Evaluating eight smoking metrics for modelling survival in non-small cell lung cancer.
[INTRODUCTION] Smoking is a strong modifiable prognostic factor for lung cancer survival.
- 95% CI 1.09-1.13
APA
Lam AC, Li Y, et al. (2026). Evaluating eight smoking metrics for modelling survival in non-small cell lung cancer.. Cancer epidemiology, 102, 103052. https://doi.org/10.1016/j.canep.2026.103052
MLA
Lam AC, et al.. "Evaluating eight smoking metrics for modelling survival in non-small cell lung cancer.." Cancer epidemiology, vol. 102, 2026, pp. 103052.
PMID
41861692
Abstract
[INTRODUCTION] Smoking is a strong modifiable prognostic factor for lung cancer survival. We compared eight smoking metrics to determine which metric best models the relationship between smoking exposure with overall survival (OS) and lung cancer-specific survival (LCSS). These metrics included cigarettes-per-day, smoking duration, pack-years, square-root pack-years, logcig-years, the comprehensive smoking index, age-of-initiation, and years-since-quit.
[MATERIALS/METHODS] This retrospective, pooled analysis included 25 International Lung Cancer Consortium studies between June 1, 1983-December 31, 2019. The performance of smoking metrics for modelling OS was compared based on 1) strength and significance in adjusted Cox-proportional hazard models and 2) linearity based on the goodness-of-fit assuming the log-hazard varies linearly with each smoking metric (i.e. the hazard ratio is constant across different values of the smoking metric) compared to models using non-linear splines. This process was repeated across clinicodemographic subgroups and for LCSS.
[RESULTS] In total, 28,702 lung cancer patients were included (median age 64 [IQR: 57-71]; 53% male). Logcig-years (log(cigarettes/day+1)·years-smoked) had the highest adjusted hazard ratio per standard deviation (aHR 1.11; 95% CI: 1.09-1.13) and best goodness-of-fit when modelled linearly. Square-root pack-years had a similar effect size (aHR 1.11; 95% CI: 1.09-1.13) and had a strong linear relationship on visual assessment of spline curves. In subgroup analyses, logcig-years had a large effect size and maintained a linear relationship regardless of age, sex, stage, and histology. For lung cancer-specific survival (LCSS), logcig-years again had the highest aHR (1.09; 95% CI: 1.05-1.12) and the best linear goodness-of-fit, while square-root pack-years demonstrated the most linear relationship on visual assessment.
[DISCUSSION] Logcig-years best modelled the relationship between smoking exposure and OS as well as LCSS, and had consistent associations across clinicodemographic subgroups. Logcig-years should be considered in clinical and research applications for quantifying smoking exposure in lung cancer.
[MATERIALS/METHODS] This retrospective, pooled analysis included 25 International Lung Cancer Consortium studies between June 1, 1983-December 31, 2019. The performance of smoking metrics for modelling OS was compared based on 1) strength and significance in adjusted Cox-proportional hazard models and 2) linearity based on the goodness-of-fit assuming the log-hazard varies linearly with each smoking metric (i.e. the hazard ratio is constant across different values of the smoking metric) compared to models using non-linear splines. This process was repeated across clinicodemographic subgroups and for LCSS.
[RESULTS] In total, 28,702 lung cancer patients were included (median age 64 [IQR: 57-71]; 53% male). Logcig-years (log(cigarettes/day+1)·years-smoked) had the highest adjusted hazard ratio per standard deviation (aHR 1.11; 95% CI: 1.09-1.13) and best goodness-of-fit when modelled linearly. Square-root pack-years had a similar effect size (aHR 1.11; 95% CI: 1.09-1.13) and had a strong linear relationship on visual assessment of spline curves. In subgroup analyses, logcig-years had a large effect size and maintained a linear relationship regardless of age, sex, stage, and histology. For lung cancer-specific survival (LCSS), logcig-years again had the highest aHR (1.09; 95% CI: 1.05-1.12) and the best linear goodness-of-fit, while square-root pack-years demonstrated the most linear relationship on visual assessment.
[DISCUSSION] Logcig-years best modelled the relationship between smoking exposure and OS as well as LCSS, and had consistent associations across clinicodemographic subgroups. Logcig-years should be considered in clinical and research applications for quantifying smoking exposure in lung cancer.