(1) error representation et e vt
t 1
From original model e y     x
t  1 t 1 1 2 t 1
e y     x
t 1 t 1 1 2 t 1
Plug the value of et into original equation and then e
t 1
yt    xt  y     x vt
1 2 t 1 1 2 t 1
yt  (1  )   xt  y   x vt
1 2 t 1 2 t 1
 Summary information
• we have transformed the original model with autocorrelated
error into a new model with uncorrelated error v t.
• Now proceed to find estimates for  ,  and  that
1 2 errors T
minimize the sum of squares of uncorrelated 2
Sv   v
t
t 2 T
• S  e 2
Minimizing the sum of squares of the correlated errors e 
t
yields the least squares estimator that is not the best, and t 1
whose standard errors are not correct.
• Minimizing the sum of squares of uncorrelated errors S
v
yields an estimator that is best and whose standard errors are
correct.
• The new model is nonlinear in parameters but they have the
usual desirable properties.
• The new model contains the lagged dependent variable,
current and lagged explanatory variables.
 3. GENERALIZED LEAST SQUARES ESTIMATION

Such approach requires the conversion of autocorrelated error

into uncorrelated errors. Conventionally this is known as
Cochrane-Orcutt procedure.
Get introduced with ARDL model (Autoregressive distributed lagged)
The transformed model
yt  (1  )   xt  y   x vt.......(9.26)
1 2 t 1 2 t 1
can be further simplified into
yt   xt  x  y vt......... (9.27)
0 1 t  1 1 t 1
Which is in the class of autoregressive distributed lagged model.
The latter is a restricted version of the former with the restriction
   
1 10
If this restriction is imposed then the number of parameters
reduces from four to three and (9.27) is equivalent to the
AR(1) model.
 Features of ARDL model
• ARDL model in (9.27) can be estimated by applying least
squares if vt satisfies the usual least squares assumptions

• The presence of the lagged dependent variable means that a

large sample is required for the desirable properties of the
least squares estimator to hold, but the least squares
procedures is still valid. If v t are correlated, the least squares
estimator would be biased, even in large samples.

• The model does not have a lagged error term but they do
have a lagged dependent variable and a lagged explanatory
variable. The dynamic features of a model implied by an
AR(1) error can be captured by using instead a model with a
lagged y and a lagged x.
 How to test for Autocorrelation?

Three testing procedures may be applied:

1. Residual correlogram
2. Lagrange multiplier test
3. Durbin-Watson test.
 Residual Correlogram
The objective is to check whether the errors of the regression
model yt    xt et are correlated or not. The
1 2
AR(1) error model: et e vt
t 1
If vt are assumed to be normally distributed with zero mean,
constant variance and uncorrelated then it follows e t would be
uncorrelated if  equal to zero. Then e t=vt.
Thus, in the context of of the AR(1) error model the null and
alternatives for the test of autocorrelation are H0 :  0
and H1:  0 respectively.

Eviews file chp_9 gives us the least squares residuals êt

and those residuals lagged by one period eˆt  1
which we call r1=0.4. This value is an estimate of the

autocorrelation coefficient . Thus, one way to test for
autocorrelation is to examine whether r 1=0.4 is significantly
 Test for autocorrelation
Test statistic Z  T r ~ N (0,1) where T indicates
sample size.
1
In our example, T=34, r1=0.404
Z=2.36>1.96, therefore reject null of no autocorrelation,
conclude that the errors are autocorrelated.
The sequence of correlations rk, k=1,2, ... .. , is called the
sample autocorrelation function or correlogram of the
residuals. It shows the correlation between residuals that
are one period apart, two periods apart, three periods
apart, and so on. These correlations are estimates of the
 ,  ,  ....
population correlation function given by 1 2 3
Where,
cov(e ,e ) E(e e )
  t t k  t t k
k var(e ) 2)
t E (e
 How to use software to investigate autocorrelation?
:  0 against the alternative
• In order to test the null hypo H
H :  0 test statistic is 0 Z  T r ~ N (0,1)
1 1
T r 1.96 or T r  1.96
Reject null when 1 1 1.96 1.96
which is equivalent to
cov(et ,es ) 0 for t s r  ; r  
1 T 1 T
This can be generalized to all lags. If rk is the correlation between et
and et-k - the correlation between residuals that are k-periods apart
then rk is significantly different from zero at 5% level of significance if
1 . 96
r 1.96 or
r 
k
k T T.

Econometric software automatically provides a correlogram with the

bounds 1.96/ T
 Correlogram
the correlations up to six lags are r 1=0.404, r2=0.122, r3=0.084,
r4=-0.353, r5=-0.420 and r6=-0.161 and the bounds are
1.96/ T 0.336
Eviews_chp9 provides correlogram, corresponding to the
function yt 1 2xt  et,with correlations along horizontal
axis and lags along vertical axis. Look at correlogram series in
eviews.

If there is correlation then a reasonable strategy is to estimate an

AR(1) error model et e vt and then estimate the
t 1
derived model
yt  (1  )   xt  y   x vt
1 2 t 1 2 t 1
If the assumption of vt is uncorrelated is reasonable one then the
residuals from this regression would exhibit very small spike
or no spike in period one.
 Correlogram

Lag Length

correlation
 0.34 0.34
 2. Lagrange Multiplier Test
LM test is conducted to test for autocorrelations. The general
version is to estimate the equation y    x  eˆ vt
t 1 2t t 1
and test for the null H0 :  0
Proceeding in this way generates t=2.382, p-value=0.02.
Since p-value is less than 0.05, LM test rejects the null
hypothesis of no autocorrelation at 5% level of significance.
 Durbin-Watson Test
Durbin-Watson test is attributable to the AR(1) model
et e vt
t 1
Where, vt are independent random errors N (0, v2)

Durbin-Watson test examines any positive autocorrelation

between errors, i.e., H :  0 against H :   0
0 1
Test statistic, d 2  2r 2(1 r )
1 1
r1=0 leads to d=2, indicating that the model errors are not
autocorrelated.
If r1=1, d=0; this implies a low value of Durbin-Watson statistic
implies that the model errors are correlated.
 D-W test
• Sample correlogram and Lagrange
multiplier tests are large sample tests;
their test statistics have their specified
distributions in large samples. D-W test
does not rely on large sample
approximation. It was developed in 1950
and, for a long time, was the standard test
for null hypothesis H0:  0 in the AR(1)
error model et et  1vt .
 • The test statistic used for testing the null
against the alternative
H :  0 H :  0
0 1
ˆ ˆ 2
 (et  e )
d t 1
ˆ
e 2
 t
(eˆ  ˆ
e ) 2 2 ˆ
e 2 2 eˆ eˆ
 t  eˆ   t
d t  1  t  t  1  t 1
ˆ
e 2 eˆ 2 eˆ 2 eˆ 2
 t  t  t  t
d 2  2r 2(1 r )
1 used less
This test is 1 frequently today because its
distribution no longer holds when the equation contains a
lagged dependent variable.
 What has been covered so far
• A particular model for lagged error term- the first
order autoregressive model.
• The simple regression model with AR(1) error
term expressed as a model with lagged
explanatory variable and lagged dependent
variable- the ARDL model.
• The AR(1) model can be viewed as a special
case of ARDL model.
• AR(1) model is convenient in many ways
including its scope to address ARDL model and
test for autocorrelation.
 Why do we estimate dynamic relationship?

1. To forecast future values of dependent

variable and
2. To trace out the time path of the effect of
a change in explanatory variable on a
dependent variable.
Forecasting: Autoregressive Models.
• The forecasting of values of economic
variable is a major activity for many
institutions including firms, banks,
governments and individuals.

• Here we describe how an autoregressive

models can be used for forecasting.
Autoregressive model of order p: AR(p)

• An autoregressive model of order p,

denoted as AR(p), is given by
yt   y  y .... p yt  p vt (9.36)
1 t 1 2 t 2
• Features of this model
 Features of AR(p) model
• The current value of a variable yt depends on its
values in the last p periods and a random error that
has Zero mean; Constant variance and
Uncorrelated over time
• The order of the model p is equal to the largest lag
of y on the right hand side of the equation.
• There are no explanatory variables. The values of y t
depends only on a history of its past values and no
x’s.
• While forecasting we use the current and past
values of a variable to forecast its future value.
• The model relies on correlations between values of
the variable over time to help produce a forecast.
 Example
• Suppose you have 433 observations on CPI of
India from January1975. You are interested in
forecasting inflation. The dataset is in the file
CPI.
• You can compute inflation rate yt as
[ln(CPIt)-ln(CPIt-1)]x100
• Rate of change in a variable can be
approximated by a change in the logarithm of
that variable.
• You have now 432 observations on inflation.
 How to forecast?
• Our concern is now to estimate the model.
Inclusion of lagged dependent variable
lowers the chance of autocorrelation
• First estimate the model using p=3.
• INFLATION= 0.3130+0.4736INFLATIONt-1 -0.0461INFLATIONt-2+0.054INFLATIONt-3

• Having three lags means that the equation

is estimated for t = 4,5, … 432
• That means you have 429 estimated
values.
• If you start from t=1 then you will need
inflation0, inflation-1 and inflation-2 . These
presample information are not available.
• For all observations to be available for
estimation, we begin from t = p+1.
• Look at the correlogram. Most of the
estimated autocorrelations are not
significantly different from zero.
How to use the model for forecasting?

• The model that has been estimated

yt   y  y  y vt
1 t 1 2 t 2 3 t 3

• Suppose you have observed yT and you

want to forecast yT+1, yT+2 and yT+3.
2011:M1=yT=1.6085
2010:M12=yT-1 = 1.6348
2010:M11 = yT-2 = 0.5511
yˆ ˆ ˆ y ˆ y ˆ y
T 1 1 T 2 T1 3 T 2
• Forecasted inflation in 2011:M2 (yT+1)=1.028
 Now you have

2011:M2=yT+1= 1.028 (check the result)

2011:M1=yT=1.6085
2010:M12=yT-1 = 1.6348
2010:M11 = yT-2 = 0.5511

You want 2011:M3=yT+2

yT+2 = 0.313+0.4736yT+1-0.0461yT+0.054yT-1
Result 0.4943 (check)
Now you have yT+2
you need yT+3
 To note
• There is a difference in finding yˆT 1 and yˆT 2

• While finding yˆ , one has the values of

T 1
yT, yT-1 and yT-2 available. But in finding yˆT 2
use is made of forecasted yT+1, i.e., yˆT 1.
 Forecast error
• Forecast error at time T+1 is u1, where
u y  yˆ
1 T 1 T 1
  y  y  y v
1 T 2 T  1 3 T  2 T 1
 ˆ  ˆ y  ˆ y  ˆ y
1 T 2 T1 3 T 2
(  ˆ) (  ˆ ) y (  ˆ ) y (  ˆ ) y v
1 1 T 2 2 T  1 3 3 T  2 T 1
Forecast error for 1 month ahead
• Convention is to ignore the error from
estimating coefficients.
• That means, we accept  ˆ and  ˆ
i i

• Doing so means we can write forecast

error for 1 month ahead as u1vT 1
Forecast error for 2 months ahead
• For 2 months ahead forecast error (u2)
gets more complicated because we have
to allow for not only vT+2 but also for the
yˆ
error that occurs from using T 1 instead
of yT+1.
 u2
• We have u y  yˆ
2 T 2 T 2
y   y  y  y v
T 2 1 T 1 2 T 3 T  1 T 2
yˆ ˆ ˆ yˆ ˆ y ˆ y
T 2 1 T 1 2 T 3 T  1
• Therefore,
u y  yˆ
2 T 2 T 2
  y  y  y v
1 T 1 2 T 3 T  1 T 2
 ˆ  ˆ yˆ  ˆ y  ˆ y
1 T 1 2 T 3 T  1
 ( y  yˆ ) v  u v  v v
1 T 1 T 1 T 2 1 1 T 2 1 T 1 T 2
For three periods ahead, forecast error u3

For one period ahead u v

1 T 1
For two periods ahead u  u v
2 1 1 T 2
For three periods ahead
u  u  u v
3 1 2 2 1 T 3
 ( u v )  v v
1 1 1 T 2 2 T 1 T 3
 2v  v  v v
1 T 1 1 T 2 2 T 1 T 3
( 2  )v  v v
1 2 T 1 1 T 2 T 3
 Forecast error variance
• Since vt’s are uncorrelated, covariance
between two periods’ v would be zero.
var(u )  2 var(v )  v2
1 1 T 1

2
var(u )  var( v v



2 
) 1  v2


2 2 1 T 1 T 2 


1 

var(u )  2 var[( 2  )v  v v ]
3 3 1 2 T 1 1 T 2 T 3
[( 2  )2  2 1] v2
1 2 1
Standard error and confidence interval

• Using coefficient estimates (ˆ,ˆ ,ˆ ,ˆ ) and

1 2 3
standard error of regression ˆv we can
estimate standard errors of forecast error ˆ j

• 100(1-α)% confidence interval for a

forecast of yT+j would be yˆ tcˆ
j
T j
 Exercise
• Construct 95% confidence interval for yT+3
9.6 Finite Distributed Lagged Model
• We discussed how having past values of a
dependent variable, the future values can
be forecasted. Now turn to a different
question. Is the impact of an explanatory
variable on the dependent variable
instantaneous or lagged.
 Example
• Say, inflation is impacted by wage
changes. The relationship between wage
changes and inflation is unlikely to be a
totally instantaneous one; it takes time for
wage changes to impact on inflation.
• One way of modeling the impact of wage
changes on inflation is through a finite
distributed lag model.
 The Model
Assume
• Inflation is yt
• Wage change is xt
• The relation is linear
• Wage changes up to q months into the
past have an influence on the current rate
of inflation. Then the model appears as
yt    xt   x   x ...... q xt  q vt t q 1, ...T
0 1 t 1 2 t 2
 Assumptions
• E(vt)=0, var(vt )  v2 , cov(vt, vs)=0
• If we have T observations on the pairs
(xt, yt) then only T-q complete observations
would be available for estimation since q
observations are lost in the creation of xt-1,
xt-2, …. , xt-q.
 Example
• If T=20, q=3 then the model to be
estimated is
yt    xt   x   x   x vt
0 1 t 1 2 t 2 3 t 3

You will need to create xt-1, xt-2 and xt-3.

Since t starts from q+1=3+1=4, the above
model becomes
y    x   x   x   x v
4 0 4 13 2 2 31 4
 Manipulation
• If four periods are lagged behind, then the
model becomes
y    x   x   x   x v
0 0 0 1 1 2  2 3  3 0
• The original model with q=3
yt    xt   x   x   x vt
0 1 t 1 2 t 2 3 t 3

E( y )
t 
s
x
t s
 Full Model
yt    xt   x   x ...... q xt  q vt t q 1, ...T
0 1 t 1 2 t 2
• This model is called distributed lagged
model because the effect of changes in xt
is distributed over time.
• This is called finite distributed lag model
because after a finite number of periods,
q, changes in x will have no impact on y.
• The parameter  is the intercept and s
called distributed-lag weight or s-period
delay multiplier. E( y )
t 
s
• Interpret s x
 Interpretation of s

• Suppose x and y have been constant for

at least the last q periods and that x
increases by 1 unit. Then as an immediate
effect the increase in yt is 0 units. One
period later yt+1 will increase by 1 units,
then yt+2 will increase by 2 units and so
on up to period t+q when yt+q will increase
q
by units. In period t+q+1 the value of y
will return to its original level.
 Impact multiplier 0
• If rate of growth of wage increases by 1%
in period t then returns to its original level,
there will be a s% increase in the inflation
rate in period t+s.
• Setting s=0 yields what is called impact
multiplier 0

• Try to explain  straightway.

0
 Interim Multiplier
• If xt increases by one unit and maintained
at its new level in subsequent periods
(t+1), (t+2), . . . ., the immediate impact
would be 0
• The total effect in period (t+1) would be   
0 1
• The total effect in period (t+2) would be     
0 1 2
and so on.
Sums of the the effects from the changes in all
proceeding periods are called interim multiplier.
For example two period interim multiplier is 0  1 2
 Total Multiplier/long-run multiplier
• The total multiplier is the final effect on y of
the sustained increase in x after q or more
periods have elapsed. Total multiplier can
be expressed as q
 s
s0
 Least squares estimates of finite distributed lag model
Forgot the file that I used.

Table 9.3
infln_wage
Proxy

Variable coefficient Std. error t-value p-value

Constant 0.1219 0.0487 2.505 0.013
Xt 0.1561 0.0885 1.764 0.079
Xt-1 0.1075 0.0851 1.264 0.207
Xt-2 0.0495 0.0853 0.580 0.562
Xt-3 0.1990 0.0879 2.264 0.024

• A one percent increase in wage growth

leads to an immediate increase in the
inflation of 0.16%, a 1-month lagged
increase of 0.11%, a 2-month lagged
increase of 0.05%.....
Lag-weights or Delay multipliers and Interim multipliers

Table 9.4 Multipliers

Multipliers
Lag
Delay Interim

0 0.1561 0.1561
1 0.1075 0.2636
2 0.0495 0.3131
3 0.1990 0.5121
Use of Finite Distributed Lag Model
1. Used for finding impact multiplier, delay multiplier and
total multiplier.
2. Used for forecasting. Values of x are used in the
estimated equation to forecast future values of y. Since
future values of x are unknown, it may be required to
independently predict these values before proceeding
to forecast y, with a consequent increase in the forecast
standard error.
3. Such models can be used for policy purposes. For
example, if x is a policy variable like interest rate or
taxation, a distributed lag model can be used to
forecast the effects of different interest rate or tax
policies.
 What covered

i. A model with lagged error term: AR(1) model

ii. Autoregressive model or forecasting model
with lagged values of yt (the AR(p) model)
iii. A model with lagged values of explanatory
variable xt (a finite distributed lag model that
shows how a change in xt has multiplier effects
over a number of periods.
iv. AR(1) model also has been transformed into a
model with yt-1 and xt-1.
 Not covered
• Have not explicitly covered a model that
contains both lagged values of xt and
lagged values of yt.
 ARDL Model
• An autoregressive distributed lag (ARDL) model
is one that contains both lagged x t’s and lagged
yt’s.
• With p lags in y and q lags in x, an ARDL(p,q)
model can be written as
yt   xt  x  x ... q xt  q  y  y ... p yt  p vt
0 1 t 1 2 t 2 1 t 1 2 t 2
Merits of ARDL model over finite distributed lag model

• The ARDL model overcomes two potential

problems with the finite distributed lag
model.
 First problem of choosing q
In the finite distributed lag model it is necessary
to choose a value for q, the point in the past
beyond which it is assumed changes in x no
longer have an impact on y. Although the above
formulated model shows that there are finite lag
up to q but the ARDL model can be
transformed into one with lagged xt’s going
back to infinite past. Look at the following
model. y    x   x   x   x ....e
t 0t 1 t 1 2 t 2 3 t 3 t

yt    s xt  s et
s0
 Infinite Distributed Lag Model

yt    s xt  s et
s0
• Because it does not have a finite cutoff point,
this model is called an infinite distributed lag
model.
• The parameter is the distributed lag weight or
s
the s-period delay multiplier showing the effect
of a change in xt on yt+s.
• The total or long run multiplier showing the long-
run effect of a sustained change in x t is  s
s0
 Multiplier of an infinite distributed lag model

• The effect of a change in x gradually dies out in

which case the values of s for large s will be
small and decreasing, a property that is
necessary for the infinite sum 
  to be finite. s
s0
• Estimates for the lag weights s can be found
from estimates of the 's and  's in the model.
j k
yt   xt  x  x ... q xt  q  y  y ... p yt  p vt
0 1 t 1 2 t 2 1 t 1 2 t 2
• The precise relationship between them depends
on the values for p and q.
 Second problem of autocorrelation
• The finite distributed lag model may suffer
autocorrelation problem. The inclusion of
lagged values of the dependent variable
can serve to eliminate such correlation.
 Third edition p. 251# Fig. 9.7
• The correlogram for the residuals from the
finite distributed lag model is portrayed in
figure 9.7. the relatively large
autocorrelation at lag 1 is significantly
different from zero at a 5% significance
level, and there are several correlations at
longer lags that are marginally significant.
These results suggest that the assumption
of uncorrelated errors vt is not a
reasonable one.
• As an alternative an ARDL(2,3) model has
been estimated. p=2 means two lags of
the dependent variable and q = 3 means
current value plus three lags of
independent variable appear on the right
side of the equation.
 Choice of p and q
• There is no unambiguous way of choosing
p and q.
• Two things to consider
One. The estimated coefficients are
significant
Two. The resulting residuals are
uncorrelated.
Importance of retaining early insignificant lags

• Sometimes earlier lags may look

insignificant nevertheless they have to be
retained because of significant later lags.
• For example, xt, xt-1 and xt-2 are
insignificant but xt-3 is significant. That’s
why they are retained together with xt-3.
• This process yields a correlogram like
figure 9.8 where a few marginally
significant correlations at long lag appears,
does not suggest autocorrelation.