Notes on the identification of VARs using external

instruments
Michele Piffer∗
June 2, 2020

These notes derive and discuss the methodology proposed by Stock and Watson
(2012) and Mertens and Ravn (2013) to identify Structural VAR models.

∗
DIW Berlin, Mohrenstrasse 58, 10117 Berlin, Germany. Email: [email protected], personal
web page: https://sites.google.com/site/michelepiffereconomics/. These notes can be reproduced
freely for educational and research purposes as long as they contain this notice and are retained for
personal use or distributed for free. All errors are mine. Please get in touch if you find typos or
mistakes.

1
1 Preliminary
The identification of structural shocks within Vector Autoregressive (VAR) models
usually consists of obtaining the B matrix in the model

y t = Π(L)y t−1 + Bst

where y t is a vector of n endogenous variables and st is a vector of n structural

shocks (bold characters indicate vectors).1
Recently, Stock and Watson (2012) and Mertens and Ravn (2013) proposed an
approach to identify one or more structural shock included in st using external in-
struments, i.e. using some measure, not included in the VAR, correlated with the
unobserved shock of interest.
Suppose for instance that we want to identify a monetary shock. Then, if we
are interested in impulse responses, identification consists of obtaining the column
vector b of B that corresponds to smonetary
t . Formally,

y t = Π(L)y t−1 + bsmonetary

t + B∗ sothers
t

where smonetary
t is the structural monetary shock, sothers
t is a vector including the
non-monetary shocks and B∗ is a matrix containing the n − 1 column vectors corre-
sponding to such shocks. If instead we are interested in estimating the realizations
of the monetary shocks, then we are after the row vector at (written as a column
vector) of the matrix A = B −1 , which maps the reduced form shocks r t = Bst back
into structural shocks, i.e.
smonetary
t = a0t r t
These notes discuss and derive such an approach of obtaining b and a. The main
case considered is the case in which we want to identify only one structural shock, and
for such shock we only have one instrument. The literature has discussed also other
possibilities. For example, Mertens and Ravn (2013) identify two structural shocks
using two correlated instruments. Stock and Watson (2012), instead, identify several
shocks of interests, and for each structural shock they consider different candidate
instruments. For simplicity, the key intuition of the approach is derived here using
the most simplified model. Most of the intuitions are outlined using a VAR with
2 variables, while the extension to a VAR with n variables will be briefly discussed
during the exposition. Some derivations are available in Olea et al. (2012).
1
An informal introduction to these models is available on
https://sites.google.com/site/michelepiffereconomics/VARs.pdf?attredirects=0d=1.

2
2 Estimating Impulse Response Functions
Consider the reduced form VAR given by

y1,t = Π1 (L)yt−1 + r1,t

y2,t = Π2 (L)yt−1 + r2,t

with

r t = Bst
= b1 sa,t + b2 sb,t
    
r1,t b11 b12 sa,t
=
r2,t b2,1 b22 sb,t
   
b b
= 11 sa,t + 12 sb,t
b21 b22

and V (st ) is known to equal the identity matrix. The reduced form model delivers
a set of covariance restrictions,
Σ = BB 0
The reduced form shocks r1,t and r2,t correspond to variables y1,t and y2,t , respec-
tively. The structural shocks sa,t and sb,t , instead, are left intentionally unspecified
and do not correspond to a defined ordering of the variables, i.e. they are not
necessarily a s1,t and a s2,t shock. To understand why this distinction is relevant,
consider two examples. In a VAR with output and the federal funds, it makes sense
to name the structural shock in the structural equation of the federal funds rate as
the monetary policy shock. In a VAR with quantity and price, it makes more sense
to interpret the structural shocks as demand and supply shocks, and not as a price
and a quantity shock. We will get back to this later.
Suppose we want to conduct structural analysis on the shock sa,t . We have a
variable mt such that

E(mt sa,t ) = φ
E(mt sb,t ) = 0

Note that we do not specify the relationship between mt and sa,t , but only the
expected value of the product, i.e. the covariance, given that structural shocks have
expected value equal to 0. mt could be some complicated function of sa,t , in which
case the covariance φ would be somehow determined by the parameters in such
function. Nevertheless, as long as there is some correlation between sat and mt we
do not need to determine where this correlation comes from.

3
 The key equality to exploit is

E(r t mt ) = E[(b1 sa,t + b2 sb,t )mt ]

   
b11 b
= E(mt sa,t ) + 12 E(mt sb,t )
b21 b22
 
b
= 11 φ (1)
b21

If we had the true value of φ and the population values of the vector E(r t mt ), we
could recover the true values of b11 and b21 simply as bi1 = E(ri,t mt )φ−1 . Similarly,
define the ratio
b21
µ1 =
b11
If we had the population vector E(r t mt ) but no value for φ, we could recover the
true ratio µ1 as
E(r2,t mt )
µ1 =
E(r1,t mt )
If instead we do not have φ nor the population value of E(r t mt ) we might still aim
for an estimation of b11 , b21 and or b21 /b11 . The rest of these notes operate under this
scenario, which is the only realistic one in applied work. We assume we have data
on mt and some estimates of r t from the estimation of the reduced form model.
There are many ways of exploiting the information provided by the instrument
mt in order to estimate bij . The first step is to get µ1 .

1. Method of Moments: Given the sample equivalent

PT
ri,t mt
Ê(ri,t mt ) = t=1
T
for the unknown population moment E(ri,t mt ), we can estimate µ1 as

Ê(r2,t mt )
µ̂1 =
Ê(r1,t mt )

The poorer are our estimates for the population moments Ê(ri,t mt ), the poorer
is our estimate for µ1 .

2. Single Regressions: Given the preliminary regressions

r1,t = δ1 mt + 1
r2,t = δ2 mt + 2

4
 we can estimate µ1 as
δ̂2
µ̂1 =
δ̂1
The corresponding estimate coincides with the one from point 1 because, from
OLS estimation,
Ê(ri,t mt )
δ̂i = (2)
Ê(m2t )
Hence, the term Ê(m2t ) drops when taking the ratio δ2 /δ1 , delivering the same
estimate for µ1 .
3. 2SLS : Given the regression
r2,t = γ1 r1,t + ηt
one can estimate µ1 with γ1 , obtained using 2 Stages Least Squares estimation
with mt as instruments for r1,t . Even this estimate coincides (by construction)
with the one from point 1 (and hence also point 2) because the fitted values
from the first step would be
Ê(r1,t mt )
r̂1,t = · mt
Ê(m2t )
so the second step yields
Ê(r1,t mt ) Ê(r1,t mt )
Ê(r2,t · r̂1,t ) Ê(r2,t · Ê(m2t )
· mt ) Ê(m2t )
Ê(r2,t · mt ) Ê(r2,t mt )
γ̂ = = = =
2
Ê(r̂1,t ) Ê(r1,t mt )
2 Ê(r1,t mt )
2 Ê(r1,t mt )
Ê Ê(m2t )
· m2t ) Ê(m2t )
Ê(m2t )

One might wonder why engaging in doing a 2SLS estimation when the results
are identical with the other two approaches. One of the answers is that the
2SLS simplifies things when one has more than one instrument per shock of
interest. For example, if one has one shock of interest and 2 instruments,
he or she can implement a GMM estimator with the 2SLS procedure by first
regressing the VAR residuals in one equation on all instruments, and by then
regressing the remaining VAR residuals on the fitted value obtained from the
preliminary regression.
In short, given some data on an instrument mt for the structural shock sa,t , we
cannot immediately estimate the single elements of the column of the B matrix
corresponding to the shock sa,t , but we can at least consistently estimate the ratio
µ1 = b21 /b11 . Note that the covariance restrictions included in Σ have not been used
so far.

Given an estimate of µ1 , the impulse responses of variables y1,t and y2,t to the
structural shock sa,t can be computed in the following ways:

5
 1. combine the information b21 = µ̂1 b11 with the covariance restrictions Σ = BB 0
of the reduced form model, solve for b1 , compute impulse responses using

absolute impulse vector = b1 · 1

Given our knowledge of V (st ) = I in population, this gives the impulse response
to a one standard deviation shock. Note that solving for b1 instead of µ1 is not
immediate, we will discuss it further below.

2. Without solving for the full vector b1 , compute impulse responses starting from
 
1
relative impulse vector = ·1
µ̂1

This cannot be interpreted as the impact effect of a one standard deviation

shock, since it would force the effect of a one standard deviation shock to have
an effect on y1,t equal to 1. Instead, we can interpret it as the effect of a shock
of the size that implies an increase of 1 in y1,t . One can give a shock of size
different from 1 so that the impulse vector generates, say, a decrease of y2,t of
2. For example, if one of the variables included is the federal funds rate, the
shock can be calibrated to imply, say, a decrease of the federal funds rate by
100 basis points. Note that using the above impulse vector is equivalent to
using directly the impulse vector
 
δ̂
inconsistently-estimated impulse vector = 1
δ̂2

from point 3 above and scaling again the size of the shock as desired. We refer
to this vector as the inconsistently-estimated impulse vector since it delivers
φ
b1 up to the scale factor E(m 2) .
t

Before continuing, it can be useful to remark that the definition of µ1 was arbi-
trary. In principle, we could compute impulse responses also using the ratio
b11
µ2 = = µ−1
1
b21
and then  
µ̂2
relative impulse vector =
1

To estimate µ2 we can compute the ratio Ê(r 1,t mt )

Ê(r2,t mt )
, compute the ratio δ̂δ̂1 or estimate
2
the coefficient γ2 in r1,t = γ2 r2,t + ηt using mt as an instrument for r2,t . In principle,
these ratios are equally valid, and (in population) simply deliver the inverse of the
ratios computed above for µ1 . Nevertheless, in applied work we might want to
reduce the possibility of dividing by a number that is zero in population. Since we

6
do not know the true values b11 and b21 , we might want to define the ratio µ as to
minimize the risk of dividing by a parameter close to zero. For example, if we are
after uncertainty shocks, we might want to define µ by dividing the vector b by bi ,
where i is the equation of the measure of uncertainty included in the VAR. This
would work under the assumption that an uncertainty shock is likely to have a non
zero impact on the uncertainty measure. In general, estimating the inconsistently-
estimated impulse vector does not require taking a stand on which equation will be
used in the normalization for the definition of the relative impulse vector. If instead
we are after the absolute impulse vector, we need to choose which variable to use to
define the relative impulse vector. Asymptotically this choice is irrelevant, in finite
sample not.
The generalization of the estimation of the relative impulse vector to a VAR
models with n variables is immediate and will be omitted. The only thing that
changes is that the scalar µ becomes a vector µ.

3 On the inclusion of constant terms

In the previous section we mentioned that if we had the true population moments
E(ri,t mt ) we could compute the ratio µ1 without any estimation error. If instead
we do not have the population moments for E(ri,t mt ) but only some finite sample,
then we can compute the ratio of the sample equivalents using either the Method of
Moments approach outlined or an appropriate set of regressions. Such regressions did
not include a constant term. Let us now consider whether these regressions should
actually include a constant.
If we specify the Single Regressions approach outlined above adding a constant
we would be estimating the models

r1,t = const1 + δ1 mt + 1
r2,t = const2 + δ2 mt + 2

In principle, the ratio δ2 /δ1 is not necessarily equivalent to the same ratio computed
from the regressions in the previous section. This turns out to be irrelevant. In fact,
the estimator for δi would now be combining the data as

C ôv(ri,t mt ) Ê(ri,t mt ) − Ê(ri,t )Ê(mt )

δi = = (3)
V̂ (m2t ) Ê(m2t ) − Ê(mt )2
where again ˆ indicates the sample equivalent. Compare now equations (3) and (2).
The denominator is clearly different (the instrument mt is not necessarily in mean
zero). This is not a problem because the denominator drops when taking the ratio
δ2 /δ1 . The numerator, instead, does not differ because the reduced form shocks ri,t
are by construction in mean zero as long as the estimated VAR included a constant

7
(and as long as the regressors are not close to multicollinear, otherwise the inverse
of X 0 X would be imprecisely calculated).2 Hence the term Ê(ri,t )Ê(mt ) equals zero.
This implies that the inclusion of a constant term in the single regression approach
is irrelevant: it would affect selected estimates, but not the ratio that we are after.
This holds in finite sample and does not rely on any asymptotic result.
Consider now the other approach, i.e. the 2SLS approach. When including a
constant, the regression in the first stage is

r1,t = const1 + δ1 mt + 1

from which we compute the fitted values

r̂i,t = con̂st1 + δ̂1 mt

which we then use for the OLS regression

r2,t = const2 + γ1 r̂1,t + ηt

The derivations will be skipped, but the results from the previous section hold, i.e.
the estimate of γ1 coincides with the estimate of the ratio δ2 /δ1 . It follows that,
again, the inclusion of a constant term is irrelevant, as long as it is included in the
regressions of both the first and the second step of the 2SLS.

4 Tests on the strength of the instrument

While the inclusion of a constant does not change the estimate of the ratio µ1 , it
does affect the tests on the strength of the instruments, at least in finite sample.
Let us take a step back. In the standard application of instrumental variable
estimation, one starts from a regression of the type yi = βxi + i and instruments
xi with zi . The instrument zi is valid for the endogenous instrument xi if zi is not
correlated with the error term i . The instrument is instead relevant if it is correlated
with the regressor xi . This means that the relevance condition can be tested, while
the validity condition remains untestable.
In the application of instrumental variable estimation to VARs, the validity con-
dition of the instrument(s) jointly refers usually to two subconditions. First, that
the instrument is correlated with the shock of interest, and second that it is uncorre-
lated to the remaining shocks. These two conditions are usually referred to relevance
(E(mt sa,t ) 6= 0) and exogeneity (E(mt sb,t )). Note that the relevance condition moves
from being a testable to a non-testable conditions, given that in the VAR application,
the shock of interest is not observed.
2
Pn
Remember that in the OLS regression yi = a+bxi +i , the OLS estimator minimizes i=1 (yi −
n n
a − bxi − i )2 and that the first order conditions imply −2 i=1 (yi − â − b̂xi ) = −2 i=1 ˆi and
P P
Pn Pn
−2 i=1 xi (yi − â − b̂xi ) = −2 i=1 xi ˆi . In our notation, ri,t = ˆi .

8
 While the relevance and exogeneity conditions are not testable, we can test if
there is a strong correlation between the instrument and the VAR innovations. In
the notation used so far, a strong relationship between the instrument mt and the
reduced form innovations ri,t is a necessary condition for mt to be considered a useful
tool to inspect the underlying drivers of ri,t . One way to do so is to consider the
regressions from the Single Regression approach. These regressions are replicated
here for convenience, considering the general equation i and considering both the
case in which a constant term is included and is not included:

ri,t = δi mt + i (4)
ri,t = const + δi mt + i (5)

We are after the statistical significance of δi . Since ri,t is by construction (or by

assumption) a combination of the structural shocks, a necessary condition for mt to
be a good instrument is that there is some statistical significance in the relationship
between some of the ri,t and the instrument. In principle, a t test on δi can lead
to different results depending on whether we estimate equation (4) or equation (5).
Similarly, we could use an F test on δi on, say, equation (5). This statistic would be
identical by construction to a Wald test, and identical to the square of a t statistic,
as long as these alternative statistics are considered on equation (5). They could
instead differ if we compare them to the statistics from equation (4). This difference
holds only in finite sample and disappears asymptotically.

5 On the ordering of variables and shocks

It is useful to appreciate the fact that the ordering of the shock of interest within
the B matrix is just a matter of notational convention.
Consider the following structural VAR including some measure of output yt and
the federal funds rate it , using the A specification of the structural form:
      y
1 −a12 yt lags st
= + mp
−a21 1 it lags st
| {z }
A

Lags are left unspecified to make notation less hefty. syt and smp
t stand for produc-
tivity shock and monetary policy shock, respectively. The equivalent way of writing
this model is

yt = a12 it + lags + syt

it = a21 yt + lags + smpt

Note that the monetary shock enters the equation of the federal funds rate while the
productivity shock enters the equation of output.

9
 If we rewrite the VAR in the B structural specification we get
      y 
yt lags 1 1 a12 st
= +
it lags 1 − a12 a21 a21 1 smp
t
| {z }
B
  h 1   
lags 1 a i
= + st + 12 smp
y
t
lags 1 − a12 a21 a21 1

Of course, one can reshuffle the columns of the B matrix and the corresponding
shocks:
    h a    i
yt lags 1 mp 1
= + 12
st + syt
it lags 1 − a12 a21 1 a 21
     mp 
lags 1 a12 1 st
= +
lags 1 − a12 a21 1 a21 syt
| {z }
B̃

Note that B and B̃ differ only up to the ordering of the columns. Since the reduced
form shocks have not changed, Σ has not changed, implying Σ = BB 0 = B̃ B̃ 0 . Put
it differently, the ordering of the shocks in the B form of the model is arbitrary.3
Nevertheless, it is not arbitrary in the A form, i.e. you cannot reshuffle the B matrix
and then expect an à specification in which syt enters the equation of the federal
funds rate and smpt enters the equation of output. To see this, compute the A form
of the reschuffled model and note that
      mp 
−a21 1 yt lags s
= + ty
1 −a12 it lags st
| {z }
Ã

which simply means

it = a21 yt + lags + smp

t
y
yt = a12 it + lags + st

In short, the order of the shocks in the B form is arbitrary, and reshufling the columns
of B and the corresponding shocks implies reshufling the rows of the corresponding
A model.
3
Note that
b211 + b212
    
b b12 b11 b21 b11 b21 + b12 b22
BB = 11
0
=
b21 b22 b12 b22 b11 b21 + b12 b22 b221 + b222
b212 + b211
    
b b11 b12 b22 b12 b22 + b11 b21
B̃ B̃ 0 = 12 = = BB 0
b22 b21 b11 b21 b12 b22 + b11 b21 b222 + b221

10
 The implication of this remark on the use of external instruments in VARs is the
following: the ratio µ1 , or equivalently the ratio µ2 , is not the ratio of the elements
of a particular column of B. It is just an arbitrarily-selected column of the B matrix.
Calling j the arbitrary position of the shock of interest within the vector of structural
shocks st , the column (row) vector in the B (A) matrix for the estimation of impulse
responses (structural shocks) will be the j − th column (row).

6 On the estimation “up to a scale”

So far we have argued that if we have an instrument mt for the structural shock
sa,t , we can estimate the relative impulse vector. In other words, we do not estimate
the single elements b11 and b21 (unless using in addition the covariance restrictions,
as discussed below) but only their ratio. This can be seen more directly from the
key condition underlying the analysis (equation (1)), which we rewrite here, after
arbitrarily considering the shock of interest as the first shock:

E(r t mt ) = b1 φ
Since we can get an estimate of the left-hand side, we can estimate b1 up to a scale,
i.e. up to the unknown φ on the right-hand side, which will scale up and down our
estimate of b1 .
While this might look like a limitation, it is actually a strength, because it allows
to get around a measurement error problem. To see this, consider again the model
that links the reduced-form shocks to the structural shocks
r1,t = b11 sa,t + b12 sb,t
r2,t = b21 sa,t + b22 sb,t
The identification problem consists of estimating the elements bij using the reduced
form shocks rj,t and an instrument. Suppose we have the true realizations of the
structural shock sa,t . Then the regression of ri,t on sa,t leaves the remaining compo-
nent bi,2 sb,t in the error term and delivers a consistent estimate of bi1 . The estimate,
in fact, equals
P
t ri,t sa,t
b̂i1 = P 2
=
t sa,t
P P
(b 11 s a,t + b 12 s b,t )s a,t t sb,t sa,t /T
= t P 2 = bi1 + bi2 P 2
→ bi1
t sa,t t sa,t /T

since sa,t and sb,t , being structural shocks, should be uncorrelated. In this case, we
consistently estimate the single parameters b11 and b21 rather than just their ratio.
Consider instead the case in which we have a noisy measure of sa,t rather than
the true values. In particular, let us write this measure as
mt = sa,t + t with E(t sb,τ ) = 0, ∀τ

11
Then the regression of ri,t on mt estimates the model

ri,t = bi1 mt + qt with qt = b12 sb,t − bi1 t

The estimator arranges the data as

P
t ri,t mt
b̂i1 = P =
t m2t
P P
t (bi1 mt + qt )mt qt mt
= P 2 = bi1 + bi2 Pt 2 6→ bi1
t mt t mt

This does not converge to bi1 because the correlation between mt and qt is not zero,
due to a simple manifestation of the measurement error bias. Nevertheless, the ratio
of the inconsistent estimates gives a consistent estimate of the ratio. To see this,
rewrite first the estimate as
P
t ri,t mt
b̂i1 = P =
t m2t
P
(bi1 sa,t + bi2 sb,t )mt
= t P 2 =
P t mt P
bi1 t sa,t mt + bi2 sb,t mt
= P 2
t mt

Then, compute the ratio, which will give

P P P
b̂21 b21 t sa,t mt + b22 sb,t mt b21 t sa,t mt b21
= P P → P →
b̂11 b11 t sa,t mt + b12 sb,t mt b11 t sa,t mt b11
The last convergence holds because, under the condition P that the measurement er-
ror
P is orthogonal to the other structural P shock(s), sb,t mt (or actually, P
the term
sb,t mt /T ) is zero in population. Since sb,t mt goes to zero, the term t sa,t mt
cancels out from numerator and denominator and we are left with the true ratio.
The source of the measurement error drops and we get the population ratio of the
parameters. It follows that the measurement error in mt considered here does not
impair the analysis.
It is worth noticing that other forms of measurement errors would equally allow
the estimator of the relative impulse vector to deliver a consistent estimate. For ex-
ample, the following alternative measurement error still allows us to obtain consistent
estimates of the relative impulse vector:

mt = sa,t + t with E(t sb,τ ) = 0, τ = t

On the contrary, the following measurement error would not imply consistent esti-
mates of the relative impulse vector

mt = sa,t + t with E(t sb,τ ) 6= 0, τ = t

12
7 Recovering the full B matrix
So far we have discussed a method to estimate the ratio µ of the parameters in the
arbitrary i column of the B matrix. Let us now recover the full column i. This can
be useful, for instance, to study the impulse responses to a one standard deviation
shock. The key piece of information to use is the covariance restrictions of the model.
Note that so far we have made no use of these restrictions, but only of the instrument
mt . Using this additional piece of information allows to estimate the impulse vector
b not only up to a scale, but in absolute terms. Nevertheless, the vector will still be
estimated up to a sign convention.
Consider the VAR with 2 variables studied so far. The covariance restrictions are
expressed in the system Σ = BB 0 . More formally, the system is

b211 + b212 = Σ11

b11 b21 + b12 b22 = Σ12
b221 + b222 = Σ22

In a standard VAR we aim to solve for bij using Σij . Here, instead, we make use of
the additional information included in the ratio µ, which has already been estimated.
Without loss of generality, consider the ratio µ = b21 /b11 . All we need in order to
solve for in order to recover the entire vector b1 = (b11 , b21 )0 is b11 .
This parameter can be estimated as follows.4 Substituting out b21 = µb11 , the
system becomes

b211 + b212 = Σ11 (6)

µb211 + b12 b22 = Σ12 (7)
µ2 b211 + b222 = Σ22 (8)

From equation (6), write the solution for b11 as

q
b11 = Σ11 − b212
If we solve for b212 , then we obtain |b11 | and hence two solutions for b1 , i.e.
 
1
b1 = ±b11
µ
To obtain b212 , subtract equation (6) times µ from (7), obtaining
4
I thank Joris de Wind for having made these derivations available in his paper, available here.

13
 b12 b22 − µb212 = Σ12 − µΣ11
b12 (b22 − µb12 ) = Σ12 − µΣ11
b212 (b22 − µb12 )2 = (Σ12 − µΣ11 )2
b212 = (b222 + µ2 b212 − 2b22 b12 µ)−1 (Σ12 − µΣ11 )2
| {z }
γ
−1
b212 =γ (Σ12 − µΣ11 )2

We do have (Σ12 − µΣ11 )2 , but we don’t have γ, which includes the parameter b12
that we are after. Nevertheless, we can rewrite γ as a function of known parameters
in the following way. First, note that the term b22 b12 appears in equation (7). Take
equation (8) and subtract equation (7) times µ, get

b222 − b12 b22 µ = Σ22 − µΣ12

b12 b22 µ = b222 − (Σ22 − µΣ12 )

Substitute this into γ, get

γ = b222 + µ2 b212 − 2b222 + 2(Σ22 − µΣ12 )

= −b222 + µ2 b212 + 2(Σ22 − µΣ12 )

Then, to get rid of b222 , note that it appears in equation (8). Start from equation (8)
and subtract equation (6) times µ2 , obtaining

b222 − µ2 b212 = Σ22 − µ2 Σ11

Substitute out b222 and solve for γ

γ = −(Σ22 − µ2 Σ11 ) + 2(Σ22 − µΣ12 ) =

= Σ22 + µ2 Σ11 − 2µΣ12

Having γ, solve for b212 , then solve for b11 . This allows to compute the full column of
the B matrix corresponding to the sa,t shock.

absolute impulse vector = b11 · relative impulse vector

Before continuing, it is worth noticing that b11 has been estimated up to a sign, as
should be apparent from the fact that it is the solution to a square root. This implies
that also the vector b1 is estimated up to sign. This form of indeterminacy is usually
referred to as identifying a shock up to a sign convention. Whether the shock should
be interpreted as, say, expansionary or contractionary should be inferred from the
pattern that it generates in the data. For example, given an arbitrary definition of

14
the relative impulse vector, if, when solving for b11 with the positive sign, a monetary
shock of size 1 implies a decrease of the federal funds rate, then one can generate a
monetary tightening by simply giving the same shock and by taking the solution to
b11 with the negative sign.

Let us conclude this section by showing the derivations for the case in which the
VAR has more than 2 variables. The only difference with respect to the case with 2
variables is that some operations arranged above with scalars must be rethought of
now in matrix notation. With a VAR with n + 1 variables, write the B matrix as

b11 b012
 
B=
b21 B22

b11 is a scalar for the 1,1 entry of the B matrix, b21 is an n × 1 vector of the first
column of the B matrix written as a column vector, after excluding the first entry,
b12 is an n × 1 vector of the first row of the B matrix written as a row vector,
after excluding the first entry and B22 is an n × n matrix containing the remaining
elements of the B matrix.
Write the system of equations as

b211 + b012 b12 = σ11

b11 b21 + B22 b12 = σ 21
b21 b021 + B22 B22 0
= Σ22

The external instrument will deliver the vector µ = b21 /b11 . We can substitute out
b21 = µb11 , obtaining

b211 + b012 b12 = σ11 (9)

b211 µ + B22 b12 = σ 21 (10)
b211 µµ0 + B22 B22 0
= Σ22 (11)

From equation (9), write the solution for b11 as

q
b11 = σ11 − b012 b12
If we solve for b012 b12 , then we obtain ±b11 and hence two solutions for b1
To obtain b012 b12 , subtract equation (9) times µ from (10), obtaining

B22 b12 − b012 b12 µ = σ 21 − σ11 µ

(B22 − µb012 )b12 = σ 21 − σ11 µ
b12 = (B22 − µb012 )−1 (σ 21 − σ11 µ)

15
Hence
0
b012 b12 = (σ 21 − σ11 µ)0 (B22 − µb012 )−1 (B22 − µb012 )−1 (σ 21 − σ11 µ)
= (σ 21 − σ11 µ)0 [(B22 − µb012 )(B22 − µb012 )0 ]−1 (σ 21 − σ11 µ)
| {z }
Γ

with

Γ = (B22 − µb012 )(B22 − µb012 )0

0
= B22 B22 − B22 b12 µ0 − µb012 B22
0
+ µb012 b12 µ0

Take equation (11) and subtract equation (10) times µ0 , get

0
B22 B22 − B22 b12 µ0 = Σ22 − σ 21 µ0
B22 b12 µ0 = B22 B22
0
− (Σ22 − σ 21 µ0 )

Substitute this and its transpose in Γ, get

0
Γ = −B22 B22 + µb012 b12 µ0 + (Σ22 − σ 21 µ0 ) + (Σ22 − σ 21 µ0 )0
0
Then, to get rid of B22 B22 , start from equation (11) and subtract equation (9)
premultiplied by µ and postmultiply it by its transpose, obtaining
0
B22 B22 − µb012 b12 µ0 = (Σ22 − σ11 µµ0 ) (12)

The transpose of this gives

0
B22 B22 − µb012 b12 µ0 = (Σ22 − σ11 µµ0 ) (13)

given that most terms are symmetric. Substitute this in Γ, get

Γ = −(Σ22 − σ11 µµ0 ) + (Σ22 − σ 21 µ0 ) + (Σ22 − σ 21 µ0 )0

= Σ22 + σ11 µµ0 − σ 21 µ0 − µσ 021

8 Estimating structural shocks

So far we have reduced the identification problem to the estimation of the column in
the B matrix corresponding to the shock of interest. We said that this could be done
by computing the relative impulse vector, and by using the covariance restrictions
of the model to recover the scaling factor up to a sign convention and up to a
scale reflecting normalizations. Nothing was said regarding how to estimate the
corresponding realizations of the shock.
If we could recover the full B matrix, we could recover the structural shocks
simply as

16
 st = B −1 r t
= Ar t

Since the procedure discussed so far does not deliver the full B matrix but only
a column of it, this cannot be done. Nevertheless, we could recover the shocks of
interest if we had only one row of A, as long as it is the row corresponding to the
shock of interest.
More precisely, the identification of the impulse response corresponding to shock
sa,t gave us the vector b, where the position of this vector within the B matrix is
irrelevant. What we are after is the row vector a of the A = B −1 matrix correspond-
ing to the position of the b column in B. When we have a, we can estimate the
structural shocks of interest as

sa,t = a0 · r t
a gives the weights we should use to combine the reduced form shocks in r t in order
to obtain the shock of interest.
The following procedure, developed for example in Olea et al. (2012), can be used
when having one instrument for one shock of interest. The first step is to estimate
φ, where we had that
φ = E(mt sa,t )
We started the notes discussing the equality

E(r t mt ) = bφ

In this expression, E(r t mt ) can be computed from the data. b, instead, can be
estimated using the procedure discussed until now. This means that we can recover
an estimate φ̂ using any of the equations in the above equality, and taking the ratio
of the element in E(r t mt ) over the corresponding element in the b vector.
Then, consider the regression of the instrument mt on the reduced form shocks
r t . Note that this is the opposite regression of the one we considered in Section 2 to
estimate the relative impulse vector. The model is

mt = r 0t ι + error

By construction the OLS estimator gives

17
 ι̂ = (R0 R)−1 R0 m
 PT r r 0 −1 PT r m
t=1 t t t=1 t t
=
T T
 PT s s0 −1 PT r m
t t 0 t=1 t t
= B t=1 B
T T
 −1
→ BB 0 E(r t mt )
= B 0−1 B −1 bφ
 
0−1 1
=B φ
0
 
0 1
=A φ
0
= aφ

In the first equation, R is a T × n matrix containing the estimated reduced form

shocks. The derivations then use the fact that the variance-covariance matrix of
−1
the structural
  shocks equals the identity matrix, and that B B = I. This implies
1
B −1 b = where b is for convenience the first column of B. a is the first row of
0
the A = B −1 matrix, rewritten as a column vector. More precisely, a is the general
row vector, written as a column vector, corresponding to the shock of interest.
Within the approach outlined above, a potentially practical way of deriving a is
to rewrite the equations above as

(R0 R)−1 R0 m → aφ
 PT r r 0 −1
PT
t=1 t t t=1 r t mt
→ aφ
T T
Σ−1 Ê(r t mt ) → aφ

which implies

a = Σ−1 Ê(r t mt )φ−1

The above methodology allows to estimate the a vector when we have one instru-
ment for the shock of interest. When one has g instruments things change, since we
would not have a single condition E(r t mt ) = bφ, but g conditions E(r t mj,t ) = φj ,
j = 1, 2, ..., g. In Section 2 we briefly mentioned that the 2SLS approach allows to
combine all these moment conditions to obtain a single estimate of the b vector.

18
If we then want to estimate the corresponding a vector we cannot proceed as just
discussed, since this would require making use of only one of the moment conditions
in E(r t mj,t ) = φj . What we can do instead is to start from the estimated b vector,
and combine it with the covariance restrictions of the model.
More precisely, we make use of the following result regarding the inverse of a
partitioned matrix. Given
 
B11 B12
B=
B21 B22
the inverse of B is
−1 −1 −1
B21 )−1 B21 )−1 B12 B22
 
−1 (B11 − B12 B22 −(B11 − B12 B22
B = −1 −1 −1
−(B22 − B21 B11 B12 )−1 B21 B11 (B22 − B21 B11 B12 )−1
In the application developed here, we have solved for the first column of the B matrix
and need to derive the first row of its inverse. Using again the notation from Section
7, write the B matrix as

b̂11 b012
 
B=
b̂21 B22
where .̂ indicates the elements that we have. We want to solve for the first row of
A = B −1 as a function of, at most, b̂11 , b̂21 , Σ.
Applying the rule for the inverse of a partitioned matrix, we get that the a vector,
written as a column, equals
 
0 −1 −1 1
a = (b̂11 − b12 B22 b̂21 ) 0−1 (14)
−B22 b12
0−1
This means that if we can solve for B22 b12 , or equally, for b012 B22
−1
, we get a, irre-
spectively on how many instruments were used to obtain the first column of the B
matrix. Note that we do not need to solve for b12 and B22 separately.
The term b012 B22−1
can be computed starting from the system of equations in
Section 7. In particular, start from

b̂211 + b012 b12 = σ11

b̂11 b̂21 + B22 b12 = σ 21
0 0
b̂21 b̂21 + B22 B22 = Σ22
This can be rearranged as

b012 b12 = σ11 − b̂211

B22 b12 = σ 21 − b̂11 b̂21 (15)
0 0
B22 B22 = Σ22 − b̂21 b̂21 (16)

19
Equation (15) can be rearranged as

B22 b12 = (σ 21 − b̂11 b̂21 )

−1
b12 = B22 (σ 21 − b̂11 b̂21 )
0−1 0−1 −1
B22 b12 = B22 B22 (σ 21 − b̂11 b̂21 )
0−1 −1
An expression for B22 B22 can be obtained by inverting equation (16). Combining
these results gives
0
b012 B22
−1
= (σ 21 − b̂11 b̂21 )0 (Σ22 − b̂21 b̂21 )−1
0
where we have made use of the fact that both terms in Σ22 − b̂21 b̂21 are symmetric.
Solving for a then requires substituting out the transpose of b012 B22
−1
in (14).
The solution for a does not depend on the sign convention taken for the last n − 1
columns of the B matrix, i.e. when flipping sign of any of the columns of B other
than the first one. On the contrary, flipping sign of all elements in the first column
of B implies opposite sign of every element in the first row of A, i.e. in the vector a.

20
References
Mertens, K. and M. O. Ravn (2013). The dynamic effects of personal and corporate
income tax changes in the united states. The American Economic Review 103 (4),
1212–1247.

Olea, J. L. M., J. H. Stock, and M. W. Watson (2012). Inference in structural vars

with external instruments. Technical report, Technical report.

Stock, J. H. and M. W. Watson (2012). Disentangling the channels of the 2007-2009

recession. Brookings Papers on Economic Activity.

21