Professional Documents
Culture Documents
Structural Vector Autoregressions With Imperfect Identifying Information
Structural Vector Autoregressions With Imperfect Identifying Information
Structural Vector Autoregressions With Imperfect Identifying Information
net/publication/360910830
CITATIONS READS
0 80
2 authors, including:
Christiane Baumeister
University of Notre Dame
60 PUBLICATIONS 3,629 CITATIONS
SEE PROFILE
All content following this page was uploaded by Christiane Baumeister on 01 June 2022.
∗
The traditional approach to identification
Baumeister: Department of Economics, Univer-
sity of Notre Dame (email: cjsbaumeister@gmail.com).
is to place enough restrictions on the struc-
Hamilton: Department of Economics, University of Cal- tural parameters in (2) to be able to estimate
ifornia at San Diego (email: jhamilton@ucsd.edu). the remaining structural parameters directly
1
2 PAPERS AND PROCEEDINGS MONTH YEAR
from the OLS regression estimates in (1). and Hamilton (2015) to the dataset in Kil-
One common approach is to assume a recur- ian (2009) with a particular prior for A in
sive form to the structural model in which which the diagonal elements of A are nor-
the upper-triangular elements of A are all malized at unity. We used degenerate priors
zero and the structural shocks ut are un- for the upper-triangular elements centered at
correlated with each other and across time. zero and with zero variance to represent the
In this case the remaining structural para- inference of an analyst who was absolutely
meters can be easily estimated from the re- certain that these three parameters had to
gression estimates and a Cholesky factoriza- be zero. We used priors with fat tails and
tion of the variance matrix of the forecast a huge variance for the lower-triangular el-
errors in (1). This approach to identifica- ements, thus treating essentially any value
tion amounts to assuming, for example, that for these parameters as reasonable. The re-
oil producers respond to changes in economic sults of that Bayesian inference are reported
activity or oil prices only with a lag, meaning in the second column of Figure 1. Not sur-
that the short-run price elasticity of oil sup- prisingly, this is identical to the first column.
ply is zero. An example of this approach The Cholesky approach to identification can
is the analysis in Kilian (2009). The first be given a Bayesian interpretation. The
column of Figure 1 reproduces his structural Bayesian prior that is implicit in Cholesky
estimates of the dynamic effects on oil pro- identification claims to know with certainty
duction, economic activity, and the oil price that the upper-triangular elements of A are
of a one-standard-deviation decrease in u1 , all zero but has no useful information at all
along with 68% and 95% confidence inter- about the lower-triangular elements.
vals. These results suggest that a decrease
in oil production is followed by a decrease in C. A Bayesian generalization of the
economic activity and increase in oil prices. traditional approach
F 1. E . C 1: C -
C . C 2: B C . C 3: B "
C .
metric sense. The error bands in the the second and third columns of Figure 1)
first two columns reflect uncertainty that re- provide the optimal estimate of the dynamic
sults because we do not know the true val- structural effects given both the finite data
ues of the reduced-form parameters in (1) set and our imperfect understanding of the
perfectly but had to estimate them from structure.
a limited dataset. If we had more ob-
servations, that uncertainty would become This Bayesian interpretation invites an-
smaller, and would eventually vanish with other natural question — are the three upper-
an infinite sample size. By contrast, the triangular elements of A the only aspects of
bands in the third column represent both the economic structure about which we have
uncertainty about reduced-form parameters any prior information? For many applied
and uncertainty about how we give those pa- researchers, these zeros were imposed as a
rameters a structural interpretation. The matter of necessity — I need 3 zeros for iden-
first source of uncertainty would disappear tification, and well, here are three. In the
as the sample size gets larger, but the second example here, we do have good information
would not. In general we would have a set that the short-run price elasticity of supply
of possible values for the structural parame- is indeed small. But we also have good
ters, known as the identified set, that would information that the short-run price elastic-
remain as possible answers even if the sample ity of demand is small, which tells us some-
size was infinite. Baumeister and Hamilton thing about the (3,1) element of A. We
(2015) characterized the identified set and also know from economic theory that sup-
noted that the Bayesian approach allows us ply should respond positively to a price in-
to rank observationally equivalent outcomes crease and demand respond negatively, and
based on their prior plausibility. Baumeis- know something about plausible effects of oil
ter and Hamilton (2018) showed if we want prices on economic activity. By bringing in
to minimize the absolute error of our infer- information from a variety of sources — all of
ence, the medians of the posterior distribu- it imperfect, none of it telling us the value
tion (represented by the solid blue lines in of any structural parameter with certainty
— we can offset some of the loss in accuracy
4 PAPERS AND PROCEEDINGS MONTH YEAR
tion about other aspects of the structure. Useful Prior Information.” Econometrica,
This allows us to perform inference in a sys- 83(5): 1963—1999.
tem in which we openly acknowledge doubts
Baumeister, Christiane, and James D
about both the validity of instruments and
Hamilton. 2018. “Inference in Structural
about our other identifying information and
Vector Autoregressions When the Iden-
to incorporate these doubts into statistical
tifying Assumptions Are Not Fully Be-
summaries of what we can conclude from the
lieved: Re-evaluating the Role of Mon-
observed data.
etary Policy in Economic Fluctuations.”
II. Conclusions Journal of Monetary Economics, 100: 48—
65.
Many researchers treat identifying as- Baumeister, Christiane, and James D
sumptions as a necessary evil, seeing the Hamilton. 2019. “Structural Interpre-
task to be to make enough of them to tation of Vector Autoregressions with
get sharp answers to questions of interest. Incomplete Identification: Revisiting
We suggest that researchers instead begin the Role of Oil Supply and Demand
by thinking carefully about the meaning Shocks.” American Economic Review,
of the structural parameters, looking for 109(5): 1873—1910.
information that we can obtain about
these from other datasets, model cali- Kilian, Lutz. 2009. “Not All Oil Price
brations, or economic theory. The next Shocks Are Alike: Disentangling De-
step is to summarize that knowledge in mand and Supply Shocks in the Crude
the form of a prior density p(A) that Oil Market.” American Economic Review,
accurately reflects both the information 99(3): 1053—69.
and our doubts, for example, using large Mertens, Karel, and Morten O Ravn.
variances for features we honestly know 2013. “The Dynamic Effects of Personal
little about. Once prior information has and Corporate Income Tax Changes in the
been represented in this way, it’s simply United States.” American Economic Re-
a matter of plugging their subroutine view, 103(4): 1212—47.
to calculate p(A) into the code posted at
https://sites.google.com/site/cjsbaumeister/ Nguyen, Lam. 2019. “Bayesian In-
research. ference in Structural Vector Au-
Identification is not an either-or decision toregression with Sign Restric-
of whether the researcher should use hard tions and External Instruments.”
restrictions, sign restrictions, or proxy vari- https://sites.google.com/view/nguyenlam/
ables. They can all be used together, even if research.
we have doubts about each one of them. A Rubio-Ramirez, Juan F, Daniel F
Bayesian perspective that unifies all the dif- Waggoner, and Tao Zha. 2010.
ferent approaches into a common framework “Structural Vector Autoregressions: The-
makes it possible to take advantage of the ory of Identification and Algorithms for
strengths of each method while incorporat- Inference.” The Review of Economic Stud-
ing honest doubts about each into the final ies, 77(2): 665—696.
statistical summary.
Stock, James H, and Mark W Wat-
REFERENCES son. 2012. “Disentangling the Channels of
the 2007-2009 Recession.” Brookings Pa-
Baumeister, Christiane, and Aram pers on Economic Activity, 43(1): 81—156.
Derdzyan. 2021. “Structural Changes in
Uhlig, Harald. 2005. “What Are the Ef-
Energy Markets.” Unpublished.
fects of Monetary Policy on Output? Re-
Baumeister, Christiane, and James D sults From an Agnostic Identification Pro-
Hamilton. 2015. “Sign Restrictions, cedure.” Journal of Monetary Economics,
Structural Vector Autoregressions, and 52(2): 381—419.