To evaluate the C2H program, we implemented a quasi-experimental, parallel group evaluation using a cluster-matched control design with 24-month follow-up (ClinicalTrials.gov ID#: NCT05206994). Given that intervention sites were preselected before the evaluation, we used propensity score matching to identify 11 control sites, for a total of 22 communities enrolled in the study. An allocation ratio of 1:1 was used. Using an SN sampling design, data collection was conducted with youth program participants and two layers of their SN to assess the diffusion of the intervention effects to socially connected peers within each community. Additionally, school-based data collection of all high school students at the primary high school in each community was conducted via the California Healthy Kids Survey. This research evaluation was conducted in partnership with CDPH, VALORUS (the TTA provider), the 22 community-based organizations, a community research advisory board (RAB) made up of staff from the C2H-implementing NGOs, and a youth research advisory board (YRAB) made up of youth from the intervention and control communities.

Given preselection of C2H sites, a natural experiment using control matching, rather than randomization, was the most rigorous design possible. Matching was conducted with the goal of nearing exchangeability [36]. A youth development program called 4-H was selected to be the control because it (1) recruits teams of prosocial youth interested in helping their community who are supported by adult volunteers, much like C2H, (2) operates in nearly every county across California, allowing for enough variability to optimize matching, and (3) was active during the same time as C2H implementation. The 4-H clubs engaged youth in prosocial learning projects (eg, agriculture, computer science, and art), with adult volunteers as allies to those efforts, yet had no programming related to SV or DV. More details about the 4-H program can be found online [37]. Participants could withdraw from the program at any time. Propensity score matching used a complete list of all local 4-H programs with at least 8 registered youth ages 14‐24 years and not located in a county with an agency implementing C2H or receiving CDPH funding for SV prevention. Propensity scores were assigned based on zip code–level characteristics relating to SV or DV: percent Black population, percent Hispanic population, percent completed bachelor’s or graduate school, percent population under the federal poverty line, criminal rape rate per 100,000 people, presence of a higher education institution, and percent rural population. The 4-H club with the highest propensity score matched to each of the C2H sites was screened by a member of the research team via a short phone call with the adult club facilitator. Screening was conducted for the following inclusion criteria: (1) at least 8 registered youth ages 14‐24 years; (2) stability of club over the past 6 months; (3) in a geographic zip code with no rape crisis center implementing C2H; (4) no violence prevention programming or education offered by the 4-H club; and (5) adult club facilitator willingness to engage in this study. Ultimately, in terms of propensity scores, 5 of C2H sites were matched with the best matched 4-H program, 4 with the second, third, or fifth best match, and 3 that proved difficult to match, which ended up with the 30th, 36th, and 45th best match (out of 69 clubs; see the Data Analysis section for details on analytical approach to address baseline differences). The 4-H clubs with higher propensity scores declined to participate or were ineligible due to an unstable club membership or leadership (discovered upon discussing with the club leader), concerns about parent perceptions of the topic of this study, or being unresponsive to recruitment efforts. Ultimately, the way self-selection played a role in the final control sample likely improved exchangeability between intervention and control communities in that all of the intervention C2H groups had stable club membership and leadership, had low parental concerns about the topic of this study, and had responsive leaders to recruitment efforts.

A provider RAB and a YRAB were recruited and engaged to inform the development of study protocols, survey instruments, study implementation, as well as ensure the feasibility and acceptability of this study. RAB and YRAB members are compensated for their time. The RAB is made up of key staff from CDPH Injury and Prevention Branch, TTA providers, including the C2H founder and staff from California’s SV coalition (ValorUS), and staff from the 11 C2H local programs. Initially, the RAB met bimonthly to deepen the conceptualization of the theory of change and mechanisms through which C2H was reducing SV and DV in their communities, to guide development of study protocols to ensure coordination and synchrony with C2H implementation, and to provide feedback on outcome and process data collection instruments. Since data collection began, the RAB met on a quarterly basis to provide feedback on implementation and to inform the refinement of study protocols. The YRAB, made up of 7‐10 youth participants of C2H or 4-H programs, engaged in parallel activities throughout the development of study protocols and instruments. The RAB and YRAB will be engaged in the interpretation of results and development of recommendations for future implementation of C2H.

Research ethical and safety protocols for this study adhered to recommendations issued by the World Health Organization for research on violence against women [38]. The current study protocol and analysis plan are registered on ClinicalTrials.gov (NCT05206994).

Three data sources were used in this evaluation: (1) process data from program staff, (2) participant SN data, and (3) data from the California Healthy Kids Survey.

Process data were collected via monthly surveys with program staff and quarterly consultations with program staff and TTA providers to monitor intervention fidelity. For C2H programs, a phase-specific checklist that reflected the guidelines for implementation of each phase was used to measure fidelity, quality, and frequency of interaction with the target community, C2H participant engagement, and barriers and facilitators to implementation. For control programs, qualitative process data were collected with adult volunteer leaders every 6 months during the intervention period, which included participant-level program dosage estimates, any unanticipated exposure to violence prevention programming, and exogenous changes in the community and schools around SV that may have impacted knowledge, awareness, and/or behavior.

This study also used the California Healthy Kids Survey (CHKS) as a mechanism for collecting evaluation data. CHKS collects school-climate data from in-school youth attending public high schools across California. Researchers and Local Educational Agencies are able to develop and build custom data collection modules as part of the CHKS system. In summer 2021, the research team developed a custom module focused on SV prevention to be included in CHKS, which was made up of an abbreviated set of measures used in the SN data collection, and WestEd integrated it into their survey system. Before and during the 2021 to 2022 and 2022 to 2023 academic years (timing guided by SN data collection for each C2H or control group) and with guidance from WestEd, the research team conducted school recruitment by contacting school district leaders and school principals from schools in the targeted zip codes of study communities via email and telephone on multiple occasions with a description of the research and incentives for schools (US $300 gift card, fees covered). For schools or districts that responded, we conducted an informational call about the purpose of the module, the risks and benefits to the school, and implementation logistics, including that implementation of the module would be requested again in 2 years. We then coordinated with WestEd to ensure the opted-in schools received the module, and schools scheduled a date for implementation. Voluntary opt-out parental consent that included specific language about the SV module was obtained by providing informed consent to students’ parents via school parent portals or websites before implementation, per WestEd protocols. Schools implemented the CHKS core and optional SV modules during one class period with all 9th- and 11th-grade students, which took about 50 minutes. Informed assent and data were collected via an online survey that was administered by teachers and/or school staff with training and protocols provided by WestEd. Students received no incentives for participation. The survey was anonymous and did not collect identifying information from students. A data use agreement with the research team’s institutions, WestEd and the California Department of Education, guided the secure transfer of data from the relevant school years to the research team, and are stored on a protected and secure server.

All schools that participated in baseline CHKS data collection in 2021‐2022 or 2022‐2023 were recruited for 24-month follow-up CHKS data collection in the academic year that included the 24-month follow-up period for each specific study community. Similar recruitment methods to those used in baseline recruitment were used, as many of the administrators have changed over the two years. Implementation, informed consent, and data transfers under new data use agreements were also conducted in the same way as baseline.

We estimated that we would have a sample of 6 C2H schools and 6 control sites at 24-month follow-up for CHKS administration, with an average of 500 students per school, which would provide data from an estimated 6000 students. We assumed control students would have the same past 12-month incidence of SV at follow-up as what has been observed to date in baseline, which would provide sufficient power to detect a 14.5% reduction in SV (reduced from 13.8%, 414/3000, to 11.4%, 343/3000) in the treatment group. It is reasonable that such a reduction in SV is possible, given previous studies [40,41]. We acknowledged that power calculations may vary depending on actual intercluster correlations and the number of schools.

All primary and secondary outcome measures for this evaluation are summarized below, with additional details in the Multimedia Appendix 3: measures.

Instruments for the SN and CHKS data collection and the SN protocol were concurrently pilot tested twice (summer and fall 2021) to ensure feasibility and acceptability. A mix of youth from RPE programs, a 4-H state leadership group, and from high schools in this study’s site communities were recruited for pilot testing (n=24 pilot egos), oversampling LGBTQ+ youth in the second pilot. Recruited pilot testers (ie, pilot egos) took the full SN survey (which contains identical items as the CHKS instrument) online, participated in the SN recruitment process, and then provided feedback on the survey and SN recruitment process during a 60-min talk-back session. In the first pilot test, pilot participants provided feedback on an initial list of name generator questions for SN sampling to help with refinement and consolidation into the final three questions. In the second pilot test, the final name generator questions were tested, as well as the newly developed measures related to LGBTQ+ youth.

Additional key changes made based on pilot testing included adding recorded videos of youth explaining this study to the assent form for alters (because they do not receive an in-person orientation such as egos) and adding the option to return later to provide peers’ contact information. To make inviting peers to participate in a survey easier, we also added a text message template into the survey instructions for participants to use to notify their nominated peers about the survey.

SV and DV behaviors were measured by assessing past 12-month and lifetime incidence of virtual or online sexual harassment, in-person sexual harassment and sexual assault, and physical, emotional, and sexual DV (Multimedia Appendix 3: measures). Participants were also asked to indicate if the person who did this to them or who they did this to was a dating or romantic partner. Two items were included for virtual or online sexual harassment; one modified from the Digital Dating Abuse Scale [42] and the other modified from the evaluation of Coaching Boys into Men (CBIM) Victimization Scale [16,40]. Two items were asked for in-person sexual harassment and SV; one modified from the Teen Health and Technology Survey conducted in 2010 to 2011 [41], and the other modified from the CBIM Victimization Scale [16,40]. Two DV items were included, also modified from the CBIM Victimization Scale [16,40]. Measurement of the use of SV by a youth followed the same structure, with inverse prompts and response options. For these items, participants indicate if the behavior was “in-person” or “online.” Items were modified from the CBIM intention to intervene and SV intervention behaviors scales [16,40].

A measure to assess experiences with forms of SV and DV targeting LGBTQ+ youth was developed by the research team. Survey participants who selected an LGBTQ+ sexual and/or gender identity received a module (via skip patterns) that contained items assessing forms of violence targeting LGBTQ+ youth, and those selecting an expansive gender identity (ie, any gender except cisgender boy or girl) received an additional set of items assessing forms of violence targeting gender expansive youth. Items included in these assessments were adapted from two existing scales developed for adults: the Identity Abuse Scale [43] and the Transgender-Specific Intimate Partner Violence Scale [44]. They were reviewed by both practice and research experts on this topic and pilot tested with LGBTQ+ youth.

Secondary outcomes were selected based on our theory of change, including constructs from the guiding theoretical models, TPB and BNI.

Behavioral attitudes related to SV were assessed via a modified version of the Illinois Rape Myth Acceptance Scale, adapted by the evaluation of Green Dot [45,46], consisting of 7 statements related to rape. Previous exploratory and confirmatory factor analyses have demonstrated both reliability (Cronbach α=.94) and construct validity for the Illinois Rape Myth Acceptance Scale in various US settings [45], and preliminary baseline analyses indicated evidence of strong reliability in our sample (Cronbach α=.93). Responses to items were summed and averaged to create a score.

Perceived behavioral control was assessed in the form of confidence in the ability to engage in active bystander behavior. Items for this construct were modified from the intentions to intervene and SV intervention behaviors scales from the evaluation of CBIM [16,40]. Participants were presented with eight SV behaviors and asked, “how confident are you that you could say or do something to stop a friend or another peer who was…” This scale demonstrated evidence of strong reliability in preliminary baseline analyses (Cronbach α=.93).

Behavioral intention to actively intervene to prevent SV was assessed using a modified version of the Intentions to Intervene Scale from the evaluation of CBIM [16,40]. In the modified version for this evaluation, participants were asked how they would react to witnessing four different SV behaviors. Response options include “I would laugh or go along with it,” “I wouldn’t say or do anything,” and “I would say or do something to let them know that I think acting like that is not okay.” Preliminary baseline data indicated evidence of reliability (Cronbach α=.85).

Bystander intervention behaviors were assessed using a modified version of the CBIM SV intervention behavior scale [16,40]. Participants were asked to indicate whether they had witnessed each of eight SV behaviors perpetrated by their peers in the past 12 months and how they responded. This scale demonstrated evidence of reliability in preliminary baseline analyses (Cronbach α=.83).

Social capital, aligned with the construct in the TPB, was assessed in terms of community connectedness and social cohesion because of the way the C2H model aims to build trust, relationships, and collaboration among community members as a mechanism for reducing SV. Community connectedness was measured using four modified items from the Hemingway Measure of Adolescent Connectedness (previously found to be internally reliable at Cronbach α=.73) [47,48]. Participants were asked to identify the community they felt most connected to, and then asked about their connection to this community using two modified items from the Sense of Belonging Scale [49]. Social cohesion was measured using two modified items from the Neighborhood Social Cohesion Scale by Sampson et al [50] and six modified items from the Brief Sense of Community Scale by Peterson et al [51]. These eight items asked about perceived social cohesion in the specific geographic location of the C2H or control community to which the participant was linked. Preliminary baseline data demonstrated reliability (Cronbach α=.88).

Subjective norms related to perceived acceptability of SV and related to perceived likelihood of community members intervening to prevent SV were assessed. These items were developed by the research team based on principles suggested by Cislaghi and Heise [52]. Subjective norms regarding acceptability of SV behaviors were measured via modification of four items of the CBIM Intentions to Intervene Scale [16,40], focused on participants’ self-selected community to which they feel most connected. In preliminary baseline analyses, these items showed evidence of reliability (Cronbach α=.89). Subjective norms regarding perceived likelihood of community members intervening to prevent SV were measured with three additional items that included reference to the geographic community. These items were developed using the Shared Concern Scale by Lippman et al [53] (originally developed for the issue of HIV, with an α of .85). In preliminary baseline analyses, these items showed evidence of reliability (Cronbach α=.92).

The primary outcomes, SV and DV, were analyzed as binary variables, constructed as lifetime and past 12-month indicators of experiencing that form of violence. Secondary outcomes, most of which are assessed as scales, were continuous variables based on mean scores across included items.

We used baseline SN data to understand the clustering of social norms and behaviors within this study’s population (aim 1). To do this, we have described the network characteristics of the sample and assessed how SV social norms and experiences of SV and DV are associated among nominator-nominee dyads. We used dyadic observations from the baseline data for each pair of socially connected youth (dyad; eg, ego and alter 1, and alter 1 and alter 2) to restructure the dataset. We calculated whether having a peer who adheres to a particular norm or behavior was predictive of corresponding responses from paired versus nonpaired individuals. This was done using logistic regression analysis (outcome: yes or no) using a generalized estimating equation to account for correlation among dyads. We assessed for effect modification of these relationships based on the types of relationships (eg, school or other peer, or family), the mode of communication, gender concordance, and racial concordance. This analysis provided us with pertinent insight into what norms, attitudes, and behaviors were clustered among socially connected people, and across what relationships, creating a map for understanding opportunities for SN-based intervention. This manuscript is in preparation for publication.

Our primary analysis for main effects (aim 2) will be designed to assess the effects of the intervention on the prevalence of past 12-month SV and DV (sexual assault, online and in-person SH, DV, and LGBTQ+-specific SV) at 24-month follow-up. Our secondary analysis for secondary outcomes (aim 2) will be designed to assess the effect of the intervention on SV attitudes, behavioral intentions to intervene to prevent SV, reported intervention behavior, perceived behavioral self-efficacy to prevent SV, social norms of the acceptability of SV, and community connection and cohesion. First, we will conduct descriptive analyses to explore differences at baseline between the intervention and control groups and factors associated with loss-to-follow-up to understand how these factors may contribute to possible bias in our results. Then, we will assess change over time and by group via adjusted hierarchical difference-in-differences regression analyses to account for unmeasured and measured baseline differences between intervention and control groups, controlling for sampling level (ego, alter 1, and alter 2), accounting for clustering by community, and confounders identified using a dialectical acyclic graph [54]. The difference-in-differences approach does not rely on the strength of the match between intervention and control groups, but rather accounts for baseline differences, offsetting the impact of the three poor intervention-control matches that exist in the sample. Furthermore, we will adjust for factors associated with treatment group assignment to further account for baseline differences. In sensitivity analyses, we will use modern difference-in-differences approaches, such as group weighting approaches proposed by Callaway and Sant’Anna [55], to account for recent developments in understanding potential violations of parallel trends. Subanalyses will be conducted to assess for interaction at each sampling level (ego, alter 1, and alter 2), by gender, LGBTQ+ identity, race or ethnicity, and triangulating with process data. Process data effect modification analyses will assess whether implementation quality, program dosage, and characteristics of program sites modify the observed effect of the intervention on primary and secondary outcomes. Sensitivity analyses will be conducted to assess the variability of effects related to potential violations of the parallel trend assumption, low response rates if they unexpectedly occur, and bias due to potential differences in intervention and control response rates. In addition to the intention-to-treat analysis described above, we will conduct an as-treated analysis based on individual-level dosage estimates from process data. Finally, we will conduct a sensitivity analysis to account for loss-to-follow-up using inverse probability of censoring weighting [56] based on baseline data, should significant baseline differences exist between those lost and those retained. All analyses will be complete case analyses.

Additionally, using SN data, we will test for social diffusion of intervention via SNs (aim 2) via multiple methods. To test the effect of social influence on attitudes, norms, and behaviors, we will use the dyadic dataset of nominator-nominee pairs to test for the effect of social contacts’ attitudes, norms, or behaviors at 24-month follow-up, controlling for both of their own attitudes, norms, or behaviors at baseline and demographics related to these outcomes at baseline. We will explore a variety of model specifications during our analyses and conduct diverse robustness checks given the quasi-experimental design. As we will know whether network nominations are reciprocal (ie, if ego A nominated first tier alter A, did first tier alter A nominate ego A in return?), we can evaluate the possibility of omitted variables or confounding events explaining the associations by examining how the type or direction of the social relationship between ego and alter affects the association between ego and alter (using a “network directionality” test) [57]. Additionally, we will also conduct a sensitivity analysis in which we estimate the bias in the association between ego and alter that might be caused by an omitted variable that is correlated with the prevalence of the outcome variable in both ego and alter [58]. This class of omitted variables includes those that explain friendship formation based on a trait (homophily) and confounding environmental factors that may affect the ego and alter relationships.

Network effects likely will not be uniform across subgroups. For this reason, we will conduct a series of subgroup analyses, looking at potential moderators of potential network effects based on what we learn about the SN characteristics from baseline data. By running a series of interaction models on our main effect analyses, we will consider whether diffusion differs depending on gender, LGBTQ+ identity, the strength of the connection, and other potential moderators such as age, ethnicity, experiences of violence, and program exposure.

The CHKS is representative of in-school youth attending public schools within the target communities of the C2H and control programs (86 per every 100 youth, 86%, participates at study site schools). We estimate that we will have 6 intervention and 6 control schools who participate in both baseline and 24-month follow-up, each with 500 students on average. Difference-in-differences hierarchical models will be used to estimate the impact of the C2H program on primary and secondary outcomes, relative to control program communities, controlling for confounders and accounting for clustering at the school level. If statistical power allows, we will assess for modification due to gender, LGBTQ+ identity, race or ethnicity, implementation quality, and program site. Sensitivity analyses will be conducted to assess the variability of effects due to any potential violations of the parallel trend assumption and bias due to potential uneven inclusion across intervention and control schools.