Summers (2004) Bayesian Evidence on the Structure of Unemployment

Summary

Summers (2004) applies Bayesian unit root inference to the 16-country OECD unemployment dataset of Papell, Murray and Ghiblawi (2000), extending their frequentist structural-break analysis. Using the Wang-Zivot (2000) Gibbs sampler for multiple breaks in level, trend and variance — extended by an independence reversible-jump Markov chain Monte Carlo (MCMC) algorithm (Metropolized Carlin-Chib, MCC) that treats the number of breaks as unknown — the paper finds virtually no support for unit-root hysteresis. Twelve of 16 countries are clearly stationary; France is the only clear unit-root case; Italy and Spain are borderline. Conclusions are robust across five priors on the AR parameter.

Key Claims

• No unit-root hysteresis: 12 of 16 OECD countries have posterior AR parameters clearly below 1.0 under all five priors; $\geq$94% of each posterior lies below $\rho = 0.95$.
• France the exception: Posterior median $\rho \approx 0.98$; 24–32% of mass above unity (prior-dependent). Italy and Spain borderline (median $\rho \approx 0.90$, 1–4% above unity).
• Prior robustness: Unlike the stylized AR(1) example of Sims-Uhlig (1991), OECD unemployment posteriors are nearly identical across five priors — N(1,1000), N(1,0.5), Lubrano Beta, Beta(10,2), and Berger-Yang (1994). The data are strongly informative about $\rho$.
• Prior reweighting: Posterior draws from N(1,1000) are reweighted by the ratio $f_j(\rho)/f_N(\rho)$ (Geweke 1999 importance sampling) to avoid re-running MCMC for each alternative prior.
• MCC algorithm for number of breaks: Wang-Zivot (2000) condition on the number of breaks $n$ and use the Bayesian information criterion (BIC) to select it. Summers treats $n \in \{0,\ldots,5\}$ as unknown, running independence reversible-jump MCMC (Green 1995 special case): propose $n_1$, draw parameters from full conditionals, accept with standard Metropolis ratio.
• Break model posteriors differ from Bai-Perron: Only Norway's modal break count matches the Bai-Perron (1998) result. Belgium, Finland, Sweden favour 4–5 breaks; France, Australia, Norway favour 0–1 breaks.
• Bounded series argument: Since unemployment is bounded in $[0,1]$, it cannot have a unit root — the Bayesian results are consistent with this theoretical prior.

Concepts Introduced or Extended

• Unit Root Inference
• Change-Point Analysis

Entities Mentioned

• Christopher Sims
• Harald Uhlig

Quotes

"Since the unemployment rate is a bounded series, it cannot have a unit root."

My Take

Clean application paper that makes two genuine contributions: the MCC treatment of an unknown number of breaks (extending Wang-Zivot), and the demonstration of prior robustness for real macro data. The Sims-Uhlig (1991) literature tends to present prior sensitivity as an inherent feature of unit root Bayesian inference; Summers shows this is specific to stylized low-information AR(1) settings — with real OECD data and structural breaks accommodated, the posterior is robust regardless of the prior. The bounded-series argument provides useful economic intuition but does little formal work. The discrepancy between the Bayesian and Bai-Perron break counts (Table 1) is unexplained and worth further investigation.