Measurement and Modeling of Within-Person Variability in Cannabis Protective Behavioral Strategies: A Novel Approach Using Scale Development, Daily Data, and Machine Learning Methods

Abstract

Cannabis is the most widely used controlled substance, with rates highest amongst young adults and trends increasing in both prevalence and frequency of use. As frequent and heavy cannabis use is associated with a variety of short- and long-term unwanted physical and psychosocial outcomes, increasing cannabis use among young adults is an urgent public health concern and there is a need for approaches to help individuals reduce use and/or use-related harms. Research examining use of protective behavioral strategies (PBS; strategies an individual can use before, during, after, or instead of using cannabis to reduce use or consequences) demonstrates negative associations with cannabis use and consequences and suggests a buffering effect against risk factors for cannabis outcomes, highlighting PBS as a promising means of reducing cannabis use and harms. However, research on PBS-focused interventions is mixed and the majority of PBS research has consisted of cross-sectional, between-person retrospective designs. This is problematic as emerging PBS research suggests both between- and within-person (i.e., daily) variability in PBS use and need for personalized approaches consistent with precision medicine frameworks. To progress research on cannabis PBS, the present study utilized a mixed-methods, sequential, daily data design and advanced methods to develop (Aim 1(a)) and provide preliminary validation for (Aim 1(b)) a measure of daily cannabis PBS, and (Aim 2) develop machine learning algorithms capable of identifying individual and daily psychosocial and contextual factors most likely to be associated with cannabis use, cannabis consequences, and subjective PBS effectiveness on days individuals report using a given PBS. Data for Aim 1(a) included cross-sectional and daily survey responses from young adult undergraduate students across four studies. Data for Aim 1(b) and Aim 2 included 7 weeks of daily surveys completed by 55 young adults in Washington state. Aim 1(a) analyses utilized literature review of empirical studies, responses to previously established PBS items, qualitative data provided by participants, and expert feedback to inform scale development. Aim 1(b) analyses examined descriptive statistics and correlations to provide initial validation of the measure. Aim 2 utilized random forest models predicting cannabis use (i.e., hours high), cannabis consequences, and subjective effectiveness of PBS for each participant. Results provide initial support for the 21-item Daily Protective Behavioral Strategies Scale and demonstrated initial evidence of Simpson’s Paradox for cannabis PBS. Additionally, machine learning models with variable importance metrics were examined for most PBS for a subset of participants. Many machine learning models were unable to be built due to insufficient power and, while some models performed well in the test dataset, the majority of models had poor performance. Findings from the present study offer to advance research on cannabis PBS by providing the first daily measure of cannabis PBS and offer to expand understanding of PBS at the daily level. Further, findings provide initial examination of the potential utility of machine learning methodologies for day-level substance use behaviors and offers to inform approaches for future research seeking to examine similar questions and/or develop related prevention and intervention methods. Future directions for cannabis PBS research are discussed.