Semi-parametric estimation of direct and indirect effects in the COMBINE (Combining Medications and Behavioral Interventions for Alcoholism) Study

The investigators of the COMBINE (Combining Medications and Behavioral Interventions for Alcoholism) Study aimed to determine whether naltrexone, a drug alleged to reduce cravings for alcohol, combined with a behavioral intervention (CBI) assumed to change stress and coping behaviors, improves drinking outcomes more than either alone. After 16 weeks, only naltrexone alone and CBI alone significantly increased percent days abstinent (PDA) in models controlling for baseline PDA and site of treatment administration. Unexpectedly, the naltrexone + CBI combination did not offer any advantage over either naltrexone alone or CBI alone. To understand moderating and mediating factors, and to help explain the Combination's lack of improvement over each monotherapy, controlled and natural direct effect analyses were performed using targeted maximum likelihood estimation (TMLE). TMLE offers several advantages over traditional direct effect analytic approaches such as double-robustness and allowance of treatment moderation by potential mediators. Cravings and stress were examined as theoretically informed mediators/moderators. Controlled direct effect results show that naltrexone, CBI, and the Combination all work best when cravings are high, while none work when cravings are low. Similarly, naltrexone and the Combination work better when stress is high. Natural direct/indirect effect results show that all three treatments' effects are at least partially mediated by cravings, and that craving reduction explains 47-63% of treatment effects. Furthermore, naltrexone appears to affect cravings earlier while CBI works later. Taken together, the set of results suggests the possibility of a threshold effect; if naltrexone reduces cravings early on and CBI is not effective when cravings are low, then the Combination's lack of improvement over either monotherapy should not be surprising.