Methodology
This analysis shows that micro-simulations of complex patterns of health impacts following scale-up of malaria control in endemic African settings can be emulated by fitting fairly simple regression models to the simulated outputs. The emulations can use data on malaria endemicity and baseline intervention coverage to project the impact of alternative scale-up strategies for specific locations.
The high explanatory power of regressions attests to good internal validity against OpenMalaria simulations (Table
2; Additional file
2). The somewhat lower R
2s for mortality outcomes may be explained by random noise in simulation results for mortality, since deaths are much rarer than positive infection status or cases. The imputation (or dropping) of zero simulation outcomes, the (logit) transformations applied to all outcomes before regression modelling, the re-scaling of health outcomes which did not naturally fall in the range 0–1 in order to allow logit transformation, and the implementation of predictor variables simulated in discrete steps as continuous in the regression (needed for predictions for a range of provinces) may have led to sub-optimally specified models. However, external and internal validities suggested that the potential bias introduced by this is minimal.
Added value for programme planning projection tools
The impact functions thus developed for the Spectrum programme planning tool considerably improve on earlier malaria planning tools (notably the LiST child survival model), by: (i) predicting morbidity reductions, which accrue faster and are in the long term proportionally larger than mortality reductions; (ii) simulating different age groups, with proportional burden reductions in adults not much less than in young children; (iii) capturing variations in impacts over time, including partial rebounds. These rebounds result from the achieved endemicity reductions and consequent declines in acquired immunity, and become apparent from around 7 years after scale-up, in particular for mortality for people older than 5 years, as previously described in dynamic simulation studies of ITNs and SMC scale-up [
23,
37,
38].
The incorporation of dependence of health impacts on baseline endemicity is another improvement. Modeled burden reductions are proportionally larger in settings with lower baseline malaria infection prevalence rates, and less seasonality in malaria transmission. This is consistent with observations from ITN trials [
33] (Fig.
4) and with models of the dynamics induced by various malaria interventions [
27,
39‐
41]. The absolute health gains—in terms of cases and deaths averted for a given coverage increase—are generally larger for higher-endemic settings, due to the larger baseline burden compared to lower-endemic settings. Existing programme planning tools, in contrast, have typically assumed fixed burden reductions at any time after intervention scale-up, in all countries and areas of Africa irrespective of endemicity.
Furthermore, the regression models capture non-linearity in the incremental health impact from progressive coverage increases, with some degree of saturation (diminishing returns) at high coverage levels. They also capture the synergy apparent in dynamic modelling studies [
17] of higher-endemic settings between impacts of CM and ITNs.
Consistent with dynamic model-based assessments [
30], impacts for a given population effective coverage level are larger for CM than for ITNs and IRS. However, it is often easier to achieve high-level coverage for vector control interventions (often delivered through vertical programmes, as campaigns) than for effective CM (through complex multi-layer health systems), so this ranking does not imply that CM is necessarily a better investment than vector control. The Spectrum-Malaria programme planning tool, by linking the current statistical effectiveness predictions with its costing module OneHealth Tool, will enable evaluation of both impacts and costs of malaria interventions and their trade-offs in short- and longer-term.
The models did not consider age differences in ITN and CM coverage, but these are likely to have only secondary effects since the burden reductions are partly driven by transmission effects which depend—especially in the longer term—mainly on average population-wide coverage and not just the coverage in people directly accessing the intervention.
There remains a need to refine these impact functions to incorporate drug and insecticide resistance, and extend them to impacts on
Plasmodium species other than
P. falciparum, such as
P. vivax malaria (for countries with high prevalence of this species), which has very different dynamics from
P. falciparum [
42].
Consistency with effectiveness data
These predictions of vector control impacts were generally consistent with best available data, as also used by WHO, the Roll Back Malaria partnership and international malaria donors [
4,
34‐
36]. In particular, the predicted proportional burden reductions in young children following ITN scale-up were generally in line with those observed in cluster-randomized trials and other field studies and evaluations.
Also, the statistical predictions were consistent with recent ecological estimates of average ITN field impacts across sub-Saharan Africa based on synthesis of climatic, entomological, epidemiological and programmatic data across Africa, including larger proportional burden reductions at lower baseline
PfPR [
2]. For malaria-related mortality in under five-year old children, the predicted 36–64 % reduction within 2 years for settings resembling the ITN trials in Kenya and Ghana, is similar to the estimate used in the LiST model of child survival of a fixed 55 at 100 % household ITN ownership (irrespective of endemicity or seasonality) [
10,
11]. Predicted longer-term impacts were somewhat higher than the LiST time-fixed 55 % reduction, which reflects additional long-term transmission dynamic effects.
The external validation against trial data is complicated by imprecision and measurement challenges in ITN trial data: (i) The observed
case incidences vary and are potentially biased across the trials by intensity of active case surveillance and treatment access, and the parasite density threshold used in case definition. (ii) The
infection prevalence reductions depend strongly on heterogeneity in transmission [
43]. If transmission is concentrated in a small subset of highly exposed individuals, then interventions will have little effect on prevalence, while if exposure is rather homogeneous, prevalence may be considerably reduced by the same intervention package. This can lead to deviations from the expected impacts in specific locations. (iii) For
mortality, because of limitations in attributing child deaths (mostly in rural homes without medical confirmation) to malaria through verbal autopsy based on mostly a-specific symptoms, ITN trials focused on observed reductions in all-cause under-5 mortality. OpenMalaria simulated direct and indirect malaria-related deaths, but not other-cause deaths, and extrapolations from OpenMalaria predicted direct and indirect mortality reductions to either all-cause under-5 or reductions in malaria-attributable mortality involved some uncertain assumptions (Additional file
3).
The ITN and IRS type and effectiveness modelled were judged the most relevant generic representation of ITN and IRS as recommended for African programmes [
44,
45]. The relative efficacies of ITNs and IRS reflect measurements from experimental hut studies using the insecticide Actellic CS recommended by the WHO’s Pesticide Evaluation Scheme for African programmes (for settings with mosquitoes fully susceptible to the insecticides used [
18,
46]). These data drive the typically larger proportional burden reductions for ITNs than for IRS, which align with the recommendation of the WHO 2015 Global Technical Strategy to prioritize ITNs for vector control. However, absolute and relative effectiveness of ITN and IRS will vary among settings depending on local insecticide product choice (with the duration of protection sometimes less than the assumed 12 months), and site-specific insecticide resistance and ITN decay and usage and adherence patterns [
47]. Current simulations and statistical functions did not capture these features, although they can be simulated in OpenMalaria through reduced effects on mosquito survival, and extent of personal protection [
18]. There is also variation across Africa in the tendency of vector species to bite humans and to bite indoors. Highest ITN and IRS effectiveness and impact are expected in settings where highly anthropophilic (human biting) and indoor biting mosquitos (e.g.
Anopheles gambiae, more so than
Anopheles arabiensis) are responsible for most of the transmission.
For SMC, based on seven relevant trials in highly seasonal settings, the WHO has estimated a 75 % reduction in malaria case incidence, and considerable child mortality reduction within the year of implementation, but with unknown longer-term impacts [
21]. The external validity of model results for SMC, therefore, remains to be assessed following future programme evaluations upon large-scale, long-term implementation.
The authors are not aware of any field studies of the effect of prompt and effective treatment of uncomplicated disease on incidence of severe malaria and mortality. Correspondingly, the models of the impact of scale-up of CM on burden are highly uncertain in OpenMalaria. Projections of impacts of scale-up of CM at country level face further uncertainty in estimates of effective coverage of treatment, which are typically available only from caregiver’s recall of treatment of febrile children under 5 years of age without stratification between malarial and non-malarial fevers, and without clear distinction of ineffective and effective treatment regimens [
48,
49].