Selective Mortality

Definition

Selective mortality (also called survivorship bias in health research) is the distortion that occurs in longitudinal or cross-sectional health studies when differential mortality — the tendency for sicker individuals to die sooner — systematically removes the least-healthy members from an observed sample. Because health is only measurable for living respondents, any analysis restricted to survivors understates true health deterioration, attenuates health disparities by socioeconomic status (SES) and race, and makes observed health appear to improve more than it actually does for the original cohort.

Key Ideas

Survivor selection mechanism: Individuals in poor health face higher mortality. Those who die exit the panel. The surviving sample therefore contains a progressively healthier fraction of the original cohort as follow-up time increases.
Cross-sectional age profile distortion: Age-specific poor-health prevalence in a living sample conflates two opposing forces — true health deterioration (steep) and selective removal of the sickest via death (attenuating). The cross-sectional slope understates the true individual aging slope.
Attenuated SES and race gradients: Low-SES and minority individuals face higher mortality and are disproportionately removed from surviving samples, compressing the health-disadvantage gradient observed among survivors.
Initial conditions problem: Individuals who enter a survey at older ages are already positively selected survivors of prior mortality, analogous to Heckman's (1981) initial conditions problem in dynamic panel models.

How It Works

The Age Profile Decomposition (Heiss 2011)

Heiss (2011) estimates a joint latent autoregressive order-1 (AR(1)) health and mortality model on Health and Retirement Study (HRS) waves $1$ – $7$ ( $n = 29{,}113$ ) and decomposes the cross-sectional age profile of poor/fair self-rated health (SRHS) into a direct aging effect and a survivorship selection effect (Table 8, average partial effects in percentage points):

Age group	Direct aging effect	Survivorship selection	Net cross-sectional
$50$ – $59$	$+1.98$	$-0.79$	$+1.18$
$60$ – $69$	$+2.91$	$-2.28$	$+0.63$
$70$ – $79$	$+7.94$	$-5.87$	$+2.07$
$80$ – $89$	$+13.95$	$-12.28$	$+1.67$
$90{+}$	$+24.55$	$-28.43$	$-3.87$

At ages $90{+}$ , survivorship selection is so powerful that cross-sectional health appears to improve even as the true aging effect is $+24.55$ percentage points (pp) per $2$ -year age increase.

Gender Paradox

Women have better latent health (direct SRHS effect $-2.64$ pp vs. men), but their lower mortality means more unhealthy women survive to be observed ( $+2.55$ pp selection effect). In the surviving sample, the female health advantage nearly disappears (net $-0.08$ pp). Studies restricted to survivors will dramatically understate the female health advantage.

SES Gradient Compression

Low-education individuals have worse SRHS (direct $+15.49$ pp) and higher mortality (selection $-1.82$ pp), so the education-health gradient is understated by $\approx 12\%$ in survivor-restricted analyses.

Persistence of Health and the Selection Problem

The latent AR(1) health model of Heiss (2011) estimates $\rho = 0.941$ – $0.946$ — meaning health is highly persistent (does not mean-revert quickly). This makes selective mortality especially severe: an initially poor-health individual who survives to wave $7$ is far healthier than the typical person who started poor and died. The surviving poor-health group at wave $7$ is thus substantially positively selected.

Disability Insurance (DI) Panels

The same mechanism operates in DI beneficiary panels. If the sickest beneficiaries die first:

Observed average beneficiary health improves over time even with no individual improvement
Duration-specific mortality tables (see Select and Ultimate Mortality Tables) partially account for this, but longitudinal analyses of "surviving beneficiaries" remain biased
The $44.6\%$ $8$ -year disability persistence documented by Heiss et al. (2007) would appear much lower without mortality correction

Detection and Correction

Inverse Probability Weighting (IPW)

Reweight surviving observations by the inverse probability of still being alive, estimated from a mortality model. Used by Contoyannis et al. (2004a) and Jones et al. (2006). Limitation: assumes selection only on observable covariates.

Joint Health-Mortality Model

Heiss (2011) proposes estimating a single latent health process that enters both the SRHS and mortality equations. This allows selection on unobservables (shared frailty) — the same unobserved health state that makes an individual sicker also makes them more likely to die. This is a stronger correction than IPW.

Conditioning on Survival

Condition the likelihood on survival to initial sampling age. Heiss (2011) implements this using a reweighting approach analogous to importance sampling from Bayes' rule on the initial latent health state.

Why It Matters

Health trend research: Apparent improvements in population health over time in a longitudinal study may be partly or entirely compositional — reflecting the death of sicker individuals rather than genuine health gains.
DI beneficiary analysis: Declining beneficiary mortality over time (documented in DI Beneficiary Mortality) may reflect selective exit via death, not improved functional capacity. See Morbidity-Mortality Distinction.
Policy evaluation: An intervention that reduces mortality alongside improving health will appear less effective in a survivor-restricted analysis — the intervention's mortality benefit creates positive survivorship selection that attenuates the measured health benefit.
Compression of morbidity debate: Studies finding morbidity compression at older ages must rule out survivorship bias before claiming genuine health improvement; the steep survivorship selection at ages $80{+}$ documented by Heiss (2011) raises the bar for such claims. See Compression of Morbidity.

Open Questions

How large is selective mortality bias for functional disability (not just SRHS) in the DI beneficiary population?
Can two-process latent models (one for lethal health conditions, one for chronic-but-non-lethal conditions) better capture the distinct dynamics of mortality-predictive vs. non-mortality-predictive health?
Do mortality-corrected SES-health gradients change enough to alter policy recommendations on targeting?