Examination of social smoking classifications using a machine learning approach

Abstract

Introduction: Idiosyncratic definitions of social smoking proliferate in the literature, making cross-study comparison challenging. This project investigated and differentiated four distinct classifications of social smoking using traditional modeling techniques as well as a multilayer perceptron artificial network, a novel machine learning approach suited for heterogeneous, multidimensional data.

Methods: One hundred thirty-three adults recruited from a college in the Pacific Northwest and from Amazon Mechanical Turk, age 18 to 25 (48% men; 37% women; 8% nonbinary; 73% white; 24% Hispanic or Latinx), completed a set of self-report measures assessing common variables associated with cigarette use. Participants also completed a well-validated audio simulation (Smoking-Simulated Intoxication Digital Elicitation) depicting social smoking contexts and reported their willingness to use cigarettes or alcohol in these contexts.

Results: Across three of the four social smoking definitions, social smokers consistently scored lower on measures of dependence, frequency, quantity, willingness to smoke, and all use motives than nonsocial smokers. The area under the curve for all four models ranged from excellent to outstanding discrimination within the training set. Frequency of days smoked in the past month was the most important predictor for three of the classification models with a relative importance of 100%.

Conclusion: The social smoking definitions demonstrated great variability across common cigarette use variables between groups, except for one. The machine learning approach successfully differentiated all four classifications. Recommendations are made for which social smoker classifications to use in subsequent research to maximize appropriate endorsement by the target population.

Keywords: Classification; Machine learning; Self-report; Social smoking.