Sign Restriction Fry and Pagan

NCER Working Paper Series
Sign Restrictions in Structural Vector

Autoregressions: A Critical Review
Renée Fry
Adrian Pagan
Working Paper #57

July 2010
Sign Restrictions in Structural Vector
Autoregressions: A Critical Review
Renee Fryy and Adrian Paganz
July 28, 2010
Contents
1 Introduction 2
2 The VAR Representation and Two Simple Structural Mod-

els 6
2.1 VAR and SVAR Representations . . . . . . . . . . . . . . . . 6
2.2 Two Simple Structural Models . . . . . . . . . . . . . . . . . . 7
2.2.1 A Market (Demand/Supply) Model . . . . . . . . . . . 7
2.2.2 A Small Macro Model . . . . . . . . . . . . . . . . . . 8
3 Identifying Shocks 9
3.1 The Parametric Approaches . . . . . . . . . . . . . . . . . . . 9
3.2 Sign Restrictions . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.2.1 Givens Matrices . . . . . . . . . . . . . . . . . . . . . . 11
3.2.2 Householder Transformations . . . . . . . . . . . . . . 13
3.3 Sign Restrictions in the Market Model . . . . . . . . . . . . . 13
We are greatful to the editor and referees for some very helpful suggestions regarding
the structure of this review. Fabio Canova and Matthias Paustian also gave us useful
comments - the latter greatly clarifying the results of Section 5.1. Pagan’s early work on
this topic was supported by ESRC Grant 000 23-0244. Fry’s research was supported by
ARC Grant #DP0664024.
y
Australian National University
z
University of Technology, Sydney and Queensland University of Technology
1
4 Some Basic Issues with Sign Restrictions 15
4.1 Model and Structural Identi…cation . . . . . . . . . . . . . . . 15
4.2 Identifying a Single Shock . . . . . . . . . . . . . . . . . . . . 21
4.3 The Origin of Sign Restrictions . . . . . . . . . . . . . . . . . 22
5 Some Advanced Issues with Sign Restrictions 25

5.1 Can We Recover Correct Impulse Responses from Sign Re-
strictions? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.2 Other Summative Representations . . . . . . . . . . . . . . . . 29
5.3 Permanent and Transitory Shocks . . . . . . . . . . . . . . . . 30
6 Conclusion 31
7 References 32
Abstract
The paper provides a review of the estimation of structural VARs
with sign restrictions. It is shown how sign restrictions solve the para-
metric identi…cation problem present in structural systems but leave
the model identi…cation problem unresolved. A market and a macro
model are used to illustrate these points. Suggestions have been made
on how to …nd a unique model. These are reviewed, along with some of
the di¢ culties that can arise in how one is to use the impulse responses
found with sign restrictions.
1 Introduction
Structural Vector Autoregressions (SVARs) have become one of the major
ways of extracting information about the macro economy. One might cite
three major uses of them in macro-econometric research.
1. For quantifying impulse responses to macroeconomic shocks.
2. For measuring the degree of uncertainty about the impulse responses
or other quantities formed from them.
3. For deciding on the contribution of di¤erent shocks to ‡uctuations and
forecast errors through variance decompositions.
To determine this information a VAR is …rst …tted to summarize the data
and then a structural VAR (SVAR) is proposed whose structural equation er-
rors are taken to be the economic shocks. The parameters of these structural
2
equations are then estimated by utilizing the information in the VAR. The
VAR is a reduced form which summarizes the data; the SVAR provides an
interpretation of the data. As for any set of structural equations, recovery of
the structural equation parameters (shocks) requires the use of identi…cation
restrictions that reduce the number of "free" parameters in the structural
equations to the number that can be recovered from the information in the
reduced form.
Five major methods for recovering the structural equation parameters
(identifying the shocks) are present in the literature. Four of these explicitly
utilize parametric restrictions. These involve the nature of the structural
equations. Parametric restrictions on these equations can vary according
to whether particular variables appear in the latter (Cowles Commission),
whether there is a recursive causal structure (Wold (1951), Quenouille (1957)
and Sims (1980)), and whether shocks have known short-run (Gali (1992)) or
long-run (Blanchard and Quah (1989)) e¤ects. In each case the parametric
restrictions free up enough instruments for the contemporaneous endogenous
variables in the structural equations, thereby enabling the parameters of
those equations to be estimated. Recently, a …fth method for estimating
SVARs has arisen that employs sign restrictions upon the impulse responses
as a way of identifying shocks - Faust (1998), Uhlig (2005), Canova and De
Nicoló (2002). Applications of this method have been growing, as seen in the
papers listed in Table 1. The table is a sub-set of published studies and adopts
a taxonomy which distinguishes between cases where only sign restrictions
are used (often there are mixtures of sign and parametric restrictions), the
type of shock (permanent or transitory), the number of shocks identi…ed, and
whether the sign restrictions come from a formal model or not. Consequently,
it is worth examining this literature in more detail, and the aim of our paper
is to exposit how the method works and to identify some of the di¢ culties
that can arise in its application.
3
Table 1 Summary of Empirical VAR Studies Employing Sign Restrictions
Fluctuations Peersman (2005) STNI

Rü¤er et al (2007) STNI
Sanchez (2007) STNF
Ex Rate An (2006) STOI
Farrant/Peersman (2006) STNF
Lewis (2007) STNF
Bjørnland/Halvorsen (2008) MTNI
Scholl/Uhlig (2008) STNI
Fiscal Policy Mountford/Uhlig (2005, 2008) STNI
Dungey/Fry (2009) MPTNI
Housing Jarociński/Smets (2008) MTNI
Vargas-Silva (2008) STOI
Monetary Policy Faust (1998) STOI
Canova/De Nicoló (2002) STOF
Mountford (2005) STNI
Uhlig (2005) STOI
Ra…q/Mallick (2008) STOI
Scholl/Uhlig (2008) STNI
Technology Francis/Owyang/Theodorou (2003) MPTOI
Francis/Owyang/Roush (2005) MPTOF
Dedola/Neri (2006) SPTOF
Chari/Kehoe/McGrattan (2008) MPTNF
Peersman/Straub (2009) STNF
Various Hau and Rey (2004) STNF
Eickmeier/Hofmann/Worms (2009) STNI
Fujita(2009) STOI
Restriction Type: S= Sign only, M = Mixed
Shock Types: P=Permanent, T= Transitory
Number of Shocks: O=One only, N=Numerous
Restriction Source: F=Formal, I= Informal
In practical work it is often found that a combination of all the meth-
ods mentioned above need to be employed in order to be able to identify all
the shocks of interest. We emphasize that which of the …ve methods men-
tioned above is used in practice does not depend on the data, but rather
on the preferences of the investigator and/or of those who wish to utilize an
4
SVAR to study some issue. These preferences may well be incompatible as
some users may feel that certain types of restrictions are more plausible than
others. Prima facie it does seem likely that long-run and sign restrictions
would be regarded as less restrictive than the other approaches, but without
a speci…c context there can be no basis for recommending any particular ap-
proach. Each has di¢ culties and these need to be understood when making
an informed judgement on their utility. Although it is likely in practice that
a mixed set of restrictions will be employed, because the literature on sign
restrictions is more recent than that on the other methods it is convenient
to simply assume that only sign restrictions are being employed.
Section 2 introduces the most common summative model (a VAR) and
two structural models used in later analysis - a simple demand-supply model
(called the market model), and a basic macroeconomic model determining
output, interest rates and in‡ation (called the macro model). Section 3 then
examines how the …ve approaches described above would identify the shocks
of the market model. Only a brief account of the parametric approaches
is provided in order to allow for a more detailed description of the sign re-
striction methodology. By using the market model it is shown that sign
restrictions do implicitly impose parametric restrictions. In section 4 we out-
line some di¢ culties that can arise when implementing sign restrictions and
the various solutions that have been proposed to them. One example is how
one is to respond to the fact that a unique set of impulse responses is not
available. This arises because, while the sign restrictions solve the struc-
tural identi…cation problem by providing su¢ cient information to identify
the structural parameters, they leave unresolved what Preston (1978) called
the model identi…cation problem - the latter referring to the fact that there
are many models with identi…ed parameters that provide the same …t to the
data. Another is what one does if only the e¤ects of a single shock such as
technology or money is of interest? Section 5 addresses a range of more com-
plex questions involving whether the methodology can recover the correct
impulse responses and how one is to handle both permanent and transitory
shocks. Finally, section 6 concludes.
5
2 The VAR Representation and Two Simple
Structural Models
2.1 VAR and SVAR Representations
Most of the literature we deal with assumes that the data can be represented
by a VAR (for simplicity we will make it of …rst order)
zt = A1 zt 1 + et ; (1)
where zt is an n 1 vector of variables and et is a set of errors that have zero
expectation, constant covariance matrix and no serial correlation. From
this an interpretation of the data is provided through a SVAR
B0 zt = B1 zt 1 + "t ; (2)
where "t are shocks that have zero mean, no serial correlation, constant
variances and no correlation between the individual shocks i.e. E("it "jt ) = 0:
Comparing (1) and (2) gives B0 et = "t i.e. the structural shocks "t we seek
to measure are linear combinations of the VAR errors et : The latter can be
estimated by the VAR residuals e^t : To estimate the structural (economic)
shocks, "t ; then requires that one construct an appropriate set of weights
^0 ) on e^t : Clearly the VAR is the reduced form of the structure set out in
(B
the SVAR.
The solution to the VAR(1) is the moving average (MA) form
zt = D0 et + D1 et 1 + D2 et 2 + ::: (3)
where Dj is the j’th period impulse response of zt+j to a unit change in et
(D0 = In ): It follows that the MA form for the SVAR is
zt = C0 "t + C1 "t 1 :::;
with the j’th period impulse response of zt+j to "t being Cj = Dj B0 1 = Dj C0
as C0 = B0 1 : It is important to note that, since A1 can be estimated by
regressing zt on zt 1 ; and so does not require a structural model speci…cation,
Dj can be estimated from that information (in the …rst order case Dj = Aj1 ):
Hence, if one knows C0 one can …nd all the Cj without stipulating a structural
model. For this reason we will sometimes set A1 = 0 in our illustrations of
the various approaches, as that facilitates a focus upon how C0 is determined
by each of them. Moreover, a failure to accurately estimate C0 will mean
that further Cj will be estimated inaccurately.
6
2.2 Two Simple Structural Models
2.2.1 A Market (Demand/Supply) Model
We take the case of a simple model comprising a demand and a supply
function with associated shocks. This will be termed the market model.
Speci…cally the SVAR system will be
qt = pt + qq qt 1 + qp pt 1 + "Dt (4)
pt = qt + pq qt 1 + pp pt 1 + "St ; (5)
where the shocks have expected values of zero and are assumed uncorrelated
with standard deviations of D and S respectively. Hence, in terms of the
qt
SVAR discussion above, zt = : The reduced form of the market model
pt
is a VAR(1) with the form
qt = aqq qt 1 + aqp pt 1 + e1t (6)

pt = apq qt 1 + app pt 1 + e2t : (7)
Since the equations (4) and (5) are essentially identical for arbitrary para-
meter values, at this point there is nothing which distinguishes the demand
("Dt ) and cost shocks ("St ); and the task is to introduce extra information
that does enable us to identify these. It would seem likely that most re-
searchers would agree with the sign information in Table 2 for the impact of
positive shocks upon the contemporaneous variables (a positive movement in
"St will mean a negative supply side shock).1 Since the patterns are distinct
this suggests that we are likely to be able to identify separate shocks. Indeed
it is clearly going to be a requirement that shocks have distinct sign patterns
in their e¤ects on variables if we are to isolate them separately.
1
Although in later analyses we will always exhibit the pattern matrix in response to
positive shocks, it needs to be recognized that a pattern for C0 of would also
+
be consistent with demand and cost shocks, although with negative signs. So one needs to
+ + +
allow for this in any search. Of course and would also be acceptable.
+
Clearly the need to check for the complete set of compatible sign restrictions will grow as
the number of shocks increases.
7
Table 2: Sign Restrictions for Market
Model (Demand/Supply) Shocks
variablenshock Demand Supply

pt + +
qt + -
2.2.2 A Small Macro Model

A small macro model that is used a lot involves an output gap (yt ); in‡ation
( t ) and a policy interest rate (it ): In terms of (1) the system variables are
zt0 = yt t it : A …rst order SVAR model for these variables would then
be
yt = zt0 1 y + yi it
+ y t+ "yt (8)
0
t = zt 1 + i it + y yt + " t (9)
it = zt0 1 i+ iy yt + i t + "it : (10)
The three shocks will be monetary policy ("it ), a demand shock ("yt ) and
a cost-push (supply) shock (" t ). For simplicity the shocks will be treated
as having no serial correlation, so that the reduced form is a VAR(1) of the
form
yt = zt0 1 y + e1t (11)

0
t = zt 1 + e2t (12)
it = zt0 1 i + e3t : (13)
The signs of the contemporaneous e¤ects to positive shocks will most likely
be those in Table 3. Again these are distinct and this enables the separation
of the three structural shocks.
Table 3: Sign Restrictions for Macro Model Shocks
variablenshock Demand Cost-Push Interest Rate

yt + - -
t + + -
it + + +
8
3 Identifying Shocks
3.1 The Parametric Approaches
In order to contrast the sign restriction approach to other methods of iden-
tifying shocks, let us think about how one might estimate the market model
using the types of parametric restrictions distinguished in the introduction
(we ignore the …rst possibility of constraining some coe¢ cients of lagged
values to zero). These restrictions are designed to identify the structural
equations and hence the shocks.
(i) If the system is assumed to be recursive e.g. is set to zero; then OLS
can be applied to the supply equation, since qt is a function of "Dt and this
is uncorrelated with "St : Three unknown parameters are left and there are
three pieces of information to estimate them with - the estimated variances
of pt ; qt and the covariance of pt and qt :
(ii) A restriction that (say) a demand shock has no long-run e¤ect upon
the price would imply that pq = ; and so the supply curve would become
a function of qt and pt 1 : This implies that there is one less structural
parameter to estimate in the supply curve and qt 1 is then freed up to be
used as an instrument for qt : Once the supply equation is estimated the
demand equation can be found by using the residuals ^"St as an instrument
for pt :
(iii) An assumption that the short run e¤ect of a demand shock upon
prices is zero would imply that (1+ ) = 0 in the reduced form (VAR) equation
for pt
"Dt "St
pt = 1 qt 1 + 2 pt 1 + + ; (14)
(1 + ) (1 + )
where j are functions of ij and ; : Consequently, the VAR residual for
pt would not involve "Dt and so can be used as an instrument for pt in the
demand curve.
Thus in all cases the identi…cation problem is solved by reducing the
number of parameters to be estimated to three and by making available
suitable instruments for estimation.
3.2 Sign Restrictions

Now a set of n estimated shocks e^t will be available from the model we choose
to be the summative one. For the market model the VAR shocks e1t and e2t
9
are from (6) (7); while in the macro model the shocks ejt are in (11)-(13). By
combining them in an appropriate way we can produce candidate structural
shocks ^"t that are uncorrelated. Now there will be many such combinations.
Some of them will produce impulse responses that have the correct signs,
while others won’t. Thus in the market model case there will only be some
weights which produce shocks that respect the patterns in Table 2. So our
…rst task is to select an algorithm that gives a set of weights: Once one has
these we can check if they are "successful", in the sense that the impulse
response functions C^j for the corresponding structural shocks agree with the
postulated sign information. If they are not successful we will discard them
and "draw" another set of weights.
Now the critical constraint needed in designing an algorithm to do this is
that the generated weights must be such as to ensure that the constructed
structural shocks ^"t are uncorrelated. Suppose we begin by …rst estimating
a recursive VAR e.g. in the market model we could act as if was zero.
In that case, after estimation, we would have a set of shocks v^t such that
e^t = B ^0 1 v^t ; where B ^0 is a lower triangular matrix, as this characterizes a
2
recursive system. By design these shocks, v^t ; are uncorrelated. However,
rather than work directly with such shocks it is desirable to work with shocks
that have unit variance, and this can be done by dividing each of the v^kt by
its standard deviation. Hence, let S^ be the matrix that has the estimated
standard deviations of the v^t on the diagonal and zeros elsewhere. Then e^t =
B^0 1 S^S^ 1 v^t = T^ ^t ; where ^t = S^ 1 v^t are now regarded as structural shocks.
These shocks possess unit variances and can be thought of as coming from a
structural system T 1 zt = T 1 B1 zt 1 + t . These ^t shocks will be termed
our base set. Notice that they are just a re-scaled version of the v^t ; so their
nature has not changed.3
2
Although the residuals v^t could be thought of as structural shocks, ^"t ; we want to make
the point that they are just shocks to begin the search process with, and there is no need
to regard the recursive system as a plausible structure. It is the fact that shocks found
from a recursive system are uncorrelated which makes them useful. Later we mention
other ways of initiating the search.
3
Numerically it is generally more e¢ cient to estimate ^t by estimating the covariance
matrix of the residuals e^t ; ^ ; and then applying a Cholesky decomposition F 1 ^ F 0 1 = In
to form ^t = F 1 e^t rather than …rst estimating a recursive system: The ^t constructed in
this way will have unit variances and be uncorrelated: This is a useful way of proceeding
since all that is needed to implement it is the estimated covariance matrix of the errors
in the equations of the summative model. It also means that the summative model need
not be a VAR. It could be a Vector Error Correction Model or a state space model and
10
Now we form combinations of the ^t using a matrix Q i.e. ^t = Q^t : The
^t are candidates for "named" structural shocks e.g. "supply" and "demand".
They need to be uncorrelated and so Q must be restricted: The appropriate
restriction is that Q is a square matrix such that Q0 Q = QQ0 = In ; since
that means
e^t = T^Q0 Q^t
= T^ ^t ;
and cov( t t 0 ) = Qcov(^t ^0t )Q0 = In : Thus we have found a new set of
shocks, ^t ; with the same covariance matrix as ^t (and which will reproduce
the var(zt )), but which will have a di¤erent impact (T^ ) upon et and the
variables zt : It is this ability to create a large number of candidate shocks
with varying impulse responses that is the basis of sign restriction methods.
It is clearly very simple to construct all these shocks using programs that do
matrix operations once we have a method for forming a Q with the property
Q0 Q = QQ0 = In : There are many such Q0 s and we will refer to each as a
"draw".
How does one …nd a Q matrix? There are actually quite a few ways
of doing this. The two most popular utilize Givens and Householder trans-
formations (the latter is the basis of the QR decomposition used in many
ill-conditioned regression problems), but this does not exhaust the possibil-
ities. We provide an account of each of these and the relationship between
them in the following sub-sections.
3.2.1 Givens Matrices

In the context of a 3 variable VAR (the macro model) a 3 3 Givens matrix
Q12 has the form 2 3
cos sin 0
Q12 = 4 sin cos 0 5
0 0 1
i.e. the matrix is the identity matrix in which the block consisting of the …rst
and second columns and rows has been replaced by cosine and sine terms and
lies between 0 and .4 Q12 is called a Givens rotation. Then Q012 Q12 = I3
we use that fact in later sections.
4
In general Qij is formed by taking an n n identity matrix and setting Qii
ij = cos ;
ij ji jj
Qij = sin ; Qij = sin ; Qij = cos ; where the superscripts refer to the row and column
of Qij :
11
using the fact that cos2 + sin2 = 1: There are then three possible Givens
rotations for a three variable system; the others being Q13 and Q23 : Each of
the Qij depends on a separate parameter k . In practice most users of the
approach have adopted the multiple of the basic set of Givens matrices as Q
e.g. in the three variable case we would use
QG ( ) = Q12 ( 1 ) Q13 ( 2 ) Q23 ( 3 ):
It’s clear that QG is orthogonal and so shocks formed as t = QG t will be

uncorrelated and their impact upon zt will be T^ = T^Q0G :
Now, the matrix QG above depends upon three di¤erent k : Canova and
De Nicoló (2002) suggested that one make a grid of M values for each of
the values of k between 0 and ; and then compute all the possible QG :
Of course all of these models distinguished by di¤erent numerical values for
k are observationally equivalent in that they produce an exact …t to the
variance of the data on zt .5 Only those QG producing shocks that agree with
the maintained sign restrictions would be retained.
As an example we look at the macro model described in Cho and Moreno
(2006) estimated with some data on the US output gap, in‡ation and the
Federal Funds rate. As described above, begin with a recursive model impos-
ing yi = 0; y = 0; i = 0: OLS on each of (8)-(10) then gives structural
equation residuals that are uncorrelated. More potential structural shocks
can subsequently be found by combining these residuals (after re-scaling to
make them have unit variances) with Q matrices. Two of these Q matrices
(from the Givens approach) are given below. They have the property that
Q0 Q = I3 :6
2 3 2 3
:4551 :3848 :8030 :0444 :8431 :5359
Q(1) = 4 :5853 :8089 :0559 5 ; Q(2) = 4 :8612 :2395 :4482 5 :
:6710 :4446 :5933 :5062 :4815 :7155
(15)
When 2 Q(2) is used 3the generated structural shocks have a sign pattern
+ +
for C0 of 4 + + + 5 ; which disagrees with the restrictions in Table 3.
+ + +
5
It is assumed in the analysis that the zt have been mean corrected before the VAR is
…tted.
6
The fact that we retain only four decimal places above means that Q0 Q is not exactly
I3 :
12
In contrast, Q(1) does produce a set of impulse responses that is consistent
with the table. Hence, employing the sign restriction methodology only the
impulses found using Q(1) would be retained.
3.2.2 Householder Transformations

The alternative method of forming an orthogonal matrix Q is to generate
some random variables W from an N (0; I3 ) density (for a three variable
VAR) and then decompose W = QR R; where QR is an orthogonal matrix
and R is a triangular matrix. Householder transformations of a matrix are
used to decompose W: The algorithm producing QR is often called a QR
decomposition. Clearly QR = I corresponds to the matrix used in recursive
orderings. Since many draws of W can be made, one can …nd many QR :
Rubio-Ramírez et al. (2005) seem to have been the …rst to propose this, and
they have argued that, as the size of the VAR grows, this is a computationally
e¢ cient strategy relative to the Givens approach. In Fry and Pagan (2007)
we show that the methods are equivalent, so the main factor in choice would
be computational speed. As the system grows in size we would expect the
Householder method to be superior.
3.3 Sign Restrictions in the Market Model

So how do sign restrictions resolve the structural identi…cation problem in
the market model? As noted previously the key problem is how to identify
the initial impulse responses C0 : For illustrative purposes, it is convenient
to suppress the dynamics by setting B1 = 0 and to study the solutions for
C0 alone. Information on the signs expected for the elements of C0 is given
in Table 2. In the next step a recursive system is set up and estimated.
Of course the market model is generally not recursive, but this is simply a
mathematical device to generate a set of shocks v^t that are uncorrelated. We
therefore assume the following recursive system
pt = v1t (16)
qt pt = v2t : (17)
Three parameters are estimated in this system - = cov(qt ; pt )=var(pt ) and

the two variances of vjt , 2j : E¤ectively, this means that the variances of pt
and qt and their covariance are used for estimation. A base set of impulses
13
will then be found from jt = j vjt ; where j is the inverse of the standard
deviation of vjt :7 By de…nition 1t = 1 v1t = 1 pt and 2t = 2 v2t =
2
(qt pt ):
These base impulses are then used to construct new shocks jt by using
a Givens rotation as the weighting matrix. Since there is only one Givens
cos sin
matrix in the two variable case, Q = ; the transformed
sin cos
system becomes
1 2
1t pt cos (qt pt ) sin
= 1 2 ;
2t pt sin + (qt pt ) cos
Letting 1 = cos and 2 = sin the two equations can be written as

1 2 2
( 1 + 2 )pt 2 qt = 1t (18)
1 2 2
( 2 1 )pt + 1 qt = 2t ; (19)
with impulse responses of (pt ; qt ) to jt being
1 2 2 1 2 2
1 + 2 2 1 1 1 2
1 2 2 =G = 1 2 1 2 :
2 1 1 det(G) 2+ 1 1+ 2
Because j are …xed by the data the sign of the impact of the shocks upon pt
and qt will be dependent on sgn( 1 ) and sgn( 2 ); and these can be positive
or negative depending upon the values taken by : Consequently there may
be many impulse responses which satisfy the sign restrictions, each of which
is indexed by a value of : Note that, even though we started with a recursive
system, we will generally not have one as varies.
It is useful now to observe that, given , the number of unknown para-
meters in (18)-(19) has been reduced to three ( ; 1 ; 2 ); just as happened
with the parametric methods. This reduction means that a unique set of
parameters can be recovered for a given system of equations (and a speci…ed
), and so the structural parameter identi…cation problem has been solved.
7
A di¤erent way to get jt would be to …t a recursive system with var(pt ) 1=2 pt and
2 1=2
{var(qt )(1 pq )g qt as regressors, where pq is the correlation between pt and qt :
14
4 Some Basic Issues with Sign Restrictions
4.1 Model and Structural Identi…cation
Now in the discussion of the previous section we only retain those shocks
whose impulses agreed with the postulated signs. But it is clear that there
may be many impulse responses that satisfy these sign restrictions. Thus,
when using Givens’matrices, it is unlikely that there will be a single value
of that will produce the requisite sign restrictions. Figure 1 shows the
large range of impulse responses one gets by applying the contemporaneous
sign restrictions of Table 3 to the macro model data that we used earlier
when illustrating the e¤ects of choosing two values for Q. It is noticeable
that, even though the initial response of output to the interest rate has been
forced to be negative, there are some cases where that response becomes
positive very quickly. Each value of produces a new model constituting a
new set of structural equations and shocks. Consequently, although we have
converted any given system of equations (consistent with a given ) to one
that has a structure that is identi…ed, we have not identi…ed a unique model.
The di¤erence between structural and model identi…cation was emphasized
by Preston (1978).
What should one do about this multiple models problem? One response is
15
Figure 1: Impulse Responses for Macro Model: Median, MT and 5,95 Per-
centiles
to try to summarize the information in the graphs in some way e.g. reporting
a central tendency and the magnitude of the spread of responses. Thus, if
(k)
the impulse responses Cj that satisfy the sign restrictions are computed,
where k indexes the di¤erent values of ; various percentiles such as the 5%,
50% and 95% might be reported. This is done in Figure 2.
It may seem as if this is emulating the approach when one presents per-
centiles of a distribution from either a Bayesian or bootstrap experiment.
But it is important to recognize that the distribution here is across models.
It has nothing to do with sampling uncertainty. Referring to this range as
if it is a con…dence interval (something that is very common in the applied
literature) is quite false. All you get is a glimpse of the possible range of
responses as the model varies. Of course, this might be valuable. Often we
16
do present information about how our answers change as models are var-
ied. Leamer (1978) did this in a regression context with his extreme bounds
analysis. But it should not be imbued with probabilistic language. Even if
the VAR parameters A1 and were known with certainty, there will be a
question of how one proceeds whenever there are many . There is of course
a greater range when one accounts for the uncertainty in A1 and ; as is of-
ten done in this literature. Examples are Uhlig (2005) and Peersman (2005),
where Bayesian methods are applied to estimate the summative model, but
it does not help to understand the model identi…cation issue by confusing
these two sources of variation.
Do any di¢ culties arise in interpreting (say) these summary measures?
Let us illustrate the issues by considering the median of the impulse re-
sponses, by far the most popular choice for capturing the central tendency.
Suppose there is a single variable and two shocks, and that we have impulse
(k) (k)
responses C11 and C12 ; where k indexes the models (values for ): Ordering
(k ) (k)
these into ascending order enables one to …nd the medians C111 = medfC11 g
(k ) (k)
and C122 = medfC12 g: But k1 may not equal k2 ; and so the model that
(k)
produced the impulse response that is the median of fC11 g may not be the
(k)
same as that for fC12 g: Presenting the medians may be likened to presenting
the responses to technology shocks from an RBC model, and the monetary
shocks from a monetary model, and it is hard to believe that this is a rea-
sonable approach. Clearly, this comment applies to other percentiles so that
the extreme values which are being reported may come from very di¤erent
models.
Another piece of information presented in many papers is a variance de-
composition of var(zt ) into the contributions from various shocks. Often this
is done using the medians of the impulse responses. Is this correct? Now
a variance decomposition requires that the shocks be de…ned in the same
way, necessitating a common value of be used, since only in this case will
the shocks be uncorrelated by design. If shocks are not uncorrelated then a
variance decomposition does not make much sense. This issue is not resolved
by another common practice that computes the fraction of the variance ex-
plained by the j 0 th shock (j = 1; ::; n) in the k 0 th model (k = 1; :::; M )
(k) (k)
, j ; j = 1; ::; n; and then reports the n medians of f j gk=1;:::;M . Since
in general these medians will not come from the same model there is noth-
(k)
ing which ensures that the medf j g sum to one across all shocks i.e. the
variance is exhaustively accounted for.
17
Now it needs to be said that the issue of model identi…cation is always
present and is not speci…c to sign restrictions. Thus, if one used a recursive
system to get structural identi…cation, there are many other such systems
(orderings) that will yield the same VAR and give the same …t to the data.
Each structure coming from a given ordering is parametrically identi…ed but,
as all of the orderings exactly replicate the data, there is no unique model.
Only if one is prepared to consider that there is a single recursive model that
is tenable as the data generating process will this occur, and that is rare. For
example, Kilian and Murphy (2009) work with recursive systems involving
oil prices but resile from being dogmatic about any one recursive system.
Indeed, one often sees comments to the e¤ect that re-ordering the equations
did not modify the conclusions much.
Why then should we pay any more attention to this model identi…cation
issue for sign restrictions than for other ways of identifying VARs? Some
insight into this comes from examining the two possible recursive versions of
the market model - that given in (16) and (17) and the other being where
pt and qt are inter-changed in these equations. Although observationally
equivalent the two models can be treated as di¤erent views about how the
market operates. In one case quantity is treated as predetermined, and so
prices reconcile supply and demand, while the other has price being set and
quantity doing the adjustment. A choice between these might be made using
institutional knowledge that is di¢ cult to put into a VAR framework. But,
in the sign restriction approach to the market model there is no equivalent
interpretation, as the restriction employed for identi…cation essentially ties
the supply and demand elasticities together. Nevertheless any solution to
the multiple models problem has to be the same as for recursive models,
namely the introduction of extra information that enables one to discriminate
between them.
What sort of extra knowledge might be used? There is no one way of
doing this in the literature. One possibility is to continue to add on sign
restrictions which relate to longer lags in the impulses. Thus one can see
from the curves in Figure 1 that, imposing a negative e¤ect of an interest rate
shock upon output and in‡ation for ten periods rather than one period, would
eliminate many of the 1000 models in that …gure. To formally understand
why this might work, and the limitations to it, we examine the relation
between impulses noted earlier viz. Cj = Dj C0 : Because Dj can be estimated
from the VAR and does not need structural information, restrictions on Cj
translate into indirect restrictions on C0 ; and so Cj > 0 requires C0 to be such
18
that Dj C0 > 0: This may well reduce the number of possible C00 s (models)
that jointly satisfy sign restrictions on the higher order impulses as well as on
C0 . If the restrictions on Cj were quantitative then the indirect restrictions
implied on C0 would certainly narrow the possible values of C0 : However,
the same e¤ect does not necessarily hold for qualitative restrictions. For
example, if all elements of C0 are positive, and so too are the estimated D1
from a VAR(1) …tted to the data, then a restriction that the elements of
C1 are positive adds nothing to what has already been assumed about the
signs of C0 : There appears to be a belief in the literature that adding on
sign restrictions for longer impulse responses, Cj ; j > 0; provides stronger
identifying information, and this seems to stem from the Monte Carlo study
in Paustian (2007). However, as is clear from the connections that exist
between the Cj and C0 noted above, nothing guarantees this.
Quantitative information about the likely magnitude of the impulse re-
sponses is sometimes invoked in order to reduce the set of models. Thus
Kilian and Murphy (2009) argue that some estimates generated of the short-
run supply elasticity of oil and the initial impact of oil prices upon activity are
implausible, and so models producing them should be discarded. A second
group of methods looks at setting up a criterion based upon the magnitude of
impulses and minimizing it with respect to : Faust (1998) and Uhlig (2005)
do this. Uhlig’s criterion is to give a high weight to "large" standardized
impulses over "small" ones. Thus he says "Given a choice among many can-
didate monetary impulse vectors...it might therefore be desirable to pick the
one which generates a more decisive response of the variables" - Uhlig (2005,
p 414). The exact form of the penalty function varies with the application.
Thus, in general all one can say is that a value for is found by minimizing
(k)
a criterion that is a function of the magnitude of the impulse responses Cj .
Provided it is clear that this is being done, it is simply a matter of deciding if
the supplementary quantitative criterion is acceptable, although one needs to
recognize that non-sign information is being invoked to get a unique model.
A di¤erent approach to selecting a single value of by minimizing a cri-
terion is that in Fry and Pagan (2005). This begins with the observation
that researchers seem to be attracted to the idea of presenting the median as
a good summary of the central tendency of impulse responses across mod-
els. Our second observation was that the median responses may come from
di¤erent models, potentially making them impossible to utilize in exercises
such as variance decompositions. So it seems logical to …nd a single model
whose impulse responses are as close to the median values as possible. We
19
will term this the median target method (MT). The MT solution is to choose
that value of (k) that produces impulses which are as close to the median
responses as possible. To devise a criterion to do this it is necessary to recog-
nize that the impulses need to be made unit-free by standardizing them.
This is done by subtracting o¤ their median and dividing by their standard
deviation, where these are measured over whatever set of models has been
retained as satisfying the sign restrictions. These standardized impulses are
then placed in a vector (l) (in a two variable case is 4 1 as there are four
impulses) for each value (l) : Subsequently we choose the l that minimizes
M T = (l)0 (l) ; and then use that (l) to calculate impulses: Whether this
strategy produces a unique l is an empirical question, although in applica-
tions we have made it turns out to do so. In the event that the median shocks
are uncorrelated then we would …nd that the median responses would be se-
lected by this criterion. So a di¤erence between the median responses and the
M T selected responses essentially indicates that the shocks associated with
the median impulses are correlated. Consequently, the M T methodology can
be regarded as a diagnostic device.
Figure 2 also shows the median impulses and those coming from the MT
approach. Clearly major di¤erences in the e¤ects of an interest rate shock
upon output emerge when it is insisted that the shocks must come from a
single model. A comparison of the median with the adjusted measure for
other shocks does not reveal as great a di¤erence, except perhaps in the
initial impact of monetary policy on in‡ation. In Fry and Pagan (2005) we
found that applying the MT method to the data in Blanchard and Quah
(1989) produced very little di¤erence when assessing the impact of demand
and supply shocks. A number of other papers also report that the results
are not too dissimilar e.g. Rü¤er et al (2007), Canova and Paustian (2010).
It does seem to us however that it is more satisfactory to ensure that the
impulses come from the same model rather then getting them from di¤erent
models, even if in some speci…c instances the adjustment does not produce
major changes. At the very least one needs to check that a failure to insist
that shocks come from a single model has not created any distortions. The
adjustment is simple to compute. One might also observe that, although the
discussion above has been about the median, it also applies to any of the
"percentile" measures.
20
4.2 Identifying a Single Shock
In the description above it was assumed that n shocks were to be found.
But sometimes only a single shock is of interest and therefore only one shock
is isolated. Examples are Ra…q and Mallick (2008), An (2006), and Uhlig
(2005), although there are many others where the number of shocks identi-
…ed is greater than one but less than n. Dealing with the single shock case
we might still utilize the n n Q-matrices above to produce n uncorrelated
structural shocks, but only focus upon one of them. Two issues arise here.
Firstly, in some papers one has the impression that it is only necessary that
the weights used for constructing the structural shocks be an n 1 vector q
that has unit length. Uhlig’s papers often state it in this way e.g Scholl and
Uhlig (2008, p 5). If q is not selected from a Q that is orthogonal then the
resulting shock need not be uncorrelated with the remaining (unidenti…ed)
n 1 shocks. To the extent that one does not need this property for analysis
then there is no problem, but if one is trying to perform a variance decom-
position it is mandatory: Our reading of a number of papers in the literature
is that q was not selected in a way to ensure orthogonality.
A second problem arises from the following scenario. Suppose that there
are two variables and we believe that one shock has a positive initial e¤ect on
the …rst variable. However we are unwilling to describe either its e¤ects on the
second variable or to set any signs for the initial e¤ects of the second shock.
+ ?
This scenario would generate signs for C0 of ; where ? means that no
? ?
sign information is provided. It is clear that this is not enough information to
+ ?
discriminate between the shocks. Indeed, even the pattern would
+ ?
not su¢ ce, since it is possible that the impulse responses found from a draw
+ +
of Q might be ; and then we are faced with the fact that both
+ +
shocks have the same sign pattern. In any …nite number of draws one may
not encounter this, but that is just fortuitous. Hence a problem arises if there
is a failure to specify enough information to discriminate between shocks. We
will refer to this as the multiple shocks problem, as distinct from the multiple
models problem that was mentioned earlier.
As an illustration of the multiple shocks problem suppose we look at
the macro model when Q(2) in (15) is used as the weighting matrix to form
shocks. Suppose it is desired to identify only a single shock - demand -
21
using the sign restrictions from the macro model. Then, when Q(2) is used
to construct three shocks, two of these would produce the right signs, and
so both are potential demand shocks. However, we cannot accept both as
demand shocks given that they are in the same model. It is only if one
describes the signs patterns for all of the shocks that it is possible to rule
out the use of Q(2) : Consequently, if only a single shock is to be isolated
(more generally any number less than n) some information will need to be
provided on what strategy was used to deal with this issue. At the moment
little mention is made in many published articles using sign restrictions. The
problem does seem to come up quite a bit e.g. Rü¤er et al (2007) mention
that it occurs in their study, although they o¤er no comment on what they
did about it, and Mountford (2005) also alludes to it.
4.3 The Origin of Sign Restrictions

Generally these have been rather informal, although increasingly they have
been drawn from DSGE models. Thus the New Keynesian (NK) policy model
with the form
yt = 1y yt 1 + 1y Et (yt+1 )+ 1i (it Et ( t+1 )) + "yt

t = 2 t 1+ 2 Et ( t+1 ) + 2y yt + " t
it = 3 it 1 + 3y yt + 3 Et t+1 + "it ;
is often invoked as a small macro model. Assuming that there is no serial

correlation in the shocks, the solution is a VAR(1) in zt0 = yt t it ; with
the VAR(1) coe¢ cients A1 ; being a function of the NK model parameters
: Maximum likelihood estimates of can be found with the same data as
used to …t the VAR(1) for the sign restriction work. Using the MLE’s of ,
the resulting impulse responses are in …gure 3, along with those coming from
the MT method of producing a unique set of impulses under sign restrictions.
There are some very large quantitative di¤erences. Indeed, the NK impulses
often lie well outside the range coming from the 1000 models produced with
sign restrictions. It might be wondered why this is the case as the NK model
implies a VAR for the variables. But it is a restricted VAR. Hence, if the
sign restriction impulses are CjSR = C0SR D ^ jV AR ; then those from the NK
model will be governed by CjN K = C0N K D ^ N K : Consequently, there can be
j
two reasons for the di¤erence between impulses from the two approaches - a
discrepancy between the initial e¤ects C0SR and C0N K and a di¤erence between
22
the estimated VAR coe¢ cients Dj .8 All the sign restriction models keep Dj
…xed at the OLS estimates of the VAR(1). In contrast, the NK model says
there are restrictions on the VAR parameters, and these are imposed in the
ML estimation. Hence, if the NK model restrictions are incorrect, there will
be a bias in D ^ jN K ; which will show up as di¤erent impulse responses to the
unrestricted ones. This points to a rationale for just using the NK model as
a source of sign restrictions rather than exploiting its stronger implications
for VAR coe¢ cient relations. Of course the sign restrictions may depend
upon the structural model coe¢ cient values fed into the NK model, and so
it has been proposed that the model be simulated for a wide range of these,
retaining only those signs that are robust to the parameter variations.The
sign restrictions in Table 3 are likely to be broadly consistent with those found
by simulating models such as NK above for a range of parameter values.
As we will see in the next section this strategy of using theoretic models to
produce sign restrictions has been increasingly used in the literature. Canova
and Paustian (2010) examine it in some detail, simulating data from a DSGE
model and then seeing if the correct impulse responses would be recovered.
They …nd that it recovers the shocks reasonably well, provided that enough of
these are used and all shocks are identi…ed. In contrast Jääskelä and Jennings
(2010) found that they could not recover the correct impulse responses from
an NK model of a small open economy, despite using many sign restrictions.
The model-based approach to producing sign restrictions seems a useful
way to proceed, as it does not commit the user to the DSGE model, but has
the advantage that it restricts the informal approach in a fashion that prob-
ably commands reasonable assent. A lot depends on why one is performing
the VAR analysis. If one is trying to "discover" what the data says about re-
lations then imposing sign restrictions from (say) the NK model above would
not appeal as much, since one would never see (say) that interest rates had
a positive impact on in‡ation in the data. "Puzzles" like this are sometimes
the source of productive theorizing and so one should be careful about pre-
determining outcomes. Of course one check on this is available from the
draws that yielded impulse responses that didn’t agree with the sign restric-
tions. If there are a large number of these then one might well conclude that
the data is not in favour of the model used to generate the sign restrictions.
8
In this case the coe¢ cients of t 1 in the VAR equation for t are .4 (NK) and .9
(unrestricted). The covariance matrix of the VAR errors also shows some di¤erences - the
correlation between the output gap and in‡ation VAR equation errors being .4 in the data
and .1 implied by the NK model.
23
Figure 2: Comparison of the impulse responses of the small macro model
(MT method - solid line), and a New Keynesian model (dashed line).
24
Sometimes the number rejected is very high e.g. Kilian and Murphy (2009)
report only 30860 "successful" models from 1.5 million draws. Peersman and
Straub (2006) used the rejected information in this way to assess whether
the New Keynesian model was a good description of the data.
5 Some Advanced Issues with Sign Restric-

tions
A number of questions and issues arise with sign restrictions that deserve
comment. First, do sign restrictions recover the correct impulse responses?
This question is considered in the next sub-section by using the market model,
and it is concluded that, while they can be potentially recovered up to an
unknown scaling factor, the standard strategies for dealing with the multiple
models problem may mean that the true impulse responses are not isolated.
But even if there is no certainty that the correct impulses can be found it
is desirable to maximize the chances of doing so and we therefore examine
some recommendations that have been made about how to do this. Second,
if we wish to align the summative model with theoretical models it is often
necessary to recognize that, whilst the latter generally imply a VAR in all
the model variables, only a sub-set of variables are actually used in mod-
elling, and these may not follow a VAR i.e. using a VAR as the summative
model would be incorrect. We show how this complication can be dealt with
fairly easily. Finally, we ask what one does if there are both permanent and
transitory shocks in the system? Again a VAR is not the correct summative
model and it needs to be replaced by a VECM. Hence we spend some time
indicating how to adapt the methods proposed earlier in connection with
VARs to the VECM case.
5.1 Can We Recover Correct Impulse Responses from

Sign Restrictions?
In the literature one sometimes gains the impression that the answer to the
question posed above is in the a¢ rmative. But it is a tricky question to give
an answer to, as it depends on what type of experiment you wish to perform
with the impulse responses. To see why, note that, in the market model with
no dynamics, the VAR equations for pt and qt would be
25
"St
qt = (1 )"Dt (20)
(1 + ) (1 + )
"Dt "St
pt = + : (21)
(1 + ) (1 + )
Now, because "St and "Dt are uncorrelated, dividing by their standard devi-
ations would produce some base shocks jt that have unit variance, namely
1 1
1t = D "Dt ; 2t = S "St : This leads to a re-writing of (20) and (21) as
1 2t
qt = D qt = (1 ) 1t (22)
(1 + ) (1 + )
1
1 1t 2t
pt = S pt = + ; (23)
(1 + ) (1 + )
where = DS : Now the impulse responses to "jt have exactly the same signs
as those for jt but the magnitude of the latter depends upon the ratio of
the standard deviations of the cost and demand shocks ( ); and not on their
separate values. Moreover, it is clear from (22) and (23) that the impulse
responses for a unit shock to jt describe the e¤ects on qt and pt and not on
qt and pt : Fundamentally, the di¢ culty is that "Dt = D 1t and "St = S 2t ;
meaning that the jt are not equal to the demand and supply shocks, but are
scaled versions of them. Another way of describing the signi…cance of this
is that the impulse responses to unit shocks in jt indicate the responses of
qt and pt to one standard deviation shocks in "Dt and "St : Hence, in partial
answer to the question of this sub-section, correct impulse responses to one
standard deviation demand and cost shocks should be recoverable using sign
restrictions (provided the summative model is correct).
When would we be happy to have just one standard deviation responses
i.e. those provided by sign restriction information? Two cases come to mind.
One is if we are looking at when a peak in impulses occurs e.g. whether there
is over-shooting in exchange rates as in Scholl and Uhlig (2008), since the
location of the peak is invariant to any positive scaling factor for the impulses.
Another would be when variance decompositions are being computed, since
here what is needed are impulse responses to one standard deviation shocks.
When would we be less enthusiastic about the impulses found with sign
restrictions? In many policy-related contexts we want to answer questions
26
such as "what would be the responses to a 100 basis point interest rate
shock", or, in the market model context, to a unit shock in demand? In the
latter case we would need to know the standard deviation of the demand
shocks "Dt in order to work out an answer from the sign restriction impulse
responses, as these only provide the impact of one standard deviation shocks.
But, because the magnitude of the standard deviation of the demand shocks
is not an estimable quantity with just sign restrictions, we cannot construct
impulse responses to answer questions like those just posed (unless of course
D = 1 ): In this important sense the sign restriction approach would not
recover the required impulse responses.9
The discussion above has centered on whether the true impulses would
be in the range of models identi…ed by sign restrictions. Leaving aside the
issue that one might need to generate a very large number of these models
to ensure that, we are still left with the problem of which one to select. As
we have mentioned earlier, the "median" impulses are often presented. But
there is nothing which says that the true impulses would coincide with the
median. One feels that often the "median" impulses are thought of as "most
probable", but, as we pointed out earlier, the range of impulses is due to
multiple models and not any uncertainty coming from data.
To emphasize that multiple models create problems in deciding on the
values of the true impulses, we perform the following experiment. Suppose a
structural macro model is used to generate a very large sample of data, and
the model is designed so that it has impulse responses that agree with the sign
restrictions in Table 3. In the analysis of section 3 alternative impulses were
found by re-combining those for the shocks in an arbitrary recursive model.
We termed these the base shocks. Instead, let us take the base shocks to be
those from the structural macro model itself. These certainly qualify, being
uncorrelated, although they would not be available in practice. Nevertheless,
if we were fortunate enough to know them, the impulse responses generated
by sign restrictions will be combinations of the true ones, with weights given
by Q: When Q = I we will get the true impulse responses. So where in the
range of models do the true impulse responses lie? In our scenario the true
9
It might be that the di¤erences between the impulse responses seen in …gure 3 come
from the fact that the standard errors of shocks estimated from the NK model are inaccu-
rate due to speci…cation problems with that model, and so these cannot be compared with
the one unit shock responses in the jt found from the sign restrictions. As mentioned the
latter would be equivalent to doing one standard deviation shocks from the "true" model
(provided it is described by a VAR(1)).
27
impulse responses were chosen to obey the Table 3 sign restrictions and were
close to those of the empirically estimated NK model. Computing a range of
estimates by choosing di¤erent Q; it was found that the true impact impulse
responses of output and in‡ation to an interest rate shock lay at the 12.5 and
.4 percentiles, far from the median. Hence, the implication of this experiment
is that, even if the true impulses lie in the range of models generated by the
sign restrictions, we do not know where in the range they are. All we can
say is that, if the range is very narrow, then we should get a good indication
of what the true impulses are. Otherwise we cannot know.
Paustian (2007) performs Monte Carlo experiments on models where sign
restrictions on a set of (primary) variables are imposed to identify the shocks,
and then the impulses to other (secondary) variables are checked to see if
they have the correct sign. He draws two conclusions from the experiments.
Firstly, it is likely that the correct signs for the impact of the shocks on the
secondary variables will be found if the identi…ed shocks have a dominant
in‡uence on the primary variables. Secondly, the more shocks that are iden-
ti…ed the greater is the likelihood that the correct signs will be recovered.
This leads him to conclude that sign restrictions can reliably recover some
qualitative features of impulse responses under certain conditions.
The results he gets can be explained. Because the reduced form VAR
shocks are et ; and the structural ones are "t ; the connection between them
is et = B0 1 "t : If there are no lags and n = 3; the …rst "VAR" equation will
have the following relation between its error and the structural shocks:
e1t = b11 12 13
0 "1t + b0 "2t + b0 "3t ; (24)
where bij 1 11
0 are the coe¢ cients of B0 : If "1t was known, then b0 (the impact
response) could be consistently estimated by regressing e1t on "1t ; since the
omitted regressors "2t ; "3t are uncorrelated with "1t : However, "1t is not known
and sign restrictions involve combining the VAR errors ejt with weights to
extract an estimate "it . Such an estimate can be written as a combination of
the "jt :
"1t = 1 "1t + 2 "2t + 3 "3t : (25)
From (25) it is clear that a regression of e1t on "1t will produce a biased
estimator of b11
0 owing to the simultaneous presence of "2t ; "3t in the regressor
and the error term of the equation. Of course this bias will decline as the
variance of "1t increases relative to the variance of b12 13
0 "2t + b0 "3t ; and this is
the …rst conclusion Paustian reaches.
28
To see the second we just need to note that, if a second shock is identi…ed,
the regression becomes one of e1t on "1t and "2t : There is no certainty, but it
is likely that the biases will be smaller now than before. If it was the case
that "2t had been correctly estimated then it would have been eliminated
from the error term of the regression, leaving only "3t . So it is likely that,
as we estimate more shocks using sign restrictions, the bias will be reduced.
Again however this is not a general result as it depends upon the extent to
which the shocks have been correctly extracted.
5.2 Other Summative Representations

As mentioned in the introduction to the section, if we try to align theory-
inspired interpretative models (such as DSGE models) with the summative
model, we often encounter the situation that there are variables in the former
that are not observable, and so the latter model is …tted with a smaller
number of variables.10 Let the observable variables (data) be zt and the
larger set in the theoretical model be zt+ : Then, it has been known for a
long time, see Wallis (1977) and Zellner and Palm (1974), that a VAR in
zt+ becomes a Vector Autoregressive Moving Average (VARMA) in zt : Thus
a VAR will not represent the data precisely if it should be generated by a
theoretical model with latent (unobserved) variables, although, if one makes
the order su¢ ciently high, it might be argued to be approximately correct.
Basically this implies that the impulse responses from the theoretical model,
Cj+ ; will not be equal to those from an approximating VAR, unless the order
is in…nite. As shown in Kapetanios et al (2007), this di¤erence can be very
large for some shocks and models and so one needs to exercise care in using
information from theory-consistent models to identify shocks in VARs.11 Of
course it is possible that this problem is less of an issue for the signs of the
responses than it is for the magnitudes i.e. the signs of Cj+ and Cj may agree
even if the magnitudes don’t.12 Fundamentally the problem is that a VAR is
10
Technology is an obvious example of a variable in a DSGE model that is rarely present
in an estimated VAR. But it is also the case that researchers often treat the capital stock
as unobservable and so it is omitted from the list of variables in the empirical VAR.
11
Using a model that was a smaller version of The Bank of England Quarterly Model
but a VAR with only a standard set of six variables, Kapetanios et al (2007) found that
a VAR(50) and thirty thousand observations were needed to recover the true impulse
responses.
12
Indeed this seems to be supported by the simulations in Canova and Paustian (2007),
although it may just re‡ect the particular context they are working in.
29
not the correct summative model.13 As an alternative one might estimate a
VARMA process or a VAR with some latent variables, but mostly researchers
have dealt with the latent variable problem by expressing the theoretic model
in a state space form (SSF)
zt = Hzt+ (26)
zt+ = M zt+ 1 + G"+
t ; (27)
and then estimating this. Readily available computer programs such as

Dynare are designed to do so. Thus the role of a theory-inspired model
is to provide the variables in zt+ ; and the order of the VAR associated with
them, while the empirical investigator selects zt . In the DSGE model M and
G will be functions of the model parameters (H simply selects variables
and so generally doesn’t depend on ): The appropriate summative model
therefore is the SSF (26)-(27), but with M being treated as unrestricted and
G"+ t being replaced by some errors et : Once estimated the residuals e^t can
be combined together using the Q matrices dealt with earlier to produce new
shocks t and then passed through the estimated SSF to …nd the impulse
responses for these new shocks. As noted in footnote 1 it will generally be
easiest to produce initial t shocks that are uncorrelated by performing a
Choleski decomposition upon e^t ; but an alternative approach would be to
make M triangular, and to then estimate the resulting SSF by a MLE. A
program such as Dynare would enable one to do this e¢ ciently.
5.3 Permanent and Transitory Shocks

If there are as many permanent shocks to be identi…ed as there are observ-
able variables then this would imply that there is no co-integration between
the variables. Therefore, the appropriate summative model is a VAR in dif-
ferenced variables. Hence it is only how the data is measured that changes,
allowing sign restrictions to be easily imposed by working on the residuals
from the di¤erenced-variables VAR. Sometimes one sees such a summative
model in the sign restrictions literature: Examples are Jarociński and Smets
(2008) and Farrant and Peersman (2006). But it needs to be stressed that
all shocks have to be regarded as transitory for this summative model to be
13
There are even cases where there is no invertible MA representation, and so no VAR
exists.
30
correct. When there are both transitory and permanent shocks there is co-
integration, and so the summative model will be the Vector Error Correction
Model (VECM)
0 +
zt+ = zt 1 + et :
Correspondingly, a Structural VECM (SVECM) of the form
B0 zt+ = (B0 ) 0 zt+ 1 + "t ; (28)

would be used to interpret the data. Because this is a relatively simple exten-
sion of the standard approach, it does not need any extensive development.
It is only if there are latent variables that special issues arise. Hence, if zt
rather than zt+ is observed (and zt has less variables than zt+ ); the problems
identi…ed in the preceding sub-section arising from latent variables occur
again, although they can be solved in the same way, namely via a state space
form. This situation arises in many DSGE models in which the permanent
shock is technology while all other shocks are transitory.
Care does need to be exercized in …nding the structural shocks. There are
standard formulae for converting the VECM residuals e^t into n r permanent
e^Pt and r transitory e^Tt uncorrelated shocks, where r is the degree of co-
integration. It would then be necessary to re-combine these with Q matrices
to produce new uncorrelated permanent and transitory structural shocks ^Pt
and ^Tt . In doing so we need to recognize that one cannot combine the
permanent shocks to produce a transitory shock. Consequently it is simplest
to work with QP ; QT (each being Givens or QR) such that ^Pt = Qp e^Pt and
^Tt = QT e^Tt : To initialize the sequence one could begin with a recursive SVAR
in which n r of the structural shocks are designated to be permanent and
r to be transitory. The methodology outlined in Pagan and Pesaran (2008)
illustrates how such a system can be constructed and estimated.
6 Conclusion
When sign restriction work …rst began it was mainly about the identi…ca-
tion of a single shock. Since then it has become popular to identify multiple
shocks. Moreover, the range of applications has grown from the initial focus
on monetary policy. Given that sign information is rather weak we suspect
that it is best to utilize the restrictions in conjunction with parametric restric-
tions and that seems to be an emerging tendency as well. A number of other
31
themes also seem to be developing. One is that contemporaneous restric-
tions might be preferable to imposing restrictions on longer lags. Another
is that DSGE models are a useful way of …nding out likely sign restrictions,
particularly as the number of variables in the VAR grows. We have tried to
show these tendencies in the review, and have also argued that more care
often needs to be taken in devising the model that is to summarize the data,
a clear statement about whether shocks are permanent or transitory should
be provided, and an account of how the multiple models and multiple shocks
problems were dealt with must be present in the research. In some instances
the latter have not been well understood and often the responses to them
have been not well documented.
Table 1 provides a quick summary of the studies that appear in the litera-
ture, characterized by a number of the items mentioned above - viz. whether
there are a mixture of sign and other restrictions, namely whether there are
permanent shocks, how many shocks are identi…ed and whether the source of
the restrictions comes from informal ideas or from a formal theory-oriented
model. This provides a quick overview of the diversity of the studies. As
well we classify them according to the main issue being dealt with such as
the isolation of the e¤ects of technology shocks, monetary policy shocks, …s-
cal policy shocks etc. It is apparent from this table that sign restrictions
have become of increasing interest to applied researchers seeking information
about a large range of phenomena.
On balance, we do feel that sign restrictions have provided a useful tech-
nique for quantitative analysis. There are a number of instances in which
variables are simultaneously determined and it is hard to justify any para-
metric restrictions to resolve the identi…cation problem. A classic example is
that of interest and exchange rates. In these cases sign restrictions appeal.
In other situations, such as isolating monetary policy, it seems more likely
that using institutional knowledge to provide parametric restrictions would
be a better way to proceed. This points to the fact that combinations of
restrictions are likely to be what we will need to adopt in the future to carry
out good applied work.
7 References
An, L. (2006), "Exchange Rate Pass-Through: Evidence Based on Vector
Autoregression with Sign Restrictions", manuscript, University of Kentucky.
32
Bjørnland, H. and J. Halvorsen (2008), "How Does Monetary Policy Re-
spond to Exchange Rate Movements? New International Evidence", Norges
Bank Working Paper #2008/15.
Blanchard, O.J. and D. Quah (1989), "The Dynamic E¤ects of Aggregate
Demand and Supply Disturbances", The American Economic Review, 79,
655-673.
Canova, F. and G. De Nicoló (2002), "Monetary Disturbances Matter
for Business Fluctuations in the G-7", Journal of Monetary Economics, 49,
1131-1159.
Canova, F. and M. Paustian (2010), "Measurement with Some Theory:
Using Sign Restrictions to Evaluate Business Cycle Models", Economics
Working Papers, Department of Economics and Business, Universitat Pom-
peu Fabra.
Chari, V.V., P.J. Kehoe and E.R. McGrattan (2008), "Are Structural
VARs with Long-Run Restrictions Useful in Developing Business Cycle The-
ory?", Journal of Monetary Economics, 55, 1337-1352.
Cho, S. and A. Moreno (2006), "A Small-Sample Study of the New Key-
nesian Macro Model", Journal of Money, Credit and Banking, 38, 1461-1481.
Dedola, L. and S. Neri (2006), "What Does a Technology Shock Do? A
VAR Analysis with Model Based Sign Restrictions", ECB Working Paper
#705.
Dungey, M. and R.A. Fry, (2009), "Identifying Fiscal and Monetary Policy
in a Structural VAR", Economic Modelling, 26, 1147-1160.
Eickmeier, S., B. Hofmann and A. Worms (2009), "Macroeconomic Fluc-
tuations and Bank Lending: Evidence for Germany and the Euro Area",
German Economic Review, 10, 193-223.
Farrant, K. and G. Peersman (2006), "Is the Exchange Rate a Shock Ab-
sorber or a Source of Shocks? New Empirical Evidence", Journal of Money,
Credit and Banking, 38, 939-961
Faust, J. (1998), "The Robustness of Identi…ed VAR Conclusions about
Money", Carnegie-Rochester Series on Public Policy, 49, 207-244.
Francis, N., M.T. Owyang and J.E. Roush (2005), "A Flexible Finite-
Horizon Identi…cation of Technology Shocks", Board of the Governors of the
Federal Reserve System, International Finance Discussion Papers #832.
Francis, N., M.T. Owyang and A.T. Theodorou (2003), "The Use of Long-
Run Restrictions for the Identi…cation of Technology Shocks", Review, Fed-
eral Reserve Bank of St Louis November-December, 53-66.
33
Fry, R.A. and A.R. Pagan (2005), "Some Issues in Using VARs for Macro-
econometric Research", CAMA Working Paper #19.
Fry, R.A. and A.R. Pagan (2007), "Some Issues in Using Sign Restrictions
for Identifying Structural VARs", NCER Working Paper #14.
Fujita, S. (2009), "Dynamics of Worker Flows and Vacancies: Evidence
from the Sign Restriction Approach", Journal of Applied Econometrics, forth-
coming.
Gali, J. (1992), "How Well does the ISLM Model Fit Postwar U.S. Data?",
Quarterly Journal of Economics, 107, 709-735.
Hau, H. and H. Rey (2004), "Can Portfolio Rebalancing Explain the
Dynamics of Equity Returns, Equity Flows and Exchange Rates", American
Economic Review, 94, 126-133.
Jääskelä and Jennings (2010), "Monetary Policy and the Exchange Rate:
Evaluation of VAR Models", Discussion Paper , Reserve Bank of Australia
Reseach Discussion Paper ( forthcoming)
Jarociński, M. and F.R. Smets (2008), "House Prices and the Stance of
Monetary Policy", Review, Federal Reserve Bank of St Louis, July/August
2008, 339-365.
Kapetanios, G., A.R. Pagan and A. Scott (2007), "Making a Match: Com-
bining Theory and Evidence in Policy-Oriented Macroeconomic Modelling",
Journal of Econometrics, 136, 565-594.
Kilian, L. and D. Murphy (2009), "Why Agnostic Sign Restrictions are
Not Enough: Understanding the Dynamics of Oil Market VAR Models",
manuscript, University of Michigan.
Leamer (1978), Speci…cation Searches and Ad-Hoc Inference for Nonex-
perimental Data, John Wiley, New York
Lewis, V.J. (2007), "Productivity and the Euro-Dollar Real Exchange
Rate", Review of World Economics, 143, 324-347.
Mountford, A. (2005), "Leaning into the Wind: A Structural VAR Inves-
tigation of UK Monetary Policy", Oxford Bulletin of Economics and Statis-
tics, 67, 597-621.
Mountford, A. and H. Uhlig (2005), "What are the E¤ects of Fiscal Pol-
icy Shocks", Discussion Paper 2005-039, SFB 649, Humboldt-Universitat,
Berlin.
Mountford, A. and H. Uhlig (2008), "What are the E¤ects of Fiscal Policy
Shocks?", NBER Working Paper #14551.
Pagan, A.R. and M.H. Pesaran (2008), "Econometric Analysis of Struc-
tural Systems with Permanent and Transitory Shocks", Journal of Economic
34
Dynamics and Control, 32, 3376-3395.
Paustian, M. (2007), "Assessing Sign Restrictions" The B.E. Journal of
Macroeconomics, 7, 1-37.
Peersman, G. (2005), "What Caused the Early Millennium Slowdown?
Evidence Based on Autoregressions", Journal of Applied Econometrics, 20,
185-207.
Peersman, G. and R. Straub (2006), "Putting the New Keynesian Model
to a Test", IMF Working Paper, 06/135.
Peersman, G. and R. Straub (2009), "Technology Shocks and Robust
Sign Restrictions in a Euro Area SVAR", International Economic Review,
50, 727-750.
Preston, A.J. (1978), "Concepts of Structure and Model Identi…ability
for Econometric Systems" in A.R. Bergstrom et al., Stability and In‡ation
(Wiley, 1978), 275–297.
Quenouille, M.H. (1957), The Analysis of Multiple Time Series, Gri¢ n’s
Statistical Monographs and Course No 1 (Gri¢ n, London).
Ra…q, M.S. and S.K. Mallick (2008), "The E¤ects of Monetary Policy on
Output in EMU3 A Sign Restriction Approach", Journal of Macroeconomics,
30, 1756-1791.
Rubio-Ramírez, J.F., D.F. Waggoner and T. Zha (2005), "Markov-Switching
Structural Vector Autoregressions: Theory and Application", Working Paper
#2005-27, Federal Reserve Bank of Atlanta.
Rü¤er, R., M. Sanchez and J-G Shen (2007), "Emerging Asia’s Growth
and Integration: How Autonomous are Business Cycles", ECB Working Pa-
per #715.
Sanchez, M. (2007), "What Drives Business Cycles and International
Trade in Emerging Market Economies?", ECB Working Paper #730.
Scholl, A. and H. Uhlig, (2008), "New Evidence on the Puzzles: Re-
sults from Agnostic Identi…cation on Monetary Policy and Exchange Rates",
Journal of International Economics, 76, 1-13.
Sims, C.A. (1980), "Macroeconomics and Reality", Econometrica, 48, 1-
48.
Uhlig, H. (2005), "What are the E¤ects of Monetary Policy on Output?
Results from an Agnostic Identi…cation Procedure", Journal of Monetary
Economics, 52, 381-419.
Vargas-Silva, C. (2008), "Monetary Policy and US Housing: A VAR Im-
posing Sign Restrictions", Journal of Macroeconomics, 30, 977-990.
35
Wallis, K.F. (1977), "Multiple Time Series and the Final Form of Econo-
metric Models", Econometrica, 45, 1481-1497.
Wold, H.O. (1951), "Dynamic Systems of the Recursive Type- Economic
and Statistical Aspects", Sankhya, 11 (1951), 205-216.
Zellner, A. and F. Palm (1974), "Time Series Analysis and Simultaneous
Equation Econometric Models", Journal of Econometrics, 2, 17-54.
36
List of NCER Working Papers

No. 56 (Download full text)
Mardi Dungey and Lyudmyla Hvozdyk
Cojumping: Evidence from the US Treasury Bond and Futures Markets

Martin G. Kocher, Marc V. Lenz and Matthias Sutter
Psychological pressure in competitive environments: Evidence from a randomized natural
experiment: Comment

Adam Clements and Annastiina Silvennoinen
Estimation of a volatility model and portfolio allocation

Luis Catão and Adrian Pagan
The Credit Channel and Monetary Transmission in Brazil and Chile: A Structured VAR
Approach
Vlad Pavlov and Stan Hurn
Testing the Profitability of Technical Analysis as a Portfolio Selection Strategy

Sue Bridgewater, Lawrence M. Kahn and Amanda H. Goodall
Substitution Between Managers and Subordinates: Evidence from British Football

Martin Fukac and Adrian Pagan
Structural Macro‐Econometric Modelling in a Policy Environment

Tim M Christensen, Stan Hurn and Adrian Pagan
Detecting Common Dynamics in Transitory Components

Egon Franck, Erwin Verbeek and Stephan Nüesch
Inter‐market Arbitrage in Sports Betting

Raul Caruso
Relational Good at Work! Crime and Sport Participation in Italy. Evidence from Panel Data
Regional Analysis over the Period 1997‐2003.

No. 46 (Download full text) (Accepted)
Peter Dawson and Stephen Dobson
The Influence of Social Pressure and Nationality on Individual Decisions: Evidence from the
Behaviour of Referees

Ralf Becker, Adam Clements and Christopher Coleman‐Fenn
Forecast performance of implied volatility and the impact of the volatility risk premium

Adam Clements and Annastiina Silvennoinen
On the economic benefit of utility based estimation of a volatility model

Adam Clements and Ralf Becker
A nonparametric approach to forecasting realized volatility

Uwe Dulleck, Rudolf Kerschbamer and Matthias Sutter
The Economics of Credence Goods: On the Role of Liability, Verifiability, Reputation and
Competition

Adam Clements, Mark Doolan, Stan Hurn and Ralf Becker
On the efficacy of techniques for evaluating multivariate volatility forecasts

Lawrence M. Kahn
The Economics of Discrimination: Evidence from Basketball

Don Harding and Adrian Pagan
An Econometric Analysis of Some Models for Constructed Binary Time Series

Richard Dennis
Timeless Perspective Policymaking: When is Discretion Superior?

Paul Frijters, Amy Y.C. Liu and Xin Meng
Are optimistic expectations keeping the Chinese happy?

Benno Torgler, Markus Schaffner, Bruno S. Frey, Sascha L. Schmidt and Uwe Dulleck
Inequality Aversion and Performance in and on the Field

T M Christensen, A. S. Hurn and K A Lindsay
Discrete time‐series models when counts are unobservable

Adam Clements, A S Hurn and K A Lindsay
Developing analytical distributions for temperature indices for the purposes of pricing
temperature‐based weather derivatives

Adam Clements, A S Hurn and K A Lindsay
Estimating the Payoffs of Temperature‐based Weather Derivatives

T M Christensen, A S Hurn and K A Lindsay
The Devil is in the Detail: Hints for Practical Optimisation

Uwe Dulleck, Franz Hackl, Bernhard Weiss and Rudolf Winter‐Ebmer
Buying Online: Sequential Decision Making by Shopbot Visitors

Richard Dennis
Model Uncertainty and Monetary Policy

Richard Dennis
The Frequency of Price Adjustment and New Keynesian Business Cycle Dynamics

Paul Frijters and Aydogan Ulker
Robustness in Health Research: Do differences in health measures, techniques, and time
frame matter?

Paul Frijters, David W. Johnston, Manisha Shah and Michael A. Shields
Early Child Development and Maternal Labor Force Participation: Using Handedness as an
Instrument

Paul Frijters and Tony Beatton
The mystery of the U‐shaped relationship between happiness and age.

T M Christensen, A S Hurn and K A Lindsay
It never rains but it pours: Modelling the persistence of spikes in electricity prices

Ralf Becker, Adam Clements and Andrew McClelland
The Jump component of S&P 500 volatility and the VIX index

A. S. Hurn and V.Pavlov
Momentum in Australian Stock Returns: An Update

Mardi Dungey, George Milunovich and Susan Thorp
Unobservable Shocks as Carriers of Contagion: A Dynamic Analysis Using Identified
Structural GARCH

No. 21 (Download full text) (forthcoming)
Mardi Dungey and Adrian Pagan
Extending an SVAR Model of the Australian Economy

Benno Torgler, Nemanja Antic and Uwe Dulleck
Mirror, Mirror on the Wall, who is the Happiest of Them All?

Justina AV Fischer and Benno Torgler
Social Capital And Relative Income Concerns: Evidence From 26 Countries

Ralf Becker and Adam Clements
Forecasting stock market volatility conditional on macroeconomic conditions.

Ralf Becker and Adam Clements
Are combination forecasts of S&P 500 volatility statistically superior?

Uwe Dulleck and Neil Foster
Imported Equipment, Human Capital and Economic Growth in Developing Countries

Ralf Becker, Adam Clements and James Curchin
Does implied volatility reflect a wider information set than econometric forecasts?

Renee Fry and Adrian Pagan
Some Issues in Using Sign Restrictions for Identifying Structural VARs

Adrian Pagan
Weak Instruments: A Guide to the Literature

Ronald G. Cummings, Jorge Martinez‐Vazquez, Michael McKee and Benno Torgler
Effects of Tax Morale on Tax Compliance: Experimental and Survey Evidence

Benno Torgler, Sascha L. Schmidt and Bruno S. Frey
The Power of Positional Concerns: A Panel Analysis

Ralf Becker, Stan Hurn and Vlad Pavlov
Modelling Spikes in Electricity Prices

A. Hurn, J. Jeisman and K. Lindsay
Teaching an Old Dog New Tricks: Improved Estimation of the Parameters of Stochastic
Differential Equations by Numerical Solution of the Fokker‐Planck Equation

Stan Hurn and Ralf Becker
Testing for nonlinearity in mean in the presence of heteroskedasticity.

No. 7 (Download full text) (published)
Adrian Pagan and Hashem Pesaran
On Econometric Analysis of Structural Systems with Permanent and Transitory Shocks and
Exogenous Variables.

No. 6 (Download full text) (published)
Martin Fukac and Adrian Pagan
Limited Information Estimation and Evaluation of DSGE Models.

Andrew E. Clark, Paul Frijters and Michael A. Shields
Income and Happiness: Evidence, Explanations and Economic Implications.

Louis J. Maccini and Adrian Pagan
Inventories, Fluctuations and Business Cycles.

Adam Clements, Stan Hurn and Scott White
Estimating Stochastic Volatility Models Using a Discrete Non‐linear Filter.

Stan Hurn, J.Jeisman and K.A. Lindsay
Seeing the Wood for the Trees: A Critical Evaluation of Methods to Estimate the
Parameters of Stochastic Differential Equations.

Adrian Pagan and Don Harding
The Econometric Analysis of Constructed Binary Time Series.

Sign Restriction Fry and Pagan

Uploaded by

Copyright:

Available Formats

Sign Restriction Fry and Pagan

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sign Restriction Fry and Pagan

Uploaded by

Copyright:

Available Formats

NCER Working Paper Series

Sign Restrictions in Structural Vector

Working Paper #57

2 The VAR Representation and Two Simple Structural Mod-

5 Some Advanced Issues with Sign Restrictions 25

Fluctuations Peersman (2005) STNI

qt = aqq qt 1 + aqp pt 1 + e1t (6)

variablenshock Demand Supply

2.2.2 A Small Macro Model

yt = zt0 1 y + e1t (11)

variablenshock Demand Cost-Push Interest Rate

3.2 Sign Restrictions

3.2.1 Givens Matrices

QG ( ) = Q12 ( 1 ) Q13 ( 2 ) Q23 ( 3 ):

It’s clear that QG is orthogonal and so shocks formed as t = QG t will be

3.2.2 Householder Transformations

3.3 Sign Restrictions in the Market Model

Three parameters are estimated in this system - = cov(qt ; pt )=var(pt ) and

Letting 1 = cos and 2 = sin the two equations can be written as

with impulse responses of (pt ; qt ) to jt being

4.3 The Origin of Sign Restrictions

yt = 1y yt 1 + 1y Et (yt+1 )+ 1i (it Et ( t+1 )) + "yt

is often invoked as a small macro model. Assuming that there is no serial

5 Some Advanced Issues with Sign Restric-

5.1 Can We Recover Correct Impulse Responses from

5.2 Other Summative Representations

and then estimating this. Readily available computer programs such as

5.3 Permanent and Transitory Shocks

B0 zt+ = (B0 ) 0 zt+ 1 + "t ; (28)

You might also like