Pedroza and King 2002 — Revisiting Demographic Methods

Summary

Pedroza and King reformulate the Lee-Carter (LC) mortality forecasting model in two progressively richer statistical frameworks: first as an ordinary least squares (OLS) model (showing the singular value decomposition (SVD) approximation is equivalent to least-squares under Gaussian errors), then as a state-space model jointly estimated via Gibbs sampler. They show that standard LC forecasts are technically incorrect because they ignore the last-period residual, and they document that the point forecasts are nonetheless nearly identical to a simple random walk fit separately to each age group. The Bayesian state-space version produces proper uncertainty intervals but still fails out-of-sample, particularly for ages 15–30, pointing to the constant age-profile assumption as the core structural limitation.

Key Claims

LC as OLS: The SVD step that produces $\hat{\beta}_a$ and $\hat{\kappa}_t$ is algebraically equivalent to minimizing squared log-rate errors under Gaussian errors — the LC approximation is the maximum likelihood estimate (MLE) under homoskedastic Gaussian noise.
State-space unification: The observation equation ( $r_{at} = \alpha_a + \beta_a \kappa_t + \epsilon_{at}$ ) and state equation ( $\kappa_t = \kappa_{t-1} + \theta + u_t$ ) can be jointly estimated via Gibbs sampler, eliminating LC's two-stage sequential procedure and yielding proper posteriors for all parameters.
Incorrect LC forecasts: In the state-space form, the correct one-step-ahead forecast uses $\epsilon_{aT}$ (the last residual). Standard LC ignores it. The bias in point predictions is nonetheless small because the per-age LC drift parameter $\delta_a = \beta_a \theta$ closely tracks the per-age random-walk drift $\phi_a$ .
LC ≈ random walk per age group: Proved algebraically and confirmed empirically (U.S. male cardiovascular mortality): drift parameters from LC and from age-group-specific random walks are nearly identical, so LC's point forecasts carry no additional information over simple per-age extrapolations.
Bayesian advantages: Posterior distributions for all parameters ( $\alpha$ , $\beta$ , $\kappa$ , $\theta$ , $\sigma^2$ , $\tau^2$ ), automatic missing-data imputation, and posterior predictive checks — without extra effort.
Poor forecast performance across the board: Even the corrected Bayesian model fails for U.S. males and females (1991–2000) and Japanese males, especially at ages 15–30 where U.S. mortality diverged from trend (AIDS, drug/accident hump). The arm-shaped age profile is preserved, but level forecasts miss.
Root cause is structural: The constant $\alpha_a$ , $\beta_a$ assumption (age sensitivity fixed at historical mean) is the binding constraint — not the frequentist vs. Bayesian estimation choice.

Concepts Introduced or Extended

Lee-Carter Model — state-space reformulation, incorrect-forecasts critique, LC ≈ per-age random walk equivalence

Entities Mentioned

Quotes

"We conclude that the Lee-Carter method needs improvements to give better predictions."

"Thus the forecasts as given by Lee and Carter are not of the right form."

"The lack of measurements of uncertainty is a growing concern in the field of demography where almost all methods used do not provide standard errors for parameters or forecasts of the mortality rates."

My Take

A technically rigorous methodological paper whose most useful finding is the empirical failure: the Bayesian state-space version also forecasts poorly, and the failure is pinned squarely on the time-invariant $\beta_a$ structure rather than estimation method. This directly motivates Meseguer's Bayesian vector autoregression (BVAR) approach, which abandons the single-factor structure entirely. The equivalence of LC point forecasts to per-age random walks is under-appreciated: it implies LC's value lies not in its forecast paths but in its compact uncertainty quantification and cross-age coherence — properties a Bayesian state-space formulation preserves and improves.