Study design and setting

We used data from the ‘Cohort of African people Starting Antiretroviral therapy' (CASA) project, a prospective, ongoing, multi-site study of a cohort of HIV-infected patients who started ART in seven urban and rural HFs located in Tigray, the northeastern region of Ethiopia. The HFs participating in the study include two health centres (Alamata Health Centre, Mekelle Health Centre), one university hospital (Ayder Referral Hospital), three general hospitals (Alamata General Hospital, Mekelle General Hospital and Lemlem Karl General Hospital of Michew) and one primary hospital (Mehoni Primary Hospital). The ART units in the health centres are run by two or three nurses, the primary hospital ART unit has three or four nurses and in some cases one health worker with a master's degree in infectious diseases, and the general hospital ART unit and the referral hospital ART unit have three or four nurses and one medical doctor.

Participants

HIV-infected adults who fulfilled the WHO criteria for starting ART at the participating HFs were eligible for enrolment in our cohort. All individuals over 14 years of age who had initiated ART between 1 January 2013 and 31 December 2014 and who agreed to provide their home address and a telephone contact number were selected for this study. Follow-up data were available until December 2015.

Ethical considerations

Ethics approval was obtained from the Health Research Ethics Review Committee of Mekelle University College of Health Science (reference number: ERC 0129/2012). All patients provided written informed consent. For patients aged 14–18, the informed consent form was signed by adult relatives acting as guardians (immediate families, e.g. father or mother or next of kin) and not by the patients themselves.

Outcome measures

The operational definitions of the outcomes used in this study are as follows:

• Loss to follow-up: patients who missed visits to the same HF for more than 3 months after the last scheduled visit.

• Stop ART medication: patients known to have stopped ART for any reason (personal or clinical decision).

• Mortality: patients recorded as dead on the patient's exit form.

• Transfer out: patients formally transferred to another HF. Transferred-out patients were not considered as retained in care because official documentation to certify that they were still on ART was not available. Transferred patients were considered as retained in the initial HF until the date they were transferred out.

• Retention in care: patients who were alive and receiving ART at the same HF after ART initiation (this does not include patients who were recorded as lost to follow-up (LTFU), who discontinued ART, who were deceased or who transferred out).

• Attrition from care: this is the opposite of retention and included patients who were LTFU, had discontinued ART, were recorded as deceased or had transferred out.

Independent variables

The following demographic and clinical variables at enrolment were considered as potential predictors of attrition: gender, age, religion, educational status, WHO clinical stage (from stage I to stage IV), haemoglobin level, body mass index (BMI) calculated as weight (kg)/height (m)² (underweight ≤18.5 , normal 18.6–25, overweight >25), initial ART regimen, CD4 cell counts (within 6 months of starting HIV treatment) and the presence of active tuberculosis (TB) infection with Mycobacterium tuberculosis where the diagnosis had been made by sputum smear and chest x-ray.

Data source

The baseline visit and follow-up visits were carried out according to the national guidelines and the standard of care of the participating HFs. Data, including demographic and clinical information as well as the patient's current status (known to be dead, transferred-out, discontinued or LTFU) were systematically collected by ART nurses using forms specifically designed for this study (enrolment form, follow-up form, exit form).

Each HF used desktop software developed in the .Net framework (Microsoft) especially for this study to enter the demographic, clinical and laboratory data of participating patients. Data were first entered by a case manager (CM) from each site and then merged and managed in a central database. It was important for data completeness, reliability and validity to standardise the data collection procedures for all participating sites. Staff members responsible for collecting and entering study data received appropriate training and data monitoring was conducted throughout the study. The project implementation office in Ayder Hospital together with the project coordinating centre (Istituto Superiore di Sanità) in Italy were responsible for oversight of data quality and generating queries (missing data, inconsistencies in the data, potential errors) for each participating HF. The project implementation office staff were responsible for conducting regular site visits to assist local personnel with the queries and ensure procedures were followed appropriately.

Tracing of patients who failed to return for a scheduled visit

Routine HIV/AIDS care is provided in the participating HFs according to national guidelines. Support for PLWH is regularly provided by local community-based organisations and community health workers (CHWs). In the present study, the active participation of CHWs was enhanced to trace patients as described below. At each participating HF, patient data were entered by a study CM. At the beginning of each week (Monday), the CM prepared a list containing the details of patients who had missed their scheduled visit for more than 1 month. These lists were then given to trained CHWs who traced the patients using telephone interviews, face-to-face contacts or proxy information (family members or friends). Traced patients were asked why they had missed the scheduled visit. CHWs were trained to provide basic information on HIV and ART and on the importance of taking drugs regularly without interruption. Patients were asked to return to the HF and resume ART. At the end of each week (usually on Friday), CHWs returned the list to CMs together with the following information gathered during the week1: whether or not the patient had been traced,2 how the patient had been traced (by phone, face to face, through other people),3 the current vital status (alive or death) of the traced patient,4 patient intent to return to the HF, and the reasons for his/her decision.

Statistical methods

The primary outcome measure was retention in care 2 years after ART initiation. The Kaplan–Meier method was used to estimate the probability of retention in care, mortality, loss to follow-up, transfer out, and stopping ART medication at different time points. Univariate and multivariate Cox proportional hazards models with robust sandwich estimates, to account for within-HF correlation, were used to identify factors associated with attrition. All available demographic and clinical variables were considered as potential predictive factors and were included in the univariate analysis: type of HF, gender, age (14–25, 26–50, >50 years), educational status (no education, primary, secondary, tertiary), religion (Orthodox Christian or other religion), BMI (≤18.5, 18.6–25, >25), WHO clinical stage (I/II or III/IV), CD4 cell count (<200 or ≥200 cells/µL), haemoglobin level (≤10 or >10 g/dL), active TB, and initial treatment regimen (efavirenz-based or nevirapine-based). Predictor variables which had statistical significance in the univariate analysis (p<0.2) were included in the multivariate analysis. An interaction term between CD4 cell count and haemoglobin level was also added in the multivariate model. Finally, variables with p<0.05 (statistically significant) were identified as predictors of attrition in the final model. Follow-up of patients who stopped ART medication, were LTFU or had transferred out was censored at the date of the last visit to the HF. Time to death was censored at the date of recorded death. Follow-up of retained patients was censored at the end of the study period (31 December 2015). A sensitivity analysis to estimate each transferred patient's probability of being retained in care was also performed using a multiple imputation technique. Statistical analyses were performed using SPSS software, version 21.0 (SPSS, Chicago, IL) and the SAS statistical package, version 9.2 (SAS Institute, Cary, NC).