2.2.2. Demographics

Demographic variables were collected at baseline and used in the analysis as follows: age (in years; continuous), race (Black/White; binary), ethnicity (Hispanic/non-Hispanic; binary), and gender (male/female; binary). Education was measured as a categorical variable with these options: middle school or less, some high school/no diploma, high school diploma/GED, junior college, technical/trade/vocational school, some college, college graduate, or graduate/professional school. Categories of race, ethnicity, gender, and educa-

tion were established in the primary outcomes paper for CTN-0049 [60]. Southern/nonsouthern residence was a binary variable determined by the study site location [61]. Sites in Atlanta, Baltimore, Birmingham, Dallas, and Miami were considered southern sites. Sites in Boston, Chicago, Los Angeles, New York, Philadelphia, and Pittsburg were considered non-southern sites.

#### 2.2.3. Psychiatric History

Participants were classified as having a psychiatric history if either of two criteria were met: (1) an initial hospital intake (at time of enrollment) with a primary diagnosis and/or any comorbid diagnoses that included terms or conditions such as "suicidal ideation", "psychosis", "schizophrenia", "bipolar disorder", "PTSD", "hallucinations", "mood disorder", and "altered mental state," or (2) participant self-report that they "saw a professional for the primary purpose of getting help for psychological or emotional issues in the past 6 months".

#### 2.2.4. Barriers to Care

An analysis of barriers to care was guided by a socioecological framework described by Mugavero et al. (2013) to examine engagemen<sup>t</sup> in HIV care across multiple levels of healthcare access [62]. Building upon earlier models of healthcare utilization, this framework categorizes healthcare utilization factors into four categories: (1) Individual factors, which may include demographics, personal health beliefs, past experiences, and coping skills; (2) Relationship factors, which may include connections with family, friends, and medical providers; (3) Community/health system factors, which may include community-level poverty, social norms, and the local health service infrastructure; and (4) Policy factors, which may include treatment guidelines, service coordination, and funding. This study specifically examined 23 barriers to care at the first two levels (individual and relationship factors) and health system factors at the third level. Addressing community factors and policy-level barriers was beyond the scope of this research. All measures were assessed at baseline.


"Have you ever been in a relationship where a sexual partner threw, broke, or punched things?" Participants who answered affirmatively to any of the items were scored a 1. Others were scored a 0.


measuring support over the last 4 weeks [72]. Each item was scored on a scale from 1 to 5, for a total score range from 5 to 25. Lower scores represented lower social support. A sample item is, "How often was someone to love and make you feel wanted available to you during the past 4 weeks if you needed it?" (1 = none of the time, 2 = a little of the time, 3 = some of the time, 4 = most of the time, 5 = all of the time).


#### *2.3. Statistical Analyses*

First, a latent profile analysis (LPA) was conducted to identify subgroups of individuals with similar barriers to care. LPA is a latent variable modeling technique that identifies unobserved subgroups of individuals within a population based on responses to a set of observed variables; it assumes that individuals can be categorized by patterns of responses that relate to profiles of personal and/or environmental attributes [76]. LPA, rather than a Latent Class Analysis, was used in this analysis, as it can accommodate both categorical and continuous indicators [77].

The current study included the 23 barriers to care previously described. Profile solutions were evaluated based on several standard fit indices, including Akaike information criteria (AIC), adjusted Bayesian criteria (BIC), model entropy, Lo–Mendel–Rubin test, and the bootstrapped likelihood-ratio test. Additionally, the clinical meaningfulness, interpretability, and sample size of each class were considered in the selection of the final model. Latent profile plots were created to visualize differences between the profiles. Differences

in latent profiles by gender, race, and southern/non-southern residence were assessed using a likelihood-ratio test with a significance level of α = 0.05.

Next, structural models were constructed to test how the relationship between the intervention groups (PN, PN+CM, and TAU) and the four distal outcomes of interest differed by profile. Model construction followed a 3-step approach [66,67]. In step 1, LPA was performed; age, gender, southern/non-southern residence, and treatment group were included as covariates using the auxiliary option in Mplus. In step 2, a new latent profile variable was created by incorporating the classification error obtained from the step 1 logits for classification probabilities. This classification method is preferred over other methods such as classify–analyze or pseudo-class draw approaches because it accounts for uncertainty in latent profile assignment and reduces bias [78]. In step 3, the distal outcome was regressed on the intervention variables, controlling for the covariates and comparing effects across the latent profiles. This process was repeated for each outcome. Odds ratios were used to interpret the effect of latent profile on each outcome.

## **3. Results**

#### *3.1. Characteristics of the Study Population*

Select demographic, clinical, psychosocial, and healthcare access factors of the 801 study participants are summarized in Table 1. The sample was mostly male (67.4%), Black (82.5%), and Non-Hispanic (89.0%) with a mean age of 44.2 years. There were slightly more participants enrolled from southern sites (59.2%) than from northern sites. The average time since HIV diagnosis was 11.8 years. Most participants reported a history of being in HIV care (82.9%) and being on antiretroviral therapy (77.2%) at some point in their lives, but approximately two-thirds of the sample had a CD4 count of less than 200 cells/μL at enrollment. About one-third of participants reported injection drug use in the last 12 months, and 55.3% had a history of substance use treatment in the 6 months prior to enrollment. Approximately 22.0% of the study sample had a recorded psychiatric history. The overall baseline mean of psychological stress as measured by the BSI-18 was 22.5 (16.1 SD). Based on established BSI thresholds, there were 39 individuals with minor elevation, 17 with moderate elevation, and 11 with marked elevation [79].

**Table 1.** Characteristics of the CTN-0049 study sample (*n* = 801).



**Table 1.** *Cont.*

Range, mean (std dev) shown for continuous variables; *n* (%) shown for dichotomous variables. Abbreviations: BSI = Brief Symptom Inventory, AUDIT = Alcohol Use Disorders Identification Test, DAST = Drug Abuse Severity Test.

Many participants had achieved at least a high school education (60.2%), but most (77.4%) reported an annual income less than \$10,000, and only 11.6% were employed. Most individuals reported unreliable transportation (90.3%), not having a case manager (70.0%), and low levels of social support (mean = 14.7 out of 25). Many participants, however, had health insurance (67.4%) and moderate to high levels of health literacy (mean = 9.0 out of 12). There were no differences in the distribution of baseline characteristics across treatment groups, which was expected due to randomized treatment assignment. The reliabilities of measurement scales are shown in Table 2. All scales had a Cronbach alpha > 0.70, indicating adequate reliability.

**Table 2.** Reliability of continuous scales used to measure barriers to care.


Abbreviations: BSI = Brief Symptom Inventory, AUDIT = Alcohol Use Disorders Identification Test, DAST = Drug Abuse Severity Test.

#### *3.2. LPA Results*

Models of two to five profiles were considered for the LPA. The five-profile solution was ruled out because the best likelihood value could not be replicated after 2000 random starts. Among the remaining models, multiple fit statistics (Table 3) and interpretability indicated that a three-profile solution best fit the data. The sample-size adjusted BIC score (69,367.41) was lower in the three-profile solution than the two-profile solution (indicating a better fit), while maintaining a high entropy (0.863). The Lo–Mendell–Rubin adjusted likelihood-ratio test, however, showed that the four-profile solution did not significantly improve fit above the three-profile solution (*p* = 0.166). The three-profile solution also presented a logical substantive interpretation, adequate class distinction, and adequate sample sizes. Therefore, the three-profile solution was selected as the best model.

**Table 3.** Latent Profile Enumeration using 23 indicators of barriers to care.


Abbreviations: AIC = Akaike information criteria, aBIC = adjusted Bayesian information criteria, LMR-A = Lo–Mendell–Rubin adjusted likelihood-ratio test, BLRT = bootstrapped likelihood-ratio test.

A comparison of the three profiles is described in Table 4 and displayed in Figure 1. Standardized means are shown for continuous variables, and proportions of item endorsement are shown for dichotomous variables. The first profile had relatively low barriers to care. Values for all barriers were the lowest for this profile except for lack of case management, low income, and not having insurance. This profile comprised half of study participants (50.3%) and was labeled "Lower Barriers (LB)." The second profile, which described 35.7% of the study sample, generally exhibited higher barriers to care compared to the first profile and was characterized by having a higher probability of reporting a history of abuse (67.3%) and intimate partner violence (65.6%). This profile was labeled "Higher Barriers with Abuse and Violence (HB-AV)." The third profile, which comprised 14.0% of the study sample, was quite close to the second, with similar values across most of the barriers. The main distinguishing features of this profile were an even higher likelihood of having a history of abuse (74.8%) and intimate partner violence (65.6%) and a high likelihood of having experienced discrimination (std mean = 5.41). This profile was labeled "Higher Barriers with Discrimination, Abuse and Violence (HB-DAV)." This three-profile solution was further analyzed for differences by key demographic characteristics including gender, race, and southern/non-southern residence (see Supplementary Materials.).

**Figure 1.** Visualization of Three-Class Latent Profile Analysis solution using 23 indicators of barriers to care.


**Table 4.** Standard means and proportions of continuous and categorical indicators by profile.

#### *3.3. Structural Model Results*

Estimates for the final three-profile LPA are shown in Table 5. After controlling for race, gender, and southern/non-southern residence, structural models indicated that there were significant effects of the PN and PN+CM interventions on being engaged in care at 6 and 12 months and viral suppression at 6 months. However, these associations were only observed for certain profiles. The greatest effects were seen for the Lower Barrier (LB) profile, where the PN+CM group was associated with higher likelihood of being in care at 6 months (β = 1.37, OR = 3.94, *p* < 0.001), being virally suppressed at 6 months (β = 0.687, OR = 1.99, *p* = 0.15), and being in care at 12 months (β = 0.881, OR = 2.41, *p* = 0.019), compared to the TAU group. The PN-only group also had a significant effect on viral suppression at 6 months (β = 0.610, OR = 1.85, *p* = 0.035) and a marginally significant effect on being in care at 6 months (β = 0.660, OR = 1.93, *p* = 0.054), compared to the TAU group. The Higher Barriers with Abuse and Violence (HB-AV) profile had higher odds of being engaged in care for both the PN+CM group (β = 1.25, OR = 3.49, *p* = 0.001) and the PN group (β = 0.981, OR = 2.67, *p* = 0.018) compared to the TAU group, but there were no significant associations with the other distal outcomes of interest. The interventions did not have any significant effects for those with the Higher Barriers with Discrimination, Abuse, and Violence (HB-DAV) profile. Additionally, there were no significant intervention effects on viral suppression at 12 months for any of the latent profiles.


**Table 5.** Effect of patient navigation interventions on engagemen<sup>t</sup> in care and HIV viral suppression by latent profile.
