Summary
Geweke and Keane (2000) model the life-cycle earnings process of male household heads in the Panel Study of Income Dynamics (PSID; 1968–1989) using Bayesian inference via Gibbs sampling. The central contribution is allowing non-Gaussian shocks — specified as a mixture of three normal distributions (seven free parameters per shock) — which substantially improves fit to observed earnings quintile transition probabilities and materially changes estimates of the education premium and the black-white lifetime earnings gap. A data-augmentation algorithm handles partial and interrupted panel records, more than tripling the usable sample from the 1,728 men observed at age 25 to 4,766 men with 48,738 person-year observations.
Key Claims
- Non-Gaussian shocks resolve the persistence puzzle: The normal model (Lillard-Willis first-order autoregressive [AR(1)] structure) overstates short-run persistence of poverty and underfits observed quintile sequences. The mixture-of-normals model correctly predicts transition probabilities the normal model misses — e.g., for young black men, three-year poverty sequence
−−−: actual 18.1%, mixture 18.1%, normal 27.8%; three-year non-poverty +++: actual 53.8%, mixture 53.2%, normal 35.5%. The mixture model achieves near-exact fit for both races and both the young-men and full samples.
- Annual variance mostly transitory; lifetime variance mostly permanent: In any given year, approximately 60–70% of unexplained log earnings variance is due to transitory shocks with serial correlation ρ≈0.65. Over a lifetime, transitory components average out: the posterior shows approximately 60% of unexplained lifetime earnings variance is attributable to permanent individual characteristics (τi and the permanent component of εi1) uncorrelated with race, education, or age.
- Low early earnings strongly predicts lifetime earnings: Quintile position at age 30 predicts future earnings powerfully even conditioning on race and education. Mixture model (full sample): whites with <12 yr education in bottom quintile at 30 spend 70.5% of remaining years (31–65) in the bottom quintile; not in bottom quintile at 30 → only 29.2%. Mean present value (PV) of lifetime earnings (ages 31–65) is $88,400 (56%) lower for those in the bottom quintile at age 30. The divergence is greater in the mixture model than the normal model.
- Education premium substantially larger in mixture model: Full sample, mixture model: each additional year of education → +$12,985 in expected PV lifetime earnings (1967 dollars); college vs. high-school premium = $51,940 (39.5% of mean PV = $131,586). Normal model: $41,112 (31.6%). Mixture model also predicts a sharper rightward shift in the lifetime earnings distribution for college-educated relative to high-school-educated — especially visible for black men.
- Racial earnings gap substantially smaller in mixture model: Mixture model (full sample): blacks' mean PV lifetime earnings $19,214 (14.6%) less than whites, ceteris paribus. Normal model: $30,724 (23.6%) less. The mixture model implies a greater association between education and earnings and a lesser association between race and earnings, because non-Gaussian shocks capture heterogeneity that the normal model incorrectly attributes to race.
- Key structural parameters: Lagged earnings coefficient γ≈−0.12 (small, negative); transitory shock autocorrelation ρ≈0.65; individual heterogeneity στ≈0.37. Second-period autocorrelation ρ~≈0.34. First-period permanent shock loading ϕ≈0.24. Marital status probit autocorrelation λ≈0.93; lagged earnings raises marital status probability (coefficient ≈0.18 in full sample mixture model).
- Variance decomposition: Mixture model attributes 38.7% of lifetime earnings variance to permanent unobserved heterogeneity; normal model attributes 44.6%. Observed covariates (race, education, parental background) explain 33.7% in mixture vs. 26.3% in normal model. The residual unobserved heterogeneity dominates observed covariates in explaining variation in lifetime earnings.
- Low-education men have longer but fewer low-earnings spells: For whites with <12 yr education (full sample), mixture model: 3.49 expected spells in bottom quintile over lifetime, mean spell length 5.45 years, fraction of lifetime in bottom quintile 47.6%. Normal model: 4.90 spells, 4.28 years mean length, 51.9% fraction. The mixture model implies poverty is more concentrated — fewer longer spells — while normal model implies more frequent briefer spells.
Concepts Introduced or Extended
- Earnings Dynamics — establishes non-Gaussian permanent/transitory decomposition of life-cycle earnings on PSID; documents that early earnings strongly predict lifetime outcomes; shows normal-model AR(1) misrepresents quintile transition probabilities.
Entities Mentioned
Quotes
"Low earnings at a specific age, like 30, is a strong predictor of low earnings later in life, even conditioning on race, education, and age."
"The mixture model implies a greater association between education and earnings and a lesser association between race and earnings than does the normal model."
"In a given year, 60% to 70% of the variation in the log of earnings not explained by covariates is accounted for by transitory components whose serial correlation is relatively weak, about 0.7 from year to year. But over a lifetime transitory components (by definition) average out. The posterior distributions show that 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."
My Take
The core contribution is showing that the shape of shock distributions — not just their first two moments — matters enormously for earnings mobility estimates. The non-Gaussian mixture resolves a long-standing puzzle (Lillard-Willis overstated persistence) and yields a substantially larger education premium and smaller racial gap. The Bayesian Markov chain Monte Carlo (MCMC) approach with data augmentation is a methodological innovation that allows use of short and interrupted panel records. Limitations: the model misfits the age-45 earnings distribution (likely requiring age-varying shock variance); results cover only male household heads 1968–1989; and the reduced-form structure cannot distinguish causal mechanisms. The finding that early earnings predict lifetime outcomes within a person's own career complements the intergenerational mobility literature (Chetty et al.) but operates through a distinct mechanism — permanent individual heterogeneity rather than parental transmission.