Autoregressive Distributed Lag (ADL) Model

Yi-Yi Chen

The regressors may include lagged values of the dependent variable and current and lagged values of one or more explanatory variables. This model allows us to determine what the eects are of a change in a policy variable. 1. A simple model: The ADL(1,1) model yt = m + 1 yt1 + 0 xt + 1 xt1 + ut , where yt and xt are stationary variables, and ut is a white noise. The White-noise process: A sequence {ut } is a white-noise process if each value in the sequence has a mean of zero, a constant variance, and is serially uncorrelated. The sequence {ut } is a white-noise process if for each period t, E(ut ) = E(ut1 ) = = 0 E(u2 ) = E(u2 ) = = 2 t t1 E(ut uts ) = E(utj utjs ) = 0, for all u. 2. Estimation: If the values of xt are treated as given, as being uncorrelated with ut . OLS would be consistent. However, if xt is simultaneously determined with yt and E(xt ut ) = 0, OLS would be inconsistent. As long as it can be assumed that the error term ut is a white noise process, or more generally-is stationary and independent of xt , xt1 , and yt , yt1 , , the ADL models can be estimated consistently by ordinary least squares. 3. Interpretation of the dynamic eect: We can invert the model as the lag polynomial in y as
2 2 yt = (1 + 1 + 1 + )m + (1 + 1 L + 1 L2 + )(0 xt + 1 xt1 + ut ).

The current value of y depends on the current and all previous values of x and 1

u.

yt = 0 , xt

This is referred as the impact multiplier. The eect after one period yt+1 = 1 + 1 0 , xt The eect after two periods yt+2 2 = 1 1 + 1 0 , xt The long-run multiplier (long-run eect) is
0 +1 11

if |1 | < 1.

4. Reparameterization: Substitute yt and xt with yt1 + yt and xt1 + xt . yt = m + 0 xt (1 1 )yt1 + (0 + 1 )xt1 + ut , yt = 0 xt (1 1 )[yt1 m 0 + 1 xt1 ] + ut . 1 1 1 1

This is called the error correction model (ECM). The current change in y is the sum of two components. The rst is proportional to the current change in x. The second is a partial correction for the extent to which yt1 deviated from the equilibrium value corresponding to xt1 (the equilibrium error). 5. Generalizations: The ADL(p, q) model: A(L)yt = m + B(L)xt + ut , with A(L) = 1 1 L 2 L2 p Lp , B(L) = 0 + 1 L + 2 L2 + + q Lq .

The general ADL(p, q1 , q2 , , qk ) model: A(L)yt = m + B1 (L)x1t + B2 (L)x2t + + Bk (L)xkt + ut , If A(L) = 1, the model does not contain any lags of yt . It is called the distributed lag model.