Patterns and Predictors of Incident Return to HIV Care Among Traced, Disengaged Patients in Zambia: Analysis of a Prospective Cohort

Supplemental Digital Content is Available in the Text.


INTRODUCTION
Re-engagement in care is a critical but poorly understood step in the HIV care cascade globally. [1][2][3][4] Dynamic movement of patients in and out of care is prevalent, 2,4,5 making care interruptions part of the natural history of HIV treatment. 4,6 These interruptions put patients at risk of poor health outcomes [7][8][9] and the onward transmission of HIV. [10][11][12] They threaten achievement of the global 95-95-95 targets. 13 However, return to care is a positive patient behavior that has the potential to improve treatment outcomes. Especially as the burden of undiagnosed disease continues to diminish and time on treatment for the average patient increases, understanding how quickly disengaged patients return to care, what factors facilitate return, and ways to encourage more rapid return represents an important scientific agenda with a potentially significant magnitude of effect and public health relevance. 14 HIV policy, service delivery, and monitoring must recognize and account for dynamic patient movement in the HIV care cascade. 3,6,15,16 However, most extant literature focuses on the traditional, linear steps, including testing, linkage, antiretroviral therapy (ART) initiation, and viral suppression. 17 Much less is known about patient re-engagement after a careseeking absence. To date, the few return-to-care studies have been primarily retrospective and examined demographic and clinical characteristics alone. These studies suggest that between one-third and one-half of patients with a gap in care have a return visit, 3,18 whereas studies including patient tracing observe return ranging from 20% to 70%. [19][20][21] Factors associated with return in studies from East and Southern Africa include older age, 22 lower CD4 count, 19,22 female gender, 20,22 health facility outreach, 19,20,22 ART use, 19,22 and latent patient factors related to poverty and poor care quality. 21 Several additional factors were identified in a North American context, but not explored in studies conducted in African countries, including mental health concerns, secure housing, and substance use. 18 Several qualitative studies have explored patients' experiences, identifying factors such as reduced stigma and social support as important for care engagement, 21,[23][24][25] but few studies measure these factors quantitatively to examine their association with return. There is a lack of prospective analyses of re-engagement that assess the effect of a comprehensive set of potential patientrelated, clinical, and social influences on return.
To improve the understanding of re-engagement in HIV care and treatment in sub-Saharan Africa, our study prospectively identified incident return to HIV care and time to return among a representative sample of traced, lost to followup (LTFU) patients confirmed to be disengaged from care from 31 facilities across 4 provinces in Zambia. We conducted a risk factor analysis identifying predictors of return from a range of factors at the individual, social, and facility levels. This analysis can inform future research and intervention development through patient re-engagement risk stratification and hypothesis generation around reengagement support opportunities.

Study Background and Procedures
This analysis is nested within a larger study, "Better Information for Health in Zambia" (BetterInfo). 26,27 Better-Info enumerated all LTFU adult patients at 31 sampled study facilities who had at least one HIV care visit between August 1, 2013, and July 31, 2015. Patients were determined to be LTFU if they were .90 days late for their last scheduled appointment and had a subsequent unknown care status. Approximately 10% of LTFU patients were then randomly sampled for BetterInfo study tracing. As described elsewhere, 26,27 sampled patients were traced by a peer educator using paper medical record review, phone calls, and in-person visits to ascertain if the patient was (1) deceased, (2) alive and in-care, or (3) alive and out of care. All contacted, disengaged patients were verbally encouraged to return to care, and although not systematically applied, in some cases tracers accompanied the returning patient to the facility or met them for their return visit. On in-person patient contact, tracers obtained voluntary written informed consent and used tablet computers to administer a survey recording care status, demographic, social, behavioral, and household characteristics and reported barriers to care engagement. No medical care was administered during the tracing interaction. Better-Info study surveys were administered in Nyanja, Bemba, Tonga, or English based on patient preference. Our nested study then extracted approximately 2 and a half years of follow-up HIV visit data after the cohort closed using electronic medical records (EMR) linked through unique patient identifiers (see Figure 1, Supplemental Digital Content, http://links.lww.com/QAI/B562).

Study Population
Our analysis included all out-of-care patients identified through BetterInfo tracing who (1) confirmed that they did not have an HIV care visit since the last one identified in their medical record, (2) completed the study survey at the time of tracing, and (3) were interviewed after their estimated date of disengagement (.90 days from the last scheduled appointment based on the paper medical record review, as recorded in the study database).

Measurements
Our study outcome, return to care, was obtained from facility visit dates in the EMR follow-up data extraction. Potential predictors of return (see Figure 2, Supplemental Digital Content, http://links.lww.com/QAI/B562) including clinical characteristics at the time of LTFU (eg, CD4 count, time in HIV care, and facility type) and gaps in care of .90 days before the BetterInfo study-identified gap were gathered from the patient's EMR at LTFU. All demographic (eg, age and marital status), social (eg, HIV status disclosure and stigma), behavioral (eg, alcohol use and travel), and household (eg, wealth and violence tolerance) factors potentially predictive of return were taken from the patient survey administered by the tracer. Missing survey items were taken from the EMR if available (eg, age and marital status). Most potential predictors were measured using closed-ended yes/ no, multichoice, or Likert scale questions. To capture patient reasons for disengagement, changes needed to return, and return intentions, however, tracers asked the open-ended questions, "Why did you stop going to any clinic for HIV care?" and "What would have to happen for you to come back to care at any clinic?," listened to the response, and recorded tick marks in as many predefined response options as were consistent with the patient's reply. Predefined subcategories developed through previous research 28 included "structural" (eg, transport and work issues), "psychosocial" (eg, need encouragement, family, and disclosure issues), "clinic" (eg, poor care quality and wait too long), and "medical" (eg, felt well and too many pills) options, each of which had 4-13 detailed response options. The response category "other" captured responses that did not fit under the predefined options.

Potential Predictors of Return
Analysis of possible predictors of return to care was guided by an adapted social ecological conceptual framework 29 of incident patient return to HIV care developed using extant literature 2,[18][19][20][21][22][23][24][25]28,[30][31][32] and contextual knowledge (see Figure 2, Supplemental Digital Content, http://links.lww.com/ QAI/B562). To model potential predictor variables, we first assessed the distribution of categorical variables, excluding variables where $97% of responses were the same. We assessed the relationship between continuous variables and return (on the log odds scale) using the LOWESS plots. Time from enrolment to disengagement was dichotomized at 18 months based on the LOWESS plot. From our 18 stigma questions developed to be consistent with draft and final HPTN 071 33 stigma questions, we used exploratory and confirmatory factor analysis to identify 4 stigma subscales with adequate internal consistency: internalized (Cronbach a = 0.70), anticipated (Cronbach a = 0.87), experienced (Cronbach a = 0.72), and resilience (single question: "I confronted, challenged, or educated someone stigmatizing and/or discriminating against me"). Stigma subscale scores were summed from item responses. For internalized stigma, patient responses were dichotomized as low versus high at the median scale score. Experienced stigma was dichotomized as none versus any, and anticipated stigma was broken into approximate tertiles. Household wealth was estimated from ownership of 14 possible household items using the Demographic and Health Survey wealth index approach 34 and broken into tertiles. Household violence tolerance scores were summed, with one point for each positive response to the 2, yes/no questions previously used in HIV research in Zambia, "If someone in the household misuses money is it acceptable to beat him/her?" and "In my household if a wife comes home late without permission of the husband, she will be beaten." 35 Alcohol use was analyzed using the AUDIT-C 36 binge drinking question. For "patient reasons for disengagement" and "needs to return," participants were analyzed as "yes" for a particular subcategory of "reason for stopping" or "need to return" if $1 detailed response option was selected for that patient under the specified subcategory. Subcategories were not mutually exclusive. 28 We used descriptive statistics to assess missingness. If a participant was missing data on binge drinking but replied that they drank "$5-6 drinks on a typical day" on a separate AUDIT-C question, 36 their binge value was set to "yes." For stigma subscales, we imputed the mean of available subscale items for a missing subscale item if at least 2 subscale items were available. We used multiple imputation with chained equations and 10 imputed data sets to account for remaining missing predictor data in the multivariable model.

Disengaged Patient Characteristics
We described the disengaged study population by potential predictors of return and used Kaplan-Meier methods to estimate the cumulative incidence of and time to return.

Return to Care
We used Cox proportional hazards regression to estimate incident return to care. The time origin was the date of disengagement from care (90 days from the last appointment or 180 days from the last visit if the appointment date missing). The time scale was days since disengagement. Study entry was the date of in-person tracer contact, the point at which a patient was confirmed to be out of care. The event, incident return to care, is defined as the first HIV visit date of any type (ie, clinical, pharmacy, or laboratory) on or after the date of in-person tracing contact. Patients were censored at database closure.
We first examined the complete case, a univariate association of each potential predictor with return to care. The final multivariable model was informed by theory (see Figure 2, Supplemental Digital Content, http://links.lww.com/QAI/B562), including the following variables: gender, age, CD4 count at the last visit, time in HIV care, past care gaps, past facility outreach for return, facility type, mobility (having to travel for .1 month in the past year), and having a psychosocial reason for disengagement or psychosocial need to return. We additionally included factors with a univariate association significance of P , 0.05. We examined variance inflation factors to assess multicollinearity and examined Schoenfeld residuals and adjusted log-log plots for each covariate to assess the proportional hazards assumption.

Supplemental Analyses
To better understand disengaged patients, we descriptively compared LTFU patients successfully traced and determined to be out of care with those found to be in-care.
To better understand return within a shorter time period, we conducted a supplemental analysis for incident return to HIV care within one year of disengagement, following the same analytic approach outlined above. To support a smaller model (more appropriate for fewer outcomes), the final supplemental multivariable model included only sex, age, and variables with a univariate association significance of P , 0.05. Acknowledging the important role of theory in a risk factor analysis, we also ran a multivariable model for return by one year with the theory-driven variables described above as a sensitivity analysis.

Disengaged Patient Characteristics
There were 556 patients identified through tracing as disengaged and included in our study sample (see  Table 1). The first supplemental analysis showed that, compared with LTFU patients

Predictors of Return to Care
Univariate analyses indicate that disengaged, traced patients were significantly (P value ,0.05) more likely to return to care if they had been contacted more than 3 times by the facility after the past missed visits and if they had challenged, confronted, or educated someone stigmatizing them once in the past year. Patients were significantly less likely to return if they sought care from an urban health center or hospital, compared with a rural health center, or were from the richest wealth tertile ( Table 2).
Independent predictors of incident return to HIV care from the multivariable model with P values at or below 0.01 level included having had a previous gap in care (aHR: 1.95, 95% CI: 1.23 to 3.09) and the patient having challenged, educated, or confronted someone stigmatizing them once in the past year (aHR: 2.14, 95% CI: 1.25 to 3.65; more than once aHR: 0.65, 95% CI: 0.33 to 1.27). Patients were less likely to return to care if they sought care from an urban health center (aHR: 0.68, 95% CI: 0.48 to 0.96) or a hospital (aHR: 0.52, 95% CI: 0.33 to 0.82) compared with a rural health center (Table 2). Although the overall P value of the wealth tertile was 0.01, the hazard ratio estimates and confidence intervals did not show a consistent direction of association between increased wealth and return. (wealthiest aHR: 0.71, 95% CI: 0.47 to 1.08; middle tertile aHR: 1.27, 95% CI: 0.89 to 1.80) ( Table 2).

Supplemental Analysis: Predictors of Return by one Year
An estimated 51.4% (95% CI: 33.2 to 72.5) of participants returned by one-year postdisengagement. In the multivariable model built, based on significant predictors from univariate analyses, statistically significant (at the 0.01 level) independent predictors of incident return within 1 year of disengagement included being 45 years or older and having used herbal remedies in the past 6 months (see Table 2, Supplemental Digital Content, http://links.lww.com/QAI/ B562). Patients were less likely to return by 1 year if they reported a psychosocial or clinic-related reason for stopping care (see Table 2, Supplemental Digital Content, http://links. lww.com/QAI/B562). The sensitivity analysis using the theory-driven model showed consistent results for the age and psychosocial reasons variables and identified no other significant predictors of return. Estimate precision was poor in these models due to limited events.

DISCUSSION
With sufficient follow-up time, a high proportion of disengaged, traced patients, 73%, returned to care across 4 provinces in Zambia. More action needs to be taken, however, to hasten return. Among those patients returning to care, the median time spent disengaged was 19 months. Our data show that the rate of return is higher soon after disengagement. Earlier efforts to facilitate return may be more  22 However, more rapid return soon after disengagement may also indicate that patients who do not return quickly may require targeted support to come back to care.
Interventions to support patient resilience to stigma and to limit stigma in the social environment may facilitate increased re-engagement. Our data indicate that, compared with not confronting stigma at all, confronting stigma once in the past year facilitates re-engagement. This is consistent with existing literature on the relationship between coping, resilience, and improved health outcomes. 37,38 However, we do not see a traditional dose-response relationship because challenging stigmatizers multiple times does not further increase return. We theorize that the repeat confrontation of stigmatizers may represent a more hostile social environment or chronic stress, limiting any positive effect the ability to respond to a stigmatizer may bring. Research has shown that the effect of HIV stigma on health is worse in the context of low perceived community support 39 and that the pathways through which resilience to stigma operates in the context of chronic stress are complex. 40 Future re-engagement research should include stigma and resilience measures and test effectiveness of resilience interventions to improve return to care. 41,42 Despite tracing, 27% of disengaged patients did not return to care by the end of study follow-up. Our data suggest that disengaged patients from urban health centers and hospitals are at a higher risk than rural patients of remaining disengaged and may require targeted interventions. Greater likelihood of return among those at rural health centers may be consistent with the more personal relationship-based care often available in rural, compared with urban and tertiary care centers. Existing research supports the importance of health care worker-patient relationships in patient engagement. 43,44 In addition, urban versus rural patients may have different needs driving engagement. Past research has shown differences, for example, in which differentiated service delivery models for HIV treatment access are preferred between urban and rural patients. 45 More research is needed to understand the mechanisms underlying facility-level difference in re-engagement and how to best address them to support return.
The finding that previous care gaps predict incident re-engagement adds additional urgency to the need to conceptualize care engagement as a dynamic process 4,5,10,14 and the need for effective interventions to support continuity of care. Although complex factors are likely associated with both having a previous care gap and a patient's subsequent re-engagement, our findings suggest that investment in supporting patient return after one care gap may pay future re-engagement dividends. The greater than 2.5-fold increase in the hazard of return among disengaged patients who were repeatedly contacted by the clinic beyond the standard of care is consistent with this suggestion and other retention literature. 46,47 Together these results warrant further investigation into the mechanisms through which extended outreach may support return, such as relationship development, and outreach effectiveness evaluation.
Our analysis suggests that factors predictive of return by one-year postdisengagement are more proximal to the patient care experience than predictors within the full study period. This suggests that effective interventions early on may need to target different mechanisms than interventions for people who remain disengaged for a longer time. In addition to older age and the use of herbal remedies in the 6 months before the survey, independent predictors of return by one year included not reporting a clinic-related complaint (eg, poor quality of care, lack of respect, and spending too much time at the facility) or a psychosocial reason (eg, clinic attendance creating conflicts, risking disclosure, being told to stop by someone influential, depression, and forgetting or seeking alternative care) for stopping care. Although self-treatment with herbal remedies may indicate illness-driven care seeking, finding ways to reduce clinic and psychosocial barriers, such as improving patient clinic experiences 43,44,48 and engaging social support, 49,50 may be important to encourage return sooner after disengagement.

LIMITATIONS
Despite intensive tracing efforts, we were not able to obtain an updated vital or care status on 25% of the sampled patients. If disengaged patients not successfully traced are systematically different from those found, the estimates may be biased. By using EMR data to compare, patients we found were more likely to be from rural health centers and from provinces other than Lusaka, indicating that our estimates may over-represent rural experiences. These 2 groups were similar on other demographics (data not shown). Our study was only able to identify return among patients whose return care visit was documented using the same unique patient number in the 4 study provinces. It is possible that patients returned as a "new" patient under a new unique patient number or to a facility outside of the study area, potentially underestimating return. Patients in urban or tertiary care settings may have more health facility options due to higher facility density, which may make them more likely to have an undocumented return under a new patient number. Predictors were largely collected using survey responses, which are subject to self-report error, recall, and social desirability biases. As study observation began after disengagement, we assume that survey-measured predictors are time invariant in the interim. Because of the poor documentation of mortality in the EMR, we were unable to look at the competing risk of death.

CONCLUSIONS
The most appropriate models of HIV care engagement show dynamic engagement patterns that demand multifaceted flexibility and support for retention, as is true for many chronic diseases. 51,52 Return to care after disengagement is a critical yet under-researched step of the HIV care cascade. Our findings suggest that patients in urban and tertiary care settings may need additional return support and that efforts to improve patient resilience and outreach after any care gap may facilitate return. Other important re-engagement influences may include positive patient experience at the clinic, having a supportive psychosocial environment, not being in the wealthiest population tertile, and older age. Future reengagement research should include measures of these predictors to investigate their mechanisms of effect and evaluate their causal effect on return to care.