Geweke-Keane (2000) An Empirical Analysis of Earnings Dynamics Among Men in the PSID: 1968–1989

Summary

Geweke and Keane fit a dynamic panel earnings model to 4,766 male Panel Study of Income Dynamics (PSID) household heads (1968–1989), replacing the Gaussian shock assumption with a 3-component mixture of normals. A Bayesian Gibbs sampler with data augmentation handles 45 earnings parameters, individual heterogeneity, and missing panel periods jointly. The central finding is that non-Gaussian shocks substantially change inference: the mixture model implies a $253K lifetime college premium (vs. $201K under normality), a $94K racial earnings gap (vs. $150K), and much stronger low-income persistence for low-education groups.

Key Claims

• Dynamic earnings equation: $y_{it} = c \cdot y_{i,t-1} + (1-c)\beta'x_{it} + (1-c)q_i + (1-c)\phi \cdot e_{i1} + e_{it}$, decomposing persistence into an autoregressive (AR) lag, unobserved heterogeneity $q_i$, and a permanent loading $\phi$ on the first-period shock $e_{i1}$.
• Three-component mixture of normals for shocks $e_{i1}$ and $g_{it}$; mixture likelihood unbounded without regularization → proper Gamma priors on precisions regularize it (Kiefer-Wolfowitz 1956 consistency motivates the approach).
• Posterior shocks are left-skewed and leptokurtic — shapes that are invisible under normality.
• ~60% of unexplained lifetime earnings variation attributable to permanent unobserved individual characteristics $q_i$; majority of year-to-year variation is transitory.
• Mixture model predicts bottom-quintile age-30 man has 81% probability of remaining in bottom quintile at age 55–65 (white, <12 years education); normal model understates this persistence.
• Present value of lifetime earnings (ages 31–65) for bottom-quintile age-30 individual is 56% lower in expectation than for top-quintile.
• Normal model overstates education return in cross-section (12%) while mixture gives 3%; mixture allocates more variation to heterogeneity $q_i$, less to education slope.
• Racial earnings gap: normal model $149,933 lower for blacks; mixture $93,764 — 37% smaller gap.

Concepts Introduced or Extended

• Mixture of Normals
• Gibbs Sampler

Entities Mentioned

• John Geweke
• Michael Keane

Quotes

"About 60% of the variation of lifetime earnings that is not explained by education and race is attributable to permanent individual characteristics that are unobserved and uncorrelated with education, age, and race."

"While the normal model predicts that present value of lifetime earnings is $149,933 less for blacks in 1998 dollars (ceteris paribus), the mixture model predicts a substantially smaller black/white differential of $93,764."

My Take

The paper's core contribution is showing that distributional assumptions — routinely left unchallenged in panel earnings models — have first-order implications for policy-relevant statistics like lifetime earnings gaps and poverty persistence. The 3-component mixture is parsimonious and the computational burden is substantial (~332 sec/iteration for the full PSID sample), but the payoff in model fit (especially for quintile transitions) justifies it. The authors note higher-order mixtures and age-specific heteroscedasticity as natural extensions. The paper is also a methodological reference for handling unbounded mixture likelihoods via proper Bayesian priors, and for data augmentation in unbalanced panels.