Heiss Borsch-Supan Hurd and Wise 2007 — Pathways to Disability Predicting Health Trajectories

Summary

Heiss, Börsch-Supan, Hurd, and Wise estimate a latent-health model on all four Health and Retirement Study (HRS) cohorts (1992–2002, n ≈ 25,000) that jointly addresses state dependence, unobserved heterogeneity, and misclassification error in observed health reports. The central finding is that observed health transitions severely understate true health persistence because selective mortality removes the sickest individuals from the panel; a latent Ornstein-Uhlenbeck (OU) process with high year-to-year autocorrelation ($\rho \approx 0.98$ for self-rated health) recovers the true trajectory. Work disability independently predicts mortality (odds ratio [OR] = 1.76 after controlling for other health dimensions), and survival to age 70 is 91.6% for those in excellent health at 50 but only 48.5% for those in poor health — a 43-percentage-point (pp) gap shaped by the joint dynamics of health deterioration and selective mortality.

Key Claims

• Disability → mortality (OR = 1.76): Work disability predicts death with OR 1.76 even after controlling for self-rated health (SRH), which is itself a separate strong predictor. The two health dimensions carry independent mortality information.
• Poor/fair SRH → work disability (OR = 4.53): Cross-dimension predictive power is large; SRH is not interchangeable with work disability but is a strong leading indicator of functional limitation.
• High latent-health persistence: 1-year latent health autocorrelation $\rho = 0.9817$ (SRH dimension) and $0.9663$ (work-disability [WD] dimension). Over 20 years, autocorrelation remains substantial: $\approx 0.7$ (SRH) and $\approx 0.5$ (WD).
• Selective mortality bias: Simple Markov transition models dramatically underpredict persistence of poor health over 8 years because those in poor health are most likely to die — leaving a surviving panel that appears healthier than the true conditional trajectory of the original cohort.
• Survival gap: Conditional on being alive at 50, survival to age 70 is 91.6% (excellent health at 50) vs. 48.5% (poor health at 50) — a 43 pp gap.
• 8-year disability persistence: 44.6% of individuals with work disability in 1992 remained disabled in 2000. Without accounting for selective mortality, this figure appears much lower because the sickest individuals exit via death.
• Socioeconomic status (SES) gradient: Low-education individuals (≈ 8 years of schooling) in poor health at age 50 face approximately 4× the poor-health prevalence of high-education individuals (≈ 16 years) over the same horizon, even holding initial health constant.
• Gender paradox: Women have better latent health than men but worse observed SRH conditional on the same latent health value — consistent with systematic gender differences in subjective health reporting thresholds.
• Ornstein-Uhlenbeck heterogeneity: Unobserved permanent heterogeneity follows a mean-reverting OU process, capturing both individual fixed effects and gradual health drift over the life cycle.
• Three modeling problems: State dependence (health history shapes future health beyond current state), unobserved heterogeneity (individual frailty not captured in observed covariates), and misclassification error (observed health categories do not map cleanly onto true latent health).

Concepts Introduced or Extended

• Selective Mortality
• Morbidity-Mortality Distinction
• DI Beneficiary Mortality

Entities Mentioned

• Florian Heiss
• Axel Börsch-Supan
• Michael Hurd
• David Wise
• Health and Retirement Study

Quotes

"The simple Markov model predicts an 8-year disability persistence rate far below what we observe — because those in poor health are disproportionately removed from the panel by death."

"Work disability and self-rated health carry independent predictive power for mortality."

My Take

The paper's core contribution is methodological: it demonstrates that failing to account for selective mortality in a longitudinal health panel produces systematically optimistic pictures of health trajectories. If the sickest individuals die and exit the panel, the surviving observed population looks healthier than the true conditional trajectory of those who were alive at baseline — biasing any inference about how health evolves with age.

For the disability insurance (DI) literature specifically, the OR = 1.76 for work disability → mortality is the most policy-relevant finding: it provides direct evidence that self-reported functional disability (not just SRH) is a genuine independent mortality risk factor. This reinforces the conceptual core of DI Beneficiary Mortality — that beneficiaries' mortality excess reflects real health severity, and that declining beneficiary mortality need not imply improved work capacity.

Limitations: working paper version (Mannheim Research Institute for the Economics of Aging [MEA] Discussion Paper 131-2007, version May 2006), so may not reflect final published form. Simulation-based maximum likelihood estimation is computationally intensive. The OU process imposes parametric health dynamics that may not capture sudden acute health events.