Using Decision Tree Analysis to Identify Risk Factors for Relapse to Smoking

Abstract

This research used classification tree analysis and logistic regression models to identify risk factors related to short- and long-term abstinence. Baseline and cessation outcome data from two smoking cessation trials, conducted from 2001 to 2002 in two Midwestern urban areas, were analyzed. There were 928 participants (53.1% women, 81.8% White) with complete data. Both analyses suggest that relapse risk is produced by interactions of risk factors and that early and late cessation outcomes reflect different vulnerability factors. The results illustrate the dynamic nature of relapse risk and suggest the importance of efficient modeling of interactions in relapse prediction.

RÉSUMÉ

Esta investigación utilizó el análisis llamado “classification tree analysis” y modelos de regresión logística para identificar factores de riesgo relacionados con la abstinencia a corto y a largo plazo. Se analizaron datos de línea base (baseline) y de resultado (outcome) de dos estudios de investigación clínica para dejar de fumar llevados a cabo desde el 2001 hasta el 2002 en dos áreas urbanas del Medio Oeste. Hubo 928 participantes (53.1% mujeres, 81.8% blancos) con datos completos. Ambos análisis sugieren que el riesgo de recaída es producido por la interacción de los factores de riesgo y que los resultados de cesación temprana y tardía reflejan diferentes factores de vulnerabilidad. Los resultados ilustran el carácter dinámico del riesgo de recaída y sugieren la importancia del modelado eficiente de las interacciones en la predicción de las recaídas.

RESUMEN

Cette recherche a utilisé des analyses de classifications par arbre décisionnel et des modèles de régression logistique pour identifier les facteurs de risques liés à l'abstinence tabagique à court et long terme. Les caractéristiques de base et les résultats sur l'arrêt tabagique provenant de deux études portant sur la désaccoutumance au tabac ont été analysés. Ces études, menées de 2001 à 2002, dans deux regions urbaines du Midwest des Etats-Unis comprenaient 928 participants (53.1% de femmes, 81.8% de blancs) avec des données complètes. Les deux analyses suggèrent que les risques de rechute sont dus à l'interaction de facteurs de risques et que les résultats sur l'arrêt tabagique à court et long terme reflètent différents facteurs de vulnérabilité. Ces résultats illustrent la nature dynamique du risque de rechute et suggèrent l'importance d'une modélisation efficace des interactions dans la prédiction de la rechute.

Keywords :

• relapse
• prediction
• classification tree
• regression
• smoking

THE AUTHORS

Megan E. Piper, Ph.D., is an Assistant Professor in the Department of Medicine at the University of Wisconsin School of Medicine and Public Health and a lead researcher at the University of Wisconsin Center for Tobacco Research and Intervention (UW-CTRI). Dr. Piper completed her Ph.D. in Clinical Psychology in 2006 at the University of Wisconsin-Madison. Dr. Piper's research interests include understanding and measuring tobacco dependence, identifying which smoking cessation treatments are most effective and how cessation treatments work. In addition, her research addresses tobacco dependence and cessation treatment effects in specific populations such as women and smokers with psychiatric comorbidities such as depression and anxiety.

Wei-Yin Loh, Ph.D., is Professor, Department of Statistics, University of Wisconsin, Madison. He received his Ph.D. from the University of California, Berkeley, in 1982, and was a visiting junior faculty research fellow at the IBM T.J. Watson Research Center in 1986-87. Dr. Loh is a Fellow of the American Statistical Association and the Institute of Mathematical Statistics. He currently serves as associate editor of the ACM Transactions on Knowledge Discovery in Data. Dr. Loh received the Benjamin Smith Reynolds Award for excellence in teaching in 1999 and the U.S. Army Wilks Award for excellence in statistics research and application in 2007. His main research interest is in decision tree algorithms for statistical modeling. He is the author of the GUIDE algorithm.

Stevens S. Smith, Ph.D., is Associate Professor, Department of Medicine, University of Wisconsin School of Medicine and Public Health. In addition, he has been a clinical investigator at the University of Wisconsin Center for Tobacco Research and Intervention (UW-CTRI) since 1992. After completing his Ph.D. in psychology in 1990 at the University of Wisconsin-Madison, Dr. Smith completed a research fellowship at the National Institute on Drug Abuse Addiction Research Center (ARC) in Baltimore, Maryland. At the ARC, he conducted psychological and molecular genetic research on vulnerability to addiction. At UW-CTRI, Dr. Smith has been involved in numerous studies investigating behavioral and medication treatments for nicotine dependence, measurement of nicotine withdrawal and dependence, development of culturally-appropriate smoking cessation treatment for American Indian smokers, and efficacy of telephone tobacco quit lines. In addition to conducting clinical research, Dr. Smith maintains an outpatient clinical practice as a Licensed Psychologist.

Sandra J. Japuntich Dr. Japuntich is a post-doctoral fellow in the Massachusetts General Hospital Tobacco Research and Treatment Center and Mongan Institute for Health Policy, Harvard Medical School. She received her Ph.D. in Clinical Psychology in 2009 from the University of Wisconsin-Madison, working under the supervision of Timothy Baker, Ph.D. at the UW Center for Tobacco Research and Intervention. Her research interests include measurement of nicotine dependence and related processes (e.g., relapse, craving, withdrawal), clinical trials of behavioral and pharmacological interventions for smoking cessation, and broadly disseminable treatments and intervention models for smoking cessation.

Timothy B. Baker, Ph.D., is a Professor of Medicine in the University of Wisconsin School of Medicine and Public Health. His principal research goals are to increase understanding of the motivational bases of addictive disorders and to develop and evaluate treatments for such disorders. He is also highly interested in developing and using technological advances to deliver effective treatments for addictive disorders and cancer. For many years he was a faculty member and Director of Clinical Training in the Department of Psychology at the University of Wisconsin. Dr. Baker has served as the Editor of the Journal of Abnormal Psychology, is the Principal Investigator of the University of Wisconsin Transdisciplinary Tobacco Use Research Center award (NIDA), has contributed chapters to multiple Reports of the Surgeon General, and is the recipient of the James McKeen Cattell Award from the Association for Psychological Science.

Notes

1 This comparison of regression and classification tree models with regard to interaction effects should not be taken to mean that subnode branching within classification tree models is mathematically equivalent to multiplicative interactions in regression models. The classification tree model illustrates interactions by showing that a variable predicts the dependent variable only among individuals who meet a certain threshold on a different variable. For instance, Figure 1 shows that treatment condition is related to abstinence status only if individuals smoke within 30 min of waking (FTND1). Thus, we can say that there is an interaction between FTND1 and treatment condition. Note that the interaction effect involves identifying the interaction variables (FTND1 and treatment condition) and identifying a threshold value for FTND1. Traditional regression approaches model interactions only through cross-product terms. However, both models indicate a predictive relation that differs as a function of another variable in the model.

2 It should be noted that the data presented here were used in the analyses of Weiss et al. (2008).

3 It should be noted that we conducted analyses using RPART (Atkinson & Therneau 2000), an R implementation of CART. The CART trees were much bigger than the GUIDE trees.

4 While acknowledging the limitations of accuracy determination in derivation samples, it is of interest to note that in the current research the smaller classification tree models were similar in accuracy to the larger regression models. This is consistent with other research, suggesting the comparable accuracy of the two approaches (Lim et al., 2000; Perlich et al., 2004).