Modeling Smoking History: A Comparison of Different Approaches

• Published in: American journal of epidemiology Vol. 156; no. 9; pp. 813 - 823
• Main Authors: Leffondré, Karen, Abrahamowicz, Michal, Siemiatycki, Jack, Rachet, Bernard
• Format: Journal Article
• Language: English
• Published: Cary, NC Oxford University Press 01.11.2002
• Subjects: Abbreviation: AIC; Akaike’s Information Criterion; Biological and medical sciences; Case-Control Studies; Epidemiologic Methods; Humans; Lung Neoplasms - epidemiology; Male; Medical sciences; Middle Aged; multicollinearity; multivariate analysis; neoplasms; proportional hazard; Proportional Hazards Models; Quebec - epidemiology; Research Design; smoking; Smoking - adverse effects; Smoking - epidemiology; time-dependent covariate; Tobacco, tobacco smoking; Toxicology; Quebec; Methodology; Tobacco smoking; Case control study; Covariate; Epidemiology; Modeling; multivariate analysis; Bronchopulmonary; Cigarette; Time dependence; epidemiologic methods; multicollinearity; smoking; Bronchus disease; time-dependent covariate; Exposure time; proportional hazard; case-control studies; Human; Lung disease; Respiratory disease; Malignant tumor; Statistical method; Risk factor; neoplasms
• ISSN: 0002-9262; 1476-6256; 0002-9262
• DOI: 10.1093/aje/kwf122

The impact of cigarette smoking on various diseases is studied frequently in epidemiology. However, there is no consensus on how to model different aspects of smoking history. The aim of this investigation was to elucidate the impact of several decisions that must be made when modeling smoking variables. The authors used data on lung cancer from a case-control study undertaken in Montreal, Quebec, Canada, in 1979–1985. The roles of smoking status, intensity, duration, cigarette-years, age at initiation, and time since cessation were investigated using time-dependent variables in an adaptation of Cox’s model to case-control data. The authors reached four conclusions. 1) The estimated hazard ratios for current and ex-smokers depend strongly on how long subjects are required to not have smoked to be considered “ex-smokers.” 2) When the aim is to estimate the effect of continuous smoking variables, a simple approach can be used (and is proposed) to separate the qualitative difference between never and ever smokers from the quantitative effect of smoking. 3) Using intensity and duration as separate variables may lead to a better model fit than using their product (cigarette-years). 4) When estimating the effects of time since cessation or age at initiation, it is still useful to use cigarette-years, because it reduces multicollinearity.