back to square one: identification issues in dsge models fabio canova, luca sala marc goñi – 19...
TRANSCRIPT
Back to square one: Identification issues in DSGE models
Fabio Canova, Luca Sala
Marc Goñi – 19 th April
Motivation
In the last years DSGE models have greatly evolved with the
objective of better forecasting and deriving policy implications
- Development in the specification of DSGE
- Comparing models with the data through the ability to match
conditional dynamics in response to structural shocks
However, this inference depends crucially on identification, which
has been partially ignored
This paper investigates identificability issues in DSGE models in
the class of minimum distance estimators
Introduction Generics of id Population id Sample id Diagnosting id Conclusion
Literature Review
Choi and Phillips (1992), Stock and Wright (2000), Rosen (2006), Kleibergen and Mavroidis (2008)
Beyer and Farmer (2004), Moon and Schorfheide (2007)
Christiano et al (2006), Fernandez-Villaverde et al (2007), Chari et al (2008)
Introduction Generics of id Population id Sample id Diagnosting id Conclusion
Outline
1. Generics of Identification
2. Population Identification
Christiano et al (2005), Smets and Wouters (2003)
3. Sample Identification
4. Dealing with Identification
Introduction Generics of id Population id Sample id Diagnosting id Conclusion
Identification
Identification is the ability to draw inference about the parameters of the model from the data
Identification requires the objective function to have
- A unique extreme at the true parameter
- Sufficient curvature in all the relevant dimension
The mapping from structural parameters to the objective function is usually non-linear or doesn’t have a closed form solution
Introduction Generics of identification Population id Sample id Diagnosting id Conclusion
Problems
• Under Identification If the objective function is independent of certain structural parameters
• Partial Identification If the parameters enter the objective function only proportionally and they cannot be separately analyzed
• Weak Identification If the objective function does not have enough curvature in all the relevant dimensions
This problems can induce Observational Equivalence, i.e, that different models with different theoretical implications become indistinguishible
Introduction Generics of identification Population id Sample id Diagnosting id Conclusion
Source of the problems
1. Location of the true parameters
2. Choice of the objective function
Consider the optimality conditions of a DSGE model
The unique stable RE solution is
In State Space representation
Introduction Generics of identification Population id Sample id Diagnosting id Conclusion
The Likelihood function provides a natural upper bound to identification of the information available in the data.
Using the Kalman filter and assuming normality for the errors
And, thus, an identification upper bound is
Which, compared to a minimum distance objective function
Introduction Generics of identification Population id Sample id Diagnosting id Conclusion
3. Mapping of Structural parameters and sample objective function
Solution mapping
Linking the parameters and the coefs. of the solution
Parameters disappear from the solution, do not have independent variability
Moment mapping
Links the coefs of the solution with the function of interest
Selection of a particular Impulse Response may poorly identify coeficients
Objective function mapping
Links the function of interest with the pop. objective function
Function may not have a unique minimum or may not display enough curvature
Data mapping
Links the pop. Objective function with the sample objective function
Estimated VAR responses may not reflect population ones
Introduction Generics of identification Population id Sample id Diagnosting id Conclusion
A simple example
• Solution mapping
Parameter a1 , a3 , a5 disappears from the solution
Introduction Generics of identification Population id Sample id Diagnosting id Conclusion
A simple example
• Moment mapping
The impulse response takes the form of
Even if we pick responses to all shocks, some parameters remain underidentified
a2a4 respond jointly to e3
Introduction Generics of identification Population id Sample id Diagnosting id Conclusion
A simple example
• Objective mapping
Choosing a MD objective function
Weak identification problems
Introduction Generics of identification Population id Sample id Diagnosting id Conclusion
Solutions
1. Calibration
Calibrate some of the parameters based on micro-evidence…
Problem: If the calibrated parameters are partially identified small calibration differences may shift the estimates
Introduction Generics of identification Population id Sample id Diagnosting id Conclusion
2. Bayesian Methods
Estimate structural parameters with Bayesian techniques
If the parameter space is not variation free, Id. Problems can be detected by setting a more diffuse prior and checking if the posterior becomes more diffuse
However, this can be driven by restrictions
Bayesian methods plus tight prior produce well behaved posteriors even when the objective function behaves poorly
Introduction Generics of identification Population id Sample id Diagnosting id Conclusion
3. Serially Correlated Disturbances
Allowing for serially correlated disturbances maintains the forward looking coefficients in the solution
However, separating internal and external propagation parameters might be difficult.
Introduction Generics of identification Population id Sample id Diagnosting id Conclusion
Introduction Generics of id Population identification Sample id Diagnosting id Conclusion
Even when the true model is known, identification problems may make inference not feasible
Consider a standard DSGE (Christiano et al, Dedola and Neri, Smets and Wouters). The analytical mapping between structural parameters and the objective function is no longer available
Instead examine the slope of the distance function in a neighborhood of the true parameter
Population Idenification
Model
Introduction Generics of id Population identification Sample id Diagnosting id Conclusion
Introduction Generics of id Population identification Sample id Diagnosting id Conclusion
Let
and
For each parameter in compute the elasticity of the distance function to it by varying it while holding fixed the rest of the parameters to the true value.
Preliminary evidence for weak identification: Although distance functions have a unique minimum at the true parameter, variation within the neighborhood is small
Preliminary identification evidence
Introduction Generics of id Population identification Sample id Diagnosting id Conclusion
Check how severe weak identification actually is and to test if it can yield observational equivalence.
1. Construct the distribution of the distance function
Pick randomly 100000 vector parameters from before and
compute the distance between its Impulse responses and 5
benchmark IR
True models: benchmark model with monetary shocks,
either p, w stickiness or indexation out and benchmark with
monetary and technology shocks
2. Pick draws in the 0.1 percentile of the distribution
Size of Identification Problems
Introduction Generics of id Population identification Sample id Diagnosting id Conclusion
Case 1: Large intervals make difficult to infer how important p, w stickiness and indexation are
Case 1-4: As intervals are similar, it is difficult to infer which friction matters (observational equivalence)
Case 5: Same results when more shocks are added
So, weak and partial identification problems are severe and may induce observational equivalence
Introduction Generics of id Population identification Sample id Diagnosting id Conclusion
These poorly identified DSGE model may be good for forecasting but not optimal for policy inference
In the presence of such a weak and partial identification problems, one needs to bring information external to the dynamics to be able to interpret the estimates
Remark
Introduction Generics of id Population id Sample identification Diagnosting id Conclusion
What are the effects of these population identification issues when the analysis has to be conducted with sample data, rather than populational.
Simulate 500 time series from the true model
Estimate a 6-variable VAR with 6 lags, identify monetary shocks and construct the data based Impulse Responses
-Avoid non-invertibility and correctly identify shocks
Estimate theoretical parameters minimizing the distance with the VAR based Impulse Responses
Sample problems
Introduction Generics of id Population id Sample identification Diagnosting id Conclusion
-Mean estimates do not depend on sample size
-Standard errors and biases decrease with sample size, but
are large
-Bimodal distributions of parameter estimates with peaks at
the boundaries
That is, standard asymptotic approximations seem not reliable
Results
Introduction Generics of id Population id Sample identification Diagnosting id Conclusion
-Model based responses fall in the range of VAR-based responses confidence bands
-Mean estimates are statistically and economically different from the true ones (observational equivalence)
That is, the technique of showing model responses in VAR-based confidence bands may lead to wrong inference in the presence of population identification problems
Results
Introduction Generics of id Population id Sample identification Diagnosting id Conclusion
Under certain conditions, asymptotic methods to compute estimates are robust to Identification problem (Stock and Wright, Kleibergen and Mavroeidis)
Rosen (2006) methodology yields similar results here
Alternatives
Introduction Generics of id Population id Sample id Diagnosting identification Conclusion
Consider the following mapping of structural parameters to sample objective function
1.First order Taylor expansion
To translate information from the function to the parameter we need to be invertible
Thus, if the rank is not full: under identification
If the eigenvalues are small: weak and partial identification
Theory
Introduction Generics of id Population id Sample id Diagnosting identification Conclusion
2. Split the problem in two
First find the θ that minimizes the distance between data and model VAR parameters (solution mapping) and then find the reduced form parameters parameters that make the IR close
Check the rank and the eigenvalues of and
Introduction Generics of id Population id Sample id Diagnosting identification Conclusion
3. Split the problem in three
When only one estimate is available, the problem can be splitted into a solution mapping, moment mapping and data mapping
Compute F and G by calibrating θ and then performing sensitivity analysis
Compare fixed and estimated parameter for identificability issues (as before)
Introduction Generics of id Population id Sample id Diagnosting identification Conclusion
Methods to test for the rank and the size of the eigenvalues
1.Anderson (1984)
Estimates of eigenvalues have asymptotic normal distribution.
Therefore, test if the smallest eigenvalue is different from 0
We can normalize the test by using the ratio of the sum of the
smallest eigenvalues over the sum of all eigenvalues
2.Concentration Statistics
Measures the curvature of the objective function around θo
For large values of accept that the objective function
has an optimum at 0
Practical issues
Introduction Generics of id Population id Sample id Diagnosting identification Conclusion
3.Standard errors not useful
Relatively small std. errors can coexist with identification
issues
Identification analysis should preceed estimation
Standard errors do not infere if the problem is model or
data based
Introduction Generics of id Population id Sample id Diagnosting identification Conclusion
Applying this methods to our example
-G’G matrix has one eigenvalue representing 99.9% of the trace
-G’F’FG smallest 13 eigenvalues account for 0.001% of the trace
Thus, solution mapping (GG) is the source of the problem
-Gs’Gs (mapping structural coefs with LOM coefs) has one
eigenvalue representing 99.9% of the trace
Thus, the source of the identification problem is the insensibility of the Law of Motion coeficients to the structural parameters
An application
Introduction Generics of id Population id Sample id Diagnosting identification Conclusion
Model reespecification
-Reparametrize commonly used functions
-Choice of different pivotal points around which log linearize
-Use Higher order approximations
Solution
Introduction Generics of id Population id Sample id Diagnosting identification Conclusion
-Identification problems ignored for a long time, with consequences in policy recommendations
-Tools given here to be applied before structural estimation
-Montecarlo Methods may help
-When choosing the objective function, choose the most informative
-Bringing more data only helps if the problem is data based
-Robust methods exists but only give you intervals
Conclusion