DISTRIBUTED LAG MODELS
A simple regression equation designed to explain variations over time in the mean value of the
dependent variable is given by:
Yt = + Xt + u t ,
where u t is a random variable with zero mean, and X t is either non-stochastic or if stochastic
is independent of u t .
In setting up the regression equation in this way one is assuming that the current value of Y
may depend on the current value of X but not on the past values of X. A more general
formulation which would allow for the current as well as the past values of X to affect Y, can
be written as:
Yt = + 0Xt + 1Xt −1 + 2Xt −2 +...+mXt −m + u t (1)
For simplicity we shall limit ourselves to the case of non-stochastic X.
The regression equation (1) extended by the relevant assumption concerning X and u t is called
a distributed lag model because the influence of the explanatory variable on E[Yt ] is
distributed over a number of lagged values of X. The number m may either be fixed or infinite.
(1) can (at least in principle) be estimated by the principle of least squares. Of course, if m is
large, we may not have enough observations to estimate all the parameters. However, if we do
have enough observations , we are likely to encounter high degree of multicollinearity, which
would have detrimental effect on the standard error of estimated coefficients. As a matter of
fact, a distributed lag model has rarely been estimated in as general a form as that specified in
(1). Most frequently, some restrictions are placed on the regression coefficients
0 , 1, 2 ,...,m so that the number of regression parameters become substantially reduced.
In practical applications, these restrictions have been of two kinds-
• One resulting from the requirement that the 's should be declining in GP (Geometric
Progression) and the other from the requirement that the
• 's should first be increasing and then degreasing.
Page 1 of 9
GEOMETRIC LAG MODEL
The most popular form of a distributed lag is geometric lag characterized by
( )
Yt = + 0 X t + X t-1 + 2 X t-2 +... + u t , (2)
where
i = 0 i , i = 0,1,2,... , 0 1.
Here the effect of X on E Yt extends indefinitely into the past (i.e., m → ), but the
coefficients decline in a fixed proportion so that the effect of the distant values of X eventually
becomes negligible.
(2)
( )
Yt-1 = + 0 X t-1 + X t-2 + 2 X t-3 +... + u t-1 ,
Yt - Yt-1 = (1 − ) + 0 X t + ( u t - u t-1 )
Yt = (1 − ) + 0 X t + Yt-1 + t ,
where
t = u t - u t-1
The model can be extended to the case of k independent variables X1 , X 2 ,..., X k
Yt = (1 − ) + 0 X t1 + 1X t2 + ...+k −1X tk + Yt-1 + t ,
where
t = u t - u t-1
This model provides a good approximation of business and industrial processes. A number of
theoretical explanation of why this should be so have been advanced.
Adaptive Expectations
For ease of exposition, we will restrict ourselves to the case where a dependent variable Y is
influenced by a single independent variable X.
Page 2 of 9
Yt = + X*t +1 + u t (5)
Since X*t+1 is not directly observable, we have to state how it is determined, i.e., some
assumption must be made about the way in which expectations are formed. Some intuitively
plausible possibility is to assume the process of adaptive expectation formation:
X*t +1 = (1 − ) X t + X*t , 0 1. (6)
According to this expression the expectation at time t+1 is the weighted average of the
expectation at previous time period and the actual value in the current time period.
(5) and (6) can be amalgamated to yield a model linking observable Yt and Xt .
To achieve this we substitute t-1 for t in (5) :
Yt −1 = + X*t + u t-1 (7)
( )
Yt - Yt −1 = (1 − ) + X*t +1 - X*t + ( u t - u t-1 )
= (1 − ) + (1 − ) X t + ( u t - u t-1 )
= (1 − ) + 0 X t + t
Yt = (1 − ) + 0 X t +Yt −1 + t
This is precisely the lagged dependent variable model.
Partial Adjustment
Consider again the case of a relationship between a dependent variable Y and a single
independent variable X. Suppose Yt be the inventory level at time t and X t be the sales at
time t. It may be desired to bring the actual level of Y to its desired level Yt* (say):
Model : Yt* = + X t
Such an attempt is partially successful due to technological constraints or institutional rigidities
or persistence of habit etc.. Since Yt −1 is the actual level in period t-1, an adjustment Yt* -
Yt −1 is required to achieve the desired level in period t.
Page 3 of 9
Suppose however the actual adjustment that can be made is on the average only a proportion
(1 − ) of the desired adjustment. We can then write
Yt - Yt −1 = (1 − ) Yt* − Yt −1( ) +u t , 0 1
= (1 − ) ( + X t − Yt −1 ) + u t
= (1 − ) + 0 X t - (1 − ) Yt −1 + u t
Yt = (1 − ) + 0 X t + Yt −1 + u t , t=1,2,...,m.
Thus, we have again arrived at the lagged dependent variable model.
Autocorrelated Errors In Regressions With Lagged Dependent Variables
Yt = (1 − ) + 0 X t +Yt −1 + t ,
where
t = u t - u t-1 ,
errors are assumed to follow first order autoregressive process, viz., t = t −1 + a t .
Durbin’s h-test:
This test is used to test for the presence of autocorrelation in the regression model with a lagged
dependent variable.
The regression equation is estimated by OLS method and the usual Durbin-Watson statistic d
is computed and autocorrelation parameter is estimated using
r = 1 − d / 2.
Durbin’s h-test statistic is then given by:
1/2
n
h = r ,
(
1 − nsc2 )
where
2
sc2 = ,
( Yt−1 -Yt−1 ) (1 − r )
2 2
Page 4 of 9
where
r 2 is the correlation coefficient between Yt and Yt-1 ,
n is the number of observations,
s c is the estimated standard error of the estimator of the parameter multiplying the lagged
dependent variable.
Under the hypothesis of no autocorrelation in the regression errors, this statistic has
approximately a standard normal distribution in moderately large samples.
Decision criteria:
• For testing
H0 : = 0
against
H1 : > 0
reject H0 if Cal. h > z ,
where z is the right tailed critical value of standard normal variate at level of
significance.
• For testing
H0 : = 0
against
H1 : < 0
reject H0 if Cal. h > -z ,
• For testing
H0 : = 0
against
H1 : 0
reject H0 if Cal. h > z /2
Page 5 of 9
Estimation Of Regression Model With A Lagged Dependent Variable
Suppose it is required to estimate a regression model with a lagged dependent variable and
errors are assumed to follow first order autoregressive process , i.e.,
Yt = + 0 X t1 + 1X t2 + ...+k X tk + Yt-1 + t ,
where
t = t −1 + a t .
Yt − Yt-1 =
(1 − ) + 0 ( X t1 − X t-1,1 ) + 1 ( X t 2 − X t-1,2 ) + ...+
k −1 ( X t k − X t-1,k ) + ( Yt-1 − Yt-2 ) + a t .
Select a grid of possible values for the autocorrelation parameter . For each chosen value
regress
( Yt − Yt-1 ) on ( X t1 − X t-1,1 ) , ( X t 2 − X t-1,2 ) ,..., ( X t k − X t-1,k ) , ( Yt-1 − Yt-2 )
By OLS method. The fitted regression is chosen with the smallest sum of squared errors:
S = Yt* - Y( )
2
ˆ* ,
t
t
where
Yt* = Yt − Yt-1 ,
is the chosen value,
Ŷt* =
ˆ (1 − ) + ˆ 0 ( X t1 − X t-1,1 ) + 1 ( X t 2 − X t-1,2 ) + ...+ˆ k −1 ( X t k − X t-1,k ) + ( Yt-1 − Yt-2 ) .
Page 6 of 9
ALMON APPROACH TO DISTRIBUTED LAG MODELS
The Almon Polynomial Lag
Geometric lag model is based on the assumption that the coefficients decline geometrically
as the lag lengthens. This assumption may be too realistic in practice.
In some cases s' may first increase and then decrease , in other cases s' may follow cyclical
pattern. In such cases geometric lag scheme of distributed lag model will not work. In such
cases we express i as a function of I (length of the lag i) and fit some suitable curve to reflect
the functional relationship between the two. This approach has been suggested by Stirley
Almon.
Consider the model:
Yt = + 0Xt + 1Xt −1 + 2Xt −2 +...+k Xt −k + u t
Which may be written more compactly as:
k
Yt = + i X t −i + ut .
i =0
The lag scheme is quadratic ,i.e.,
i = a 0 + a1i + a 2i 2
or third degree polynomial
i = a 0 + a1i + a 2i 2 + a 3i3 .
More generally
i = a 0 + a1i + a 2i 2 + a 3i3 + ... + a mi m .
To explain how te Almon scheme works, let us suppose that
i = a 0 + a1i + a 2i 2 , (8)
Then
Page 7 of 9
( a 0 + a1i + a 2i2 ) X t −i +
k
Yt = + ut
i =0
k k k
= + a 0 X t −i +a1 iX t −i +a 2 i 2 X t −i + u t
i =0 i =0 i =0
= + a 0 Z0t + a1Z1t + a 2 Z2t + u t
In the Almon’s scheme:
Y is regressed on the constructed variables Z, not on the original X variables. The estimates of
and a i thus obtained will have all the desirable statistical properties provided u t satisfies all
the assumptions of the CLRM. Once the as’ have been estimated the original s' can be
estimated from (8) as follows:
ˆ 0 = aˆ 0 ,
ˆ 1 = aˆ 0 + aˆ 1 + aˆ 2 ,
ˆ 2 = aˆ 0 + 2aˆ 1 + 4aˆ 2 ,
ˆ = aˆ + 3aˆ + 9aˆ ,
3 0 1 2
ˆ k = aˆ 0 + kaˆ 1 + k 2aˆ 2 .
Before we apply the Almon technique we must resolve the following practical problems:
• Max. length of the lag ‘k’ must be specified in advance. In practice one hopes that ‘k’
is reasonably small. For example, for annual data of 10 years ‘k’ should be not more
than 2 or 3. In any event the researcher must decide the maximum value of ‘k’.
• Having specified ‘k’ , the degree ‘m’ of the polynomial must be specified. Suppose we
must decide between 2nd and 3rd degree polynomial. For the 2nd degree polynomial the
estimating equation is given by:
Yt = + a 0 Z0t + a1Z1t + a 2 Z2t + u t .
For the 3rd degree polynomial
Yt = + a 0 Z0t + a1Z1t + a 2 Z2t +a3Z3t + u t ,
where
k
Z3t = i3X t −i .
i =0
Page 8 of 9
After running regression if we find a 2 is statistically significant but a 3 is not, we may
assume that the second degree polynomial provides a good approximation, provided
the problem of multicollinearity is taken into account. This is because it may be possible
that a 3 is significantly insignificant not because true a 3 is zero but because the sample
in hand does not allow us to assess the individual impact of Z3 on Y.
References
Dougherty, C.(2002). Introduction to Econometrics, Oxford University Press, 2nd ed..
Granger, C. W. J. and P. Newbold (1986) Forecasting economic time series, Academic
Press, New York, 2nd ed.
Page 9 of 9