Bridging the etiologic and prognostic outlooks in individualized assessment of absolute risk of an illness: application in lung cancer

Abstract

Assessment of individual risk of illness is an important activity in preventive medicine. Development of risk-assessment models has heretofore relied predominantly on studies involving follow-up of cohort-type populations, while case–control studies have generally been considered unfit for this purpose. To present a method for individualized assessment of absolute risk of an illness (as illustrated by lung cancer) based on data from a ‘non-nested’ case–control study. We used data from a case–control study conducted in Montreal, Canada in 1996–2001. Individuals diagnosed with lung cancer (n = 920) and age- and sex-matched lung-cancer-free subjects (n = 1288) completed questionnaires documenting life-time cigarette-smoking history and occupational, medical, and family history. Unweighted and weighted logistic models were fitted. Model overfitting was assessed using bootstrap-based cross-validation and ‘shrinkage.’ The discriminating ability was assessed by the c-statistic, and the risk-stratifying performance was assessed by examination of the variability in risk estimates over hypothetical risk-profiles. In the logistic models, the logarithm of incidence-density of lung cancer was expressed as a function of age, sex, cigarette-smoking history, history of respiratory conditions and exposure to occupational carcinogens, and family history of lung cancer. The models entailed a minimal degree of overfitting (‘shrinkage’ factor: 0.97 for both unweighted and weighted models) and moderately high discriminating ability (c-statistic: 0.82 for the unweighted model and 0.66 for the weighted model). The method’s risk-stratifying performance was quite high. The presented method allows for individualized assessment of risk of lung cancer and can be used for development of risk-assessment models for other illnesses.

Appendix

Formula for calculation of the comprehensive smoking index (as per Ref. [16]):

$${\text{CSI}} = (1 - 0.5^{{{\text{dur}}*/\tau }} )(0.5^{{{\text{tsc}}*/\tau }} )\ln \left( {\text{int} + 1} \right),$$

where τ = 26, σ = 1 in males and 0.7 in females, tsc = time since cessation (in years), dur = total duration of smoking (in years), int = average smoking intensity (in cigarettes smoked per day), tsc* = max(tsc − σ, 0) and dur* = max(dur + tsc − σ, 0).