Inference and Speci?cation Testing in DSGE Models with Possible Weak Identification

This paper considers inference and model diagnostics for log-linearized DSGE models allow- ing an unknown subset of parameters to be weakly (including un-) identified. The framework allows for latent state variables, measurement errors and also permits analysis using only part of the spectrum, say at the business cycle frequencies. The latter is important because DSGE mod- els are often designed to explain business cycle movements, not very long-run or very short-run ?uctuations. For inference, we first characterize weak identi?cation from a frequency domain perspective and propose a score test for the structural parameters based on the frequency domain maximum likelihood. The construction heavily exploits the structures of the DSGE solution, the score function and the information matrix. In particular, we show that the test statistic can be represented as the explained sum of squares from a complex-valued multivariate linear regression, where weak identification surfaces as (imperfectly) multicollinear regressors. Then, we prove that asymptotically valid inference can be carried out by inverting this test statistic and using Chi-square critical values. Next, we suggest procedures to construct uniform confidence bands for the impulse response function, the time path of the variance decomposition, the individual spectrum and the absolute coherency. For model diagnostics, we propose a family of frequency domain misspecification tests that are robust to weak identification. They can be used to test for misspecification in the mean, in the spectrum as well as misspecification within a band of frequencies. A simulation experiment using a calibrated model suggests that the tests have adequate size even in relatively small samples. It also suggests that it is possible to have informative confidence sets in DSGE models with unidentified parameters, particularly regard- ing the impulse responses functions. Although the paper focuses on DSGE models, the methods developed are potentially applicable to other dynamic models with well defined spectra, such as the stationary (factor-augmented) structural vector autoregression.