Related Experiment Video
Updated: Oct 19, 2026
Real-World M3-BREATHE: Toward Multimodal Mobile Monitoring of Behaviour, Respiration, and Exposures for Treatment and Health Evaluation
Published on: June 5, 2026
Paradoxical PM2.5-longevity estimates in global country panels are sensitive to differential trend control: a
1School of Industrial Technology, Universiti Sains Malaysia, Gelugor, Malaysia.
Background:
Country-year panels are widely used for environmental health research, but heterogeneous national trajectories can produce ecological coefficients that conflict with established health evidence.
Methods:
We assembled 4,406 country-year observations from 186 economies in the World Bank World Development Indicators for 2000-2023. Annual life expectancy was regressed on population-weighted ambient PM2.5 exposure using economy and year fixed effects, with adjustment for gross domestic product per capita, health expenditure, and urbanization. The original sensitivity analyses examined alternative weights, covariance estimators, exposure lags, sample restrictions, and temporal transformations; a future-exposure negative control tested residual temporal structure. Economy-specific trends and rolling five-year differences were added after the conventional estimate showed an unexpected positive sign.
Results:
The conventional adjusted model associated a 10 μg/m3 higher PM2.5 exposure with 0.406 additional years of life expectancy (95% confidence interval, 0.144 to 0.668). The coefficient changed to -0.023 years (95% confidence interval, -0.175 to 0.129) after adding economy-specific linear trends and to -0.006 years (95% confidence interval, -0.085 to 0.073) in first differences. PM2.5 measured 2 years after the outcome was positively associated with current life expectancy in a separate-lead model (0.296 years; 95% confidence interval, 0.042 to 0.549), but the +2-year lead was attenuated and imprecise after adjustment for current exposure (0.152 years; 95% confidence interval, -0.039 to 0.342; p = 0.119). Regional estimates also differed in direction.
Conclusion:
The positive conventional coefficient was not robust to trend-focused specifications, and the future-exposure pattern raised concern about unresolved temporal structure without constituting a stand-alone falsification test. The pattern is consistent with unresolved differential trajectories, but the analysis does not establish that trend heterogeneity is the sole source of bias. Aggregate PM2.5-longevity studies require explicit estimand definition, trend diagnostics, change-based estimators, covariance sensitivity checks, and temporal negative controls before interpreting coefficients as health effects.
