Study location

Neno district is a rural and impoverished district in south-west Malawi, with an estimated population of approximately 165 000 in 2017.14 Located along the border of Mozambique in a remote and mountainous region, Neno District Hospital is not accessible by tarmac roads. The majority of the population is subsistence farmers and 95% of households are without electricity.15 The socioeconomic status of Malawi is poor in general, where nationally the average income is US$1 per day.16 HIV prevalence in Malawi’s south-west zone was as high as 15% in 2010,15 the highest in the country; 2014 national statistics indicate a 10% average national prevalence.17 Malawi was listed by WHO as one of the 30 countries with high TB/HIV burden,18 and malnutrition rates (prevalence) in children under 5 years are around 4%.19 Maternal mortality in the country is also the highest in the world at 439 per 100 000 live births, and Neno district data show that only 57% of married women are on modern family planning.20 NCD care, while listed in the Essential Health Package in Malawi’s Health Sector Strategic Plan II, is extremely limited in the rural areas, and population data suggest that up to a third of Malawian adults suffer from hypertension and over 5% from diabetes.21

Study design: stepped-wedge, cluster-randomised trial

The implementation of the new CHW model is designed so that it may be evaluated as a stepped-wedge, cluster-randomised trial (SW-CRT) (figure 2). The stepped-wedge study design was selected for a number of reasons. First, the training of CHWs needs to be staggered due to training capacity constraints. Second, all sites in Neno will receive the intervention. And third, the stepped-wedge RCT design permits estimation of the causal effects of the intervention.22 23 We plan to conduct a long-term follow-up at 3 years postimplementation making use of routinely collected data to determine whether any effects observed in the primary analysis persist over time. Additionally, we plan to conduct an economic evaluation of the programme in the future.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Household model stepped-wedge roll-out. CHW, community health worker; SCHW, senior community health worker.  

Sample selection

Eleven intervention sites were clustered into six groups based on population size such that each group had a manageable number of CHWs to train. The order of implementation for these six sites was randomised by a third party. In the SW-CRT study design, each cluster crosses over from the control group to intervention group until all groups have received the intervention. The six clusters were created from the 11 sites based mostly on geographical location and on catchment area sizes (mean population 21 848, range 12 515 to 27 426). There are no exclusion criteria, as the study aims to assess whether CHWs assigned to more households and with a more integrated care system (rather than just HIV and TB), will be able to more effectively reach households in need of health support. Households can refuse the CHW’s service; however, based on our 10 years experience with CHWs working in Neno, this is a rare occurrence.

Recruitment, hiring and supervision of CHWs

Each CHW will serve an average of 30 households (range 20–40). Household is defined by the family unit and the head of household, encompassing all of the dependents that individual is providing for. This ratio was chosen accounting for household spacing and difficult terrain in Neno; the anticipated CHW workloads including monthly home visits and longitudinal support for approximately 12 patients with chronic illness and six pregnant women; expert advice from CHWs and CHW programme leaders; and the need to have CHWs in every village living in the communities they serve. In addition, a previous pilot study tested a ratio of one CHW to 20 households, and this study will use a higher ratio in order to test a less costly model. The ratio will vary slightly based on the geography and household spacing of given catchment areas. Each of Neno’s 138 villages will be assigned enough CHWs to meet this requirement. The CHW cadre in Neno will grow from 1000 CHWs to approximately 1200. Each CHW will receive a monthly stipend of Malawian Kwacha (MWK) 15 000 or approximately US$21. Additionally, all CHWs will be supervised at the village level, with one senior CHW (SCHW) assigned per village. For larger villages (>300 households), two or more SCHWs will be assigned. If the villages are small and proximate, one SCHW will oversee multiple villages. SCHWs will also be responsible for 15 households in addition to their supervision duties. After rolling out the Household Model, we estimate 185 SCHWs will be assigned. SCHWs will be paid a monthly stipend of MWK23 000 or approximately US$32.

Each community will be briefed about the Household Model programme and the CHW management team will call village heads and local community structures together to discuss the criteria for hiring CHWs and the number of new CHW hires needed in the area. The village heads and community structures will nominate twice as many candidates as the number required.

The CHW selection criteria are:

• Personal requirements:

  • Acceptance: approval by the community in which they will work.

  • Time: ability to perform tasks without sacrifice of other commitments, with an average commitment of 20 hours per week.

  • Literacy: ability to read data forms and educational resources.

  • Influence: ability to motivate others.

  • Communication skills: ability to effectively communicate pertinent health education messages to all household members.

  • Good health: ability to travel long distances on foot.

  • Compassion: ability to empathise, relate and support household members.

• Programme requirements:

  • A ratio of households to CHWs that is feasible given the number of household members, patients in the catchment area and expected accompaniments required in the health facility.

  • Supervision and mentorship close to the community to ensure CHW performance.

Following nomination, candidates will be invited to take a written exam to assess reading and writing skills, as well as basic health knowledge. The top candidates who pass this exam are invited to attend the CHW training.

SCHWs will be selected based on the following criteria:

• Typically, they are previously employed as a CHW.

• Living within the same village or community where they will work.

• Mature, respected and have a history of confidentiality.

• Demonstrated leadership and/or problem-solving skills during the CHW training.

• Ability to influence others.

• Demonstrated above-average understanding of the curriculum.

• Strong data aggregation skills.

• Should be someone that the other CHWs will respect and work well with.

• Should be flexible and adaptable to changes.

• Preference will be given to those with stronger English skills after meeting the above criteria.

Training CHWs

All CHWs in Neno will be reassigned and trained in the Household Model in a staggered roll-out over 2 years. The maximum number of trainees per group is capped at 60 participants, with some trainings occurring with two groups of CHWs per catchment area.

CHWs will receive a 5-day foundational training. The purpose of the foundational training is to orient CHWs to the following: the new programme and its objectives; their cross-cutting roles; their specific tasks for each primary health focal area; the data collection and documentation requirements; and the knowledge, skills and attitudes needed to successfully perform the aforementioned tasks. The training is structured around the core home visits, which is the key point of interaction between CHWs and their clients.

Evaluation of CHWs

Along with a rigorous evaluation plan for the recruitment of CHWs and SCHWs, trainings and daily performance of the CHWs will be evaluated throughout the roll-out and continued existence of the Household Model.

CHWs will be evaluated on health knowledge through pretests and post-tests at each training session. A pass mark for each test will be set using the Ebel method of standard setting. A panel of 8–12 PIH and MOH leaders will be established for this process, with these experts required to have a good knowledge of the Household Model, as well as an understanding of the knowledge a CHW needs to be successful in his or her role. The group will collaboratively agree on a definition of the knowledge expected of a ‘minimally competent’ CHW which will be used as the reference point for determining the pass mark. Furthermore, CHWs will be continuously evaluated based on training attendance, as well as specific knowledge and skills.

Supervision meetings will be held regularly to ensure timely communication between different levels of staff is maintained. The following supervision meetings are core to the Household Model: CHW management meetings; site supervisor meetings; SCHW meetings; general meetings; and village-level CHW meetings. Additionally, SCHWs and site supervisors will conduct monthly supervision home visits, and monthly spot checks will be performed in order to: assess the performance of CHWs and SCHWs; identify areas of support; and identify gaps where the site supervisor or SCHW mentor can provide more clarity or support to CHWs.

Community sensitisation and mobilisation

PIH’s dedicated community education and sensitisation team, the Community Programme (CP) team, will be responsible for working with the CHW and clinical teams to keep the community leaders, stakeholders and community members informed and updated about the Household Model. This will encompass education about the role of CHWs in the community, specifically as they shift from an HIV/TB focus to broader household health. These educational sessions will also include the objectives and aims of the model, the timeline, and any changes necessary as the CHW and clinical teams make adjustments along the way. CP will also be a key recipient of information from the community and its leaders, be it challenges, questions for clarification, concerns or successes.

Patient and public involvement

The development of the Household Model design was rooted in the needs of the population in Neno, based around: the burden of disease; health outcomes data; and feedback from existing patient recipients of the CHW programme. The original shift to household assignments, piloted in two catchment areas, was co-designed by clinical and CP staff in consultation with community members, patients with HIV and community leaders in order to tackle maternal health issues. Focus group interviews with these stakeholders were held to refine the programme details. This was piloted over 18 months, with lessons learnt incorporated into the design of the Household Model.

Patients were involved as community members in the focus groups in the original design, with community forums and community leaders consulted in the programme expansion into the eventual Household Model.

The launch of the Household Model was conducted through multiple forums including the District Executive Council, the health facility management staff, and in public forums through community leaders. The eventual results of this trial will be distributed in the same manner.

The burden of the intervention for the randomisation was not assessed by patients, as this is a cluster randomised trial, with randomisation at health facility level, not by patient level. All clusters will eventually receive the intervention. Stepwise scale-up was essential because of logistical and training needs, so the order of implementation was randomised.

Outcomes to be measured

Tables 1 and 2 describe the primary and secondary outcomes to be measured at the cluster level by this study.

MOH data for family planning, malnutrition, TB, antenatal care and maternity are already collected from every health centre and hospital through monthly, standardised reports. These reports summarise aggregated data from standard MOH registers. Registers are populated with information captured from individual patients’ health passports, where clinical providers document health details from their facility visit including diagnoses, prescriptions and follow-up instructions. Reports are entered for each facility into the MOH District Health Information Software II (DHIS2). All data are available on a monthly basis, except for TB data which are captured during quarterly supervisory visits.

The PIH electronic medical record (EMR) captures patient-level data for clients enrolled in HIV and/or NCD care, which includes hypertension, asthma, diabetes, epilepsy and mental health. Data are captured from standardised MOH master cards for each patient. Aggregate data by health facility from the EMR are produced in automated reports generated by the PIH staff.

For each health facility, data on the outcomes will be extracted from the DHIS2 and EMR for the period from October 2016 to July 2019.

To our knowledge, there are currently no other ongoing interventions or planned interventions that we expect will occur at the district level that would systematically compromise our results. PIH is the only non-governmental organization (NGO) working with the MOH across all district health centres and supporting the MOH in all of the disease focus areas described here. There are currently plans for one NGO to implement volunteer-based community sputum collection points across five catchment areas that is funded by the Global Fund. Additionally, in three non-study catchment areas along with one Household Model intervention site, the PIH plans to implement a 2-year randomised controlled trial to test models that promote household usage of treatment for potable water. This intervention will not address any conditions to be measured in the Household Model; however, it risks task oversaturation for CHWs involved in the water treatment study. Over time, we will track any new interventions and we will also measure unexpected influences like natural disasters, national stock-outs, or fuel shortages during the intervention that may impact our results.

Additionally, we will track key programme changes that may affect the primary outcomes. We expect an overall increase in enrolment given the plans to strengthen and add routine, integrated screening at outpatient departments (SHARF) and continued screening in the community (SHARC). Thus, as this CHW Household Model intervention is rolled out, the size of the cohorts with chronic conditions—namely HIV, TB, hypertension and diabetes—is anticipated to concurrently increase. Furthermore, routine health activities in the district will continue, such as periodical mass screening campaigns for nutritional disorders. There are some parts of the district where CHWs have previously been instrumental in these periodic campaigns, therefore we will continue to document these efforts so that they can be considered in the analysis. Though nutrition is the most extreme example, CHWs have been long-standing members of the healthcare team in Neno, periodically working with MOH employees to encourage family planning, antenatal care, HIV testing and TB screening, though never in a systematic, structured or longitudinal way. Additionally, the study staff will keep track of other forces that often influence trends in a health system without significant resiliency. For instance, over the past few years Neno district has seen floods, drought, famine, fires and a typhoid epidemic, with corresponding health system responses.

Household qualitative assessment

In addition to the quantitative data collection described above, we will conduct short semistructured qualitative interviews with a purposively selected sample of around 40 CHW programme recipients. The interviews will be undertaken in November 2017, aiming for a group of respondents with diversity in key characteristics, including socioeconomic status, household location, sex of the household head and length of participation in the programme (eg, families who started receiving CHW support before the roll-out of the Household Model, and families who first received CHW support more recently). Interviews will be conducted until saturation has been reached: in other words, new interview participants will be sought within each cluster until the interview responses are producing no new major themes or ideas. The online supplementary file provides the draft interview guide and informed consent script to be used during recruitment. Transcripts will be analysed for emerging themes using grounded-theory methods, and a codebook will be created to assist with data analysis. Depending on these findings, the team may undertake repeat interviews (with different participants) at the end of the roll-out to determine consistency across catchment areas.

Data will be collected using field workers familiar with the study area. All raw data will be kept in a locked cupboard of a locked office, and all transcribed and translated data will be kept in password-protected files.

Data analysis plan

Data will be extracted from the DHIS2 and EMR records at the cluster level. Our analysis will follow standard guidelines for the analysis of stepped-wedge, culster randomized control trials-,22 which specify a model at the cluster level for each time period. The implementation order is randomised, and the model captures the difference between clusters and trends over time, which allows us to estimate the treatment effect of the intervention. Each cluster contains between one and three health facilities. As such, we specify a model for the cluster-time cell means. Given the outcome is restricted to lie between 0 and 1, we specify a generalised linear model. We will consider different link functions: probit, logit, linear and a β-regression model. To select among these link functions, we will use the Watanabe-Akaike Information Criterion and graphical posterior predictive mode checks, comparing the observed data to samples from the posterior predictive distribution. To allow for extreme 0 and 1 observations in the β-regression model we will use the transformation Missing \left or extra \right . 

The model for the primary analysis, for cluster j at time t is

Where y_jt is the outcome, D_jt an indicator for whether cluster j had the intervention at time t, θ_t is the time fixed effect, α_j is a normally-distributed cluster random effect with variance \sigma^2_\alpha and ƒ(.) is a link function. The model parameters will be estimated for all outcomes. An expert elicitation workshop was conducted to quantify beliefs about treatment effects on the primary outcomes. The elicited distributions will serve as both the prior distribution for the treatment effect parameters and in the power calculation (see below). Weakly informative prior distributions will be specified for the other variables on the basis of previous studies.24 Standard deviations will be assigned a half-t(4) distribution. The remaining parameters will be assigned N (0, 5² ) priors.

Sensitivity analyses will be conducted with respect to the choices of prior distribution. An uninformative N(0,10²) prior distribution will be used for the treatment effect, and half-normal (0, 3) distributions for the within-cluster and between-cluster variance.

Power calculation

We conducted a Bayesian assurance analysis—analogous to a frequentist power calculation.25 26 We determined the probability that there will be at least a 95% posterior probability that the treatment effect parameter is >0(or <0, where appropriate). We conducted a series of simulations27 of trial data to simulate observations from the posterior distribution of the treatment effect parameter. The assurance is calculated on the basis of a design and analysis stage:25 26 for the design stage we assumed weakly informative N (0, 5² ) priors on the treatment effect parameter and we simulated data using parameter distributions for the primary outcomes elicited from experts. For all the outcomes, except HIV/ART retention in care, there was >95% probability  that there will be at least 95% posterior probability that the treatment effect parameter is greater than zero. Only a small negative change was expected in HIV/ART retention in care, we estimated that there is a 75% probability that there will be at least 95% posterior probability that the treatment effect parameter is less than zero.