Engagement Profiling

A person-centered approach to messaging can help motivate individuals to complete goal-oriented behaviors like engagement with a lifestyle management app [17]. This approach involves (1) tailoring delivery based on the person’s behavior profile within the app and (2) focusing messaging on targetable behaviors to motivate users to complete small, manageable actions toward their goal (ie, the goal gradient hypothesis in decision-making) [18]. We performed an LPA on the participants in the daily data set to determine the subgroups of participants’ behaviors. LPA identifies latent clusters of individuals based on continuous variables [19]. The contributions of multiple variables (ie, the facets that explain the unobserved profile of a user) contribute to the outcome experienced by a user. We used the covariates with the highest global mean Shapley values from the gradient-boosted forest model for the LPA for 2 reasons. First, these variables offer the most explanatory power behind the probability of engagement with the dDPP, allowing us not to assume a priori the behaviors that contribute to the usage of the dDPP. Second, profiling users of a digital app such as this dDPP can be more complicated than traditional approaches to consumer profiling, given the interaction between a user’s health and app engagement. To determine the minimum usage data after enrollment into a dDPP to start profiling participants, we conducted LPAs after 2 weeks and iteratively added days until 3 weeks of engagement. We used the profiles from the timestamp with the lowest Bayesian information criterion (BIC), the established goodness-of-fit metric for LPA. We used the mclust package in RStudio (version 2022.12.0+353; Posit Software, PBC) to run the LPAs [20].