ECON 706 Midterm: Prelim (Diebold 706) Aug 2012-1
Econ 706 Prelim
August 2012
All questions are weighted equally. Good luck!
Consider Bayesian analysis of a simple Markovian dynamic model:
yt=φyt−1+εt
εt∼iid N(0, σ2)
|φ|<1.
(1) Derive the natural conjugate prior (NCP) for φ|σ. What is the posterior mean of φ|σand
how is it related to the prior mean, the MLE, the prior precision, and the MLE precision?
From this point onward, use the NCP for φ|σ(and, when relevant, for σ|φ) unless explicitly
instructed otherwise.
(2) What is the key benefit of using NCPs? What is the key cost? Maintaining use of NCPs,
how might you attempt to represent prior ignorance regarding φ|σ? What complication
arises? If you could use non-NCPs, how might you attempt to represent prior ignorance
regarding φ|σ? What complication arises?
(3) Why might you be interested in posterior inference not on φ|σ, but rather on φ? Design
a Markov chain Monte Carlo algorithm (in this case, a Gibbs sampler) to sample from the
joint posterior distribution of (φ, σ). Be precise. How would you assess convergence to steady
state of your Gibbs sampler? How would you use the Gibbs draws from the joint posterior
to sample from the marginal posterior of φ?
(4) Is the Gibbs sequence for φlikely to be serially uncorrelated? If not, what is the likely
form of the serial correlation? If you know the spectral density at frequency zero of the
Gibbs sequence of φ,fφ(0), how would you use it to assess the accuracy of your estimate of
the posterior mean of φbased on the Gibbs sequence?
(5) If instead fφ(0) is unknown, how would you estimate it consistently using an autore-
gressive model-based estimator? What determines the bandwidth, what conditions must it
satisfy, and how might you select it in practice?
