Applied mathematical modelling to inform national malaria policies, strategies and operations in Tanzania

The concept of using mathematical modelling for strategic planning of infectious disease control is not new [1‐3]. Multiple examples exist for a wide range of infectious diseases [4‐8] and specifically for malaria [2, 9‐12]. Mathematical modelling uses available information to generate data-driven simulations of transmission dynamics and control for specified settings [2, 9, 13, 14]. The model predictions can quantify with some uncertainty the expectations of the impact of interventions for different areas. The exploration of alternative scenarios aids in decision-making and facilitates a more strategic approach in the selection of interventions [15‐18]. More than ever, it is crucial to make the best use of existing country data and analytical tools [19] because: (1) there is an increasing complexity with the expanding available malaria control tools as a result of effective research and development, (2) the local epidemiology is becoming more heterogeneous as a result of massive ongoing control efforts, and (3) resources, especially funding, are not increasing. Hence, global and local policymakers, as well as funders, increasingly recognize the value of mathematical modelling as a strategic tool to support decision-making [1, 12] (Table 1). In addition, growing stakeholder coordination and the need to use evidence will lead to more strategic questions about priorities and combination of interventions. In parallel, more and better quality data become available in endemic settings, enhancing the value of modelling [2].

These developments clearly call for a more sustainable and in-depth relationship between modellers, NMCP managers and donors. Given the historical difficulty of linking modelling and strategic planning, intensified technical support, closer interactions and capacity building within-country NMCPs are required. This case study presents such close collaboration between modellers, donors and the NMCP managers, providing a unique and effective example of modelling for strategic planning.

Mathematical models have been applied for various countries at varying resolutions, examples include sub-Saharan African (SSA) countries [15, 20‐22], Ghana [23, 24], Kenya [25, 26], Mozambique [27, 28], Nigeria [16, 17, 29], Uganda [30], South Africa [31], Zambia [32‐35], and the Asia–Pacific Region [36, 37]. In those examples, modelling was used to investigate relevant transmission dynamics, intervention effectiveness or for stratification. While sometimes useful for global policy writing, there have been fewer examples where mathematical modelling has been applied in a country at the required operational unit and accompanied with a national policy dialogue. Exceptions are Zambia [35], Ghana [24], South Africa [31], Cambodia and Thailand (Mahidol Oxford Research Unit (MORU)), Mozambique, Sri Lanka, Phillippines, Benin (Swiss Tropical and Public Health (Swiss TPH)). In Kenya, Tanzania and Uganda, a decision support tool has been developed in communication with local stakeholders, to link research and policy for “guiding the selection of more effective, evidence-based control strategies” [14, 38]; however, no country-wide application could be found.

In 2016, a team of modellers from Swiss TPH were invited by the Global Fund to Fight AIDS Tuberculosis and Malaria (GFATM) to provide support to the Tanzanian NMCP for preparing the upcoming funding request [39]. After this initial undertaking ended in early 2017, the NMCP and the Swiss TPH team suggested to continue modelling which then could be made an intrinsic part of the on-going planning processes of the NMCP. The sections below describe the non-technical process of applying mathematical modelling, its added value, challenges and lessons learned. The development of the modelling approach is described in [40] and the results of modelling application are included in the Supplementary Midterm Malaria Strategic Plan 2018–2020 [41].

The Swiss TPH has a long-established relationship with the NMCP in Tanzania. In 2002, the Swiss Agency for Development and Cooperation (SDC) launched the NETCELL project to provide technical and strategic support to the NMCP, with the Swiss TPH as implementing partner [42]. Since its launch, NETCELL contributed to the strengthening of the NMCPs capacities to plan, coordinate, and implement malaria control interventions, in particular insecticide-treated bed nets (ITNs) [43, 44]. The NETCELL team collaborates with the Ministry of Health, Community Development, Gender, Elderly and Children (MoHCDGEC), the President’s Office, Regional Administration and Local Government (PO-RALG), UK Department for International Development (DfID), United States Agency for International Development (USAID), Worldbank, GFATM, among others [42]. The Swiss TPH modelling team closely worked with the NETCELL team, which in turn facilitated the interactions between the modellers and the NMCP programme members. The NETCELL project has recently been renewed under the financing of the SDC and has many more years to provide continuous support to the MoHCDGEC. Another important regional partner was the KEMRI-Wellcome Trust Programme, who managed DFID funded projects (INFORM and LINK) [45] to provide spatial epidemiological analytical support using nationally available malaria data for subnational decision making in Tanzania and other NMCPs across Africa [46, 47].

Strategic planning in Tanzania is based on a strong malaria monitoring and surveillance system, including high-quality district health information system (DHIS2) data [48], entomological surveillance [49], resistance monitoring [50], demographic and health surveys, and malaria indicator surveys [51‐55]. Since 2014, nationwide annual school malaria parasitaemia surveys also bring high-quality and high-resolution cross-sectional data to the NMCP database [56]. The Tanzanian epidemiological data show nowadays a highly heterogeneous malaria transmission and burden throughout the country [51‐56]. The National Malaria Strategic Plan (NMSP) for 2015–2020 acknowledged that diversity of malaria transmission and disease burden within the borders of Mainland Tanzania, but largely adopted a uniform approach to disease management and prevention nationwide [57]. An increase in national average prevalence from 9.5 to 14.8% between 2012 and 2015–16 [52, 53], led to the questions of whether the current NMSP would technically be feasible to achieve the national target, of a prevalence of less than one per cent in 2020. In line with this question arose the issue of optimizing intervention mixes according to endemicity and key epidemiological parameters. As a result, a decision was made by the NMCP to work on a supplementary malaria midterm strategic plan aiming at optimal intervention mixes in different epidemiological strata to ensure optimal impact for available resources [41]. A timeline describing the events leading to the supplementary malaria midterm strategic plan is shown in Fig. 1.

In 2016, two workshops were held in Dar es Salaam to introduce the concepts of modelling, to assess available data sources and data owners, and to discuss input parameters and model assumptions. Following these two workshops, an extended phase was used for the model calibration. In 2017, the results of the initial models were fed into a midterm-review of the strategy, which concluded that the national prevalence target of less than one per cent by 2020 would not be achievable. Indeed, the modelling results suggested that the current NMSP objective could not be achieved unless a much more aggressive intervention mix was put in place. Unfortunately, that was neither operationally feasible nor financially doable. At a malaria expert meeting held in February 2018, with national and international stakeholders, the modelling results were presented alongside with the empirical view of the NMCP on the country context. At that meeting, it was decided to (1) gather all available data for risk stratification at council level, and (2) put together a more detailed plan for improved targeting of interventions at council level.

In May 2018, the modelling team was invited alongside NMCP staff and the NETCELL team to join a strategic planning workshop. During that meeting, the NMCP and stakeholders stratified the councils according to malaria risk (Thawer et al. pers. commun.) and discussed the allocation of appropriate interventions targeted to the strata. The previously calibrated transmission model (using OpenMalaria) was then used interactively during the work session by simulating requested alternative scenarios and directly answering questions from the country programme. Finally, selected outputs of the model were included as an additional set of evidence in the revision of the strategic plan launched in February 2019 [41].

Mathematical modelling allowed primarily a technical assessment of the national malaria targets. Once calibrated, predictions of the likely impact of current and potential future interventions at council level could be provided. Beyond the simulation results, the process in itself was useful to inform policy. Modelling did not only use and process quantitative data, but also expert opinions, programme experiences, and local knowledge. Together, these created a platform for an in-depth interdisciplinary dialogue. Presenting model assumptions and the comparisons of the predicted versus expected impact triggered controversial as well as constructive comments. Controversial or unexpected predictions led to a critical review of the data, model structure, assumptions made, as well as the planned intervention scenarios. The ongoing engagement between modellers and practitioners enabled knowledge transfer and established a long-term interest in modelling. The former one was demonstrated by a developed critical but more appreciative view which replaced an initial misconception about modelling (i.e. “why to use modelling when you have data” changed to “why is the model different from the data, and how would the predictions change if…”). The interaction and close collaboration were also of great benefit to the modellers, as the local knowledge and data were invaluable for model improvements leading to more context-specific modelling.

Moreover, statistical modelling and traditional descriptive analyses were performed to describe temporal and spatial trends based on empirical data and not on dynamics of malaria transmission as the mathematical model used has. Indeed, dynamic transmission model use available data to inform parameters to simulate malaria transmission and burden based on an understanding of the transmission dynamics, while statistical models only infer relationships based on collected data, without necessarily understanding the system. Discussions with partners on data for input parameters and major model assumptions were highly relevant to understand and inform the main drivers of malaria transmission. As a direct illustration, the prevalence predictions from the geospatial model provided by KEMRI-WT were discussed between partners including the NMCP and decided to be used to calibrate the transmission model for council prevalence.

A number of challenges affected the accuracy of the modelling outputs and timeliness of the project. First, there was no previous experience for country modelling available at that level of detail that could have guided the process and the type of required outputs. Second, methodological challenges led to extended times for model calibration and complicated uncertainty estimates around the predictions [40]. Uncertainty resides in model predictions. This uncertainty can be due to data quality and accuracy for model parameters, or due to model structure and random variability. The advantage of a simulation model would be to assess the impact of this uncertainty on the predictions. However, given the fact that this framework is representing each council of the entire country, the computational power becomes challenging. As a result, assessment of uncertainty was kept to its minimal, only accounting for random variability by using multiple runs for the historical simulation period, and accounting for uncertainty in transmission intensity by fitting a range of transmission intensities to prevalence estimates. Third, gaps in communication and understanding slowed down the process, requiring much more frequent and in-depth engagement between the modelling team and NMCP staff than had been anticipated. Fourth, challenges also included the busy schedule of the NMCP staff, as well as tight deadlines expected by external donors. Moreover, building capacity within the NMCP without a dedicated modelling person within the NMCP or at least within a local institution was challenging and the NETCELL advisory team was invaluble to bridge that gap. However, in order to sustain the modelling support on the long term, the analytical capacities within the NMCP need further strengthening i.e. through additional personnel with quantitative skills, training and increasing experience as the modelling application continues. The first phase of the project has been to set up a framework and ensure engagement with and usefulness for the programme, the second phase will be to transfer knowledge by training in-country modellers. Lastly, it took time to build trust between all partners, to be able to understand the strengths and limitations of the models. The main key challenges and their implications are summarised in Table 2.

Once modelling activities were understood and adopted by the NMCP (and not perceived only as an academic exercise), the modelling process was used systematically as a way to think about the data. Furthermore, model strengths and limitations became better understood by the NMCP and partners, making the entire effort more productive. Ultimately, the whole process fed into the strategic planning process through interactive presentations and discussions. This exchange allowed for an additional layer of thoughts and interpretation and was found to be essential for the model to be meaningful and appropriate at the end Additional file 2. To achieve this, multiple interactions, workshops and demonstration of the model were required. The NETCELL team made up of technical experts understanding both programme constraints, and the basics of modelling facilitated the communication by ‘translating’ between technical language to programmatic language. The NETCELL team also ensured continuity in the process, especially in-between visits by the modelling team. Their country-specific knowledge and resources were invaluable for many aspects of the modelling. A summary of the critical elements for success identified throughout the process is provided in Table 3.