Snowball sampling was employed to access MSM who are difficult to reach via probability frames and to trusted peer networks for online recruitment about HIV-preventive behaviors. This approach is commonly used in studies of stigmatized or hard-to-reach populations due to feasibility and rapid data collection. Recognized limitations include selection bias, over-representation of homogeneous peer clusters, and unknown sampling probabilities. To mitigate these concerns, multiple heterogeneous “seeds” were initiated across venues and communities, namely (1) the Thai Red Cross AIDS Research Centre; (2) the Institute of HIV Research and Innovation (IHRI); (3) Community-Based Treatment (CBTx) and the Community-based Harm Reduction and Rehabilitation Service at Public Health Service Center 41 Khlong Toei, Bangkok Metropolitan Administration; (4) the Rainbow Sky Association of Thailand; and (5) the Mplus Foundation, Chiang Mai. Standardized eligibility screening was applied, referrals per participant were capped, and sample composition was monitored throughout data collection to minimize dominance of any single network.

Eligibility was limited to MSM aged ≥25 to target working-age adults in Thailand with independent income and greater autonomy in health decisions, which shape HIV-preventive behaviors and service access differently from younger student populations. This enhances relevance to community/clinical use. The subjects identified themselves as MSM at least once within the previous six months, and were willing to answer the online HIV preventive behavior scale using a Google form. A total of N = 424 respondents was recruited and randomly split into two independent subsamples for factor analysis: Exploratory factor analysis (n = 212) and Confirmatory factor analysis (n = 212). An item-to-sample ratio of approximately 20:1 was used for initial validation. With 9 items of HIV preventive behavior scale, this yields a minimum of 180 participants per analysis, exceeding this benchmark and aligning with guidance that typically recommends 10–20 participants per item for factor models and absolute N ≥ 200 for stable factor recovery and fit evaluation. ^{16 , 17}

HIV preventive behavior scale. A brief, context-appropriate scale was developed for Thai MSM to capture contemporary prevention practices not fully reflected in adult Thai measures. Items were adapted from the instruments of Pimthong and Bhanthumnavin ^{13 , 14} and refined through cognitive interviewing ¹⁸ with three experts who each had ≥5 years of experience in HIV prevention among Thai MSM. Each expert completed two interview sessions (60–90 minutes per session). Interviews assessed item comprehension, appropriateness for Thai MSM, and alignment with HIV-prevention behaviors, using think-aloud and probing techniques to identify ambiguous or double-barreled wording; revisions prioritized clear Thai phrasing and domain alignment. Qualitative analysis of the cognitive-interview data yielded three thematic categories (reported in the Results), which informed an initial 34-item pool mapped to key HIV-prevention behaviors. Content validity was quantified using the Index of Item Congruence (IOC) ¹⁹ based on ratings from another three independent experts. Items with IOC < .50 were removed or revised, and those with corrected item–total correlations < .30 were further refined to enhance content relevance and internal coherence prior to field testing. The final instrument comprised 9 items loading on two components: risk avoidance for HIV infection and self-protection behavior before and during sexual activity. All items are positively keyed on a five-point frequency scale (1 = never to 5 = routinely), referencing behaviors over the preceding six months, with higher total scores indicating higher HIV-preventive behavior. The construct is defined as self-protective actions enacted before and during sexual activity, together with effective regulation of sexual emotions, to reduce acquisition and transmission risk.

AIDS risk behavior avoidance scale, developed by Pimthong in 2011, ^{13 , 14} it measures behavioral intent or a person’s readiness to attempt to avoid actions or activities that lead to AIDS risk. By choosing to act or not to act, such as changing sexual partners often taking drugs or intoxicants watching porn, and choosing places that should or should not go, such as entertainment venues, and gay saunas, a total of 12 items, with 7 positive items and 5 negative items. The responses are rated on a 6-point scale ranging from the truest to not true at all. Scores were calculated using a total score of 12–72. If the score was high, it indicated that there was a high risk of AIDS avoidance behavior. For the psychometric characteristics, the discriminant power ranged from 4.74 to 8.93, the item correlation coefficient with the total score was .32 to.64, and the reliability for the whole version with the alpha coefficient was.83. with empirical data with a χ ² = 34.70, df = 46, p-value = .89, NFI = .94, GFI = .95, AGFI = .91, SRMR = .049, CFI = 1.00, and RMSEA = 0.0. This scale was used to examine convergent validity with the HIV prevention behavior scale.

AIDS prevention scale, developed by Pimthong in 2011, ^{13 , 14} it is a model to measure a person’s sexual behavior. By considering the method of sexual intercourse, such as using a condom with lubricant every time you have anal sex. Not swallowing partner’s semen, etc. There were a total of 10 items, 4 positive and 6 negative items. Each item had a 6-point evaluation scale ranging from very true to absolutely false by measuring 3 behavioral components. The aspects related to the risk of HIV infection during sex were 1) no risk, 2) moderate risk, and 3) high risk. The score is calculated using a total score of 10–60. If the score is high, it indicates that AIDS prevention behaviors while having sex are high. For the psychometric characteristics, the discriminant power ranged from 4.09 to 9.90, the item correlation coefficient with the total score was .20 to.62, and the reliability for the whole version with the alpha coefficient was.77. with empirical data with a χ ² = 23.52, df = 30, p-value = .70, NFI = .93, GFI = .96, AGFI = .92, SRMR = .049, CFI = 1.00, and RMSEA = 0.0. This scale was used to examine convergent validity with the HIV prevention behavior scale.

Thai Language Learning Attitude Scale, developed by Samrongthong in 2011 ²⁰ as a model to measure the attitude towards learning the Thai language. There are a total of 20 items. Answers are self-exploratory. Questions are both positive and negative items. Then choose a response from 5 levels: strongly disagree, disagree, not sure, agree, and strongly agree. There was reliability from internal consistency analysis with Cronbach’s reliability coefficient of. 85. This scale was used to examine discriminant validity with the HIV prevention behavior scale.

Qualitative data analysis

Transcripts were de-identified and content-coded by item–probe segments. Codes were iteratively grouped and synthesized into categories with brief verbatim excerpts as evidence. Categories were summarized in an item-by-category matrix to guide revision, cut-off decisions (rewording, examples, response options) or generate new items.

Quantitative data analysis

After developing the draft HIV preventive behavior scale from cognitive interviewing, we first examined corrected item–total correlations (target ≥ 0.30) and internal consistency via Cronbach’s alpha (acceptable ≥ 0.70). Construct validity was assessed in two stages. For the EFA, we used principal axis factoring with varimax rotation; factorability was verified (KMO ≥ 0.70; Bartlett’s p < .001). ¹⁶ Items were retained if primary loadings were ≥ 0.30; items with evidence of redundancy/multicollinearity (inter-item r > 0.80) were removed. For the CFA, model fit was evaluated against conventional thresholds (χ ²/df < 3.0, CFI ≥ 0.90, TLI ≥ 0.90, RMSEA ≤ 0.08, SRMR ≤ 0.08). ²¹ To examine convergent/discriminant patterns, we computed Pearson correlations with theoretically related measures (AIDS Risk Behavior Avoidance Scale, AIDS Preventive Behavior Scale) and an unrelated construct (Thai Language Learning Attitude Scale), expecting significant correlation with the former and no significant correlation with the latter.

Data were collected via a secure google form which need to use password to log-in. The landing page of the form presented an information sheet describing the study purpose, procedures, inclusion criteria (≥25 years), expected time, potential risks/benefits, voluntary nature of participation, and contacts for questions or concerns. Participants provided explicit electronic consent by ticking an “I agree to participate” box before any survey items were shown; those who did not consent were automatically exited from the survey. Participants could skip any item and withdraw at any time prior to submission without penalty.

To maintain anonymity, the survey did not request names, phone numbers, or other direct identifiers. IP addresses and device metadata were not retained in the analytic dataset. Responses were stored under randomly generated codes and exported in de-identified form. To ensure confidentiality, data were kept on encrypted, password-protected drives with access limited to the principal investigator and authorized research staff; reporting used only aggregated results and non-identifiable quotes. Data handling adhered to the Institutional Review Board (IRB) requirements and applicable national regulations. The procedures complied with the Declaration of Helsinki and guidance for research involving vulnerable populations. For the cognitive interviews with experts, participants received a formal request letter via their affiliations along with an information sheet. Written (or recorded verbal) consent was obtained before the session; permission to audio-record was sought separately. Experts could decline or stop the interview at any point, and any illustrative quotations were anonymized in dissemination.

Cognitive interviewing yielded three categories of HIV preventive behaviors among MSM: 1.

Denial and avoidance of HIV risk; behaviors that minimize or steer away from situations with elevated exposure (e.g., downplaying susceptibility, ignoring risk cues, avoiding venues/partners/substances linked to condomless sex or chemsex.

“Outdoor activities should be avoided because they can increase sexual risk” (Participant 1)

The categories that emerged from the cognitive interviews were generated through an iterative and systematic analytic process. First, all interview data were transcribed verbatim and reviewed multiple times to achieve familiarization with participants’ interpretations and response processes. Meaning units related to how participants understood, interpreted, and responded to each item were then identified and coded. Next, similar codes were grouped together to identify recurring patterns in participants’ comprehension, perceived relevance, and suggested revisions. Through constant comparison across cases, these patterns were refined into broader conceptual categories. The research team engaged in discussion to reach consensus on the boundaries and definitions of each category, ensuring that the categories accurately reflected participants’ cognitive processes and were grounded in the data. These categories subsequently informed item refinement and conceptual clarification of the measure.

Quotations under denial and avoidance of HIV risk demonstrate how participants appraise and minimize exposure (e.g., avoiding high-risk venues/contexts, recognizing partner turnover and substance-linked encounters), thereby justifying items in the risk-avoidance component. Quotations under self-protective actions before and during sexual activity specify concrete practices (e.g., condom negotiation/use, testing routines, and avoiding intoxicated sex to preserve adherence), directly informing items in the self-protection component. Quotations related to appropriate sexual response clarify mechanisms of boundary-setting and emotion regulation that support consistent prevention and sexual well-being.

The psychometric sample comprised 424 MSM reporting sex with a man in the past six months (mean age = 32.08, SD = 6.42). Content validity, assessed by three expert reviewers, showed item–objective congruence (IOC) ranging 0.66–1.00 across 19 items. Items were then screened via corrected item–total correlations (CITC) using a retention threshold of ≥0.30. Only nine retained items exhibited CITC values 0.33–0.47. Assumptions of univariate normality were acceptable (skewness and kurtosis within ±2), supporting subsequent factor analyses.

Pairwise associations among the nine items (N = 424) were examined using Pearson correlations; 35 of 36 pairs were significant at p < .05 (range r = .11–.58; no pair exceeded.90), with one nonsignificant pair (items 2 and 4). Sampling adequacy was acceptable (KMO = .78) and Bartlett’s test of sphericity was significant (χ ²(36) = 843.73, p < .001), supporting factor analysis. An exploratory analysis was conducted on a random half-sample (n = 212) using PCA with varimax rotation. Component retention followed multiple criteria: eigenvalue >1, inspection of the scree plot for an elbow, simple structure/interpretability, and consistency with the a priori content framework. Items were retained on a component when primary loadings ≥.40 and cross-loadings < .30 ^{16 , 21} items failing these thresholds were reviewed for content coherence and communality.

The initial solution yielded three components with eigenvalues >1. Component 1 (denial/avoidance of HIV risk) showed loadings .75–.83 for its three items. Component 2 (self-protective actions before/during sexual activity) showed loadings .46–.73 for items 5–9. Item 4 emerged as a singleton on Component 3, with insufficient item count for a stable factor and content overlap with self-protective actions. Given its acceptable conceptual alignment (pre-sex planning/safer-sex enactment), item 4 was reassigned to Component 2, yielding a two-component structure consistent with the theoretical model: (1) denial/avoidance of HIV risk and (2) self-protection before/during sexual activity. Details are in Table 1.

The two-component structure was tested in the another half-sample (n = 212) using CFA and showed adequate global fit, meeting or exceeding conventional benchmarks: nonsignificant χ ² (χ ² = 36.56, p = .06), χ ²/df = 1.46 (≤3.0), GFI = .96 (≥.90), CFI = .98 (≥.95), TLI = .96 (≥.95), RMSEA = .05 (≤.06), and RMR = .07 (≈ ≤ .08), ¹⁷ supporting the adequacy of the model. At the item level, Component 1: denial and avoidance of HIV risk comprised items 1–3 with standardized loadings (β) = .60–.85, SE = .15–.20, t = 7.32–7.55, p < .01, and R ² = .36–.73. Component 2: self-protective actions before and during sexual activity comprised items 4–9 with β = .41–.74, SE = .17–.26, t = 4.15–6.29, p < .01, and R ² = .13–.54. The latent correlation between the two components was.58, indicating related but distinct constructs ( Table 2). Together, the fit indices and item-level criteria justify the allocation of items to the two reported components and the reassignment of item 4 based on statistical thresholds and construct coherence. ¹ Table 2.

Items	b	SE	t	R 2
Dimension 1, Denying and avoiding the risk of contracting HIV (F1)
Item 1 You avoid having sexual contact. When it is discovered that your partner is not protected by wearing a condom. (X1)	.60	-	-	.36
Item 2 You decline when a buddy invites you to a potentially risky sex activity, such as group sex or traveling to a meeting spot for sex. Substance Dependency. (X2)	.85	.20	7.55 **	.73
Item 3 You consistently refuse to switch partners. (X3)	.66	.15	7.32 **	.44
Dimension 2, Self-protective behavior before and during sex (F2)
Item 4 You use a condom during sexual activity. (X4)	.49	-	-	.24
Item 5 You use oil-free lubricants during sexual activity. (X5)	.37	.21	4.15 **	.13
Item 6 You avoid touching the wound immediately after sexual activity. (X6)	.74	.26	6.29 **	.54
Item 7 When you need to, you can show your sexual feelings in many ways, such as masturbation, spiritual activities, and recreational activities. (X7)	.41	.17	4.51 **	.17
Item 8 You engage in sexual action while avoiding being harmed or producing blood or lymph. (X8)	.72	.26	6.26 **	.52
Item 9 You engage in sexual activity without causing physical harm to your partner. (X9)	.63	.26	5.88 **	.39

**

p < .01.

Consistent with empirical data, the following two components of the HIV Preventive Behavior scale model were identified by CFA: Chi-square: χ ² = 36.56, p = .06, Relative Chi-square: χ ²/df = 1.46, Goodness of Fit Index: GFI = .9, 6, Comparative Fit Index: CFI = .8, Adjusted goodness of fit index: AGFI = .94, Root Mean Square Residual: RMR = .07, Root Mean Square Error of Approximation: RMSEA = .05, Root mean square residual: TLI = .96. Overall, the model matches the empirical data well (good fit).

Internal consistency from Cronbach’s reliability coefficient was used to determine that the 9-item HIV Preventive Behavior Scale’s reliability was equivalent to .77 Pearson product-moment correlation coefficients, which were used to examine convergent and discriminant validity among 424 individuals Thai MSM. There are positive correlations between the HIV Preventive Behavior Scale (SUMPHIV) with the AIDS risk behavior avoidance scale (SUML), and the AIDS prevention scale (SUMA) were statistically significant at the .01 level, with correlations of .21 and .26, respectively. There was no correlation between the HIV Preventive Behavior Scale and the Thai Language Learning Attitude Scale (SUMT) (see Table 3).

Several limitations warrant consideration. First, recruitment via an online, snowball-based survey introduces selection bias (internet access, network homophily) and unknown sampling probabilities, which can limit generalizability beyond MSM engaged with digital/community networks. Second, reliance on self-reported behaviors over a six-month window is vulnerable to recall error and social desirability, potentially inflating protective reports (e.g., condom use) and underreporting risk contexts (e.g., chemsex). Third, the sample’s age composition (≥25 years) and sexual-orientation profile may yield sample homogeneity, constraining applicability to younger MSM and other subgroups (e.g., bisexual-identified men, gender-diverse partners). Fourth, the cross-sectional design precludes test–retest reliability, temporal stability, and predictive validity (e.g., incident STI/HIV).

Future research should broaden sampling frames (e.g., respondent-driven or venue-based strategies, multiple regions) and include younger MSM to assess age-related differences. Methodological refinements include test–retest assessment, longitudinal designs for predictive validity, incorporation of social desirability scales, ecological momentary assessment to reduce recall bias, and triangulation with clinic/biomarker data (testing history, PrEP dispensing, STI diagnoses). Cross-cultural studies should examine how sexuality norms, masculinity scripts, stigma, and service access shape the two domains in other Thai regions and countries. Finally, intervention trials can use the 9-item scale to tailor content, addressing denial/avoidance (risk-cue literacy, venue/substance harm-reduction) and enhancing self-protection (condom negotiation, PrEP/PEP adherence), and to track component-specific change as part of routine program monitoring.