The finite distributed lag (FDL) model
allow for one or more variables to affect yt with a lag. For example, the model
yt=α0+δ0zt+δ1zt−1+δ2zt−2+ut
(1)
which is an FDL of order two.
To interpret the coefficients in
(1), lets zt
is a constant (or equal to c
for all time period before t ) .
At time t , let z increases by one unit to c+1 an then it reverts back at time t+1 . (We assume he
increase in z is
temporary);
..,zt−2=c,zt−1=c,zt=c+1,zt+1=c,zt+2=c,...
(2)
If we focus the effect of zt on yt and ceteris paribus, and we set the error term
each period to zero,
yt−1=α0+δ0c+δ1c+δ2c
yt=α0+δ0(c+1)+δ1c+δ2c
yt+1=α0+δ0(c+1)+δ1(c+1)+δ2c
yt+2=α0+δ0(c+1)+δ1(c+1)+δ2(c+1)
(3)
and so on.
With the permanent increase in zt after one period, yt has
increased by δ0+δ1. After two periods, yt has increased by δ0+δ1+δ2. There is no further changes in yt after two
period.
This show that the sum of the coefficients on current lagged
zt, δ0+δ1+δ2 is the long-run change in yt given a permanent increase in zt. It’s
called the long-run propensity (LRP)
or long-run multiplier. The LRP is
often interest in distributed lag models.
LRP=δ0+δ1+δ2 (4)
In Eq(1), if zt permanently
increase by one unit, then, after two years, yt will have changed by δ0+δ1+δ2. This model assume that there are no further
changes after two years. Whether this is a case is an empirical matter.
If we regress the Eq(1), we do not enough information from
Stata output to obtain the standard error (SE) for the estimated LRP=δ0+δ1+δ2 to
calculate the value of t-statistics
for LRP. We cannot just simply sum up all the SE for the LRP, (se(ˆδ0)+se(ˆδ1)+se(ˆδ2)) .
To get the value of SE for the LRP, we will use some trick.
Lets we donate
θ=δ0+δ1+δ2 (5)
as for LRP.
And then, write δ0 in
term of θ,
δ1
and δ2;
δ0=θ−δ1−δ2 (6)
Next, we substitute
for δ0, δ1 and δ1 in the Eq(1);
yt=α0+(θ−δ1−δ2)zt+δ1zt−1+δ2zt−2+ut
=α0+θzt−δ1zt−δ2zt+δ1zt−1+δ2zt−2+ut
=α0+θzt−δ1(zt−1−zt)+δ2(zt−2−zt)+ut (6)
From the last equation Eq(6), now
can obtain the ˆθ and also its value of SE by
regressing yt
on zt, (zt−1−zt)and (zt−2−zt).
For this regression, we need the
value of coefficient and SE for variable zt for the LRP.
Estimation with Stata
For the estimation , we use the fertil3.dta to estimate the model;
gfrt=β0+β1pet+β2pet−1+β3pet−2+β4ww2t+β5pillt+ut (7)
where
gfrt = general fertility rate (children born per
1,000 women child bearing age)
pet = real dollar value personal tax exemption
ww2t = United States involved in World War 2 during
1941-1945. 1= year involved, 0 = otherwise.
pillt = the year of introduced birth control pill
in year start from 1963 . 1= year start 1963, 0 = otherwise (before 1963).
To estimate
the Eq(7);
reg
gfr pe L.pe L2.pe ww2 pill
To calculate the estimated value
of LRP as in Eq(5);
display _b[pe]+_b[L1.pe]+_b[L2.pe]
To estimate the value of LRP along with their SE and the
- statistics, we need first to generate the value of (pet−1−pet) and (pet−2−pet)
gen dif1 = L.pe-pe
gen dif2 = L2.pe-pe
And now, we already to estimate the LRP as in Eq(6);
reg gfr pe dif1 dif2 ww2 pill
The regression result give ˆθ=0.1007 and must be same as (δ0+δ1+δ2)≈0.1007.
The SE for ˆθ or (δ0+δ1+δ2)=0.03 .
Therefore, the t -statistic for ˆθ is
about 3.38, so ˆθ is statistically different from zero at small significance level.
Even though none of the ˆδj
is individually significant, the LRP is very significant.