The PATHS comparison cohort (hereinafter referred to as the “comparison cohort”) constitutes the comparison group for the study. We operate under the initial assumption that the representative sample will have minimal or moderate exposure to field medicine services, making it a reasonable comparison group.

Recruitment for PATHS is conducted on a rolling basis each year during data collection for the LA County Homeless Count Demographic Survey, a representative geographic sample survey conducted as part of LA County’s point-in-time homeless count. The survey is conducted from December to March each year in approximately 550 of LA County’s 2310 census tracts, with systematic canvassing of each tract. Unsheltered adults who complete the demographic survey and have a valid mobile phone number are invited to participate in PATHS. Additional recruitment has also been conducted through dedicated canvassing exercises in hotspots of unsheltered homelessness between April and July.

For those interested in participating in PATHS, surveyors either collect their phone number through a secure web form or hand out a study card with a self-enrollment option. Subsequently, prospective participants are directed to a website with study information, self-administered consent procedures, and a screening survey to determine eligibility. Eligibility requires (1) spending at least one night in the past month in a homeless shelter or unsheltered setting; (2) residence in LA County; and (3) being at least 18 years of age. Eligible participants receive a follow-up text inviting them to complete a baseline survey the next day.

On completion of the baseline survey, they are sent a survey link each month via text or email. Participants receive an electronic gift card for each survey they complete, including US $10 for the prescreen survey, US $20 for the baseline survey, US $20 for each monthly survey, and US $20 for the 6-month follow-up (ie, biannual baseline update) surveys. All surveys are conducted on a secure mobile website accessible through a phone or computer and are currently available in English or Spanish.

To define the comparison cohort for this evaluation, we include participants who had completed a baseline or 6-month follow-up survey between August 2024 and October 2025 (the period when the field medicine cohort was recruited and enrolled) and completed a monthly survey within 2 months afterward.

The field medicine cohort was recruited through a large and representative set of field medicine providers working in LA County. To enhance the comparability of the field medicine cohort to the comparison cohort, we have replicated the recruitment methodology as much as possible. Recruitment was carried out by public health workers and case managers working with field medicine teams as well as dedicated study recruitment staff. Building on existing protocols for PATHS recruitment, the evaluation team worked with practitioners and intake staff at these providers to maximize recruitment while avoiding any disruption to service delivery or clinician rapport. Prospective participants were approached for recruitment only after they had already received field medicine services for that visit, so that study participation was not contingent on, nor perceived as a condition of, receiving care.

As in the PATHS sample, prospective participants access the prescreening questionnaire by providing their phone number or receiving a study card. A survey was developed in the Qualtrics survey software to collect phone numbers and the field medicine provider name to link a prospective participant with their enrolling provider. To ensure ease of enrollment even in settings with low wireless bandwidth, we also introduced methods to enter phone numbers of prospective participants via text message. For those enrolling via a study card, the participant can be identified with their provider based on the unique codes embedded in the study cards. Once prospective participants enter the prescreen survey, all subsequent methods for prescreen, baseline, and monthly surveys, as well as compensation schemes, are identical to those used for the PATHS comparison cohort. If a participant already captured in the existing PATHS sample is additionally enrolled via a field medicine provider (as identified by a duplicate phone number), they are migrated into the field medicine cohort.

Baseline analysis will compare the field medicine cohort to the comparison cohort on the predisposing, enabling, and need characteristics identified earlier, along with demographic and geographic location controls. When comparing the field medicine vs comparison cohorts, we will report descriptive statistics as frequencies and column percentages for categorical variables and means and SDs for continuous variables. To test unadjusted associations between reported characteristics and field medicine participation, we will use the Pearson chi-square test or Fisher exact test for categorical variables where at least one expected cell count in the cross table is less than 5 and a Kruskal-Wallis rank sum test for continuous variables. To test for adjusted baseline differences, we will test logistic regression models in which the dependent variable identifies the study arm (ie, field medicine versus comparison cohort) and demographic, predisposing, enabling, and need factors are treated as independent variables. As the PATHS survey includes a “prefer not to answer” option for all questions, we anticipate a relatively high level of missingness and, thus, will conduct all analyses using first a complete case analysis followed by a complete sample using multiple imputation methods. We presume that analyses will primarily use imputed data, while complete case analysis will serve as a robustness check. Results of baseline analyses will be used for publications and policy reports that provide critical evidence on the composition of field medicine patients, populations or areas with unmet access needs for field medicine services, and potential health conditions requiring further treatment in field medicine programs. Baseline analyses will also identify the extent of baseline imbalance in the evaluation sample, thus informing the potential need for matching or other quasi-inferential strategies under aims 2 and 3.

The second aim will report on patterns of field medicine service usage, knowledge, and perceptions across the 2 study cohorts, concentrating primarily on the field medicine cohort. The first focus will be on usage of field medicine services, including monthly ever-use and frequency of use across multiple months. The analysis will also explore types of services used and quantify the use of multiple field medicine providers, both for general descriptive assessment and for potentially informing heterogeneity analysis. This analysis will inform the impact evaluation in aim 3 in a variety of ways. First, we will quantify the extent of field medicine usage in the comparison cohort, quantifying potential contamination in the design and informing alternative uptake-based designs. Second, we will explore the frequency of service use, which will be used for descriptive purposes and for the construction of dose variables indicating frequency of field medicine contact. We will characterize the extent to which patients enrolled through a particular field medicine provider also obtained services from other field medicine providers. Finally, we will measure patient satisfaction with the services received each month. In all cases, these data will provide helpful policy-relevant descriptions of usage profiles and patient engagement. We will also conduct scientific analysis of variations in patient retention, usage, and progression of intensity of service engagement.

The final aim of the study is to evaluate the impact of field medicine services on patient health outcomes; access to health, social, and housing services; and trust in the medical profession in comparison to the comparison cohort. Building on the realist evaluation approach, we will use a variety of ways of identifying “exposure” to field medicine services as the treatment measure. The primary models will use an intent-to-treat approach, comparing individuals enrolled in the field medicine cohort to those in the comparison cohort. Alternate design will instead use a usage-based independent variable, including ever-use or frequency or cadence of use of field medicine services, thereby accounting for the fact that there may be crossover between the study arms, with some members of the comparison cohort using field medicine services more heavily than treated individuals. Additional models will disaggregate overall field medicine treatment effects into effects associated with specific providers or types of services. The analysis will include a mix of outcome measures captured either monthly or at specific 6-month (biannual) intervals (January and July of each calendar year). We describe separate statistical estimation strategies for each.

Two-sample, 2-tailed t tests will be used to determine whether the mean change in each outcome from baseline to follow-up is significantly different between the field medicine and comparison cohorts. Mixed effects models for repeated measures will assess whether the field medicine cohort shows statistically significant improvements in health care engagement, health status, and access to services at each biannual follow-up compared to the comparison cohort. All models will adjust for potential baseline differences in primary outcomes across study groups.

An advantage of repeated-measures analysis is that participants with baseline data who are missing at follow-up can still be included in the analysis. To account for the rolling recruitment of participants over a 12-month period, we will adjust models for the enrollment month. Period fixed effects will also be included to address any changes in the policy environment over the follow-up period. All analyses will be conducted in Stata software (version 19; StataCorp LLC) and R software (version 4.1.2; R Foundation for Statistical Computing), with appropriate weighting to account for our sample selection strategy. We will assess clustering of variance within sample clusters by examining the intraclass correlation and, if necessary, correct SEs for multivariate results using appropriate clustering correction methods.

Monthly measures may display considerable variation within person as well as distinct trajectories for different population groups. For these models, we will use hierarchical generalized linear models (HGLMs) that can be used to estimate a program effect in the presence of correlated errors. HGLM is well suited for modeling intraindividual changes in outcomes and interindividual differences in these changes, as well as for handling unbalanced data structures and missing within-person data. A series of 2-level HGLMs, in which monthly observations (level 1) are nested within individuals (level 2), will be used to examine the effects of field medicine on temporal changes in monthly health measures within individuals over time, controlling for any time-varying covariates and time-invariant baseline covariates (ie, sociodemographic characteristics, duration of homelessness, and coping strategies). We will include fixed and random effects for the intercept and slope for time to examine the mean initial status and rate of change, individual deviations from these means, and whether the field medicine intervention impacts initial status and rate of change.

Level 1 (within-person level):

Yit=β0i+β1iTimeit+β2i(TVCit−TVC¯i)+eit

Level 2 (between-person level):

β0i=γ00+γ01MED+γ02(TVC¯i)+γ03TICi+r0iβ1i=γ10+γ11MED+r1iβ2i=γ20

where i is the index of persons, t is the index of time per month, MED is intervention assignment, TVCi¯ represents the mean value of time-varying covariates for an individual across all time points (ie, the person-specific mean), and(TVCij−TVCi¯) represents the time-specific deviation of the variable from the person-specific mean. β_0i and β_1i represent the intercept (ie, initial status of the outcome) and slope (ie, rate of change) for individual i, γ₀₀ and γ₁₀ are the overall mean intercept and slope, and r_0i and r_1i are the individual-specific deviations from the mean intercept and slope (random effects).

To improve our confidence in the inference emerging from our analyses, we will use matching and weighting methods to account for key threats to validity, including (1) differential selection into field medicine services based on prior health and other factors, which would limit our ability to compare field medicine patients and nonpatients; and (2) selective attrition from the cohort over time, which may also potentially differ between field medicine patients and nonpatients.

Our primary approach for addressing selection on observables will be matching. We will match field medicine cohort members to comparison cohort members on demographic covariates (eg, age, gender identity, and race or ethnicity), geographic location at recruitment, baseline health status (eg, chronic physical disease, serious mental illness, substance use disorder, and patient health activation), and service access (eg, receiving government benefits and housing systems engagement). We will assess covariate balance before and after matching with standard diagnostic approaches (eg, visual plots, standardized mean differences, and variance ratios); outcome analyses described earlier (biannual and monthly models) will then be conducted within the matched sample. As a secondary and sensitivity analysis, we will construct inverse probability of treatment weights using the same set of covariates, estimated via a propensity score model predicting field medicine cohort membership. This approach retains the full analytic sample rather than restricting to matched pairs. Consistency across matching- and weighting-based approaches will strengthen our confidence in the robustness of findings; discrepancies will be reported and interpreted as a limitation of the nonrandomized design.

We note that because the field medicine cohort is enrolled directly into the existing PATHS platform and follows identical recruitment, consent, and survey procedures as the comparison cohort, both cohorts share the same requirement of phone access and willingness to complete a phone-based survey. This shared enrollment pipeline reduces differential selection on these dimensions between the 2 study arms. We discuss the implications of phone-based enrollment for the generalizability of the study population as a whole in the Discussion section.

Given the preliminary nature of this study, we may find that study retention is an important outcome in and of itself, due to the potential for differential attrition across cohorts. The importance of these selection-adjusted models will depend in large part on the extent of observable baseline differences between the cohorts. Regardless of the balance achieved on observables, our nonequivalent control group design may also introduce potential unobserved variation in the samples. Depending on the extent of this concern, we will use geographic data on the service coverage areas of field medicine providers to exogenously identify selection into field medicine service access. Additionally, we will conduct sensitivity analyses for unmeasured confounding, including tipping point analyses (E-values and robustness values) and benchmarking against observed covariates within an omitted-variable-bias framework, to quantify how strong an unobserved confounder would need to be to alter our substantive conclusions [28,29].

For the qualitative component of the study, we conducted qualitative interviews and participant observation with 9 field medicine teams. These teams are affiliated with 3 distinct field medicine programs. Program directors facilitated recruitment by sharing the contact information of team leads who were likely to be interested in participating. IDMN then contacted those leads directly to invite participation. Most teams operate on the west side of LA County, although some fieldwork has also been conducted in the southern region of the county.

IDMN directly recruited field medicine providers, including physicians, physician assistants, nurse practitioners, nurses, and community health workers (CHWs), for semistructured interviews during participant observation activities. Field medicine patients and unhoused individuals located within the service areas of participating teams but not currently receiving care were recruited for semistructured interviews using 2 recruitment strategies. First, field medicine providers directly referred potential participants—both patients and nonpatients—after inviting them to participate in the study during routine outreach. Providers were encouraged to ensure demographic variation across gender, race, and ethnicity. Once a potential participant agreed to participate in the interview, the field medicine provider introduced IDMN, who would conduct the interview, and then stepped away to allow the interview to proceed privately. If IDMN was not present in the field when a potential participant agreed to participate, the field medicine provider scheduled the interview for a date and time convenient for the participant, accompanied IDMN to the scheduled appointment, introduced her to the participant, and likewise stepped away before the interview began. Second, IDMN conducted in-person recruitment by accompanying previously shadowed teams during street outreach and personally inviting patients and nonpatients to participate in the study. When potential participants agreed to participate, IDMN conducted the interview immediately following recruitment.

Participant observation was used to examine the everyday practices of field medicine teams and to gain a deeper understanding of how care is delivered in situ. The primary focus is on the processes through which patients are identified and recruited, the nature of provider-patient interactions, and how clinical assessments and service offerings unfold in real time. Attention is given to how teams address both medical and social needs, navigate barriers to continuity of care (ie, scheduling follow-up visits and making specialty referrals), and coordinate with other service providers. We also closely observe how trust is built between providers and unhoused patients.

Semistructured interviews with patients focused on their experiences with field medicine, including how they initially connected with providers, barriers to accessing care regularly, services received, unmet needs, aspects of care they valued, and suggestions for improvement. Interviews with nonpatients explored their current health care needs, patterns of health care usage, knowledge of field medicine services, and perceived barriers to accessing field medicine. Providers were asked about perceived barriers and facilitators to delivering health care on the streets and ensuring continuity of care, strategies for navigating those challenges, the scope of medical and social services they offer, and gaps in care delivery. Interviews ranged from 30 to 90 minutes and were conducted in person for the unhoused subsample and either in person or by phone for providers. All interviews were audio recorded with participant consent. Unhoused participants received a US $40 gift card as a token of appreciation for their time.

Semistructured interview transcripts and field notes were imported into NVivo (version 15; Lumivero) for analysis. We used a thematic analysis [30], with codes developed both deductively—based on the study’s core questions (ie, experiences with field medicine, access barriers and facilitators, continuity of care, satisfaction, and unmet needs)—and inductively, through close coding of the interview transcripts and field notes to capture emergent themes. IDMN and a graduate research assistant coded the same 2 interview transcripts separately and then engaged in an iterative process of refining the codes—merging, expanding, and collapsing them—until they reached consensus on the final codebook. Once the codebook was finalized, the remaining interview transcripts were coded independently by IDMN and the graduate research assistant using NVivo software. The same software was used by IDMN and 2 undergraduate research assistants to code the field notes.

We use both methodological and data source triangulation to enhance the depth of our qualitative findings. Methodologically, we integrate data from semistructured interviews and participant observation. Interviews capture participants’ perspectives, experiences, and perceptions related to field medicine, while participant observation provides insight into real-time provider-patient interactions, care delivery processes, and the broader context in which services were delivered. Observational data will help complement, clarify, or complicate interview accounts, allowing us to identify gaps between reported and enacted practices and uncover patterns that are not always explicitly described by participants. In addition, we use data source triangulation by including multiple participant groups—field medicine patients, nonpatients, and providers—which allows us to examine the same phenomena from diverse viewpoints. This combined triangulation strategy strengthens our analysis and will lead to a more comprehensive understanding of the implementation and lived experience of field medicine care.

This study was approved by the UCLA South General Institutional Review Board and UCLA North General Institutional Review Board (protocol IRB21-001148 and IRB-22‐1985, respectively).

Participants in both the field medicine and comparison cohorts complete self-administered informed consent prior to screening, as described earlier.

For field observations, IDMN obtained oral informed consent directly from field medicine providers before observations and from patients before observing their clinical encounters. For semistructured interviews, IDMN obtained oral informed consent directly from each participant before the interview.

Survey participants are assigned a unique study identifier upon screening, independent of personally identifiable information (PII), to deidentify their study data. Although participant contact information (eg, phone numbers and email addresses) is collected during recruitment and enrollment, this information is used only to distribute subsequent study surveys, maintain contact with participants during the study period, and distribute financial incentives. Contact information is stored only on the secure data collection platform and is stripped from survey responses prior to transfer to the secure, encrypted server used for data storage and analysis. Only study team members with authorized access permissions have access to PII. As survey responses are deidentified from contact information prior to analysis, participant confidentiality is maintained during dissemination of study findings in papers, presentations, and reports.

Surveys are self-administered and completed remotely on participants’ own schedules via a secure mobile website; participants are not required to be in any particular physical location to complete them. As part of our trauma-informed survey design, to reduce distress and response burden while answering monthly surveys, all questions in the surveys include a “prefer not to answer” option, allowing participants to decline to respond to any question without penalty or exclusion from the study.

Deidentified quantitative survey data will be made available in a public data repository (eg, the Inter-university Consortium for Political and Social Research) following study completion, consistent with data sharing requirements of the study’s funders. Only deidentified data, stripped of PII as described earlier, will be shared. The publicly shared dataset will not include geographic identifiers. Geographic data will be governed by a restricted data use policy and made available only upon request through a data use agreement with the study team.

Between December 2021 and October 2025, PATHS screened 3004 respondents, of whom 1836 (61%) were eligible and completed a baseline survey. Among these, 1413 (77%) proceeded to complete at least one monthly survey and thus had been officially enrolled in the study since study launch. For this evaluation study, the comparison cohort will include 902 participants who had completed a baseline or 6-month follow-up survey between August 2024 and October 2025 (the period when the field medicine cohort was recruited and enrolled) and completed a monthly survey within 2 months afterward.

Recruitment of the field medicine cohort began in August 2024 and concluded in October 2025. During this period, we partnered with 6 field medicine providers that operate in LA County, including Healthcare in Action, Venice Family Clinic, UCLA Health Homeless Healthcare Collaborative, Clínica Romero, Bartz-Altadonna Community Health Center, and the Los Angeles County Department of Health Services. Patients were recruited through their field medicine services operations, which included mobile clinics, backpack medicine teams, static sites at parks or community centers, encampment-based care, and shelter-based or transitional housing clinics.

As of October 2025, 847 field medicine patients had been referred from our field medicine partners and screened for eligibility, with 749 (88%) meeting eligibility criteria. Of these, 542 (72%) completed the baseline survey, and 436 (80%) participants completed a subsequent monthly survey, officially enrolling in the study.

Recruitment efforts were strengthened by the development of strong, collaborative relationships with field medicine teams, whose support was instrumental in engaging potential participants and facilitating enrollment across a wide range of outreach settings. These partnerships enabled flexible, real-time coordination and helped build trust with patients receiving care in nontraditional environments.

At the same time, the study team encountered numerous operational challenges that required ongoing adaptation. Participant turnout at street-based sites was inherently unpredictable and influenced by a range of contextual factors, such as weather, safety concerns, and competing priorities faced by patients. Recruitment was further impacted by 2 significant period effects occurring within LA County: widespread wildfires in January 2025 and a wave of roaming US Immigration and Customs Enforcement (ICE) raids targeting public spaces, which were followed by civil unrest in June 2025. These events intensified law enforcement presence and fostered a climate of fear, unease, and instability in the unhoused population.

In response, the team implemented a range of adaptive strategies to maintain study continuity and protect all involved. These included shifting outreach to safer locations, modifying schedules in real time, implementing a buddy system for data collection, and strengthening communication protocols. Identifying safer locations involved monitoring local independent news sources to identify areas of LA County with ICE activity and allowing teams to determine potential areas to avoid. It was through consistent monitoring of these various sources that, for example, we became aware of an ICE raid at MacArthur Park, which disrupted scheduled field medicine services on July 7, 2025. Daily schedule adjustments were communicated through work phones and a designated team group chat to ensure timely coordination. We implemented a buddy system for data collection as well, so that together they could use situational awareness to assess the different sites we visited. Throughout this period, we also maintained open communication with service providers and partners, following their lead in identifying safety risks and adapting to site-specific procedures as needed. During law enforcement events, all in-person recruitment activities were paused; any participant present at the time was informed of the situation and prioritized for immediate safety. Recruitment activities only resumed once participants had moved to a safe location alongside staff and remained willing to continue. To further minimize risk to participants, the study does not collect any information on immigration status, and survey data are stored without directly identifying information (see Data Privacy and Distress Safeguards section).

Between October 2022 and October 2024, more than 300 patient-clinician interactions were observed during a total of 82 field visits, resulting in approximately 500 hours of participant observation. IDMN joined field medicine teams at the beginning of their shifts and accompanied them throughout their outreach rounds—typically 6 hours—by riding along and observing their interactions with patients, colleagues, and partner organizations. Team composition varied: smaller teams typically included a health care provider (eg, physician, physician assistant, nurse, or nurse practitioner) and a CHW, while larger teams might also include a medical assistant and security officer. Teams were diverse in terms of age, gender, race, and ethnicity.

Participant observation took place at various locations, including medical vans, sidewalks, parking lots, beaches, parks, and interim sites (eg, shelters, board and care facilities, and motels). During each shift, IDMN took notes on a mobile phone to capture key moments, dialogues, and reflections. That same day, these notes were expanded into detailed, narrative-style field notes that captured the flow of the day and contextualized the observations.

Between May 2024 and October 2024, IDMN interviewed 15 field medicine providers—including physicians, physician assistants, nurse practitioners, nurses, and CHWs—as well as 23 field medicine patients and 12 unhoused individuals located within participating service areas but not currently receiving care. The final sample size was determined by thematic saturation rather than a predetermined numerical target, with recruitment concluding once additional interviews no longer generated novel themes or perspectives [32].