Tuljapurkar, Li and Boe 2000 — A Universal Pattern of Mortality Decline in the G7 Countries

Summary

Tuljapurkar, Li, and Boe (TLB) apply the Lee-Carter (LC) singular value decomposition (SVD) decomposition to all seven G7 countries (Canada, France, Germany, Italy, Japan, UK, US) using 1950–1994 annual data, establishing that the single-factor LC structure is not a US artifact but a universal empirical regularity: the first singular value explains over 94% of the temporal variance in log death rates in every country. The dominant time factor $k(t)$ declines linearly in all seven countries. Extending the LC stochastic forecast methodology internationally, the authors find that stochastic median life expectancy forecasts for 2050 exceed official central forecasts by 1.3 years (UK) to 8.0 years (Japan), implying dependency ratios 6–40% higher than official projections by 2050.

Key Claims

1. Universal single-factor structure: In every G7 country, the first singular value of the SVD of $\log m(x,t)$ explains >94% of temporal variance (range 94.3%–97.5%). The dominant time factor $k(t)$ shows highly linear long-term decline with superimposed short-term fluctuations in all seven countries — confirming the LC bilinear structure is a cross-national regularity, not a US idiosyncrasy.
2. Death-adjusted $\hat{k}(t)$: Like Lee-Carter's life-expectancy re-estimation step, TLB recompute $\hat{k}(t)$ each year to match historical total deaths rather than using SVD $k(t)$ directly. This eliminates the Jensen's inequality bias and provides the launch point for stochastic forecasts.
3. Stochastic model: $\hat{k}(t+1) = \hat{k}(t) - z + \varepsilon_t$, same autoregressive integrated moving average ARIMA(0,1,0) random walk with drift as LC. Drift $z$ ranges from 0.26/yr (US) to 0.79/yr (Japan); substantial short-term variability in all countries (Table 2).
4. Stochastic medians exceed official forecasts everywhere: In 2050, stochastic median $e_0$ exceeds official central projection by: Canada +3.6 yr, France +3.5 yr, Germany +1.6 yr, Italy +3.8 yr, Japan +8.0 yr, UK +1.3 yr, US +2.5 yr (Table 3). The Japan gap is particularly striking — Japan has only one official forecast (no high/low variants).
5. Dependency ratio implication: Each 1-yr difference in $e_0$ corresponds to >5% difference in the dependency ratio (65+ ÷ 20–64). TLB's stochastic medians imply dependency ratios 6% (UK) to 40% (Japan) higher than official projections by 2050 — with accelerating divergence beyond 2050 as official scenarios asymptote.
6. $b(x)$ profiles: Broadly similar across G7 but with social/historical variation. All show higher $b(x)$ at younger ages (faster historical proportional improvement) and lower $b(x)$ at old ages — consistent with the $b(x)$ tilt described for the US. Japan's old-age data are the most accurate (no lumping at 85+); all other countries lump ages 85+ into a single class.
7. Theoretical account: Constant long-run exponential rates of decline reflect a balance between growing resources devoted to mortality reduction (raising the rate) and decreasing marginal effectiveness of those resources (lowering it). TLB expect this balance to continue absent an unprecedented knowledge breakthrough.
8. Sex differences: Treated as short-term variation relative to aggregate mortality; TLB propose modeling aggregate decline separately from relative sex-specific change rather than fitting separate LC models by sex.

Concepts Introduced or Extended

• Lee-Carter Model — international validation: single-factor structure holds in all G7 countries; cross-country b(x) comparison; death-adjusted k̂(t) as standard adjustment procedure
• SSA Mortality Forecasting — US stochastic median 82.9 vs. official 80.5 in 2050; official forecasts influenced by short-term slowdowns and conservative expert opinion
• Period Mortality — universal b(x) tilt pattern across G7; dependency ratio as key policy criterion
• Long-Term Actuarial Balance — dependency ratios 6–40% higher than official by 2050; direct implication for Social Security and pension financing

Entities Mentioned

• Shripad Tuljapurkar
• Nan Li
• Carl Boe
• Ronald Lee
• Lawrence Carter

Quotes

"In every G7 country over this period, mortality at each age has declined exponentially at a roughly constant rate."

"The demographic forecasts that shape policy analyses related to old-age support should reflect the regularity of mortality change that we describe."

My Take

The paper's primary contribution is establishing that the LC single-factor structure is a robust cross-national empirical law, not a statistical accident of US data. The fact that >94% of variance is explained by one factor in all seven major developed economies — across very different social, historical, and epidemiological contexts — is a genuinely striking result that strengthens the LC model's claim to universality. The dependency ratio finding is the most policy-relevant: official forecasts for Japan appear to have been off by 40% in their projected old-age burden, with direct implications for pension and health-care financing. The paper is also the first to demonstrate that the LC stochastic forecast approach extends cleanly to international data without modification, making it the methodological foundation for all subsequent cross-national LC work. The main limitation — noted implicitly — is that lumping ages 85+ for non-Japan countries is a structural data constraint that may bias the b(x) profiles and old-age forecasts; Japan's superior data advantage is visible in Figure 1c.