Identification analysis of DSGE models with DYNARE by
Identification analysis of DSGE models with DYNARE by M. Ratto (Joint Research Centre) with contribution of N. Iskrev, Bank of Portugal Discussion by Stephan Fahr, ECB (Usual disclaimer applies) 20 September 2011
DYNARE features MODEL parameters Forecasting Set parameters: parameter block Calibration: _steadystate. m Uniqueness: steady (1 solution) Determinacy: check Sensitivity: dynare_sensitivity Identification … Simulation Moments, IRFs, variance decomp. MLE Bayesian estimation (log-Likelihood) Shock decomposition GMM / SMM Identification … Data Consistent story telling / policy implications Addressing specific questions: single parameter, modelling block
Identification: what are the problems? Objective: consistent story telling • How important is a specific parameter for the implications of the model? Moments (mean, variance, covariance, auto-covariance), impulse responses, … • Potential problems: 1. Under-identification: parameter does not affect model moments. 2. Partial identification: only group of parameters affects model moments Simulation delivers different results for different parameter values, but data cannot pin down the specific parameter value. 3. Weak identification: parameters affects moments only little Explanatory power of parameter is small / model is rigid • Ultimately: how capable is the model (and its parameters) in replicating specific features of the data • This is not about shock identification.
Features of the toolbox I How well are parameters in DSGE models identified? • • For assessment, Identification Toolbox exploits • state equation • first and second moments of the model. • analytical derivatives • efficient computation procedures. Issues covered Non-identification based on rank deficiencies 1. Under-identification: Parameter does not affect moments 2. Partial identification: Parameters are collinear and cannot be identified seperably 3. Code identifies paramter(s) that lead to rank deficiency Identification strength based on information matrix 1. T-test type analysis using parameter value and standard deviation 2. Weak identification through sensitivity of moments or near multi-collinearity 3. Normalisation of sensitivity measures (elasticities) to be unit-free.
Features of the toolbox II Monte Carlo evaluation of parameter space • Based on previous work on Global Sensitivity Analysis • Drawing paramter combinations from prior parameter space Discarded simulations (due to non-existant solution) reported individually. Q: What are the parameter combinations? Can common features of discarded simulations be reported? Is the discarded parameter space covering a specific region? How does sensitivity evolve once approaching the boundaries of the parameter space? • Sensitivity measure based on cond. variance relative to uncond. variance How much does output vary due to parameter? Procedure combines 1. Sensitivity of output vis-à-vis parameter and 2. Parameter uncertainty due to variance of prior distrbution
Usefulness as a tool Identification toolbox addresses • Main elements of local identification problems • Easy and straightforward implementation (brief command) • Helps to better understand the inner working of the model • Many options require getting used to interpreting the output
Usefulness as a tool Some issues / nice to haves • Adding additional superfluous parameter led to error in my case. Possible bug? • Example of perfect collinearity Collinearity patterns with 2 parameter(s) Parameter [ Expl. params e_a [ alp bet e_m [ bet rho alp [ e_a bet [ e_a alp gam [ mst del mst [ gam rho [ e_m del ] ] ] ] cosn 1. 000 0. 992 1. 000 0. 989 0. 701 0. 992 • If collinearity is found, could one back out the equation? • Illustration of IRF sensitivity for parameters with largest SVD would be interesting. • Distinction if identification is based on 1 st and/or 2 nd moments. Relevant for 2 nd moment / IRF matching • How can higher-order moments help in better identifying parameters?
Next step (for the economist) What to do if identification problems exist? 1. Find other data? YES: Concentrate on moments/ IRFs to be explained NO: Identification problem is model issue, not data problem 2. Find model-external methods to identify parameters? YES, but gives rise to tensions between calibration and estimation strategy 3. Change model blocks? YES: Toolbox gives indications on which parts to work
- Slides: 8