Quantifying the contribution of Plasmodium falciparum malaria to febrile illness amongst African children

Version of Record

Accepted for publication after peer review and revision.

Download
Cite
Share
CommentOpen annotations (there are currently 0 annotations on this page).

Version of Record updated: May 16, 2018 (This version)
Version of Record published: November 1, 2017 (Go to version)
Accepted Manuscript published: October 16, 2017 (Go to version)
Accepted: October 12, 2017
Received: June 1, 2017

1. Of interest
Pathogenic mechanisms of post-acute sequelae of SARS-CoV-2 infection (PASC)

Zaki A Sherif, Christian R Gomez ... RECOVER Mechanistic Pathway Task Force

Review Article Mar 22, 2023
Further reading

Article
Figures and data
Abstract
Introduction
Results
Discussion
Materials and methods
Data availability
References
Decision letter
Author response
Article and author information
Metrics

Abstract

Suspected malaria cases in Africa increasingly receive a rapid diagnostic test (RDT) before antimalarials are prescribed. While this ensures efficient use of resources to clear parasites, the underlying cause of the individual’s fever remains unknown due to potential coinfection with a non-malarial febrile illness. Widespread use of RDTs does not necessarily prevent over-estimation of clinical malaria cases or sub-optimal case management of febrile patients. We present a new approach that allows inference of the spatiotemporal prevalence of both Plasmodium falciparum malaria-attributable and non-malarial fever in sub-Saharan African children from 2006 to 2014. We estimate that 35.7% of all self-reported fevers were accompanied by a malaria infection in 2014, but that only 28.0% of those (10.0% of all fevers) were causally attributable to malaria. Most fevers among malaria-positive children are therefore caused by non-malaria illnesses. This refined understanding can help improve interpretation of the burden of febrile illness and shape policy on fever case management.

https://doi.org/10.7554/eLife.29198.001

Introduction

Following new case management guidelines issued by the World Health Organization in 2010, individuals presenting with fever at a health clinic in sub-Saharan Africa have an increased chance of receiving a rapid diagnostic test (RDT) prior to receiving antimalarial treatment (World Health Organization, 2015). This reduces the over-prescription of antimalarial drugs to uninfected patients while ensuring patent infections are identified and the parasites cleared through treatment. However, in highly endemic areas, malaria infections are both common and frequently asymptomatic (Bousema et al., 2014), meaning that for many patients presenting at health care facilities, their fever and parasitaemia are not causally related.

Ambiguity around causality in RDT-positive fever cases is problematic from both the disease surveillance and health system perspectives. Routine case surveillance systems generally report the incidence of clinical malaria based on counts of febrile individuals with a presumed or confirmed malaria infection (World Health Organization, 2016). If the malaria infections in many of these cases are in fact asymptomatic then the resulting case reports will over-represent clinical malaria, even where RDT testing is ubiquitous, leading to an overestimation of morbidity due to malaria. Conversely, fevers caused by another pathogen but coincident with a malaria infection will suffer a systematic under-reporting under this protocol and consequently the disease burden of non-malarial febrile illness (NMFI) will be underestimated. In instances where a febrile individual has both a malaria infection and a NMFI, both infections can simultaneously be the cause of the fever. Indeed co-infections have been known to modulate both the parasite load of both diseases and the severity of the individual’s fever (French et al., 2001; Hartgers and Yazdanbakhsh, 2006).

Diagnosing and treating NMFI is challenging as symptoms of many of these diseases can be non-specific and similar to malaria, for example bacterial infections such as pneumonia (Hildenwall et al., 2016; Källander et al., 2004) and meningitis (Gwer et al., 2007). Additionally, routine diagnostic tests have not yet been developed or are not commonplace for many of these diseases (Chappuis et al., 2013). From the health systems perspective, case management may be inadequate if individuals who receive a positive RDT result but have a co-infection with another pathogen do not receive effective treatment for their fever. For these reasons, there is a pressing need to understand the contribution of malaria to febrile illness and how this varies spatially and temporally across endemic Africa.

Estimating malaria-attributable fever prevalence

The rate of fever associated with (but not necessarily caused by) a malaria infection has been shown to have declined over time in some settings, halving between 1986 and 2007 in one collation of field studies (D'Acremont et al., 2010); a trend likely driven by the declining prevalence of malaria over the latter portion of that time period. Estimating the fraction of these fevers in which the malaria infection is the causal pathogen (i.e. the malaria-attributable fraction) is challenging, and various methods have been developed. Attempts have been made to measure the malaria-attributable fraction arithmetically using case-control trials, where the difference in the rate of malaria positivity in febrile and afebrile individuals is used to calculate the attributable fraction (Ehrhardt et al., 2006; Schellenberg et al., 1994; Smith et al., 1994). Alternatively, logistic regression approaches have been developed to estimate causal fractions based on measurements of blood parasite density. This approach has been demonstrated in numerous field trials, including various settings in Kenya (Afrane et al., 2014; Bloland et al., 1999), in an area of seasonal transmission in Burkina Faso (Bisoffi et al., 2010), and in a national survey in Mozambique (Mabunda et al., 2009). Computational simulations of malaria transmission can also be adapted to estimate an upper bound on the malaria-attributable fraction of fevers at varying levels of Plasmodium falciparum prevalence (PfPR), by monitoring the proportion of parasite-positive individuals within a simulation with a symptomatic infection (here, a symptomatic infection is synonymous with a malaria attributable fever, as co-infections with other pathogens are not typically simulated) (Griffin et al., 2010; Ross et al., 2006). None of these approaches make use of the rich trove of national household survey data recording malaria infection status as well as fever history that are now available for multiple countries and years from sources such as the DHS Program (DHS Program, 2017) and the UNICEF Multiple Indicator Cluster Surveys (UNICEF, 2017).

Using household survey data to model malaria-attributable fever and non-malarial febrile illness

Household survey data on malaria infection status and two-week fever history do not allow direct attribution of malaria causality at the individual level. However, when data from multiple individuals are combined, then the causal contribution of malaria infections to fevers within the group can be explored by measuring the extent to which fevers are more common in infected versus uninfected individuals. Building on this intuitive logic, we have developed a multinomial geospatial model (described in detail in Materials and methods) that uses the georeferenced survey data on individual-level malaria infection and fever status to infer the community-level fraction of malaria positive fevers that are either caused by or coincident with underlying malaria infections, and subsequently to map these quantities across sub-Saharan Africa. This refined understanding of the contribution of malaria versus other causes to febrile illness can improve disease burden estimation and interpretation and thus inform case management policy in sub-Saharan Africa.

Results

Model overview

Using a total of 38 household surveys in 24 Saharan African countries collected between 2006 and 2014, we collated 155,369 observations of two-week fever prevalence and RDT diagnostic outcome for P. falciparum in children from 10,606 locations with two modelled predictor variables: P. falciparum prevalence in children under five years of age (PfPR_0-5); and suitability for fever without a malaria infection. The final hierarchical Bayesian model predicted the proportion of individuals with a fever directly attributable to malaria (hereafter Malaria Attributable Fever, MAF) and the proportion of individuals with a fever not attributable to malaria (hereafter Non-Malarial Febrile Illness, NMFI) within each 5 × 5 km pixel across the stable P. falciparum transmission zones on the African continent for each year 2006–2014. These maps were used to derive a series of further metrics as outlined below. Details of model validation are provided in Materials and methods and Figure 1—figure supplement 1.

Prevalence of all-cause fever

All-cause fever prevalence was calculated as the sum of the two metrics estimated by the model: prevalence of P. falciparum malaria-attributable fever (MAF) and non-malarial febrile illness (NMFI). In 2014, the prevalence of fever of any cause across the stable limits of P. falciparum transmission was 31.0%, and increased slightly over the study period, from 27.0% in 2006. Countries were highly heterogeneous in their overall fever burden, with Niger (53.4%), Gabon (44.9%) and Nigeria (42.8%) having the highest all-cause fever prevalence in 2014. The countries with the lowest fever burden in the same year were Swaziland (4.60%), Eritrea (5.3%) and Somalia (10.1%) The mapped posterior prediction of all-cause fever prevalence across Africa is shown in Figure 1.

Figure 1 with 1 supplement see all

Download asset Open asset

Predicted all-cause fever prevalence within limits of stable *P. falciparum* transmission in children under 5 years of age in 2014.
https://doi.org/10.7554/eLife.29198.002

Malaria-attributable fevers among malaria-infected children

In 2014, we estimate that less than one-third (28.0%) of all fevers in P. falciparum malaria-infected children under five were attributable to P. falciparum in any given two-week period. This fraction varied geographically, as shown in Figure 2, with the largest contribution of MAF to malaria-positive fevers in 2014 in Swaziland (86.9%), Eritrea (81.9%) and Somalia (69.9%) and the lowest in Niger (14.7%), Gabon (20.9%) and Nigeria (20.9%). The fraction also varied though time, decreasing continent-wide from 36.1% in 2006 to 28.0% in 2014. Time-series estimates for all countries are shown in Figure 3.

Figure 2

Download asset Open asset

Predicted malaria-attributable fevers as a proportion of malaria-positive fevers (children under 5 years of age, 2014).

Predictions are shown within the limits of stable *P. falciparum* transmission.

https://doi.org/10.7554/eLife.29198.004

Figure 3

Download asset Open asset

Malaria-attributable fevers as a proportion of malaria-positive fevers (children < 5 years of age, 2014).

Plotted values are the population-weighted mean for 43 sub-Saharan African countries over the study period, 2006–2014. Countries have been grouped by region to improve clarity.

https://doi.org/10.7554/eLife.29198.005

Age of the child was also shown to have an effect on the likelihood of a child developing a malaria-attributable fever. Figure 4 shows the fitted relationship between local PfPR_0-5 and the probability of a malaria-attributable fever in the past two weeks, for all children under five years of age in Figure 4a and in individuals over and under two years of age in Figure 4b. For all children, the probability of an attributable fever grew with PfPR_0-5, with a decreasing gradient as PfPR_0-5 increases. Figure 4b shows that this trend is also true for children over two years of age, but the probability of a malaria-attributable fever increases linearly with PfPR_0-5 for children under two years of age.

Figure 4

Download asset Open asset

Final fitted relationship between PfPR_0-5 and probability of a malaria-attributable fever (MAF) in the past two weeks.

(a) shows this relationship in children under five years of age, and (b) disaggregated into children under 2 years of age, and children aged 2–4 years. The probability of MAF in the past two weeks is greater for children under 2 years of age than for children above 2 years of age in areas with a PfPR_0-5 higher than approximately 0.3. Median values of the posterior distribution are shown, with shaded 95% credible intervals.

https://doi.org/10.7554/eLife.29198.006

Malaria infections within all fevers

In 2014, we estimate that 35.7% of all fevers in children under five were accompanied by an RDT-patent P. falciparum infection, and that this fraction has declined from 48.5% in 2006. The countries with the highest proportion of malaria-positive fevers in 2014 were the Central African Republic (71.0%), Equatorial Guinea (66.2%) and Guinea (65.7%) and the lowest were Ethiopia (1.4%), Botswana (1.6%) and Swaziland (2.9%). Figure 5 highlights the large disparity between the fraction of all fevers that are malaria positive (35.7% continent-wide in 2014) and the fraction of all fevers that are causally attributable to malaria (10.0% continent-wide in 2014).

Figure 5 with 1 supplement see all

Download asset Open asset

(a) Predicted malaria-positive fevers as a proportion of all fevers; (b) predicted malaria attributable fevers (MAF) as a proportion of all fevers. Both maps are shown for the year 2014, for children under 5 years of age and bounded by the limits of stable *P. falciparum* transmission.
https://doi.org/10.7554/eLife.29198.007

Contribution of non-malarial febrile illness to all fevers

Non-malarial febrile illness (NMFI) arises in two forms: (i) fevers in malaria-negative individuals, and (ii) fevers in P. falciparum malaria-positive individuals where the fever is coincident with but not caused by the malaria infection (hereafter referred to as Malaria-Coincident Fevers, MCF). The sum of these two types of febrile illness are referred to here as NMFI, and are estimated directly by the model. The fraction of all fevers that are due to NMFI increased over the study period, from 82.5% in 2006 to 90.0% in 2014. The countries with the largest fractions of NMFI to all fevers were Ethiopia (99.4%), Botswana (99.2%) and Gambia (98.5%), and the smallest were Guinea (58.2%), Equatorial Guinea (60.4%) and Central African Republic (64.6%).

The proportion of MCF within NMFI fell over the study period, from 37.6% in 2006, to 28.6% in 2014. The countries that had the highest prevalence of MCF within NMFI were the Central African Republic (55.2%), Burkina Faso (54.9%) and Equatorial Guinea (44.2%), and the countries with the lowest prevalence of MCF within NMFI were Swaziland (0.4%), Botswana (0.8%) and Ethiopia (0.8%). A corollary of this is that Swaziland, Botswana and Ethiopia were also the countries with the lowest proportion of malaria-positive individuals within the national cases of NMFI. The posterior estimation of prevalence of NMFI in children less than five years of age within the spatial limits of P. falciparum transmission is shown in Figure 6.

Figure 6

Download asset Open asset

Predicted non-malarial febrile illness (NMFI) prevalence in children under 5 years of age.

NMFI prevalence is defined as the sum of the prevalence of febrile illness without a *P. falciparum* malaria infection and the prevalence of febrile illness coincident with, but not caused by, a P. falciparum malaria infection (MCF), for children under 5 years of age and bounded within the spatial limits of stable *P. falciparum* transmission in 2014.

https://doi.org/10.7554/eLife.29198.009

Full details national-level estimates of MAF, MCF, malaria-positive fevers and NMFI can be found in Supplementary file 1. A plot detailing the relationship between all-cause fever, malaria-positive fevers, and MAF with varying P. falciparum transmission intensity in both response data and predictions can be found in Figure 5—figure supplement 1. These plots of our estimates are displayed with overlaid estimations of clinical incidence over the duration of the past zero to two, and two to four weeks, with increasing PfPR_0-5 from an ensemble of transmission models (Cameron et al., 2015). The breakdown of the contribution of NMFI, malaria-coincident fevers, and fevers without a patent malaria infection to all-cause fever varied considerably by country; this breakdown is displayed in detail for 2014 in Supplementary file 5.

Discussion

The analysis presented here shows that although the proportion of fever cases that are accompanied by an RDT-patent P. falciparum infection remains high, only approximately a third of these fevers are causally attributable to malaria. The majority of febrile illness in Africa is caused by pathogens other than P. falciparum malaria, even in areas where malaria is highly endemic, and the proportion of all fevers caused by NMFI has risen since 2006. We estimate here that over a typical two-week period in 2014 one in four children under five years old residing in the limits of stable malaria transmission will suffer a fever not caused by P. falciparum malaria, whereas only one in every 32 children will suffer a fever directly caused by P. falciparum. We show that current estimations of malaria burden in Africa based on RDT-positive cases of fever may be overestimating the burden by up to two-thirds, and the level of overestimation is likely to be highly heterogeneous between different countries (Figure 3). For example, countries such as Niger, Gabon and Nigeria, where fewer than 21% of malaria-positive fever cases are causally attributable to malaria, may be substantially overestimating the morbidity caused by malaria (and systematically underestimating the burden of other diseases that may be co-infecting and causal). Additionally, these results have implications for the effectiveness of treatment received at point of care, as a substantial proportion of RDT-positive fevers are likely to be coincident with and caused by a NMFI. Our estimations of MAF prevalence in the past two weeks are comparable to estimations of clinical incidence from an ensemble of transmission models (Figure 5—figure supplement 1) (Cameron et al., 2015) although our estimates match more closely to transmission model estimations of the prevalence of febrile illness lasting the duration of the past two-to-four weeks, rather than the past zero-to-two. For 13 of the 38 household surveys used in this analysis, the number of days since the onset of the child’s fever was available. Children in these surveys had a fever for on average 5.38 days (±3.85 days) prior to the interview. This disparity between fever duration in our estimates and transmission model estimates can perhaps be explained by the cross-sectional nature of the surveys; the transmission models estimate the full course of the fever without treatment, whereas the household surveys aim to curtail the fever at the interview by administering appropriate medication to febrile children.

The total number of fever cases accompanied by an RDT-patent P. falciparum infection declined substantially over the study period (from 48.5% in 2006 to 35.7% in 2014), linked with declining PfPR_0-5 over the same time (Bhatt et al., 2015; Smith et al., 2007). The only country with a noteworthy increase in malaria-positive fevers over the study period was Gabon, which also experienced an increase in PfPR_0-5 between 2006 and 2014.

This modelling approach makes use of household survey datasets in a novel way; no previous attempts have been made to measure childhood fever prevalence (and the malaria/non-malaria causality of the fever) from surveys recording two-week fever history. This approach has some drawbacks: it relies on two-week recall of fever from the child’s caregiver, the accuracy of which can be subject to biases such as (i) length of time since the fever; (ii) the child’s frequency of febrile episodes, (iii) the economic status of the family, and (iv) even the child’s sex (Das et al., 2012; Rockers and McConnell, 2017). We also include fever caused by P. vivax malaria as a NMFI by definition, as not all household surveys tested for P. vivax with RDT and thus only P. falciparum outcomes were used in this analysis. This is unlikely to have a major effect on most of Africa due to high Duffy negativity resulting in unstable P. vivax transmission, and very low levels of stable transmission in Madagascar and the horn of Africa (Gething et al., 2012; Howes et al., 2011). One drawback of using household survey data to generate predictions such as those presented here is that the survey data interviews children on one day within the time frame that the survey is conducted (typically two to four months). It is not necessarily the case that children who present a patent malaria infection but no history of infection over two weeks prior to their survey date have never, or will never, develop a malaria-attributable fever, and may have already done so outside of the two-week recall period of the survey interview. Additionally, co-infections may lead to a fever that would not have occurred had the child only had a monoinfection with malaria or the NMFI (French et al., 2001; Hartgers and Yazdanbakhsh, 2006).This study does not consider the effects of co-morbidity in this sense, and fevers that fall into this category are accounted for within NMFI. We also incorporate no predictor variables at resolutions below the pixel-level (e.g. individual or household level), which are likely to be significantly predictive of both MAF and NMFI prevalence. Predictors such as household wealth and other associated health indicators are available within the household survey datasets, but owing to the heterogeneity of individual level variables measured in different surveys and the lack of a common set of information directly representative of those variables outside of surveyed countries, we have not incorporated them in the current model.

The low prevalence of malaria-positive fevers at all levels of P. falciparum endemicity suggests that passive case detection alone may not be enough to halt transmission of malaria, as many infected children are afebrile at any one time (regardless of whether their fever is attributable to the malaria infection or not). Further, this study estimates MAF prevalence in children less than two years of age, and between two and four years of age. We find that the relationship between MAF prevalence and PfPR is not equivalent across all age groups (Figure 4), with children under two years of age having a higher probability of developing a malaria-attributable fever than children older than two in areas where PfPR_0-5 is greater than 30%. For older children living in high PfPR locations, a higher proportion of the population will have acquired immunity to malaria infection (Doolan et al., 2009), and therefore may be less likely to present with a malaria-attributable fever. Although there is uncertainty over how long malaria immunity lasts (Filipe et al., 2007; Wipasa et al., 2010), in areas of rapid decline of malaria prevalence that are now moving towards elimination, this residual acquired immunity may present problems for elimination if programs rely on passive case detection and treatment to halt transmission. Again, the limitation of our estimates which use a two-week window for reporting of symptoms means that, in order to assert with confidence that passive case detection would be insufficient to halt transmission, evidence for a high prevalence of never-symptomatic infections would also be necessary.

Despite all malaria-endemic countries now adopting the official policy of parasite-based diagnosis of suspected malaria cases before issuing antimalarials, a large number of suspected malaria cases still do not receive a diagnostic test before treatment (Hertz et al., 2013; World Health Organization, 2015). This proportion has decreased over the time period of this study, from 64% of suspected cases not receiving a diagnostic test in 2005 to only 35% in 2014 (World Health Organization, 2015). The reasons for this high proportion of suspected cases not receiving parasite-based diagnosis lies predominantly in resource limitation in rural health clinics in malaria endemic countries (Ghai et al., 2016). This remaining proportion of NMFI that is assumed to be caused by malaria, and thus treated with an antimalarial, exacerbates the problem of over-prescription of antimalarials. By enumerating the proportion of NMFI, this study provides evidence of the on-going need to test febrile individuals for parasitaemia, but can also be used identify different geographic and malaria-burden settings where administering antimalarials in the absence of a positive test for P. falciparum is least warranted.

This suite of maps demonstrates the need for better diagnostic tests for other pathogens causing NMFI in Africa, which will in turn lead to a better understanding of the range of diseases that make up the non-malarial fraction of all-cause fever prevalence. Whilst the spatial distribution of pathogens that cause NMFI is largely unknown, occurrence data for these diseases are becoming more widely available, particularly in the context of identifying diseases with similar symptomatic presentations as malaria. The Worldwide Antimalarial Resistance Network (WWARN) now collects clinical evidence of non-malarial febrile illness (NMFI) caused by neglected tropical diseases from peer-reviewed case studies of fever diagnoses from across the malaria-endemic world (Worldwide Antimalarial Resistance Network, 2017). The collation of these studies is designed to give an overview of the presence of fever-causing pathogens in areas of overlapping endemicity with malaria, but cannot presently be used to quantify the spatiotemporal distribution of NMFI, nor can it be used to measure the individual contribution of each disease as an underlying cause of febrile illness. For example, a major review of 146 case studies in the Mekong region of Southeast Asia (Acestor et al., 2012), showed that misdiagnosis and mistreatment of the fever was a common outcome for individuals who were co-infected with malaria and other diseases with overlapping symptoms (e.g., Dengue fever virus, Orientia tsutsugamushi and Rickettsia species). In addition to implicating potential overestimation of the burden of malaria in the past, the results of this study suggest that current WHO guidelines for integrated management of childhood illness (IMCI) is currently suboptimal for the treatment of non-malarial fever. Whilst treating positive malaria infections is the correct strategy for slowing malaria transmission, removing the malaria-attributable fevers from the population, and reducing the reservoir of infected individuals that perpetuate transmission; current procedures may lead to systematic mismanagement of NMFI. This study shows the need for increased efforts to develop and distribute routine diagnostics for NMFI-causing pathogens.

Materials and methods

Household survey data

Request a detailed protocol

Thirty-eight cross sectional, nationally-representative georeferenced surveys of malaria prevalence in children less than five years of age across 24 countries in sub-Saharan Africa were obtained from the Malaria Atlas Project database (Guerra et al., 2007; Malaria Atlas Project , 2017). These surveys originate from a variety of sources such as the DHS Program, UNICEF Multiple Indicator Cluster Surveys, and national Ministries of Health in malaria-endemic countries. Full details of the surveys used can be found in Supplementary file 2.

In these surveys, a blood sample was taken from any children present and tested for malaria parasites with both a RDT and Giemsa-stained microscopy. Here, we used the RDT-derived diagnostic outcome as an indicator of malaria positivity within the last two weeks, to lessen the possibility of a recent infection having been cleared via artemisinin combination therapy (ACT) (and therefore presenting as a negative result by microscopy). RDTs offer a more accurate representation of two-week infection than microscopy due to the persistence of histidine-rich protein II in the blood for up to 14 days after parasite clearance with ACT (Mayxay et al., 2001). Diagnostic and fever history outcomes from a total of 155,369 children that were aggregated within 5 × 5 km raster cells, resulting in a total 10,606 data points. In addition to the diagnostic and fever history outcomes, the wealth quintile (i.e. poorest, poorer, middle, richer, richest) of the individual’s household was also extracted for use as an household-level predictor of fever in the final multinomial model.

Covariates

Request a detailed protocol

Independent model-based spatiotemporal predictions of both the prevalence of P. falciparum infection (PfPR_0-5) and for the environmental suitability for background (non-malarial) fevers in children under five years of age across Africa for each year within the span of the household survey data (2006–2015) were constructed as covariates for use in the final multinomial model.

The PfPR_0-5 covariate was generated for each year using the PfPR_2-10 modelled predictions of slide-patent African PfPR from the Malaria Atlas Project (Bhatt et al., 2015), age-standardised to 0–5 years using the ‘agestand’ package in R, based on a previously described age-standardisation procedure for PfPR (Smith et al., 2007), and transformed to predict RDT-based prevalence via an microscopy-to-RDT relationship identified in earlier work (Mappin et al., 2015).

Prior to this study, no mapped estimates of environmental suitability for background fever existed for the African continent, hence a modelled estimate of environmental suitability for background fever was constructed using a boosted regression tree (BRT) model. The BRT approach generates a flexible regression model in two steps. Initially, response data are split recursively by predictor variables, where predictor variables that maximise the partitioned response data’s homogeneity are selected more frequently in subsequent steps. This is followed by a ‘boosting’ step, where the regression tree models are combined to reduce the model’s overall predictive deviance (Elith et al., 2008), and at each iterative step cross-validated against a randomly held-out subset to avoid overfitting (Bhatt et al., 2013).

Firstly, a training dataset was constructed, consisting of a subset of the household survey dataset where survey cells with ten or fewer tested individuals were removed in order to increase the reliability of observed prevalence values. This procedure reduced the final number of data points in the training dataset to 5320. Background fever prevalence was obtained at each data point through aggregation of individuals with a fever in the previous two weeks without a patent malaria infection.

A large suite of potential environmental and socio-demographic covariates was assembled, as described elsewhere (Weiss et al., 2015). Briefly, 5 × 5 km resolution surfaces across the African continent of different environmental and sociodemographic covariate types—elevation, land cover, population density, precipitation, enhanced vegetation index (EVI), land surface temperature (LST), P. falciparum temperature suitability index (TSI) (Weiss et al., 2014), tasselled cap wetness (TCW), tasselled cap brightness (TCB), nighttime lights, Global Urban Footprint (Esch et al., 2012), and national-level World Bank indicators of poverty and healthcare quality (World Bank, 2017)—were obtained from the Malaria Atlas Project database and leveraged by constructing pixel-level spatial summaries (maximum, mean, minimum, standard deviation and range). These summarisations produced a total of 986 spatially- and temporally-contemporaneous covariates from which we extracted results for each of the 5320 georeferenced response data points.

To reduce the number of covariates an exploratory BRT model was constructed using background fever prevalence as the response variable and the covariates as predictor variables using the 'gbm.step' function in the 'dismo' package in R (Elith et al., 2008). The BRT model was parameterised with a tree complexity of 5, a learning rate of 0·05, a hold-out fraction of 0·05 and an error structure within the family 'gaussian' after a suitable inverse-sigmoid transform of the observed background fever prevalences. Each of the 986 covariates was ranked by their contributions to the fitted exploratory model. To reduce redundancy, each variable within the assembled suite of covariates was tested for collinearity with every other variable. Covariates were considered collinear if a correlation of greater than 0.7 existed between the two covariates, and when collinearity was identified only higher-ranked covariate was retained. This approach left a final dataset of 174 predictor variables for each of the 5320 response data points.

The final model, using the reduced set of 174 non-collinear predictor variables, was fitted with 2000 trees, a tree complexity of 5, and a reduced learning rate of 0.01 (reducing the learning rate improves model performance [Elith et al., 2008]). A list of the 174 predictor variables and each variable’s contribution to the final model used can be found in Supplementary file 3. Predictions for the environmental suitability for background fever amongst children under five years old for each month between January 2006 and December 2014 across the area of stable malaria transmission in Africa were generated using the fitted BRT model and the 174 predictor variables. For a number of predictor variables, data were unavailable in November and December 2014, so data from the same month in the previous year were used instead. Yearly predictions for 2006–2014 were generated as a mean of each monthly prediction for use in the multinomial model. Gaps in the resulting predicted layers, caused by small gaps in the remotely-sensed covariate layers, were filled by recursively scanning through each raster cell by cell, and filling in cells with no data with the mean of the first layer of surrounding cells until no gaps remained.

Multinomial model

Request a detailed protocol

A training dataset for the multinomial model was constructed from all 10,606 points in the household survey dataset. Of these points, 25% (2,652) were selected for the holdout dataset, where probability of selection was directly proportional to the distance of the nearest neighbouring point, to avoid clustering of hold-out points in areas of dense data coverage. The remaining 75% of points (7,954) were used in the training dataset.

The observed data at each site, $i$ (a 5 × 5 km pixel), with $N_{t o t .}^{i}$ surveyed individuals may be represented as a two-by-two categorical table of counts according to fever status (febrile or afebrile) and RDT-based P. falciparum parasite status (positive or negative). The hierarchical Bayesian model we construct here thus takes a top-level likelihood with multinomial distribution having four unknown parameters, ${q_{1}^{i}, q_{2}^{i}, q_{3}^{i}, q_{4}^{i}}$ , that describe the expected proportion of counts in each category; i.e.,

[N_{f e b . & p o s .}^{i}, N_{f e b . & n e g .}^{i}, N_{a f e b . & p o s .}^{i}, N_{a f e b . & n e g .}^{i}] \sim M u l t i n o m i a l [{q_{1}^{i}, q_{2}^{i}, q_{3}^{i}, q_{4}^{i}} | N_{t o t .}^{i}] .

Since each expected proportion must lie between zero and one while their joint sum is strictly unity, the effective dimension of unknown parameters here is three, with constraint to the standard 3-simplex. Our parameterisation takes three components accordingly, which we define so as both to respect these constraints and to represent the key targets of our inference as shown in Table 1.

Table 1

Four-way infection outcome table and formulae for deriving targets of inference.

https://doi.org/10.7554/eLife.29198.010

	Pf pos.	Pf neg.
Febrile	$q_{1}^{i} = p_{p o s .}^{i} (r_{m a f}^{i} (1 - p_{b g}^{i}) + p_{b g}^{i})$	$q_{2}^{i} = (1 - p_{p o s .}^{i}) p_{b g}^{i}$
Afebrile	$q_{3}^{i} = p_{p o s .}^{i} (1 - r_{m a f}^{i}) (1 - p_{b g}^{i})$	$q_{4}^{i} = (1 - p_{p o s .}^{i}) (1 - p_{b g}^{i})$

Here $p_{p o s .}^{i}$ is the expected prevalence of malaria parasites in children under five years of age, $p_{b g}^{i}$ is the expected prevalence of background fever (within a two-week window) in the same cohort, and $r_{m a f}^{i}$ gives the expected proportion of P. falciparum positives reporting a fever attributable to malaria. By convention (and in line with our objectives) we define the prevalence of these ‘malaria-attributable fevers’, $p_{m a f}^{i}$ , as $p_{m a f}^{i} = p_{p o s .}^{i} r_{m a f}^{i} (1 - p_{b g}^{i})$ ; that is, excluding causal fevers ‘coexisting’ with a background fever.

Our representation of the local parasite prevalence in children under five years of age takes the form of a spatial generalized linear model (GLM) in which the logit transform of predicted prevalence from the MAP PfPR_2-10 space-time cube (Bhatt et al., 2015), age-standardised to PfPR_0-5 (Smith et al., 2007), is used as a linear predictor for the logit transform of the under-five year old prevalence, augmented with a latent (spatial) Gaussian random field (GRF). That is,

l o g i t p_{p o s .}^{i} = l o g i t P f P R_{M A P}^{i} + f_{x (i),}

f \sim G R F_{θ .}

An identical model is used for the rate of background fevers in this cohort, except that the predictor variable is now the logit of the environmental suitability for background fevers predicted by the BRT model outlined above. Hence,

l o g i t p_{b g}^{i} = l o g i t F E V_{B R T}^{i} + g_{x (i),}

g \sim G R F_{ϕ .}

Our model for the proportion of parasite positives with a causal malaria fever is a simple quadratic dependence on the local prevalence, motivated by the form observed in a previously described transmission model (Maire et al., 2006; Ross et al., 2006; Smith et al., 2006a; Smith et al., 2006b). Namely,

l o g i t r_{m a f}^{i} = δ + ψ \times l o g i t p_{p o s .}^{i} + ξ \times {(l o g i t p_{p o s .}^{i})}^{2} .

For computational tractability we adopt a Gaussian Markov Random Field (GMRF) approximation to the continuous GRF components of our model using the version for Matern GRFs identified by Lindgren et al. (Lindgren et al., 2011) from the INLA package (Rue et al., 2009). Implementation of this model requires construction of a mesh-based tessellation enclosing the African continent (plus Madagascar) which we also perform with the INLA package using 6280 mesh nodes. We fixed the smoothness of both GMRFs used in our model to $ν = 1 (α = 2, d = 2)$ and set (weakly informative) standard Normal priors on the logarithm of the hyperparameters for each, $κ (i . e ., 2 \sqrt{2} \times i n v e r s e r a n g e)$ and $τ (i . e ., i n v e r s e v a r i a n c e)$ . Finally, we complete our model by placing priors on the intercept and two slope coefficients of the $l o g i t r_{m a f}^{i}$ model:

δ \sim N o r m a l [- 2, {1.0}^{2}], ψ \sim N o r m a l [0, {0.25}^{2}], a n d ξ \sim N o r m a l [0, {0.25}^{2}] .

Model fitting was performed with the TMB package (Kristensen et al., 2015) for automatic differentiation, which returns functions for the likelihood plus gradient and Hessian matrices with respect to all parameters of our model after (approximate) marginalisation over the random fields via a Laplace approximation. These are then plugged into the ‘nlminb’ function in R and optimised to find the empirical Bayes solution with the local Hessian used to form approximate credible intervals as summarized in Supplementary file 4. Additional model fits to response data disaggregated by age (children over and under two years of age) were conducted to assess the age-effect on the likelihood of developing MAF with varying PfPR.

Model validation was perform via an initial fit of the above model to a sub-sample (75%) of the full dataset using a spatial leave-one-out cross validation procedure (Le Rest et al., 2014). The probability integral transform diagnostic (Angus, 1994; Diebold et al., 1997) for multinomial probabilities at the holdout sites estimated from the empirical quantiles of draws from the posterior predictive was used as a qualitative check on the calibration of our posterior uncertainties (Figure 1—figure supplement 1). All code for the multinomial model and the BRT model are available on GitHub (Dalrymple, 2017; a copy is available at https://github.com/elifesciences-publications/MAF-NMFI).

Data availability

The following previously published data sets were used

1. UNICEF MICS
(2011) UNICEF MICS Ghana 2011
Available upon request from UNICEF MICS (http://mics.unicef.org/surveys).

https://www.unicef.org/ghana/Ghana_MICS_Final.pdf

References

1. Acestor N
2. Cooksey R
3. Newton PN
4. Ménard D
5. Guerin PJ
6. Nakagawa J
7. Christophel E
8. González IJ
9. Bell D
(2012) Mapping the aetiology of non-malarial febrile illness in Southeast Asia through a systematic review--terra incognita impairing treatment policies
PLoS ONE 7:e44269.
https://doi.org/10.1371/journal.pone.0044269
- PubMed
- Google Scholar
1. Afrane YA
2. Zhou G
3. Githeko AK
4. Yan G
(2014) Clinical malaria case definition and malaria attributable fraction in the highlands of western Kenya
Malaria Journal 13:405.
https://doi.org/10.1186/1475-2875-13-405
- PubMed
- Google Scholar
1. Angus JE
(1994) The probability integral transform and related results
SIAM Review 36:652–654.
https://doi.org/10.1137/1036146
- Google Scholar
1. Bhatt S
2. Gething PW
3. Brady OJ
4. Messina JP
5. Farlow AW
6. Moyes CL
7. Drake JM
8. Brownstein JS
9. Hoen AG
10. Sankoh O
11. Myers MF
12. George DB
13. Jaenisch T
14. Wint GR
15. Simmons CP
16. Scott TW
17. Farrar JJ
18. Hay SI
(2013) The global distribution and burden of dengue
Nature 496:504–507.
https://doi.org/10.1038/nature12060
- PubMed
- Google Scholar
1. Bhatt S
2. Weiss DJ
3. Cameron E
4. Bisanzio D
5. Mappin B
6. Dalrymple U
7. Battle K
8. Moyes CL
9. Henry A
10. Eckhoff PA
11. Wenger EA
12. Briët O
13. Penny MA
14. Smith TA
15. Bennett A
16. Yukich J
17. Eisele TP
18. Griffin JT
19. Fergus CA
20. Lynch M
21. Lindgren F
22. Cohen JM
23. Murray CLJ
24. Smith DL
25. Hay SI
26. Cibulskis RE
27. Gething PW
(2015) The effect of malaria control on Plasmodium falciparum in Africa between 2000 and 2015
Nature 526:207–211.
https://doi.org/10.1038/nature15535
- PubMed
- Google Scholar
1. Bisoffi Z
2. Sirima SB
3. Menten J
4. Pattaro C
5. Angheben A
6. Gobbi F
7. Tinto H
8. Lodesani C
9. Neya B
10. Gobbo M
11. Van den Ende J
(2010) Accuracy of a rapid diagnostic test on the diagnosis of malaria infection and of malaria-attributable fever during low and high transmission season in Burkina Faso
Malaria Journal 9:192.
https://doi.org/10.1186/1475-2875-9-192
- PubMed
- Google Scholar
1. Bloland PB
2. Boriga DA
3. Ruebush TK
4. McCormick JB
5. Roberts JM
6. Oloo AJ
7. Hawley W
8. Lal A
9. Nahlen B
10. Campbell CC
(1999) Longitudinal cohort study of the epidemiology of malaria infections in an area of intense malaria transmission II. Descriptive epidemiology of malaria infection and disease among children
The American Journal of Tropical Medicine and Hygiene 60:641–648.
https://doi.org/10.4269/ajtmh.1999.60.641
- PubMed
- Google Scholar
1. Bousema T
2. Okell L
3. Felger I
4. Drakeley C
(2014) Asymptomatic malaria infections: detectability, transmissibility and public health relevance
Nature Reviews Microbiology 12:833–840.
https://doi.org/10.1038/nrmicro3364
- PubMed
- Google Scholar
1. Cameron E
2. Battle KE
3. Bhatt S
4. Weiss DJ
5. Bisanzio D
6. Mappin B
7. Dalrymple U
8. Hay SI
9. Smith DL
10. Griffin JT
11. Wenger EA
12. Eckhoff PA
13. Smith TA
14. Penny MA
15. Gething PW
(2015) Defining the relationship between infection prevalence and clinical incidence of Plasmodium falciparum malaria
Nature Communications 6:8170.
https://doi.org/10.1038/ncomms9170
- PubMed
- Google Scholar
(2013) Rapid diagnostic tests for non-malarial febrile illness in the tropics
Clinical Microbiology and Infection 19:422–431.
https://doi.org/10.1111/1469-0691.12154
- PubMed
- Google Scholar
(2010) Reduction in the proportion of fevers associated with Plasmodium falciparum parasitaemia in Africa: a systematic review
Malaria Journal 9:240.
https://doi.org/10.1186/1475-2875-9-240
- PubMed
- Google Scholar
Software
1. Dalrymple
(2017) MAF-NMFI, version 585d9e2
GitHub.

https://github.com/udalrymple/MAF-NMFI, v1.0
(2012) The impact of recall periods on reported morbidity and health seeking behavior
Journal of Development Economics 98:76–88.
https://doi.org/10.1016/j.jdeveco.2011.07.001
- Google Scholar
Website
1. DHS Program
(2017) Demographic and Health Surveys
Accessed October 12, 2017.

http://dhsprogram.com/
(1997) Evaluating density forecasts
International Economic Review 39:863–883.
https://doi.org/10.3386/t0215
- Google Scholar
(2009) Acquired immunity to malaria
Clinical Microbiology Reviews 22:13–36.
https://doi.org/10.1128/CMR.00025-08
- PubMed
- Google Scholar
1. Ehrhardt S
2. Burchard GD
3. Mantel C
4. Cramer JP
5. Kaiser S
6. Kubo M
7. Otchwemah RN
8. Bienzle U
9. Mockenhaupt FP
(2006) Malaria, anemia, and malnutrition in african children--defining intervention priorities
The Journal of Infectious Diseases 194:108–114.
https://doi.org/10.1086/504688
- PubMed
- Google Scholar
(2008) A working guide to boosted regression trees
Journal of Animal Ecology 77:802–813.
https://doi.org/10.1111/j.1365-2656.2008.01390.x
- PubMed
- Google Scholar
1. Esch T
2. Taubenböck H
3. Roth A
4. Heldens W
5. Felbier A
6. Thiel M
7. Schmidt M
8. Müller A
9. Dech S
(2012) TanDEM-X mission—new perspectives for the inventory and monitoring of global settlement patterns
Journal of Applied Remote Sensing 6:061702.
https://doi.org/10.1117/1.JRS.6.061702
- Google Scholar
(2007) Determination of the processes driving the acquisition of immunity to malaria using a mathematical transmission model
PLoS Computational Biology 3:e255.
https://doi.org/10.1371/journal.pcbi.0030255
- PubMed
- Google Scholar
1. French N
2. Nakiyingi J
3. Lugada E
4. Watera C
5. Whitworth JA
6. Gilks CF
(2001) Increasing rates of malarial fever with deteriorating immune status in HIV-1-infected Ugandan adults
AIDS 15:899–906.
https://doi.org/10.1097/00002030-200105040-00010
- PubMed
- Google Scholar
1. Gething PW
2. Elyazar IR
3. Moyes CL
4. Smith DL
5. Battle KE
6. Guerra CA
7. Patil AP
8. Tatem AJ
9. Howes RE
10. Myers MF
11. George DB
12. Horby P
13. Wertheim HF
14. Price RN
15. Müeller I
16. Baird JK
17. Hay SI
(2012) A long neglected world malaria map: Plasmodium vivax endemicity in 2010
PLoS Neglected Tropical Diseases 6:e1814.
https://doi.org/10.1371/journal.pntd.0001814
- PubMed
- Google Scholar
(2016) Multi-method assessment of patients with febrile illness reveals over-diagnosis of malaria in rural Uganda
Malaria Journal 15:460.
https://doi.org/10.1186/s12936-016-1502-4
- PubMed
- Google Scholar
1. Griffin JT
2. Hollingsworth TD
3. Okell LC
4. Churcher TS
5. White M
6. Hinsley W
7. Bousema T
8. Drakeley CJ
9. Ferguson NM
10. Basáñez MG
11. Ghani AC
(2010) Reducing Plasmodium falciparum malaria transmission in Africa: a model-based evaluation of intervention strategies
PLoS Medicine 7:e1000324.
https://doi.org/10.1371/journal.pmed.1000324
- PubMed
- Google Scholar
1. Guerra CA
2. Hay SI
3. Lucioparedes LS
4. Gikandi PW
5. Tatem AJ
6. Noor AM
7. Snow RW
(2007) Assembling a global database of malaria parasite prevalence for the malaria atlas project
Malaria Journal 6:17.
https://doi.org/10.1186/1475-2875-6-17
- PubMed
- Google Scholar
(2007) Over-diagnosis and co-morbidity of severe malaria in African children: a guide for clinicians
The American Journal of Tropical Medicine and Hygiene 77:6–13.
https://doi.org/10.4269/ajtmh.2007.77.6
- PubMed
- Google Scholar
1. Hartgers FC
2. Yazdanbakhsh M
(2006) Co-infection of helminths and malaria: modulation of the immune responses to malaria
Parasite Immunology 28:497–506.
https://doi.org/10.1111/j.1365-3024.2006.00901.x
- PubMed
- Google Scholar
1. Hertz JT
2. Munishi OM
3. Sharp JP
4. Reddy EA
5. Crump JA
(2013) Comparing actual and perceived causes of fever among community members in a low malaria transmission setting in northern Tanzania
Tropical Medicine & International Health 18:1406–1415.
https://doi.org/10.1111/tmi.12191
- PubMed
- Google Scholar
1. Hildenwall H
2. Amos B
3. Mtove G
4. Muro F
5. Cederlund K
6. Reyburn H
(2016) Causes of non-malarial febrile illness in outpatients in Tanzania
Tropical Medicine & International Health 21:149–156.
https://doi.org/10.1111/tmi.12635
- PubMed
- Google Scholar
1. Howes RE
2. Patil AP
3. Piel FB
4. Nyangiri OA
5. Kabaria CW
6. Gething PW
7. Zimmerman PA
8. Barnadas C
9. Beall CM
10. Gebremedhin A
11. Ménard D
12. Williams TN
13. Weatherall DJ
14. Hay SI
(2011) The global distribution of the Duffy blood group
Nature Communications 2:266.
https://doi.org/10.1038/ncomms1265
- PubMed
- Google Scholar
(2004) Symptom overlap for malaria and pneumonia--policy implications for home management strategies
Acta Tropica 90:211–214.
https://doi.org/10.1016/j.actatropica.2003.11.013
- PubMed
- Google Scholar
1. Kristensen K
2. Nielsen A
3. Berg CW
4. Skaug H
5. Bell B
(2015)
TMB: Automatic Differentiation and Laplace Approximation

Journal of Statistical Software 70:1–21.
- Google Scholar
(2014) Spatial leave-one-out cross-validation for variable selection in the presence of spatial autocorrelation
Global Ecology and Biogeography 23:811–820.
https://doi.org/10.1111/geb.12161
- Google Scholar
(2011) An explicit link between Gaussian fields and Gaussian Markov random fields: the stochastic partial differential equation approach
Journal of the Royal Statistical Society: Series B 73:423–498.
https://doi.org/10.1111/j.1467-9868.2011.00777.x
- Google Scholar
1. Mabunda S
2. Aponte JJ
3. Tiago A
4. Alonso P
(2009) A country-wide malaria survey in Mozambique. II. Malaria attributable proportion of fever and establishment of malaria case definition in children across different epidemiological settings
Malaria Journal 8:74.
https://doi.org/10.1186/1475-2875-8-74
- PubMed
- Google Scholar
1. Maire N
2. Smith T
3. Ross A
4. Owusu-Agyei S
5. Dietz K
6. Molineaux L
(2006) A model for natural immunity to asexual blood stages of Plasmodium falciparum malaria in endemic areas
The American Journal of Tropical Medicine and Hygiene 75:19–31.
https://doi.org/10.4269/ajtmh.2006.75.19
- PubMed
- Google Scholar
Website
1. Malaria Atlas Project
(2017) MAP - Malaria Atlas Projects
Accessed October 12, 2017.

http://www.map.ox.ac.uk/
1. Mappin B
2. Cameron E
3. Dalrymple U
4. Weiss DJ
5. Bisanzio D
6. Bhatt S
7. Gething PW
(2015) Standardizing Plasmodium falciparum infection prevalence measured via microscopy versus rapid diagnostic test
Malaria Journal 14:460.
https://doi.org/10.1186/s12936-015-0984-9
- PubMed
- Google Scholar
(2001) Persistence of Plasmodium falciparum HRP-2 in successfully treated acute falciparum malaria
Transactions of the Royal Society of Tropical Medicine and Hygiene 95:179–182.
https://doi.org/10.1016/S0035-9203(01)90156-7
- PubMed
- Google Scholar
1. Rockers PC
2. McConnell M
(2017) Child gender and parental reporting of illness symptoms in sub-saharan Africa
The American Journal of Tropical Medicine and Hygiene 96:994–1000.
https://doi.org/10.4269/ajtmh.16-0249
- PubMed
- Google Scholar
1. Ross A
2. Maire N
3. Molineaux L
4. Smith T
(2006) An epidemiologic model of severe morbidity and mortality caused by Plasmodium falciparum
The American Journal of Tropical Medicine and Hygiene 75:63–73.
https://doi.org/10.4269/ajtmh.2006.75.63
- PubMed
- Google Scholar
1. Rue H
2. Martino S
3. Chopin N
(2009) Approximate Bayesian inference for latent Gaussian models by using integrated nested Laplace approximations
Journal of the Royal Statistical Society: Series B 71:319–392.
https://doi.org/10.1111/j.1467-9868.2008.00700.x
- Google Scholar
(1994) What is clinical malaria? Finding case definitions for field research in highly endemic areas
Parasitology Today 10:439–442.
https://doi.org/10.1016/0169-4758(94)90179-1
- PubMed
- Google Scholar
1. Smith DL
2. Guerra CA
3. Snow RW
4. Hay SI
(2007) Standardizing estimates of the Plasmodium falciparum parasite rate
Malaria Journal 6:131.
https://doi.org/10.1186/1475-2875-6-131
- PubMed
- Google Scholar
1. Smith T
2. Killeen GF
3. Maire N
4. Ross A
5. Molineaux L
6. Tediosi F
7. Hutton G
8. Utzinger J
9. Dietz K
10. Tanner M
(2006a) Mathematical modeling of the impact of malaria vaccines on the clinical epidemiology and natural history of Plasmodium falciparum malaria: Overview
The American Journal of Tropical Medicine and Hygiene 75:1–10.
https://doi.org/10.4269/ajtmh.2006.75.2_suppl.0750001
- PubMed
- Google Scholar
1. Smith T
2. Ross A
3. Maire N
4. Rogier C
5. Trape JF
6. Molineaux L
(2006b) An epidemiologic model of the incidence of acute illness in Plasmodium falciparum malaria
The American Journal of Tropical Medicine and Hygiene 75:56–62.
https://doi.org/10.4269/ajtmh.2006.75.56
- PubMed
- Google Scholar
(1994) Attributable fraction estimates and case definitions for malaria in endemic areas
Statistics in Medicine 13:2345–2358.
https://doi.org/10.1002/sim.4780132206
- PubMed
- Google Scholar
(2007) Paracheck-Pf accuracy and recently treated Plasmodium falciparum infections: is there a risk of over-diagnosis?
Malaria Journal 6:58.
https://doi.org/10.1186/1475-2875-6-58
- PubMed
- Google Scholar
Website
1. UNICEF
(2017) UNICEF Multiple Indicator Cluster Surveys
Accessed October 12, 2017.

http://mics.unicef.org/
1. Weiss DJ
2. Bhatt S
3. Mappin B
4. Van Boeckel TP
5. Smith DL
6. Hay SI
7. Gething PW
(2014) Air temperature suitability for Plasmodium falciparum malaria transmission in Africa 2000-2012: a high-resolution spatiotemporal prediction
Malaria Journal 13:171.
https://doi.org/10.1186/1475-2875-13-171
- PubMed
- Google Scholar
1. Weiss DJ
2. Mappin B
3. Dalrymple U
4. Bhatt S
5. Cameron E
6. Hay SI
7. Gething PW
(2015) Re-examining environmental correlates of Plasmodium falciparum malaria endemicity: a data-intensive variable selection approach
Malaria Journal 14:68.
https://doi.org/10.1186/s12936-015-0574-x
- PubMed
- Google Scholar
1. Wipasa J
2. Suphavilai C
3. Okell LC
4. Cook J
5. Corran PH
6. Thaikla K
7. Liewsaree W
8. Riley EM
9. Hafalla JC
(2010) Long-lived antibody and B Cell memory responses to the human malaria parasites, Plasmodium falciparum and Plasmodium vivax
PLoS Pathogens 6:e1000770.
https://doi.org/10.1371/journal.ppat.1000770
- PubMed
- Google Scholar
Website
1. World Bank
(2017) World Bank Open Data
Accessed October 12, 2017.

http://data.worldbank.org/
Report
1. World Health Organization
(2015)
World Malaria Report 2015

Geneva, Switzerland: World Health Organization.
- Google Scholar
Report
1. World Health Organization
(2016)
World Malaria Report 2016

Geneva, Switzerland: World Health Organization.
- Google Scholar
Website
1. Worldwide Antimalarial Resistance Network
(2017) WWARN Non-Malarial Febrile Illness Map
Accessed October 12, 2017.

http://www.wwarn.org/surveyor/NMFI/offline-first-v1.1.jsp#0

Decision letter

Mark Jit

Reviewing Editor; London School of Hygiene & Tropical Medicine, and Public Health England, United Kingdom

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

Thank you for submitting your article "Quantifying the contribution of malaria versus other causes to febrile illness amongst African children" for consideration by eLife. Your article has been favorably evaluated by Prabhat Jha (Senior Editor) and three reviewers, one of whom, Mark Jit, is a member of our Board of Reviewing Editors. The following individual involved in review of your submission has agreed to reveal their identity: Thomas Eisele.

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission. We agree that while the manuscript is not acceptable in its current state, we believe that it has potential and would like to see a version with major revisions.

In particular, the reviewers and I agreed that you took an interesting approach to an important public health question, using a large, rich dataset. However we felt that there were methodological shortcomings in the model (or at least in the current description of it) that needed to be addressed. The main issues that need to be addressed are listed below.

1) Model validation

Although it is mentioned that the fever model has been validated (subsection “Model overview”), it is not clear how this was actually done. Part of the problem is that you have not presented any of the original data on fevers and malaria prevalence. This makes it impossible to tell the goodness of fit.

Given that eLife has no space or figure restrictions, the following data plots would ideally be shown:

- The raw 2x2 table showing all fevers versus RDT positivity (data and predictions), which is really interesting and is key to the message of the paper. This should also be plotted by country and transmission intensity.

- Model-predicted vs. observed all cause fever by pixel or by country. A fairly small subset of the available DHS and MICS surveys are used, presumably because the rest do not have any data on RDT positivity (and/or GPS?). However these other surveys do contain data on all-cause fever and therefore would be an excellent source of validation data for this aspect of the model.

- A plot of the final fitted relationship between fever prevalence, parasite prevalence and the malaria-attributable fraction (and data on the 1st two variables)

- Comparison with observed vs. predicted fever prevalence. These data are not presented, and one of the reviewers who checked could not reconcile some of their predictions with the available data online. E.g. Results subsection “Prevalence of all-cause fever”: fever prevalence in Liberia is quoted as 51.7% in 2014, but on the DHS survey website only a Liberia survey in 2013, which gave fever prevalence as 28.6%. Similarly the most recent data from Niger (DHS 2012) shows a fever prevalence of 14.2%, but the quoted value by the authors is 57.2%. Please could you check and explain the reasons for any discrepancies.

- Related to those countries with the highest and lowest prevalence of fever (subsection “Prevalence of all-cause fever”): predicted values are given for Eritrea, Niger, Botswana and Zimbabwe. But in the supplementary data table (Supplementary file 2), the list of surveys does not include any data from any of these countries. Is this correct, or is the data list incomplete? If correct it seems concerning that the least certain estimates produce the outlying predictions. This should warrant some reinvestigation of the model fit. We suggest at least comparing the predictions against available fever prevalence estimates, even if not coupled with RDT data (as mentioned above).

More generally, external validation should be performed by comparing model results to datasets not used to parameterise it. For instance you mention that case-control studies, transmission models etc. have been used to estimate the attributable fraction of malaria in fever – it would be useful to compare the model results with these studies.

For transparency and to comply with eLife policy for mathematical models, you should provide the initial and final set of model equations (including coefficients) and the code used to select the final model, either in Supplementary Materials or in a suitable online repository such as github. Major data sets used that are not already in the public domain should also be provided unless there are compelling scientific or ethical reasons not to.

2) Extending the model beyond environmental variables

Although in the fifth paragraph of the subsection “Covariates” it mentions that socio-demographic predictors of fever were considered, all the examples in the paragraph and in Supplementary file 3 are environmental. This seems like quite a large omission, especially given that fever is self-reported. As well as many varied causes of fever, cultural perceptions of fever are important here (and language, e.g. some languages do not have a specific word for fever). We were not clear how well environmental variables could be expected to predict this, especially given that you are extending fever predictions into countries for which there are no data.

Besides cultural factors, a list of suitable covariates may include socioeconomic factors such as nutrition, crowding, mother's education, access to clean water etc. Many of these are available in DHS/MICS data but it is difficult to see how they were used. Indeed, given that the final model contained 167 predictor variables, it is difficult to see why there was any human selection at all. The authors have not tried to justify any of their chosen variables based on causality arguments, so would it not make more sense to use the entire DHS dataset as predictors, and then let variable selection algorithms winnow this down?

We are also concerned about the apparent lack of variables at the household or even individual level. This may be a problem if being malaria-positive is (positively or negatively) associated with having a fever beyond what can be explained by the spatial covariates examined. For instance, within a particular town or village (with homogeneous environmental variables) there will be poorer households who are more likely to be both malaria-positive and to have non-malaria fevers. Within the household there will be further associations in distribution due to the age, gender, birth order, genetic makeup etc. of individuals. Perhaps this has been taken into account, but it is not clear how this happened.

In general, we get the impression that the ecology of the pathogen, its environment and insect host is well-described, but the epidemiological, immunological, cultural and socioeconomic determinants of disease within the human host are either less well captured or at least less well explained. Perhaps you should include someone with clinical or at least public health training in your authorship list.

3) Scope

There are a number of areas where we believe that the scope and implications of the results may be overstated.

- You need to clarify that this study aims at improving burden estimates of uncomplicated malaria, and not case management policy. WHO clearly recommends, as do all national policies in African countries with endemic malaria, that all fevers/suspected malaria presenting at facilitates in malaria endemic areas should receive a laboratory diagnosis for malaria, and if positive treated with the first-line antimalarial. This doesn't mean the attending health professional cannot go on to treat other presenting illnesses and symptoms. But even if the fever is not directly attributable to the Pf infection at that time, it should be treated. This needs to be made clear in the paper. You should stick to how these findings impact the overall epidemiology of fever illness among children in Africa, and not make recommendations or draw conclusions from this study in the discussion for malaria case management (or IMCI) policy.

- You argue that the results of their work will improve burden estimates. We find this to be somewhat of a 'straw man' attack, as to our knowledge no burden estimates have been based on an RDT positive child with a history of fever in the past 2 weeks. Neither WHO GMP, MAP nor GBD uses such a method.

- You present cross-sectional household survey data that measures a 2-week (or there about) RDT period prevalence based on persisting HRP2 antigenemia from a Pf infection, plus an overlapping fever history based on the recall by the mother/caregiver. Their primary results suggest a large proportion of these fevers are not directly attributed to the underlying Pf infection. While this seems an appropriate interpretation of the results and in line with malaria epidemiology, the cross-sectional nature of the study is a major limitation. You need to make note that the underlying Pf infection likely would have resulted in at least 1 parasite-attributable fever, likely in the first month of the infection, and additional parasite-related fevers will likely occur, especially if a new infection occurs on top of the existing infection (just based on the malaria therapy data). So the timing of the observed RDT+ and fever is important in understanding the true relationship between the underlying Pf infection, the observed fever recall, and the relationship between the underlying infection and fever at that time. Results and Discussion should take this into consideration when interpreting results throughout the paper.

https://doi.org/10.7554/eLife.29198.019

Author response

In particular, the reviewers and I agreed that you took an interesting approach to an important public health question, using a large, rich dataset. However we felt that there were methodological shortcomings in the model (or at least in the current description of it) that needed to be addressed. The main issues that need to be addressed are listed below.

1) Model validation

Although it is mentioned that the fever model has been validated (subsection “Model overview”), it is not clear how this was actually done. Part of the problem is that you have not presented any of the original data on fevers and malaria prevalence. This makes it impossible to tell the goodness of fit.

Given that eLife has no space or figure restrictions, the following data plots would ideally be shown:

- The raw 2x2 table showing all fevers versus RDT positivity (data and predictions), which is really interesting and is key to the message of the paper. This should also be plotted by country and transmission intensity.

A 2x2 table of RDT and fever positivity, from data and predictions, can now be found in an augmented Supplementary file 2. Additionally, we have added extra columns to Supplementary file 2 detailing the RDT/fever positivity breakdown by survey. The predicted national population-weighted prevalence of fever and RDT positivity can also be found in Supplementary file 1. Pie charts of national MAF, malaria-positive fever, and NMFI can also be found in Supplementary file 5.

A plot of the relationship between PfPR_0-5 and the fever metrics of interest (all-cause fever, malaria-positive fever, and malaria-attributable fever) at surveyed sites, and pixel-level predictions, can be found in Figure 5—figure supplement 1.

- Model-predicted vs. observed all cause fever by pixel or by country. A fairly small subset of the available DHS and MICS surveys are used, presumably because the rest do not have any data on RDT positivity (and/or GPS?). However these other surveys do contain data on all-cause fever and therefore would be an excellent source of validation data for this aspect of the model.

We used thirty-eight DHS and MICS surveys as both RDT result and GPS location were required for the model. Other DHS surveys do contain information on all-cause fever either at a national or ADMIN1 level collected typically over 3-4 months, often with data being collected in a non-randomised fashion (e.g. collecting data from one region at a time). Fever prevalence is likely to be highly seasonal and heterogeneous between (and within) countries due to the seasonality of diseases that contribute to all-cause fever.

For this reason, national or sub-national measurements of all-cause fever prevalence from these surveys cannot be directly compared with our estimates (or used directly to validate them), as we estimate the annual all-cause fever prevalence within an average two weeks in that year. Our model is able to bypass this restriction with georeferenced surveys by incorporating information on both the location and month of measurement within the model.

We do agree with the reviewers that these surveys are a valuable source of currently untapped information; a potential extension to this modelling framework would include predicting metrics at monthly time-steps to allow direct comparability with household surveys.

- A plot of the final fitted relationship between fever prevalence, parasite prevalence and the malaria-attributable fraction (and data on the 1st two variables)

These plots have been added as Figure 5—figure supplement 1.

- Comparison with observed vs. predicted fever prevalence. These data are not presented, and one of the reviewers who checked could not reconcile some of their predictions with the available data online. E.g. Results subsection “Prevalence of all-cause fever”: fever prevalence in Liberia is quoted as 51.7% in 2014, but on the DHS survey website only a Liberia survey in 2013, which gave fever prevalence as 28.6%. Similarly the most recent data from Niger (DHS 2012) shows a fever prevalence of 14.2%, but the quoted value by the authors is 57.2%. Please could you check and explain the reasons for any discrepancies.

For the same reasons explained in our second response to point 1, national surveys are not directly comparable to our predictions due to the differing temporal dynamics of the surveys and predictions, and the seasonality of all-cause fever.

Since improving the model with suggestions from the reviewers (described in detail in following answers), all the noted disparities have decreased. To investigate this fully, we extracted all-cause fever prevalence from 78 DHS/MICS surveys between 2006 and 2014; 72% of our predictions had a disparity of fewer than 10 percentage points from the household surveys. The majority of surveys with a disparity of more than 10 percentage points were from either the earlier or later years of the study period where less response data is available (the bulk of the surveys used were conducted in 2009-2012). As more data becomes available in later years, we expect model performance to improve.

- Related to those countries with the highest and lowest prevalence of fever (subsection “Prevalence of all-cause fever”): predicted values are given for Eritrea, Niger, Botswana and Zimbabwe. But in the supplementary data table (Supplementary file 2), the list of surveys does not include any data from any of these countries. Is this correct, or is the data list incomplete? If correct it seems concerning that the least certain estimates produce the outlying predictions. This should warrant some reinvestigation of the model fit. We suggest at least comparing the predictions against available fever prevalence estimates, even if not coupled with RDT data (as mentioned above).

The reviewers are correct that four out of the six countries we predict to have either the highest or lowest prevalence of all-cause fever are countries for which we do not have response data. In this exercise we have data from 38 household surveys, with each survey collecting data from one country across a few months. Our prediction spans 43 African countries across a 9-year time period, totalling 387 country-years. Our data therefore only covers a tiny fraction of this period so it is not unexpected that some of the highest and lowest predictions are in country-years with no data.

More generally, external validation should be performed by comparing model results to datasets not used to parameterise it. For instance you mention that case-control studies, transmission models etc. have been used to estimate the attributable fraction of malaria in fever – it would be useful to compare the model results with these studies.

We do discuss the contribution of case-control studies and transmission model estimates of MAF in the Introduction (subsection “Estimating malaria-attributable fever prevalence”).

We have overlaid the estimated clinical incidence from an ensemble of transmission models on our plot of all-cause fever, malaria-coincident fever, and malaria-attributable fever in Figure 5—figure supplement 1 (explained in more detail in our first response to point 1). As outlined in the Discussion (first paragraph), our MAF predictions match most closely with the transmission model estimates when the transmission models are predicting clinical incidence (i.e. fever) for the duration of the past 2-4 weeks (rather than the previous 0-2 weeks). The number of days prior to the interview since the onset of the child’s fever is known in 13 of the household surveys. In these surveys, the mean number of days since the onset of fever was 5.78 (SD: 3.85). Fevers in the interview are unlikely to always be detected in the final day of their natural cycle, so the close fit between our estimates and transmission model estimates of MAF within the past 2-4 weeks are comparable. The ensemble transmission model was calibrated using active case detection (ACD) studies where any fever occurring within the same 30-day period was counted as a single incident case of malaria (standard protocol for counting incident cases in ACD studies).

For transparency and to comply with eLife policy for mathematical models, you should provide the initial and final set of model equations (including coefficients) and the code used to select the final model, either in Supplementary Materials or in a suitable online repository such as github. Major data sets used that are not already in the public domain should also be provided unless there are compelling scientific or ethical reasons not to.

The model equations are given in the manuscript; we have added final model coefficients to Supplementary file 4.

The code for both the final multinomial model and the BRT covariate generation model are available at https://github.com/udalrymple/MAF-NMFI

2) Extending the model beyond environmental variables

Although in the fifth paragraph of the subsection “Covariates” it mentions that socio-demographic predictors of fever were considered, all the examples in the paragraph and in Supplementary file 3 are environmental. This seems like quite a large omission, especially given that fever is self-reported. As well as many varied causes of fever, cultural perceptions of fever are important here (and language, e.g. some languages do not have a specific word for fever). We were not clear how well environmental variables could be expected to predict this, especially given that you are extending fever predictions into countries for which there are no data.

We have revised the model to incorporate four indicators of national-level socio-demographic factors for 2006-2014, these are: percentage coverage of diphtheria, pertussis and tetanus (DPT) immunization; annual GDP growth; percentage of pregnant women who received prenatal care; and percentage with primary school education. These data were obtained from the World Bank Database and are described in full in the fifth paragraph of the subsection “Covariates” and Supplementary file 3.

We have also incorporated three additional socio-demographic variables at the same spatial resolution as the existing predictor variables (5x5km pixels) derived from the Global Urban Footprint dataset of global urban coverage; again these are described in full in the aforementioned paragraph and Supplementary file 3.

Additionally some of the variables used in the initial analysis are of a socio-demographic nature despite being remotely-sensed (accessibility, night-time lights).

Besides cultural factors, a list of suitable covariates may include socioeconomic factors such as nutrition, crowding, mother's education, access to clean water etc. Many of these are available in DHS/MICS data but it is difficult to see how they were used. Indeed, given that the final model contained 167 predictor variables, it is difficult to see why there was any human selection at all. The authors have not tried to justify any of their chosen variables based on causality arguments, so would it not make more sense to use the entire DHS dataset as predictors, and then let variable selection algorithms winnow this down?

The predictor variables used in the creation of the covariate surfaces (PfPR_0-5 and the BRT predicted surface of background fever prevalence) are required to be complete surfaces of known values across the continent, prompting the decision to use remotely-sensed variables. These variables (with the exception of the variables added in the revision, described in the answer to Q4) were the same as those used in the MAP PfPR_2-10 model (Bhatt et al., Nature 2015), with the logic that this broad suite of variables are inclusive of most currently available geospatial data informing on the distribution of diseases that are constrained by environmental conditions.

We agree with the notion that other DHS variables would be suitable predictors of MAF and NMFI prevalence; however their values are unknown in locations other than the survey sites and so cannot function as covariates in a predictive spatial model. We have addressed this in more detail in our next response.

We are also concerned about the apparent lack of variables at the household or even individual level. This may be a problem if being malaria-positive is (positively or negatively) associated with having a fever beyond what can be explained by the spatial covariates examined. For instance, within a particular town or village (with homogeneous environmental variables) there will be poorer households who are more likely to be both malaria-positive and to have non-malaria fevers. Within the household there will be further associations in distribution due to the age, gender, birth order, genetic makeup etc. of individuals. Perhaps this has been taken into account, but it is not clear how this happened.

Throughout this study we have placed greater emphasis on predicting the patterns of fever throughout Africa and relative contributions of malaria and other causes. As such we have configured our model to predict these quantities for 5x5km pixels across Africa. As in the previous response, this precluded the use in our predictive model of any individual-level or household-level covariates since these are not available on a pixel-by-pixel basis. We agree with the reviewers that at the individual level many survey covariates (such as housing type and income level) are likely to be significantly predictive of both malarial and non-malarial fever prevalence, and moreover able to explain a correlation of increasing risk for both. Including this information in our model could potentially have increased predictive power for the former effect and reduce the possibility of over-dispersed errors contributed by the second effect. However, we have not included these individual level predictor variables in our current model owing to the heterogeneity of individual level variables measured in different surveys and the lack of a common set of information directly representative of those variables outside of surveyed countries.

During model building we attempted fitting an over-dispersed multinomial model but encountered a number of numerical difficulties suggestive of an ill-conditioned Laplace approximation step. Hence we remain concerned as to the potential for some errors due to ignoring the potential for correlated risk factors within cells. We have made a new note of this in our discussion and draw some reassurance from the probability integral transform diagnostic (Figure 1—figure supplement 1) that suggests our model uncertainties are reasonably well, though not perfectly, calibrated. In order to further explore the potential extent of this effect we have examined the relationship between leave-one-out predictive error for PfPR and NMFI and the standard deviation of relative wealth (a variable consistently reported across surveys), and found no significant effect when plotted against the PIT histogram or with quantile regression at quantiles 0.025, 0.5 and 0.975.

In general, we get the impression that the ecology of the pathogen, its environment and insect host is well-described, but the epidemiological, immunological, cultural and socioeconomic determinants of disease within the human host are either less well captured or at least less well explained. Perhaps you should include someone with clinical or at least public health training in your authorship list.

Following our last two responses, we reiterate that our focus has been on continental prediction of the quantities of interest and their interaction rather than to explicitly offer new insight into the contextual factors listed above. We agree these are interesting and worthy of further investigation but propose that doing so is not an essential component of the current study.

3) Scope

There are a number of areas where we believe that the scope and implications of the results may be overstated.

- You need to clarify that this study aims at improving burden estimates of uncomplicated malaria, and not case management policy. WHO clearly recommends, as do all national policies in African countries with endemic malaria, that all fevers/suspected malaria presenting at facilitates in malaria endemic areas should receive a laboratory diagnosis for malaria, and if positive treated with the first-line antimalarial. This doesn't mean the attending health professional cannot go on to treat other presenting illnesses and symptoms. But even if the fever is not directly attributable to the Pf infection at that time, it should be treated. This needs to be made clear in the paper. You should stick to how these findings impact the overall epidemiology of fever illness among children in Africa, and not make recommendations or draw conclusions from this study in the discussion for malaria case management (or IMCI) policy.

One of the major findings of this paper is that many fevers that are coincident with a malaria infection are causally attributable to another disease. This analysis is the first, to our knowledge, to quantify this; and the spatial heterogeneity in the proportion of fevers attributable to P. falciparum malaria and other diseases means that a febrile child has a differing chance of having MAF or NMFI depending on where they are in Africa. This does have ramifications on the likelihood of missing a co-infecting disease given a positive RDT at the clinic, should the child be infected with a disease with overlapping symptoms to P. falciparum malaria.

IMCI for fever in high to low malaria risk, in situations when the child’s RDT tests positive, recommends looking also for a bacterial cause of fever from external examination of temperature, stiffness, lesions etc. However if no external symptoms are present an antibiotic will not necessarily be provided. The child will, at the end of the consultation, receive a first-line antimalarial and an antibiotic if the visual examination suggests that they have a bacterial co-infection. By claiming that current WHO guidelines for integrated management of childhood illness (IMCI) currently suboptimal for the treatment of fever, we intend to convey that the diagnosis of NMFI is far less robust than for malaria, given that it is a visual diagnosis rather than a parasite based diagnosis, and could be improved should diagnostic tests for common NMFIs become more widely available.

We have adapted the Discussion to convey this more clearly.

- You argue that the results of their work will improve burden estimates. We find this to be somewhat of a 'straw man' attack, as to our knowledge no burden estimates have been based on an RDT positive child with a history of fever in the past 2 weeks. Neither WHO GMP, MAP nor GBD uses such a method.

The reason for using RDT positivity, rather than microscopy, is the advantage of HRP2 persistence giving a more accurate indication of two-week positivity. We are unable to use microscopy results (although available from the household surveys) as microscopy results are not a reliable indicator of 2-week malaria positivity in situations where an individual has already received antimalarial treatment for their fever. Additionally, as RDTs become more widely available, they may be used in the future for burden estimates, particularly in new 'real-time' burden estimations from new routine case reporting systems (such as DHIS2).

When estimating the total malaria burden, confirmed malaria “cases”- that is, a febrile individual who seeks care at a facility that reports to a national HMIS and receives a positive P. falciparum diagnosis- are utilised to calculate the total incidence of malaria nationally. If many of these malaria cases are actually asymptomatic malaria cases with a coincident and symptomatic NMFI, then the number of these types of malaria “case” would vary due to the prevalence of NMFI rather than the prevalence of malaria. A more accurate (although clinically challenging) measurement of malaria cases would be to calculate the total number of malaria-attributable fevers, rather than those that simply present at a clinic with a positive RDT. We would expect that, in the hypothetical complete absence of NMFI, the number of individuals presenting with a fever at clinics would decline and subsequently the total number of positive RDTs within clinics would also decline.

Additionally, the treatment-seeking rate for all-cause fever is used as a further adjustment to calculate national P. falciparum incidence. It is possible that the treatment seeking rate for MAF and NMFI is not equivalent. While measuring the treatment seeking rate for MAF and NMFI is not within the scope of this paper, a better understanding of the causal contributions of MAF and NMFI to all fevers is crucial to enhancing our estimates of malaria burden in the future.

- You present cross-sectional household survey data that measures a 2-week (or there about) RDT period prevalence based on persisting HRP2 antigenemia from a Pf infection, plus an overlapping fever history based on the recall by the mother/caregiver. Their primary results suggest a large proportion of these fevers are not directly attributed to the underlying Pf infection. While this seems an appropriate interpretation of the results and in line with malaria epidemiology, the cross-sectional nature of the study is a major limitation. You need to make note that the underlying Pf infection likely would have resulted in at least 1 parasite-attributable fever, likely in the first month of the infection, and additional parasite-related fevers will likely occur, especially if a new infection occurs on top of the existing infection (just based on the malaria therapy data). So the timing of the observed RDT+ and fever is important in understanding the true relationship between the underlying Pf infection, the observed fever recall, and the relationship between the underlying infection and fever at that time. Results and Discussion should take this into consideration when interpreting results throughout the paper.

We accept these comments and have adjusted the Results and Discussion accordingly. “…[I]n any given two-week period” has been added, and we also added a discussion point on ever-symptomatic fevers in the Discussion (third paragraph).

https://doi.org/10.7554/eLife.29198.020

Article and author information

Author details

Ursula Dalrymple
1. Department of Zoology, University of Oxford, Oxford, United Kingdom
2. Big Data Institute, Li Ka Shing Centre for Health Information and Discovery, University of Oxford, Oxford, United Kingdom
Contribution
Conceptualization, Resources, Data curation, Formal analysis, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing—original draft, Project administration, Writing—review and editing

For correspondence
ursula.dalrymple@zoo.ox.ac.uk

Competing interests
No competing interests declared

"This ORCID iD identifies the author of this article:" 0000-0001-6206-3777
Ewan Cameron

Big Data Institute, Li Ka Shing Centre for Health Information and Discovery, University of Oxford, Oxford, United Kingdom

Contribution
Conceptualization, Resources, Software, Formal analysis, Supervision, Validation, Investigation, Methodology, Writing—original draft, Writing—review and editing

Competing interests
No competing interests declared
Samir Bhatt
1. Big Data Institute, Li Ka Shing Centre for Health Information and Discovery, University of Oxford, Oxford, United Kingdom
2. Department of Infectious Disease Epidemiology, Imperial College London, London, United Kingdom
Contribution
Methodology, Writing—review and editing

Competing interests
No competing interests declared
Daniel J Weiss

Big Data Institute, Li Ka Shing Centre for Health Information and Discovery, University of Oxford, Oxford, United Kingdom

Contribution
Data curation, Methodology, Writing—review and editing

Competing interests
No competing interests declared
Sunetra Gupta

Department of Zoology, University of Oxford, Oxford, United Kingdom

Contribution
Supervision, Writing—review and editing

Competing interests
No competing interests declared
Peter W Gething

Big Data Institute, Li Ka Shing Centre for Health Information and Discovery, University of Oxford, Oxford, United Kingdom

Contribution
Conceptualization, Resources, Supervision, Methodology, Writing—original draft, Writing—review and editing

For correspondence
peter.gething@bdi.ox.ac.uk

Competing interests
No competing interests declared

Funding

Medical Research Council (Doctoral Training Grant)

Ursula Dalrymple

Bill and Melinda Gates Foundation (H5R00640)

Ewan Cameron
Samir Bhatt
Daniel J Weiss
Peter W Gething

Bill and Melinda Gates Foundation (H5R00690)

Ewan Cameron
Samir Bhatt
Daniel J Weiss
Peter W Gething

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Reviewing Editor

Mark Jit, London School of Hygiene & Tropical Medicine, and Public Health England, United Kingdom

Publication history

Received: June 1, 2017
Accepted: October 12, 2017
Accepted Manuscript published: October 16, 2017 (version 1)
Version of Record published: November 1, 2017 (version 2)
Version of Record updated: May 16, 2018 (version 3)

Copyright

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

2,165

Page views
292

Downloads
23

Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Mendeley

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

Ursula Dalrymple
Ewan Cameron
Samir Bhatt
Daniel J Weiss
Sunetra Gupta
Peter W Gething

(2017)

Quantifying the contribution of Plasmodium falciparum malaria to febrile illness amongst African children

eLife 6:e29198.

https://doi.org/10.7554/eLife.29198

Share this article

Cite this article

Predicted all-cause fever prevalence within limits of stable P. falciparum transmission in children under 5 years of age in 2014.

Predicted malaria-attributable fevers as a proportion of malaria-positive fevers (children under 5 years of age, 2014).

Malaria-attributable fevers as a proportion of malaria-positive fevers (children < 5 years of age, 2014).

Final fitted relationship between PfPR0-5 and probability of a malaria-attributable fever (MAF) in the past two weeks.

(a) Predicted malaria-positive fevers as a proportion of all fevers; (b) predicted malaria attributable fevers (MAF) as a proportion of all fevers. Both maps are shown for the year 2014, for children under 5 years of age and bounded by the limits of stable P. falciparum transmission.

Predicted non-malarial febrile illness (NMFI) prevalence in children under 5 years of age.

Four-way infection outcome table and formulae for deriving targets of inference.

Author details

Ursula Dalrymple

Contribution

For correspondence

Competing interests

Ewan Cameron

Contribution

Competing interests

Samir Bhatt

Contribution

Competing interests

Daniel J Weiss

Contribution

Competing interests

Sunetra Gupta

Contribution

Competing interests

Peter W Gething

Contribution

For correspondence

Competing interests

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

Categories and tags

Further reading

Final fitted relationship between PfPR_0-5 and probability of a malaria-attributable fever (MAF) in the past two weeks.