Youssef Genaidy, Roshan Hasan, Julia Incitti, Erika L Bloom, Paul Tremblay, Marc S Mitchell
This study extends mHealth engagement research by identifying behavioral phenotypes within a sample of subscription-based app users. Using early survey, device-assessed health, and app engagement data, five phenotypes were identified. Notably, two phenotypes were at risk of low engagement (ie, nonsubscription). Although these five phenotypes may be less useful for predicting 6-month renewal, they could support early identification of users unlikely to subscribe by the end of the free trial period, possibly allowing for more timely, targeted onboarding strategies to reduce dropout and increase the likelihood of longer-term healthy habit formation.
BACKGROUND: Mobile health (mHealth) app effectiveness may be limited by low engagement. Increasing understanding of factors influencing engagement may help. Paid mHealth app subscription and renewal are two engagement metrics of particular interest to commercial app developers.
OBJECTIVE: This study aims to identify homogeneous user subgroups (ie, behavioral phenotypes) within a paid mHealth app context and examine associations with app subscription and renewal.
METHODS: A 6-month prospective cohort study was conducted with new users of the WayBetter (WayBetter Inc.) app, a subscription-based mHealth app. Through convenience sampling, users initiating a 7-day free trial between November 2023 and January 2024 were recruited. Latent class analysis (LCA) using the 3-step Bolck-Croon-Hagenaars (BCH) approach derived distinct classes, or phenotypes, from 33 indicator variables spanning three areas: (1) purpose-built survey responses (eg, sociodemographics), (2) device-assessed health data (eg, daily step count), and (3) 7-day free trial period engagement data (eg, app opens). Candidate LCA models ranging from 1 to 7 classes were generated. Model fit was assessed using several fit indices (eg, Vuong-Lo-Mendell Rubin likelihood ratio test [VLMR test]) alongside theoretical interpretability. Phenotypes were named based on distinct indicator probability profiles. Logistic regression was then used to estimate odds ratios (ORs) comparing the likelihood of subscription and renewal among users assigned to their most likely class, relative to a reference phenotype.
RESULTS: The sample included 934 users (mean age=41.53 [SD 9.65] years, 91.97% women [859/934], 72.16% with a chronic disease diagnosis [674/934]). Based on LCA fit indices (eg, VLMR test=467.16, P=.03), the 5-class LCA model was selected: (1) high trial period engagers, (2) moderate trial period engagers with multimorbidity, (3) moderate trial period engagers with lower disease burden, (4) low trial period engagers with multimorbidity, and (5) low trial period engagers with lower disease burden. Chronic disease diagnosis and total app opens emerged as the strongest drivers of class distinction. Phenotypes 1-3 had greater odds of subscribing (OR 21.31, 95% CI 8.56-53.06; OR 7.11, 95% CI 4.04-12.50; OR 8.28, 95% CI 4.26-16.08, respectively) than phenotype 4 (OR 0.82, 95% CI 0.48-1.41), compared to phenotype 5 (the reference). Additionally, renewal odds for phenotypes 1-4 were 1.06 (95% CI 0.62-1.81), 0.90 (95% CI 0.54-1.49), 0.99 (95% CI 0.58-1.69), and 0.93 (95% CI 0.48-1.80), respectively (vs reference).
CONCLUSIONS: This study extends mHealth engagement research by identifying behavioral phenotypes within a sample of subscription-based app users. Using early survey, device-assessed health, and app engagement data, five phenotypes were identified. Notably, two phenotypes were at risk of low engagement (ie, nonsubscription). Although these five phenotypes may be less useful for predicting 6-month renewal, they could support early identification of users unlikely to subscribe by the end of the free trial period, possibly allowing for more timely, targeted onboarding strategies to reduce dropout and increase the likelihood of longer-term healthy habit formation.