Sub-national stratification of malaria risk in mainland Tanzania: a simplified assembly of survey and routine data

Since 2000, there has been an unprecedented increase in funding to support the coverage of malaria interventions across Africa [1]. This renewed commitment translated into a reduction in the prevalence of malaria infection and disease burden in many parts of Africa [1, 2]. However, in recent years, progress has stalled [1, 2]. Ten countries in Africa currently account for 66% of the global malaria disease burden [3], despite increases in the distribution of effective vector control and disease management strategies. Further increases in international donor assistance are unlikely and a new model of improving investment efficiencies is required to maximize the benefits of interventions in areas likely to achieve the largest disease burden reductions. The World Health Organization (WHO) Global Technical Strategy (GTS) for malaria 2016–2030 revisited an old paradigm of stratifying sub-national malaria burden based on the analysis of past and contemporary malaria data, risk factors and the environment [4]. A major pillar of the GTS 2016–2030 is the use of accurate and timely routine data for tracking the changes in malaria epidemiology.

Since the launch of the WHO “T3” (Test, Treat, Track) initiative in 2012 [5], many African countries have increased testing rates at health facilities and are now able to provide data on malaria parasitological diagnosis performed through microscopy or malaria rapid diagnostic testing (RDT) [6]. Furthermore, countries have initiated efforts to improve their Health Management Information System (HMIS) system using the open source web-based software known as the District Health Information Software (DHIS2). Adoption of this software in many countries has facilitated the availability and access to routine malaria parasitological diagnosis data generated from health facilities which has strengthened the utilization of such data for malaria risk mapping and evaluations of intervention programmes.

Since the 1960s, the epidemiology of malaria in mainland Tanzania has been mainly described through the length of the malaria transmission seasons, urbanization, altitude and community-based parasite prevalence [7‐9]. All have highlighted the extreme diversity in the potential, and empirically defined malaria transmission intensity, within the country’s borders. A more recent assembly of 10 years of community- and school-survey parasite prevalence data was used within a model-based geospatial framework to empirically highlight the heterogeneous nature of sub-national malaria transmission intensity [10‐12], and used to describe the country’s epidemiological profile in the 2015–2020 National Malaria Strategic Plan (NMSP) [7]. However, these statistical models of opportunistic research data, or under-powered national household sample health surveys, provide only one means to define variations in malaria prevalence. To-date, other data, notably those generated from routine health information systems, have been underutilized and the use of epidemiological evidence to tailor sub-national malaria intervention strategies has been limited. These approaches should be data-driven, using all available routine and survey information and the stratification should be country-led [3, 13].

Since the launch of the Roll Back Malaria (RBM) initiative in 1998, the National Malaria Control Programme (NMCP) of mainland Tanzania has developed 3, 5-year NMSPs [7‐9]. The third NMSP covered the period 2015–2020 [7] and aimed to reduce the national malaria prevalence from 10% in 2012 to 5% in 2017 and further to less than 1% by 2020. The initial ambition of the strategy was to sustain progress and achievements through a universal coverage of existing interventions; and during the second phase (2018 to 2020), to consolidate these achievements and explore the feasibility of a malaria pre-elimination in defined areas of the country [7].

Although progress was made towards reducing national parasite prevalence from 18% in 2008 [14] to 7% in 2017 [15], a mid-term review (MTR) in 2017 [16] recognized that a more strategic allocation of limited resources was needed to ensure continued progress in the future. The MTR was followed by a consultative meeting with global and national malaria experts [17, 18]. Recommendations from this forum together in concert with the GTS 2016–2020 [4], reiterated the need to consider tailoring intervention approaches to the sub-national local context, based on epidemiological stratification. To establish epidemiological strata at operational units of programme delivery (councils), a data-driven approach was required, that maximizes the use of survey and routine data. This paper provides an outline of the methods used to assemble infection prevalence and other malaria indicators from routine data to develop a sub-national epidemiological stratification for mainland Tanzania’s 184 councils. This paper presents the first documentation of a national effort to combine multiple epidemiological indicators from different data sources to form a composite risk stratification. The process of policy development [19] and the allocation of interventions [18] following development of this malaria risk stratification are presented elsewhere.

This paper presents a novel approach to stratify malaria at sub-national level in mainland Tanzania, using a combination of routine malaria indicators from health facilities and school surveys. The resulting map stratified the burden into four epidemiological risk strata; very low, low, moderate and high plus one non-epidemiological stratum for urban councils. This was used to guide the malaria control programme in revising its malaria strategic plan in an evidence-based manner and in developing targeted intervention packages per strata [18].

There are many indicators of malaria risk that can represent sub-national heterogeneity. The precision and bias of each indicator, associated costs for collection and the level and frequency available to measure variability across space and time can affect the suitability of indicators to measure transmission [34]. Several studies have attempted to compare measures from routine sources against community prevalence to highlight the representativeness of these indicators [25, 27, 35]. However, evidence to suggest which indicator is most suitable to measure transmission is limited and a further understanding of how these vary across different transmission settings would help identify which indicators are most sensitive to council-level transmission strata and how these change over time.

While there are several approaches to malaria risk stratification that have been developed, there is no one specific approach recommended by the WHO. A review that looked at malaria risk maps developed during pre-GTS, across 47 countries [36] found that most countries rely on either API or infection rates for describing the malaria risks although a range of other indicators have also been used such as qualitative descriptions and climatic suitability. The current methodology presents a pragmatic approach that levers data from routine reporting and national survey data. Not limiting the stratification to only one data source enhances the best use of all available data, and the credibility/robustness of the resulting stratification. Importantly, through a detailed interrogation of routine data, it is possible to make reasoned council indicators to align with other survey data sources for sub-national level stratification, harnessing data from those that seek care at facilities, attend ANC and schools nationwide.

Notably, two of these indicators, the malaria prevalence in pregnant women (from ANC clinics) and among school aged children (from school surveys), not available in many countries, contributed a uniquely rich source of information into the stratification for mainland Tanzania. The high attendance rates of pregnant women at ANC makes them an easily accessible surveillance population to track malaria transmission intensity and provides a simple routine real-time measure of malaria prevalence at higher spatial and temporal resolutions than national household surveys [37]. Prevalence from ANC clinics shows a correlation with community-based childhood infection prevalence [25, 27, 38] thereby serving as a good measure to reflect malaria trends in the community. Community-based malaria parasite prevalence has been a benchmark measure of malaria endemicity since the 1950s [29, 39] and used in Tanzania as a milestone for controlling progress since 2000s [7‐9]. Since survey data obtained from national household surveys are not powered to provide information below regional levels, school-based surveys provide a rapid, cheaper alternative to household sample surveys [40, 41] and have been used in several countries during the 1960s [40, 42] to establish national malaria risk profiles. Tanzania’s investment into these two surveillance approaches was driven by the need for additional surveillance data as advocated by the GTS. While many countries do not conduct nationwide school surveys nor have a malaria surveillance established in ANC clinics, the basic principle of using other related data layers remains critical to developing a multilayered stratification. Countries might additionally include national household survey data, climatology or abiotic strata such as urban areas (as used in mainland Tanzania).

An important aspect to the methodology undertaken in mainland Tanzania is the simplicity of the design, without requiring complex modelling approaches often beyond the scope of those working within many national malaria programmes across Africa. The approach used was conservative, categorizing councils by their maximal risks over the past 2–3 years. Taking the maximum of multiple years’ data is valuable in ensuring that unstable councils prone to rebound of prevalence were not misclassified into the lower strata which improves the validity of the stratification and exposes more councils to aggressive control interventions. Statistical uncertainty is an important concept in risk mapping [43], but hard to interpret for many control programmes, and such a maximal-conservative use of data is one approach to a public health criterion avoiding “doing harm” [13].

The increasing availability of routine information from health facilities via DHIS2 offers an attractive scope for analyzing continuous epidemiological trends over time and monitoring service delivery at a frequency and level that is not possible through the national representative household surveys [44]. One of the most common criticisms for the use of HMIS data is the extent of the quality of the data reported through DHIS2, thereby leading to unreliable estimates of malaria risk [45]. However, as the reporting system in countries continues to improve, particularly following the launch of the High Burden to High Impact (HBHI) initiative that calls for improvements in HMIS system, the data will become increasingly more reliable. Recent evidence demonstrates the utility of these data, despite their inherent imperfections, for programme evaluations [46, 47].

There are obvious limitations to the use of routine data that could be improved with the use of new tools and better statistical handling of incomplete data. In the present approach, data from all health facilities were used, irrespective of their reporting rates. Additional file 1: Table S2 shows how the proportion of health facilities that can be included in the stratification varies depending on which threshold for reporting is applied. The influence on stratification when using only data from facilities with greater than 50% reporting rates is shown in Additional file 1: Figure S4. Applying a very strict criterion under which only data from facilities with complete reporting are included would mean that a small proportion of facilities could be included in the stratification. However, using a less stringent criterion, for example, including facilities with more than 50% reporting would increase the proportion of facilities that could be included in the stratification and was shown not to affect the overall strata allocation per council. Moreover, the arbitrary approach applied in setting appropriate cut-offs for classifying the routine indicators into the four risk groups questions the robustness of this approach. Defining accurate risk groups is crucial in ensuring that all councils are designated the correct strata.

Future work might include using all data with appropriate spatial interpolation techniques between missing months and missing reporting facilities [48] or consider the use of sentinel facility data with better reporting rates. Population distributions within councils are invariably uneven and assuming equivalent access to reporting facilities across a council could be improved with higher resolution population mapping, allowing for a more informed basis for facility-population catchments [49]. Furthermore, measures of incidence are influenced by a myriad of factors [50]. Novel techniques that adjust for treatment seeking behaviors have been developed and applied in malaria incidence estimation [51], however, these require complex models and simpler council-level adjustments are required for who seeks treatment from where [52]. Exploring the correlation matrices of the various routine indicators with each other and how they compare with community based prevalence is important in understanding the nature of the indicators in different transmission settings and defining robust and accurate thresholds for the classification.

Whilst the approach taken here has presumed equivalence between indicators, and a crude weighting applied to others (based on coverage), a more informed basis could be developed to maximize the relationships between indicators. In the absence of any formal guidelines to understand the representativeness, relatedness and appropriate cut-offs for individual strata, this is work planned over the next 3 years in mainland Tanzania. Meanwhile, the approach taken represents the most simplified means of handling multiple routine and survey composite data.

The stratification approach of mainland Tanzania served as a basis in guiding the malaria control programme in re-defining packages of interventions across the spectrum of malaria risk. No current guidelines exist as to which mix of interventions works best for which strata. In the absence of empirical evidence, using a data-driven approach guided by integration of impact modelling and expert recommendations, the country has developed the most suitable packages based on local context [12]. It is proposed to revise data inputs, approaches and strata every 3 years, as part of mid-term strategic reviews [18]. With increasing completeness of data, improved methodologies, and a changing impact of revised intervention, the process of stratification becomes dynamic.

Central health planning of malaria control in mainland Tanzania considers the council as the primary unit for resource allocation and policy. As the country moves towards implementing a targeted malaria control approach, a more granular stratification of malaria risk at sub-council level will become increasingly valuable in informing council health managers about their malaria situation. The wards will represent as important planning units especially when transmission intensity declines and stratification at this level will thereby support an evidence-based decentralized malaria control planning and implementation in mainland Tanzania.

Mainland Tanzania has used a simple and novel methodological approach, combining multiple routine data sources with survey data for local, real-time monitoring of malaria risk at the council level. Whilst the data quality could still be further strengthened, it was sufficient to define and reflect the malaria risk heterogeneity across administrative boundaries. Using knowledge from multiple indicators of transmission increases confidence in stratification and allows for a baseline upon which the current national strategic plan might be judged.