We conducted a document analysis to develop the program theory using 2 data sources. First, we reviewed the program documentation, including the intervention design documents, RCT protocol, and data collection materials, to identify its intervention components, extract the corresponding features, understand the assumptions of the designers on how this multicomponent digital health intervention would lead to desired outcomes, and select the corresponding outcome indicators. Second, we reviewed published research articles on behavioral science theories, with special attention paid to digital behavior change interventions (DBCIs) and chronic disease self-management. The objective for reviewing pre-existing theories was twofold: (1) to inform the coding of intervention features, determine the key mechanisms, and identify the individual characteristics that may influence the implementation process; and (2) to select appropriate preexisting conceptual frameworks to inform the development of CIMO configurations, that is, build causal relationships among the constructs [53]. We focused on microlevel theories according to the study objective, that is, to understand the implementation process of a digital health intervention among individuals.

The KiT intervention [20] was developed using a user-centered design approach that engaged adolescents and emerging adults living with T1D as well as adult and pediatric diabetes providers. Specific intervention content was informed by clinician consultation and an environmental scan of diabetes transition resources mapped to domains of the “readiness of emerging adults with diabetes diagnosed in youth” (READDY) tool with the aim of improving the transition experiences among emerging adults living with T1D (Textbox 1). In addition to providing care coordination support (eg, appointment note-keeping and reminders), KiT is designed to provide personalized T1D informational support (eg, educational content and a question and answer feature) based on an individual’s interests and self-reported confidence about their diabetes knowledge and skills using a set of “if-then” rules (Multimedia Appendix 1).

We have identified intervention components and features (Table 1). By performing deductive coding according to existing literature, we extracted several common components and features of DBCIs, including personalization [55,56], problem-solving support [57-59], reminders [55], real-time interactivity [55,60], credible sources [57], user-friendly SMS text message tone [55], and diverse forms of information (enhanced media) [56]. We classified the remaining KiT components (eg, coping with T1D and care navigation) through inductive coding.

We have developed the initial program theory (Figure 2) by integrating 3 empirically validated individual-level behavioral change theories, including “capability, opportunity, motivation, and behavior” (COM-B) [61], “health action process approach” (HAPA) [62], and “technology acceptance model” (TAM) [63].

Informed by COM-B [61], we hypothesize that KiT may improve participants’ transition outcomes (ie, transition readiness, self-management behavior, and health outcomes) by triggering changes in individual motivation (ie, self-efficacy, negative emotions, outcome expectancies, behavioral cueing) and capability (ie, knowledge, skills). T1D self-management information and suggestions, transition support information, and problem-solving support may increase participants’ T1D and transition-related knowledge and skills [64], and therefore, capability [61]. T1D self-management information and suggestions [65] and problem-solving support [59] may improve one’s outcome expectancies and self-efficacy, respectively, which are the key motivational factors of an individual’s behavioral change [66]. One’s self-efficacy may also be improved by receiving stress management strategies [67]. Meanwhile, stress management strategies may lead to reduced negative emotions [58,67], which have been recognized as one of the most prominent motivational challenges to a successful T1D health care transition [14,68]. Transition reminders may function as behavior cueing to plan for and attend clinic appointments [69]. Well-designed intervention functions, such as personalization, credible sources, and diverse forms of information, may amplify the impact of KiT’s intervention content.

The extent to which KiT can trigger changes in health care transition capability and motivation is hypothesized to be dependent on the level of user engagement [38,55,70,71]. Positive intervention outcomes (eg, improved transition readiness) may in turn incentivize increased user engagement [55], suggesting a positive feedback loop. According to TAM [63], user engagement is influenced by 2 core factors, that is, perceived usefulness and perceived ease of use [72], which may be influenced by intervention content and technological functions [73].

Intervention effectiveness depends heavily on the interaction of the intervention with users and their context [55], yet there is a paucity of contextual insights in the DBCI literature, creating a gap in our understanding of which contextual factors influence the efficacy of intervention mechanisms in T1D and how [74]. According to the HAPA framework [62], behavioral determinants, such as self-efficacy [75], outcome expectancies [66,75], risk perception [65,76], and social support [66,77], impact both outcome and the intervention implementation process itself. Contextual factors, such as the complementary or substitute relationships between KiT and other conventional self-management tools (eg, handwritten methods) [38], may also predict user engagement alongside more traditional variables (ie, participant demographics [55,70]).

Eligible participants will include English- or French-speaking emerging adults living with T1D enrolled in the RCT intervention arm who provided informed consent to be contacted for the embedded process evaluation at the time of enrollment in the RCT. We are targeting a conservative sample size of 25-30 participants for the interviews based on previous realist interviews and CNA studies [38,42,78,79].

To increase the representativeness of the interview sample, we will purposively select participants with desired and undesired proximal outcomes of user engagement. This will be achieved by identifying “high-engagers” and “low- to medium-engagers” according to the intervention usage data at month 3 of the RCT. As there is no standard classification or threshold of DBCI engagement, we have used early KiT intervention usage data to identify a feasible definition of engagement. Considering the data availability, we conservatively define “high-engagers” as those with a 100% response rate to question prompts, with all other participants classified as “low- to medium-engagers.” All low- to medium-engagers will be invited for an interview. Should resource constraints not allow us to interview all high engagers, we will purposively sample to balance representation by gender and site of recruitment. The iterative sampling will continue until we reach thematic saturation (ie, no new themes are identified in the multisource data analysis), or all interested participants have been interviewed.

To test the initial program theory developed in step 1, we will perform semistructured realist interviews in a “teacher-learner cycle” [49]. The interviews will start with some general questions about the participants’ experiences managing T1D and engaging with KiT [80,81]. Based on their level of engagement and responses, the interviewer will act as a teacher to introduce the candidate CIMO configurations to the participant for their comments. After learning about the CIMO configurations, the participant will then act as teacher to confirm, extend, or refute the hypothesized pathways with their own examples of how they reacted to specific intervention components or features, what they saw as influencing their decision-making, and how this was perceived to affect their outcomes [80,82]. For low- to medium-engagers, the interview topics will focus on CIMO configurations regarding user engagement. For high engagers, the interview topics will focus on CIMO configurations regarding user engagement as well as distal outcomes (ie, transition readiness, self-blood glucose monitoring, and blood glucose). The interview guides (Multimedia Appendix 2) will first be piloted to ensure a comprehensive assessment of all relevant CIMO constructs. Interviews will be conducted by PhD-trained research coordinators with qualitative interview training, previous experience conducting semistructured interviews, and in-depth knowledge of the initial program theory. They will also receive study-specific training from the senior study lead, who is an experienced qualitative researcher. Research coordinators will have no previous relationship with study participants or the T1D community. Interviews will be transcribed verbatim by an independent third party.

Transcripts will be coded using MAXQDA, a software for interview data analysis, and analyzed using the principles of thematic analysis strategies [82,83]. Data will be deductively coded and mapped to the predefined CIMO constructs to reflect triads (eg, context + intervention outcome, intervention mechanism outcome) and tetrads (context + intervention mechanism outcome) that would confirm, extend, or refute the current CIMO configurations [84]. Open coding will be applied when themes are identified that do not fit within the definitions of predefined CIMO constructs. The findings will be compared against the hypotheses developed in step 1, with a label of “supported,” “refined,” or “rejected” given to each hypothesized pathway based on the evidence.

Several strategies will ensure the fidelity and credibility of the interview data, such as using multiple sources of data; creating a chain of evidence that documents all elements of the study database; having both broader research team members and interview participants participate in the triangulation analysis and the return of findings (construct and external validity); examining points of convergence and divergence within and among the data set (internal validity through cross-comparative analyses); and having a stepped analysis process with an initial independent review of the data by 2 reviewers who then meet to reach consensus around the common themes.

We plan to triangulate information from realist interviews, intervention usage data, patient-reported outcomes, and a demographic survey as outlined in Table 2. Patient reported outcome and demographic information will be collected through RCT baseline and follow-up surveys. Intervention usage data will be collected by our third-party collaborator, Memotext, which delivers and manages the RCT intervention. Memotext’s system logs all incoming and outgoing messages, and they will be sharing them with the research team periodically.

We will create a data set in which each interview participant is treated as a unique case and each CIMO element is included as a variable (as defined in the initial program theory outlined in Figure 2). Where variables are continuous (eg, self-efficacy, negative emotions, transition readiness, and blood glucose), we will include them directly. Where variables are qualitative (eg, knowledge and skills, outcome expectancies, behavioral cueing), we will convert them to data-driven categorical variables that will be defined by the research team and reported in the final manuscript.

We will conduct CNA using the R package “cna” to consolidate the initial program theory using a bottom-up, data-driven approach [85]. In order to identify the crucial difference-makers of outcomes from a large number of candidate-influencing factors (ie, context, intervention, and mechanism), we will screen the exogenous factors of each outcome before developing CNA models to demonstrate the causal relationships among them.

To achieve factor reduction, we will begin with exploratory data analysis using a routine that operates within the same regularity framework as CNA. Specifically, we will apply the “minimally sufficient condition” (msc) function in the “cna” package to search across the entire data set (ie, all variables and all cases) at once to identify factors with the strongest connections to target outcomes, following the process outlined in Miech et al [86]. We will perform separate analyses in order to identify the minimally sufficient conditions for the presence of desired outcomes (ie, engagement and transition readiness) and those for the absence of desired outcomes, respectively [52,78,79]. For each outcome condition, we will run the “msc” routine 5 times at consistency thresholds of 0.95, 0.90, 0.85, 0.80, and 0.75, examining all 1-, 2-, and 3-condition configurations that meet consistency requirements; have the highest coverage score for their complexity level; and align with theory, background knowledge, case familiarity, and logic [78,87]. Through this process, we will identify a subset of factors to use in the subsequent modeling phase.

In the modeling phase, we will iteratively develop preliminary and final models based on the following criteria: overall model consistency of 0.80 or greater; overall model coverage of 0.80 or greater; and aligns with theory, experience, background knowledge, and logic. Since CNA analyzes data in a bottom-up manner, such analysis will provide empirical findings that can be used to consolidate the initial program theory.