CAPI technology is used to administer the survey. Survey components are primarily administered by study staff, and sensitive components (ie, race-based harassment, posttraumatic stress disorder, and intersectional discrimination) are self-administered by the participants. Survey measures are described in Table 2.

After the survey, trained phlebotomists collect the following biological samples from participants: (1) urine for toxicology screening for substance use; (2) one oral swab and one self-administered rectal swab to screen on site for chlamydia and gonorrhea [53]; (3) two additional self-administered rectal swabs to evaluate levels of proinflammatory cytokines and chemokines; and (4) a blood draw for screening for syphilis and for quantifying cannabis metabolites [54], HIV antibodies, and measures of PrEP usage. The biospecimen collection measures and corresponding constructs are summarized in Table 3.

Members of the testing team at CCHE are trained in trauma-informed and youth-centered care to deliver results for HIV and sexually transmitted infections (chlamydia, gonorrhea, and syphilis). Calls are conducted only after confirming the participant’s identity via full name and date of birth and are conducted in a private and confidential space, with no results left on voicemail, via SMS text messaging, or with another individual. Results are provided to participants over the telephone 1 to 2 weeks after their study visit date, with linkage to care options for those seeking treatment or state interest in PrEP. If a participant experiences distress following disclosure of results, immediate support will be provided via a referral for mental health services through CCHE, as well as established organizational safety protocols to ensure participant well-being.

fMRI scans are completed on a Philips Ingenia 3T scanner at the University of Chicago Magnetic Resonance Imaging Research Center. A scan session is conducted while participants complete the Iowa Gambling Task (IGT), a measure of R/R processing. The IGT [55] is a well-established computerized task designed to assess risk preferences using real-world decision-making. Participants view 4 decks of cards on a screen (labeled as decks A, B, C, and D) and start with a base amount of US $2000. Participants have 100 chances to select a card from the decks and are told, “With each card, you can win some money, but you can also lose some. Some decks will be more profitable than others. Try to choose cards from the most profitable decks so that your total savings will be as high as possible.” When a card is chosen, the subsequent gain or loss of money will appear, marked by red and green status bars at the bottom of the screen. Two decks are coded as being riskier, resulting in an overall loss. The other 2 are coded as safe, with fewer losses and more wins, resulting in an overall gain. Which decks are risky or safe changes between scan sessions. Participants are familiarized with the task before scanning, with a practice session conducted outside the scanner.

Following the IGT, a high-resolution T1-weighted anatomical scan is acquired for alignment and spatial standardization of the functional images across participants. Finally, a 6-minute resting-state scan is conducted with eyes fixated on a white crosshair on a dark background, using similar acquisition parameters as the IGT (gradient-echo echo-planar imaging sequences; repetition time =2 ms; echo time =3.5 ms; 39 3-mm-thick axial slices aligned to the AC-PC line with a 0.6-mm slice gap; 19.2*19.2 cm field of view; SENSE factor =2, flip angle =8°). Four images are acquired and discarded just prior to task start or resting-state sequence start.

Separate from the fMRI scan, EF is measured through multiple tests that sample across the range of processes considered EF. The Behavior Rating Inventory of Executive Function [56] (BRIEF-2 or BRIEF-A [57] for individuals aged 18 years and older or 16 to 17 years, respectively) is considered part of this battery and is administered as part of the survey. Additional EF measures include tests from the NIH Toolbox, which are administered on an iPad. We administer the Flanker Inhibitory Control and Attention Test, List Sorting Working Memory Test, Dimensional Change Card Sort Test, and the Oral Symbol Digit Test. Next, pen-and-paper tests are administered from the Delis-Kaplan Executive Function System [58]: the Trail Making and Verbal Fluency Test. The Word Reading subtest from the Wide Range Achievement Test IV is used to estimate general cognitive ability.

We code, enter, edit, link, and maintain biological data, electronic medical records, and CAPI-administered questionnaires in REDCap (Research Electronic Data Capture) [59]. Each participant is assigned a unique study ID to enable us to link subfiles. Before each participant leaves the site, their data will be reviewed for completion.

The principal exposure variable is cannabis use, measured objectively via plasma metabolites of cannabis and triangulated with self-reported cannabis use from the survey using validated measures such as the Cannabis Use Disorders Identification Test-Revised [60]. Objective measures of cannabis use are prioritized using previously validated decision rules for combining objective biomarker data with self-reported substance use, which have been used for stimulants [60] and alcohol [61] and are adapted for cannabis. This approach allows us to account for limitations of respective measures, such as inaccuracy of self-report as well as varying rates of metabolization with cannabis use. It also allows us to precisely delineate participants into potentially relevant categories of use.

HIV prevention and care outcomes are comprehensively assessed using multiple objective data sources triangulated with self-reported measures. Through expanded research ROI forms, we access detailed clinical data, including HIV and sexually transmitted infections testing history, diagnoses and treatment; HIV primary care visits; viral load measurements; antiretroviral therapy and PrEP medication records; receipt of prevention services including PrEP, nPEP, and doxycycline postexposure prophylaxis; mental health and substance treatment services; and case management attendance. This robust clinical dataset enables thorough assessment of the entire PrEP care continuum and HIV care engagement, which is triangulated with validated self-report measures from our survey instrument. Consistent with previous research showing strong correlations between self-reported PrEP use and objective clinical measures, our multimethod approach strengthens data validity by addressing the limitations inherent to any single data collection strategy. This methodology allows for precise characterization of both prevention behaviors among HIV-negative participants and care engagement patterns among participants living with HIV. This expanded research ROI was implemented beginning in March 2025, with data collection through this enhanced protocol applying to participants enrolled from that point forward.

We will also rigorously account for other substance use, including substance use disorder, as there is increasing recognition of the need to study substance use as it occurs in real-world settings, including polysubstance use [23,62]. It is difficult to determine whether neurocognition and HIV prevention outcomes are a result of cannabis use alone or a combination of cannabis and other substances. However, it is not necessary to definitively prove that neurocognition and HIV prevention outcomes are due solely to cannabis use in order to identify viable targets for biobehavioral interventions to prevent HIV infection in cannabis users. Nonetheless, to isolate the (causal) effects of cannabis use on neurocognition and HIV prevention outcomes, which is our primary objective, we will measure and control for a wide range of confounding (time-varying) covariates, including noncannabis substance use.

The fMRI data collected from the IGT will be processed using Analysis of Functional Neuro Images [63] and will include realignment of the time series, slice-time correction, high-pass filtering, and spatial normalization. Voxelwise statistical maps of responses to task stimuli (deciding to choose a risky deck, deciding to choose a safe deck, processing “win” feedback, processing “loss” feedback) are generated by estimating beta weights (activation levels) for each condition using an idealized response function. Standard atlas-based masks (Harvard-Oxford) are used to extract mean activation from bilateral R/R brain regions, including the anterior cingulate cortex, medial prefrontal cortex, orbitofrontal cortex, ventral striatum, thalamus, and parahippocampus during win or loss and risky or safe choices. IGT behavioral performance is calculated by subtracting the number of cards picked from safe decks from risky decks. These measures comprise the R/R processing construct.

EF assessment includes multiple performance-based and self-report measures. Performance-based measures include NIH Toolbox tests (Flanker Inhibitory Control, List Sorting Working Memory, Dimensional Change Card Sort Test, and Oral Symbol Digit Test), Delis-Kaplan Executive Function System tests (Trail Making Test and Verbal Fluency), and Wide Range Achievement Test IV Word Reading subtest. Self-report measures include the BRIEF-A and BRIEF-2 (for participants aged 16‐17 years). Standard scores are calculated and adjusted for age and education level per manual instructions. These measures comprise the EF construct.

Separate latent variable models will be developed for R/R processing and EF constructs, allowing for domain-specific analysis of cannabis effects on neurocognitive functioning.

The purpose of these analyses will be to determine both the cross-sectional and longitudinal associations between cannabis use and neurocognitive biomarkers. Let z_ijk and λ_ij be observed measures of cannabis use and neurocognitive performance, respectively, for the ith participant at the jth time point (j=1, 2, or 3). Since multiple measures are being collected for each construct, the subscript k is used to identify a specific measure. We will begin by developing an appropriate generalized measurement model for each construct of the form:

where θ_ij is the true underlying (ie, latent) value of cannabis use for participant i at time point j. The parameters a_k and b_k capture the discrimination and difficulty of the kth measure, respectively. Equation 1 is of the generalized linear model form [64], and a different link function g_k(.) is used for each measure to accommodate a mix of measurement types (eg, continuous, binary, ordinal, etc). This approach provides a formal way to combine multiple measures of the same construct (see the Data Triangulationsection above). A similar model will be developed for each neurocognitive construct (eg, R/R and EF); here, we represent the underlying value of neurocognitive performance by λ_ij. While this approach is quite flexible, we shall also explore other measurement models, such as those with multiple underlying dimensions (eg, factor analytic models), mixture models with discrete groups (eg, latent class models), or linear or quadratic discriminant analysis.

Next, we will fit longitudinal, mixed-effects models to the underlying values of cannabis use and neurocognitive performance, both unadjusted and adjusted for sociodemographic and other relevant covariates x_i:

where t_j=(j-1) for j=1, 2, or 3. In Equation 2, α_i and β_1i are random effects representing the baseline value of cannabis use and its change over time for the ith participant. Likewise, γ_i and η_1i ith β₂α_iη_1iα_iγ_i represent the association between cannabis use, and β_1iη_1i represents the association between changes in cannabis use and neurocognitive performance over the 18-month study period. x_iθ_ijλ_ij⁴ may be fit simultaneously using software such as the gsem command in Stata.

The primary purpose of additional analyses is to determine the prospective association between cannabis use at baseline and PrEP care continuum outcomes over the study period, as well as to explore mediation by neurocognition. As described above, we measure these outcomes both objectively and using validated survey measures. Initial analyses model each outcome separately as a function of cannabis use, adjusting for sociodemographic covariates. Generalized linear models are used to accommodate varied measurement types (eg, continuous, binary, ordinal). Since we have multiple outcome measures, naive testing of each increases the likelihood of a type I error—a problem, which we address by combining these into a multivariate outcome model and performing a joint test of the association with cannabis use across outcomes. In addition, we examine whether associations among outcome measures are adequately described by a latent variable model (eg, factor analytic or latent class model). If so, examining the association between the underlying latent variable (measuring PrEP care) and cannabis use provides a simpler description and more powerful test of the association, as well as another way to address the issue of multiple testing. After estimating the overall association between cannabis use and PrEP care, we add measures of neurocognitive performance (RR and EF) to the models. Since several of these models are nonlinear (eg, logistic regression), we use the counterfactual approach to mediation analysis to obtain appropriate estimates of the direct and indirect effects (through neurocognitive performance) of cannabis use on PrEP care [65,66]. This approach is not only applicable to nonlinear models and to situations where there are multiple mediators, but also to cases where there is an interaction between the covariate and the mediator (as would be the case if, for example, those with heavy use and poor neurocognitive function were especially likely to have worse PrEP outcomes). Since these outcomes are determined using data extracted from electronic medical records, we expect to have little missing data except for participants who move during the period or are otherwise untraceable. If necessary, we use multiple imputation [67] to address potential bias due to missing data.

Additional analyses are also conducted to determine whether self-reported motivations for using cannabis modify associations between cannabis use, neurocognition, and HIV prevention outcomes. The analyses are conducted by introducing interaction terms between actual cannabis use and motivations for use to generalized linear models predicting PrEP care and HIV transmission behaviors. Two approaches are used. First, in cases where motivation is measured with a single variable, we add and test a single interaction between this variable and cannabis use; this provides the simplest and most powerful test for a moderating effect (assuming linearity). Second, we divide the sample into subgroups of participants with homogeneous motivations and fit a hierarchical model in which the association between cannabis use and PrEP care is allowed to vary across groups. This second approach is more flexible in accommodating multifaceted motivations, and the hierarchical model will help to provide some regularization to avoid overfitting.

Calculations are conducted across a range of scenarios to identify minimum effect sizes detectable with 80% power for binary and continuous outcomes for the primary aims of quantifying associations between cannabis use, neurocognitive outcomes (aim 1), and HIV care engagement (aim 2), and to identify modifiers (ie, interactions) of these associations (aim 3). Because a significantly larger sample size is required to detect statistically significant interaction effects unless the interaction is of very large magnitude, analyses involving effect modification will be considered exploratory. However, we present minimum detectable interaction effect sizes for our given sample size. The estimates presented represent minimum detectable effect sizes estimated under simple, conservative assumptions about total analytic sample size, correlation between repeated measures, and dropout. In practice, we will make use of all available data at all time points, so actual power may be higher. The planned recruitment of 280 participants at baseline and anticipated retention of 206 participants at 2 years (all receiving scans and neurocognitive testing) provide good power for estimating both the cross-sectional and longitudinal associations between cannabis use and neural response. For example, at baseline, 280 participants provide approximately 80% power to detect a correlation of only 0.17 between cannabis use and neurocognitive performance. Conservatively assuming that we will have estimates of change (ie, β_1i and η_1i) for only those 206 retained at 2 years, we still have approximately 80% power to detect a correlation between changes in cannabis use and neurocognitive performance of 0.19. These calculations assume 2-sided tests at the .05 significance level. As noted above, we expect to retain approximately 80% of the participants to the third time point. To determine whether nonrandom dropout affects our estimates of change in neurocognitive performance, we jointly model the change in neurocognitive performance and the likelihood of attrition [68]. This approach allows us to determine whether the rate of change in neurocognitive performance is associated with the likelihood of attrition and, if so, to adjust our estimate of the mean change accordingly.

For additional analyses looking at prospective associations between cannabis use at baseline and PrEP care continuum outcomes, the proposed sample size provides adequate power to detect small to moderately sized effects. For example, we have approximately 80% power to detect a correlation of only 0.17 between cannabis use and total PrEP time and an odds ratio of 0.66 for a binary outcome such as any PrEP use. We also have reasonable power to detect an indirect effect of cannabis use through its effect on neurocognitive performance. For example, assuming binary measures of cannabis use and PrEP use with an overall odds ratio of 0.9 and a continuous measure of neurocognitive performance correlated at 0.5 with cannabis use, 280 participants provide approximately 80% power to detect an indirect effect that is 37% of the overall effect on the scale of the linear predictor [69,70].

The power to detect interaction effects is less than that to detect main effects. We base this analysis on baseline values of cannabis use and motivations for use, helping to ensure that all 280 participants can be included. A common problem with performing these types of analyses in moderate sample sizes is that statistically significant interaction effects are overestimated (ie, the “winners curse”). The hierarchical model described above provides partial pooling across subgroups with common motivations, which helps to mitigate this issue [71].