Preston and Wang 2006 — Sex Mortality Differences in the United States The Role of Cohort Smoking Patterns

Summary

Preston and Wang (2006) demonstrate that U.S. sex differences in all-cause mortality from 1948 to 2003 follow a cohort diagonal rather than a period pattern: the gap between male and female death rates is best explained by which birth cohort is being observed, not which calendar year. Using an age-period-cohort (APC) negative binomial regression with mean years of pre-age-40 smoking as the cohort covariate, they attribute most of the widening and subsequent narrowing of the sex mortality gap to differential cohort smoking histories. Controlling for smoking, the underlying mortality decline over 1948–2003 was 56%—substantially larger than the naive 48% estimate—because men's higher smoking prevalence had masked genuine improvement.

Key Claims

• Cohort organization: Sex mortality differences organize strongly along cohort diagonals in Lexis diagrams, not along period columns. The peak male/female death rate ratio occurs for birth cohorts 1903–1908, tracking through ages over time rather than appearing in any single calendar period.
• Smoking explains the cohort pattern: Including mean years smoked before age 40 (a cohort-specific behavioral variable) in an APC regression absorbs most of the cohort diagonal structure. Three independent lines of evidence support the smoking attribution: (1) the cohort pattern precedes lung cancer's rise and mirrors smoking uptake timing; (2) the male excess is concentrated in smoking-related causes; (3) the cohort of heaviest smoking women (born ~1930s) shows narrowing sex differentials on track.
• APC model results: Smoking coefficient = 0.0230 (standard error [SE] = 0.0022, p < 0.001), implying a 2.33% increase in the male/female death rate ratio per additional mean year of pre-40 smoking. Sex × smoking interaction = −0.0100 (p < 0.001), implying women face only ~57% of the mortality risk per year of smoking that men face (ratio of hazards: $\exp(0.0230 - 0.0100) / \exp(0.0230) \approx 0.57$).
• Magnitude across cohorts: Smoking raised the male/female sex ratio of death rates by +41% for the peak cohort (born 1900–1904) and only +18% for the 1945–1949 cohort — a 23 percentage point narrowing attributable to smoking alone as women's smoking declined relative to men's.
• Period distortion: Without controlling for smoking, the estimated mortality decline 1948–2003 is 48%. With smoking controlled, it is 56%. The additional 8 percentage points represents genuine underlying improvement that was masked by the rising male smoking burden during the same period.
• Future projections: As younger cohorts with lower male-to-female smoking ratios reach older ages, the sex mortality gap should continue to narrow. Men aged 50–85 in 2003 will see survival probabilities rise ~23% from cohort smoking decline alone by 2030; women only ~2%, implying convergence.
• Period forecast distortion: Period-based projections that fail to account for cohort smoking histories will overstate future sex mortality differences and understate the pace of male mortality improvement.

Concepts Introduced or Extended

• Period Mortality — shows that sex mortality differences require cohort analysis, not period analysis; documents period-distortion from smoking patterns
• SSA Mortality Forecasting — empirical foundation for why cohort smoking is a necessary covariate in mortality forecasts
• Period vs. Cohort Life Expectancy — concrete illustration of how period estimates can be systematically misleading when driven by cohort behavioral patterns

Entities Mentioned

• Samuel H. Preston
• Haidong Wang

Quotes

"Mortality differences between the sexes in the United States have followed a course that is largely explicable by reference to the different smoking behaviors of male and female birth cohorts."

"Controlling for smoking, mortality fell by 56% during 1948–2003 rather than the 48% estimated without this control."

"The peak in sex differences in mortality occurs for cohorts born approximately 1903–1908, not for any particular calendar period."

My Take

This paper is methodologically exemplary: the cohort diagonal finding is not merely descriptive but generates a falsifiable prediction (future narrowing of the sex gap as later cohorts age) that has largely been borne out. The APC framework here is cleaner than in most applications because the cohort covariate (smoking history) is observable and theoretically motivated, not just a birth-year fixed effect. The period-distortion finding — that naive analysis underestimates mortality improvement by 8 percentage points — has direct implications for any forecasting model that uses all-cause period death rates without a smoking adjustment: such models import a downward bias into estimated baseline improvement rates. The key limitation is that "mean years smoked before age 40" is a retrospective cohort summary derived from survey data with potential recall bias, and the authors assume a fixed functional form for the smoking-mortality dose-response that may not hold across cohorts with different cigarette types and smoking intensities.