The design was a multicenter, single-blind, parallel-group RCT, registered with ClinicalTrials.gov (NCT03695679). Ethical approval was obtained from the institutional review board of the University of Hong Kong/Hospital Authority Hong Kong West Cluster (UW-17029) and other universities. We followed the CONSORT (Consolidated Standards of Reporting Trials) guidelines.

Our sample size calculation was based on a primary comparison of behavioral change in the consistency with condom use. A previous study reported that the mean percentage of condom use was 68% (SD 39%) in individuals who underwent a computer-based intervention, compared with 24% (SD 35%) among those in a control condition, corresponding to a moderate to large Cohen effect size of 0.6 [29]. To conservatively detect a small Cohen effect size of 0.3 with 80% power and a maximum 5% false positive error rate using a 2-sided, 2-sample t test, a total of 352 (ie, 2 groups of 176) female university students were required. Assuming a 30% attrition rate based on a previous study using a web-based intervention [30], we calculated a total necessary sample size of 500 participants.

In total, 1503 female university students across various disciplines and years of study were screened from 5 universities with dormitory or residential halls in Hong Kong. Of these, 781 students were recruited. The eligibility criteria were (1) female university students aged 18 years or above who (2) are able to read Chinese or understand Cantonese, (3) are unmarried, (4) have been with intimate partners (ie, current and former dating partners or partners in a relationship) in the past 12 months, and (5) have not received any sexual health information (including formal face-to-face or online education or training courses related to contraceptives and sexually transmitted diseases, from a university, hospital, clinic, or nongovernmental organization) in the past 12 months. The exclusion criteria were (1) an unwillingness to complete the questionnaires at 3 time-points, (2) being pregnant or postnatal, and (3) having a psychiatric illness.

Students were reached via leaflets, campus booths, bulk emails, and posters. All interested participants received an invitation email to log into the Smart Girlfriend webpage (Multimedia Appendix 1-3). Online enrolment was used to screen students for eligibility. For eligible participants, written informed consent was obtained online, followed by a baseline questionnaire. After completing the questionnaire, the recruited participants were randomized to either the intervention group or the control group, according to a list prepared by blocked randomization (with blocks of 4), with a 1:1 randomization ratio. The block size and order of allocation were kept securely in the randomizer to avoid selection bias. The online platform conducted masking and allocation concealment according to the participants' enrolment sequence. Participants were automatically guided to the webpage associated with their allocation and were not aware of the group allocation in advance. Participants who completed all of the web-based questionnaires were given vouchers with a value of HKD $300 (USD $38.50). The privacy of all participants was ensured. Data collected from all questionnaires were stored in a protected university database.

Findings from our previous local multisite survey (n=1076) revealed that Chinese university students saw the relationship between unsafe sex practices and sexual coercion as problematic [25]. The results showed that 12% of female university students who were engaged in a dating relationship or had dating partners in the past year experienced sexual coercion within that year [25]. Therefore, we conducted an intervention to help enhance the awareness of sexual coercion among university students. Sexual issues are typically taboo subjects in Chinese culture; therefore, to minimize embarrassing situations for participants, we implemented the Dating Compassion, Assessment, reFerral, and Education (Dating CAFE) Ambassador Programme to help Chinese university students with dating violence. The intervention's development was based on the theory of planned behaviors and was conducted via 3 face-to-face workshops (totaling 7.5 hours) with 81 university students. Compared to students in the control group, we found that the students trained to be ambassadors had significantly enhanced behavioral intentions and control to help peers who were experiencing dating violence, decreased acceptance of dating violence, and increased subjective norms for helping others [31]. Moreover, we learned that discussions about dating issues were attractive and acceptable among university students. In addition, the face-to-face intervention was labor-intensive, and some students could not be enrolled due to timetable clashes. Thus, a web-based intervention could be a more cost-effective and practical approach for reaching as many eligible young people as possible.

The development of Smart Girlfriend was based on the Dating CAFE initiative. Smart Girlfriend is designed to be a sexual health literacy intervention empowering female university students with enhanced knowledge, attitudes, norms, and self-efficacy around managing sexual health, particularly condom use for safe sex practice. Based on the Health Belief Model (Figure 1) [32], we delivered Smart Girlfriend in 3 phases. In the first phase, participants were able to check their perceived susceptibility and perceived severity for individual STIs. The Harvard Cancer Risk Index [33] was used to ask 8 questions to estimate individual STI and cervical cancer risks. Respondents obtained personalized results from their answers, including factors that may increase or reduce their risk of getting STIs and cervical cancer. In the second phase, participants were given knowledge-based information about STIs and cervical cancer in text format, including relevant statistics, development, possible symptoms, and prevention methods. Some scale-based questions were asked to help the students think about the positive and negative features of condom use. Participants were able to offer personal feedback and were provided with an opportunity to reflect on which benefits or barriers mattered most to them (Figure 2).

In the second phase, participants were also prompted to take action regarding condom use. Therefore, condom use procedures and tips, as well as web links for local STI testing, cervical screening programs, and HPV vaccination programs, were provided. In addition, participants' self-efficacy in condom use was enhanced by providing information about assertiveness in sexual consent and sexual communication to avoid sexual coercion and casual sex. Three 5-minute videos were created with narrative stories about STIs and HPV infection based on different scenarios relevant to common situations for university students, about the handling of sexual consent in dormitories, engaging in sexual communication at home, and talking to a friend about worries regarding sex without a condom after a Christmas party (Multimedia Appendix 2). The Continuum of Conflict and Control (CCC) theory [34] was used to guide the stories, emphasizing that sexual coercion can occur without physical violence and with minimal fear to strengthen participants' knowledge about sexual coercion, sexual consent pertaining to condom use, and sex refusal. Participants were able to assess their own values and receive feedback on their choices. Textbox 1 shows the perceived benefits of and barriers to condom use featured in the interactive intervention.

All data were collected from the webpage. Participants were sent a reminder email and SMS text message for completing the online questionnaires at 3 time-points: baseline (T1); 3-months postintervention (T2); and 6-months postintervention (T3). The primary outcomes were the consistency of condom use with every partner, in accordance with the recommended guidelines of a systematic review of 56 studies [35], which used the percentage of male condom protected sex with every partner during the past 3 months.

Secondary outcomes were (1) knowledge, attitudes, norms, and self-efficacy of condom use, as appraised by the 25-item Multidimensional Condom Attitudes Scale (MCAS) [36]; (2) knowledge, attitudes, norms, and self-efficacy of sexual coercion and sexual consent, as measured by the 39-item Sexual Consent Scale–Revised (SCS-R) [37]; and (3) self-efficacy in sexual communication, estimated by the 20-item Sexual Communication Self-efficacy Scale (SCSES) [38]. The MCAS items were answered using a 7-point Likert scale, and total scores ranged from 7 to 175; a previous study has shown acceptable validity and reliability in the Chinese population [39]. The Cronbach alpha in this study was .84. The SCS-R contained 3 attitudinal subscales (positive attitude toward establishing consent, lack of perceived behavioral control, and sexual consent norms) and 2 behavioral subscales (indirect consent behaviors, and awareness and discussion). The SCS-R items were answered using a 7-point Likert scale, and the Cronbach alpha in this study was .67. The SCSES items were answered using a 4-point Likert scale, and total scores ranged from 20 to 80. With the exception of one of the SCS-R subscales (lack of perceived behavioral control), high scores on the scales indicated a high level of measured outcomes. The Cronbach alpha in this study was .94.

Other outcomes included participant inquiries, participant satisfaction, and participation in the intervention. Participants' inquiries were collected to understand their further needs. Satisfaction with the intervention was evaluated by recording the overall satisfaction of the intervention on a scale of 0-10. In addition, participants were asked which part of the intervention was most memorable. The higher the score, the higher the overall satisfaction with the intervention. The total recorded views of each website page (recorded via Google Analytics) and the incidence of searching for more information were measured.

Requested demographic information included sexual orientation, birthplace, dating relationship status, and history of childhood sexual coercion. Information regarding individual risk of STIs and cervical cancer (Multimedia Appendix 3) was collected at baseline, including age, age at first sexual intercourse experience, number of sexual partners during one's lifetime, frequency of condom use, history of being diagnosed with an STI, smoking status, history of giving birth, and history of having a Pap smear test. The experience of sexual coercion was measured using a 7-item subscale of the Revised Conflict Tactic Scale [40], which indicated whether it had happened and how often the behavior had occurred in the past year.

The primary outcome, consistency of condom use, was analyzed using a zero/one inflated beta (ZOIB) regression model because the raw data of the consistency of condom use were not normally distributed, exhibiting excessive zeros and ones (ie, consistency of condom use=100%). ZOIB is based on a piecewise distribution, which accounts for the probability mass at 0 or 1 and the probability density within (0,1). Bayesian-based results were obtained using Stan (Stan Development Team; 4 chains, 4000 iterations, 1500 warm-ups) [41]. Odds ratios (ORs) and credible interval (CrI) values were calculated for Bayesian-based analysis.

We adopted the intention-to-treat principle, and all study subjects were included in the analysis. Missing values at the 3- and 6-month follow-ups were replaced by the last observed value. Missing values at baseline were replaced by values from the 3- or 6-month follow-up. If there was no value obtained at any of the 3 time-points, the participant was excluded from the analysis. For other outcomes, the t test was applied for continuous data and the chi-square test was applied for categorical data. All P values were 2-sided, and P<.05 was considered statistically significant. R software (version 3.6.1; R Core Team) with “tidyr” and “brms” packages was used to analyze the data. Questions were downloaded from the inquiry system. Content analysis was performed to categorize the collected responses.

Figure 3 shows the flow of participants at baseline, grouping randomization, 3-month follow-up, and 6-month follow-up. Of 1503 screened students, 722 were excluded and 781 students were enrolled (enrolment rate: 52%), randomized into the intervention group (384/781, 49%) and the control group (397/781, 51%). Of the 781 included participants, the dropout rate was 8% (60/781) at the 3-month follow-up (intervention: 28/384, 7%; control: 32/397, 8%) and 9% (70/781) at the 6-month follow-up (intervention: 36/384, 9%; control: 34/397, 9%). Finally, 229 participants in the intervention group and 246 participants in the control group were included in the analysis of the primary outcome. A total of 362 participants in the intervention group and 384 participants in the control group were analyzed for other outcomes. Details of the quality check criteria used are provided in Multimedia Appendix 4.

The mean age of participants was 21.5 (SD 2.6) years, ranging from 18 to 32 years. Most participants were born locally (536/746, 72%), and 18% (134/746) self-reported as sexual minorities (Table 1). Approximately 10% (78/746) of participants reported child sexual abuse experiences, and 22% (164/746) reported a history of sexual coercion.

Of the 781 participants, 10% (81/781) sent inquiries via the webpage. Of these, 27% (22/81) of inquiries asked about condoms and other contraceptive methods, 28% (23/81) asked about STIs, 14% (11/81) asked about sexual behaviors, and 31% (25/81) asked about suggestions for intervention, technical issues, or other issues (Table 4). A significant difference was found for condoms and other contraceptive methods (χ²=9.78, P=.002) and STIs (χ²=6.71, P=.01) between the 2 groups. Compared with the control group, participants in the intervention group were more likely to send inquiries about STIs (intervention: 18 vs control: 5), whereas intervention group participants were less likely to send inquiries about condoms and other contraceptive methods (intervention: 6 vs control: 16). In the category of STIs, 3 participants in the intervention group asked whether there is a risk of STIs and how to prevent STIs among women who have sex with women.

A slight but statistically nonsignificant difference in satisfaction (t₆₇₇=0.15; P=.89) was found between the intervention (n=353) and control (n=326) groups, with average satisfaction scores of 6.19 (SD 2.75) and 6.22 (SD 2.78), respectively. Among participants who reported the type of content they remembered most (Table 5), almost half of the participants in the intervention group (48/118, 41%) and more than half of the participants in the control group (58/107, 54%) reported that content related to condoms was most memorable. In addition, more intervention group participants reported that content about sexual consent was the most memorable (32/118, 27%) than control group participants (21/107, 20%). There was a significant difference between the 2 groups for seeking out more information about safe sex practice at the 3-month follow-up (intervention: 120/334, 36% vs control: 92/353, 26%; χ²=7.83, P=.005) but not at the 6-month follow-up (intervention: 105/326, 32% vs control: 97/353, 28%; χ²=1.81, P=.18). Data from Google Analytics revealed that the most visited pages of the intervention were the inquiry page and the page with information about risk factors for cervical cancer.

This multicenter RCT in Hong Kong is the first study to examine the effects of an interactive web-based sexual health literacy program to promote safe sex practice among female Chinese university students. The consistency of condom use increased over time in both groups. The intervention group exhibited 5% higher consistency at the 3-month follow-up and 1% higher consistency at the 6-month follow-up compared to the control group; however, these differences were not statistically significant. Thus, the results did not reveal an effect from the intervention for increasing the consistency of condom use.

One of the possible explanations for the lack of positive effect may be due to the passive exposure of the intervention to our target audiences. In our intervention, the main components were delivered passively, including the knowledge about STIs, the narrative stories about STIs, and the introduction about sexual coercion. A meta-analysis for HIV-prevention interventions also revealed that participants in active interventions (such as client-tailored counseling and other activities to improve behavioral skills) reported a greater behavioral change improvement compared to passive interventions (such as messages to share procondom information and norms and to verbally model skills); and likewise, passive interventions in the intervention groups did not differ significantly from the control groups [42]. These findings show the positive effect of active interventions and suggest that increasing active components would be necessary for future research in promoting safe sex practice.

Another potential reason for the lack of positive effect could be associated with the nonpersonalized nature of the web-based intervention content. According to the inquiries we received from the participants, some participants asked if women who had sex with women were at risk for STIs, and some intervention content was not applicable to them. Previous studies found that personalized web-based interventions were more effective than nonpersonalized interventions, and the difference between the nonpersonalized intervention group and the blank control group was nonsignificant [43]. Thus, behavioral change may be improved if we tailor the intervention content to the participant's characteristics, such as sexual orientation.

Additionally, the imbalanced ceiling effect for 2 groups might also contribute to the insignificant intervention effect. The proportion of participants reporting 100% condom use at baseline in the intervention group was significantly higher than the control group (68% vs 59%, respectively; χ²=3.90, P=.048. Owing to the ceiling effect, the potential capacity of improvement in terms of condom use in the intervention group would be less than the control group, and the effect of intervention might be underestimated [44].

Increased MCAS scores at the 3-month follow-up indicate that the Smart Girlfriend program improved participants' knowledge, attitudes, norms, and self-efficacy of condom use temporarily; however, there was no significant long-term effect on behavior observed at the 6-month follow-up. These results indicate that the intervention could temporarily improve knowledge, attitudes, norms, and self-efficacy of condom use in young adults but cannot ultimately change their behaviors. These findings were in line with a previous study [45]. Starosta et al [46] found that a web-based intervention improved attitudes toward condoms but could not change condom use at the 3-month follow-up. Thus, it would seem that improvement occurs, but it would decrease over time if there is no further exposure in the intervention [47].

The strengths of the study include the detailed development of the intervention based on theory and previous experience. The low dropout rate in this study indicates that university students welcomed the intervention and that it served the needs of the target population. With the participation of 5 universities in Hong Kong, the study included a diversity of student population characteristics, sexual orientations, and sexual coercion experiences. ZOIB was employed in this study, given the nonnormal distribution data with excessive zeros and ones. In promoting sexual health literacy, self-reported condom use is characterized by a substantial number of zeros or ones and discrete nonzero counts [48]. Excessive numbers of zeros or ones in this dependent variable make it difficult to fit the data using traditional methods, including ordinary least squares models and mixed linear models, which may yield biased estimates of the results. The beta distribution is considered a versatile function that fits a broad range of probability distribution shapes. This study may serve as a promising reference for future studies with similar problems. A high participation rate (91%) was reported in this study, indicating a high acceptability of digital interventions among university students and suggesting that a large number of university students can be reached at a very low cost.

This study has implications for future studies. Temporary improvement was observed in the knowledge, attitudes, norms, and self-efficacy of condom use, but not observed postintervention in behavioral change; this suggests that future research should more thoroughly consider novel methods that maintain the intervention effect and increase the behavioral change effect of web-based sexual health literacy. First, an active approach for intervention delivery may lead to the greatest increase in condom use. For example, virtual reality's effectiveness for behavioral skills training and practical exercises has been demonstrated in other research fields [49]. Second, the intervention content should be more personalized, especially for our target participants, before the intervention. More specific and advanced information (eg, whether nonpenetrative sex is a risk factor for contracting STIs), more customized content (eg, for women who have sex with women), and personal online counseling through the website might be needed. Third, more attention should be paid to the organization of the external links to other websites. We provided external hyperlinks to other websites to give participants more opportunities to explore other related information. In hindsight, providing external hyperlinks might not have been ideal since they may have distracted the participants from the intervention pathway and led them away from (and discouraged the return to) our website [50]. Recent research tested a web-based basic version alongside a version with added links to external resources, and it was found that the latter version was not effective [51]. The external links could be listed in the last (or a separate) page in future revised interventions.

Our study has several limitations. First, despite our relatively large sample of participants, compared with many other sexual health interventions, our study might lack power concerning the primary outcome for the consistency of condom use. The number of sexually active participants during the 6-month follow-up was relatively lower than expected. We included those who only reported sexual activities in one of 3 questionnaires and used a conservative approach that considered sexually inactive participants as exhibiting no change. This conservative approach might bias our results. Second, there is a potential influence of the Hawthorne effect. A systematic review found that the Hawthorne effect could influence participants' behavior in RCTs [52]. In our study, self-reported data were collected, and participants knew that the program was designed to increase their consistency of condom use. Third, we did not collect information about the duration of participation. However, due to the digital intervention design, all participants received information about the whole intervention before completing the questionnaires. Finally, our results were vulnerable to differential error because self-reported data were collected. The intervention group may have misreported the consistency of condom use to a greater extent than the control group at baseline (intervention: 78% vs control: 72%), thereby biasing the estimated intervention effect. There is no alternative to self-reporting in web-based trials because collecting biological data via web-based systems is unfeasible. Differential error could not be avoided; however, our trial was conducted under conditions that maximize accurate reporting, using an anonymous web-based questionnaire that assured participants of confidentiality with an absence of investigators.

Among university students in Hong Kong, an interactive web-based sexual health literacy program resulted in a small but statistically nonsignificant increase in the consistency of condom use, as well as a significant and temporary increase in knowledge, attitudes, norms, and self-efficacy of condom use, but not in sexual coercion, sexual consent, or sexual communication, when compared with participants receiving only one webpage of condom use information. The number of participants at enrolment and the high participation rate highlight the need for sexual health literacy programs for young adults. Moreover, a future revision of this intervention should be personalized and delivered with an active approach.