private sector to know scal shocks one quarter in advance. For
notational simplicity, let us look at a two-variable VAR in T and
X, and ignore quarter-dependence:
(6) Tt 5 a1 Xt 1 A11~L!Tt21 1 A12~L! Xt21 1 et
t
(7) Xt 5 c0 EtTt11 1 c1Tt 1 A21~L!Tt21 1 A22~L! Xt21 1 etx
.
The rst equation is as before: taxes depend on output, current
and lagged, and lagged taxes, together with scal shocks. The
second equation now allows output in quarter t to depend not only
on current and lagged taxes, but also on the expectation of taxes
in quarter t 1 1, based on an information set which includes et11
t . In other words, this specication assumes that, because of imple- mentation lags, the private sector knows scal shocks one quarter
ahead, and may therefore react to them one quarter before they
are actually observable by the econometrician. To see the econometric problems this specication raises, rewrite the equation for output as
(8) Xt 5 c0Tt11 1 c1Tt 1 A21~L!Tt21 1 A22~L! Xt21 1 tx9, where the error term is dened as t
x9 [et
x 2 c0
(Tt11 2
EtTt11
)].
From the assumption that et11
t is known at time t, the term
in parentheses in the composite error term is uncorrelated with
et11
t but will typically be correlated with et11
x
. Hence, both Tt and
Tt11
in (8) are correlated with the composite error term. As in our
benchmark case, Tt is correlated with et
x
, through the effect of et
x
on output and in turn on tax revenues. Also, now Tt11
is correlated
with both components of the error term: with et11
x through
the effects of the latter on taxes in quarter t 1 1, and with et
x
,
through the effect of the latter on Tt and therefore on Tt11
. Can equation (8) be estimated by instrumental variables?
The answer is yes, if we are willing to make stronger identication
assumptions. The reasoning follows the logic of our previous
identication strategy. If we can construct a series for et
t, then et
t
and its value led once, et11
t , can be used as instruments for Tt and
Tt11
. Both are correlated with Tt or Tt11 and uncorrelated with
the two components of the composite error term. Put another
way, if we can identify et
t in the tax equation, we can then use it
and its value led once as instruments in the output equation. With this in mind, let us turn to the tax equation, and for