gh-2015_1.

qxp_Hrev_master 29/10/15 07:54 Pagina 163

Geospatial Health 2015; volume 10:306

Estimating malaria burden in Nigeria: a geostatistical modelling approach

Nnadozie Onyiri
Swiss Tropical and Public Health Institute, Basel, Switzerland

created suggests that the two main environmental covariates correlat-

Abstract ing with malaria presence were land surface temperature for day and
rainfall. It was also found that malaria prevalence increased with dis-
This study has produced a map of malaria prevalence in Nigeria tance to water bodies up to 4 km. The malaria risk map estimated from
based on available data from the Mapping Malaria Risk in Africa the spatial model shows that malaria prevalence in Nigeria varies from
(MARA) database, including all malaria prevalence surveys in Nigeria 20% in certain areas to 70% in others. The highest prevalence rates
that could be geolocated, as well as data collected during fieldwork in were found in the Niger Delta states of Rivers and Bayelsa, the areas
Nigeria between March and June 2007. Logistic regression was fitted surrounding the confluence of the rivers Niger and Benue, and also
to malaria prevalence to identify significant demographic (age) and isolated parts of the north-eastern and north-western parts of the

ly
environmental covariates in STATA. The following environmental country. Isolated patches of low malaria prevalence were found to be
covariates were included in the spatial model: the normalized differ- scattered around the country with northern Nigeria having more such

on
ence vegetation index, the enhanced vegetation index, the leaf area areas than the rest of the country. Nigeria’s belt of middle regions gen-
index, the land surface temperature for day and night, land use/land- erally has malaria prevalence of 40% and above.
cover (LULC), distance to water bodies, and rainfall. The spatial model

se
Introduction
lu
Correspondence: Nnadozie Onyiri, Swiss Tropical and Public Health
Institute, Socinstrasse 57, 4002 Basel, Switzerland. In spite of being entirely preventable, malaria has a high level of
a
Tel: +41.63.939.2785548. mortality and is the world’s most prevalent parasitic disease. It is
ci

E-mail: dozieonyiri@yahoo.co.uk caused by infection with single-celled parasites of the genus

Plasmodium, which are transmitted by the bite of Anopheles mosqui-
er

Key words: Malaria; Prevalence; Prediction; Control measures; Nigeria.

toes. Apart from the endemic tropical and sub-tropical regions, malaria
was once widespread in North America and other temperate countries.
m

Ethical clearance: there were no ethical issues involved in the study and
clearance was granted by the ethical review board of the Swiss Tropical and Today the disease occurs mostly in sub-Saharan Africa and Southeast
m

Public Health Institute, Basel, Switzerland. Asia. In 2013, 97 countries had ongoing malaria transmission (WHO,
2013). According to the World Health Organization (WHO, 2014),
co

Acknowledgements: I wish to acknowledge the following individuals who malaria is the second leading cause of death from infectious diseases
assisted in various ways in the making of this project. Firstly, let me thank in Africa after HIV/AIDS, with Nigeria and the Democratic Republic of
Professor Marcel Tanner, Director of Swiss Tropical and Public Health Congo (DRC) accounting for 40% of the global malaria deaths. Almost
on

Institute (Swiss TPH) for his guidance and support for this project. I also 20% of all deaths of children under-five in Africa are due to malaria
wish to thank immensely Dr. Penelope Vounatsou who guided the project (WHO, 2014).
every step of the way and did the statistical work. My immense gratitude also Each year, over 500 million people suffer clinical malaria episodes
N

goes to Professor Thomas A. Smith also of the Swiss TPH for his insights in
caused by P. falciparum infection alone, resulting in a conservative
both the data collection and analysis phases of the project. I also wish to
estimate of 1 million deaths annually (Guinovart et al., 2006; Vaughan
acknowledge the help of Nigerian partners in the success of this project.
Special mention must be made of Prof. Chinyere Ukaga and Prof. B.E.B. et al., 2008). Despite concerted efforts, of which the Roll Back Malaria
Nwoke of Imo University, Nigeria. Finally let me acknowledge the assistance programme (http://www.rollbackmalaria.org/) is an example, malaria
of Dr. Laura Gosoniu and Mr. Suguru Nakashima. remains a major health challenge. Between 2000 and 2012, the scale-
up of interventions helped reduce malaria incidence rates by 25% glob-
Received for publication: 7 January 2015. ally, while reaching as high as 31% in WHO’s African region (WHO,
Revision received: 7 May 2015. 2013). The global malaria mortality rate was reduced by 42% during
Accepted for publication: 8 May 2015. the same period, while the decrease in African was 49% (WHO, 2013).
During the same period, an estimated 3.3 million lives were saved by
the scaled-up malaria interventions, 3 million of which (90%) con-
©Copyright N. Onyiri et al., 2015
Licensee PAGEPress, Italy
cerned the under-five age group in sub-Saharan Africa (WHO, 2013).
Geospatial Health 2015; 10:306 In Nigeria, statistics shows that malaria accounts for 25% of the under-
doi:10.4081/gh.2015.306 five mortality, 30% of childhood mortality and 11% of maternal mortal-
ity (Okonko et al., 2009). All Nigerians are at risk of malaria and the
This article is distributed under the terms of the Creative Commons problem is compounded by the increasing resistance of malaria to
Attribution Noncommercial License (by-nc 3.0) which permits any noncom- hitherto cost-effective drugs (Okonko et al., 2009).
mercial use, distribution, and reproduction in any medium, provided the orig- Describing spatial and temporal variation in transmission and dis-
inal author(s) and source are credited.
ease risk is fundamental to epidemiological understanding and control

[Geospatial Health 2015; 10:306] [page 163]

gh-2015_1.qxp_Hrev_master 29/10/15 07:54 Pagina 164

Article

of malaria. Risk maps are, by definition, outcomes of models of disease reported by age, but with different age-groupings used in different sur-
transmission based on spatial and temporal data. These models incor- veys. Direct mapping of age-prevalence data therefore involves choos-
porate, by varying degrees, epidemiological, entomological, climate and ing a target age-group (with some flexibility in the choice of age-cate-
environmental information (Kitron, 2000). Decades of experience con- gory boundaries), and discarding data for other age-groups and for
firm that successful malaria control depends on accurate identification sites where data for the age-group are not available. A number of malar-
and geographical reconnaissance of high-risk areas (Wijeyaratne, ia distribution maps are available for Africa based on climatic and other
1999; Carter et al., 2000). In the past, malaria risk maps at different environmental predictors of malaria transmission (Craig et al., 1999;
geographical levels were largely based on expert opinion based on lim- Snow et al., 1997; Kleinschmidt et al., 2000; Rogers et al., 2002;
ited data, crude climate isolines with no clear and reproducible numer- Gemperli, 2003; Gosoniu et al., 2006; Gething et al., 2011). However,
ical definition (Craig et al., 1999). In recent years, the availability of they make little or no use of the data from field surveys of malaria
new data sources such as remote sensing (RS) and mapping tools, prevalence, which form the largest body of relevant information.
such as computerized geographic information systems (GIS) for quan- The Mapping Malaria Risk in Africa (MARA) project is a collaborative
titative analysis of spatial data, have provided an unprecedented network of key African scientists and institutions with the aim of pro-
amount of information and increased capability to describe, predict and viding empirical risk maps of malaria in Africa (Snow et al., 1999).
communicate risk and outcome of interventions (Hay et al., 2000; Initially, this involved the development of continent-wide climate-based
Kitron, 2000; Thomson and Connor, 2000; Berquist, 2001). theoretical models of climatic suitability (Craig et al., 1999) and the
Measures that might be mapped include categories of endemicity collection of parasite prevalence data to validate and/or improve these
(e.g. unstable, mesoendemic or holoendemic), vector density and models. The MARA database was established in 1996 using published
capacity, entomological inoculation rate (EIR) and incidence of dis- and unpublished malaria survey data compiled by a collaborative net-

ly
ease. However, although malaria endemicity can vary widely over only work of African scientists and institutions with the aim of providing an

on
short distances, most of these measures have only been studied in a atlas of malaria for evidence-based and targeted malaria control in
few, widely separated localities. In general, results from different sites Africa. This is the most comprehensive database on malaria data in
differ. The most broadly available measure is point prevalence assessed Africa so far, but information with respect to Nigeria, comprising 2581

se
by microscopy. Estimates of malaria prevalence at unsampled locations age-specific surveys carried out at 126 distinct locations between 1930
can be made by incorporating information from environmental covari- until 2007, is comparatively scarce. As malaria is an environmentally-
ates (Hay et al., 2000). The precision of such estimates can be further driven disease, the Swiss Tropical and Public Health Institute (Swiss
lu
improved by using spatial smoothing or geostatistical methods (Ribeiro TPH), in collaboration with its partners at the Malaria Research and
et al., 1996; Kleinschmidt et al., 2000, 2001a, 2001b; Diggle et al., 2002), Training Centre (MRTC) in Bamako, Mali initiated an updating the
a
while Bayesian geostatistical methods have demonstrated their value MARA database for Nigeria. This is important, not only because recent
ci

for mapping childhood malaria risk in the Gambia (Diggle et al., 2002) climate changes, but also given the absence of recent data from large
and generally for relating infant mortality to malaria risk (Gemperli parts of northern Nigeria. We aimed to estimate malaria prevalence in
er

and Vounatsou, 2004). Spatial statistical models have also been used to areas without survey data using rigorous statistical methods with envi-
produce malaria maps of Mali (Kleinschmidt et al., 2000) as well as for ronmental parameters as predictors. In this study, we used Bayesian
m

the whole of West Africa (Kleinschmidt et al., 2001a). All these analyses geostatistical approaches to assess the malaria-environmental rela-
modelled the prevalence data directly without taking into account age- tionship for the purpose of malaria risk mapping.
m

dependence of the malaria risk. Malaria prevalence data are usually

co
on

Table 1. Summary of fieldwork in Nigeria.

N

Location Number of surveys (years) Type of survey Number of individuals

examined
University of Port-Harcourt 3 (2000 to 2006) Community-based surveys 50 to 500
University of Calabar 1 (2004) Published community-based 594
University of Benin 4 (1998 to 2005) Published surveys 120 to 500
Imo State University 15 (1999 to 2006) Unpublished theses (5) surveys, published surveys (10) 134 to 1200
University of Nigeria 4 (1998 to 2006) Unpublished thesis work 80 to 500
University of Lagos 10 (1999 to 2004) NA 150 to 1563
Nigerian Institute for Medical Research 2 (2000 to 2006) Published surveys 165 to 350
University of Ibadan 3 (1996 and 2003) Published surveys 100 to 405
Ahmadu Bello University 8 (2000 to 2005) Unpublished theses 150 to 220
University of Abuja 1 (2000) Published survey 200
University of Jos/Federal University 5 (2000 to 2005) Published works 120 to 250
of Technology Yola (two of the surveys were malaria
drug efficacy trials)
NA, not available.

[page 164] [Geospatial Health 2015; 10:306]

gh-2015_1.qxp_Hrev_master 29/10/15 07:54 Pagina 165

Article

data collection in Nigeria is shown in Table 1.

Materials and Methods Additional survey data were obtained through online searches.
Longitude and latitude co-ordinates were determined for each parasito-
The data collected over the years by MARA up to 1999 form part of the logical survey using the Geonames geographical database
data used for this study. As data for Nigeria were sparse or relatively (http://www.geonames.org/). Each location was described as a set of
old, more data had to be collected from the year 2000 and onwards and either a first, second or third order administrative region in Nigeria.
these included both published and unpublished survey data.
Environmental data
Malaria prevalence data The following factors were considered: the normalized difference
The fieldwork was carried out from March 1 to June 30 2007 and was vegetation index (NDVI), the enhanced vegetation index (EVI), the leaf
aimed at obtaining unpublished malaria prevalence data from various area index (LAI), the amount of rainfall, the land surface temperature
Nigerian sources. The data collection was done in phases according to for day and night (LSTday and LSTnight), respectively, land use/land-cover
Nigeria’s six geopolitical zones. For the south-south zone, I visited the (LULC), elevation and distance to the nearest water body. The databas-
Ministries of Health in Calabar (Cross River State), Port-Harcourt es from which the environmental data emanate are given in Table 2.
(Rivers State) and Benin (Edo State). However, these Ministries had Environmental predictors were extracted from RS sources at spatial
no survey data in their archives except recent malaria incidence data. and temporal resolutions and shown in Table 2. To take into account of
I also visited the local universities in these states obtaining unpub- the elapsing time between the climatic suitability for malaria transmis-
lished malaria survey data from all three states with the University of sion and parasitaemia, the climatic data were gathered for different

ly
Port-Harcourt providing the greatest number. periods (up to one year) prior to the survey starting from July 2006. The
In the southeastern zone, University of Nigeria campuses at Nsukka variables LST, NDVI, EVI, LAI and land-cover were downloaded from the

on
and Enugu city (Enugu State) were visited. Imo State University pro- Moderate Resolution Imaging Spectroradiometer (MODIS)
vided a much larger number of malaria surveys than any other institu- (http://modis.gsfc.nasa.gov/), from the U.S. Geological Survey (USGS)
tion in the entire country. From this location, I got data not only from Land Processes Distributed Active Archive Center (LP DAAC). LST data

se
Imo State and the rest of the southeast zone, but also from just about were extracted as averages over 8-day periods at 1-km spatial resoluti-
every other part of the country. The visit to the southeastern zone was on. The NDVI was obtained as a 16-day average at 0.25-km spatial reso-
lu
followed by a visit to the Capital of Nigeria, Abuja, to get a national per- lution. EVI data, available from MODIS for the year 2006, represents an
spective on the malaria situation. From the Roll Back Malaria Office (a improvement on the NDVI because it corrects some distortions in the
a
division of the Federal Ministry of Health), I got only malaria incidence reflected light caused by particles in the air as well as ground cover
ci

data, which were ultimately not used for this study. However I was able below the vegetation (Weier and Herring, 2000). The EVI data product
to get some malaria prevalence data from the University of Abuja. also does not become saturated as easily as NDVI when viewing rainfo-
er

Data from the southwestern zone was collected from Ibadan (Oyo rests and other areas of the earth with large amounts of chlorophyll
State) and Lagos city. At the University of Ibadan, our data came from (Weier and Herring, 2000). The LAI data for 2006 were used. They defi-
m

the College of Medicine and the Department of Zoology. In Lagos, the ne the number of equivalent layers of leaves relative to a unit of the
data was sourced from the Nigerian Institute of Medical Research and ground and are computed daily at 1-km resulotion from MODIS spectral
m

the Department of Zoology of the University of Lagos. These two insti- reflectances for all vegetated land surface globally (USGS, 2014). The
tutions are both strong in malaria research. MODIS data, obtained at a spatial resolution of 1 km, not only provides
co

Finally, data from the North (the north-eastern, north-central and land-cover (characterising five global land cover classification sys-
north-western zones) were obtained from the University of Abuja (as tems) and land-cover type assessment, but also includes a quality con-
mentioned earlier), the University of Jos, Ahmadu Bello University trol mechanism (USGS, 2014). Altitude data were extracted from an
on

Zaria, and the Federal University of Technology in Yola. The bulk of the interpolated USGS digital elevation model (DEM) available at a spatial
data came from Ahmadu Bello University, which provided much of the resolution of 1 km.Digital maps for three different kinds of water
N

data from the northeastern and north-central zones. A summary of the bodies in Nigeria (lakes, rivers and wetlands) were acquired from the

Table 2. Sources and resolution of remote sensing data.

Data Source Spatial resolution (km2) Temporal resolution
LSTday MODIS 1×1 8 days
LSTnight MODIS 1×1 8 days
LAI (vegetation) MODIS 1×1 8 days
NDVI (vegetation) MODIS 0.25×0.25 16 days
EVI (vegetation) MODIS 1×1 16 days
Land cover MODIS 1×1 1 year
Rainfall estimate ADDS 8×8 Daily
Evapotranspiration ADDS 8×8 Daily
Elevation USGS 1×1 -
Water bodies (rivers, lakes, wetlands)HealthMapper 1×1 -
LST, land surface temperature; LAI, leaf area index; NDVI, normalized difference vegetation index; EVI, enhanced vegetation index.

[Geospatial Health 2015; 10:306] [page 165]

gh-2015_1.qxp_Hrev_master 29/10/15 07:54 Pagina 166

Article

HealthMapper database (WHO, 2010). The distance from each location assumes independence between the surveys. However, the geographi-
to the nearest water body source was calculated in IDRISI 32, which is cal location introduces correlation since the malaria risk at nearby
an integrated geographic information system (GIS) and RS software locations is influenced by similar environmental factors and it is there-
developed by Clark Labs (http://www.clarklabs.org/). Rainfall estimates fore expected that the closer the location, the more similar the way

introduced an error term (random effect) Φi at each location si,

were obtained daily at a spatial resolution of 8 km from the Aviation malaria risk varies. To account for the spatial variation in the data we

logit(pi)= Xi T ß + Φi and modelled the spatial correlation on the i

Digital Data Service (ADDS) (http://www.aviationweather. gov/adds).
The MODIS reprojection tool (USGS) was used to convert the RS data
to geo-referenced maps and ArcMap, v. 9.1 (http://www.esri.com/) was parameters, i.e. the i’s are not independent but derive from a distribu-
used as mapping tool. Additional data processing was performed in tion which models the correlation, or equivalently the covariance

mal distribution, for the Φi’s since they represent error terms and are
STATA/SE 9.2 (Stata Corporation, College Station, TX, USA). between every pair of random effects. We adopted the multivariate nor-

Statistical modelling therefore defined on a continuous scale, i.e. Φi =(Φ1,Φ2,...,Φn)T ~

Logistic regression was fitted to malaria prevalence to identify sig- N(0,∑ ). ∑ is a matrix with elements ∑ij quantifying the covariance

tively. The distribution of random effect Φ defines the Gaussian spatial

nificant demographic (age) and environmental covariates using STATA Cov(Φi,Φj) between every pair (Φi,Φj) at locations si and sj, respec-
v. 9.0. Covariates with a significance level below 0.15 were fitted into
Bayesian geostatistical logistic models using WinBUGS v. 1.4 (Imperial process.
College and Medical Research Council, London, UK) in Fortran code. To To complete the Bayesian model formulation of the geostatistical
take into account spatial heterogeneity, location-specific random models mentioned above, we needed to specify prior distributions for

ly
effects were integrated in the logistic models, assuming that they were their parameters. For the regression coefficients we adopted a non-
informative uniform prior distribution with bounds-∞ and ∞. For the
spatial parameters σ2, σ2k, ρ, and ρk we adopted inverse gamma and
distributed according to a multivariate normal distribution with vari-

on
ance-covariance matrix parameterizing the correlation structure in the
data as an exponential function of distance. gamma prior distributions, respectively, with parameters chosen to
Let Ni be the number of children tested at location si, i =1,...,n, Yi the have means equal to 1 and variances equal to 100. We estimated the

se
number of those found with malaria parasites in a blood sample and Xi parameters of the model using Markov chain Monte Carlo simulation
=(Xi1,Xi2,...,Xip)T the vector of p associated environmental predictors with Gibbs sampling (Gelfand and Smith, 1990). Starting with initial
lu
observed at location si. We assumed that Yi arised from a binomial dis- values about the parameters, the algorithm iteratively updates the
tribution Yi ~ Bin(Ni,pi) with parameter pi measuring the malaria risk parameters by simulating from their full conditional distributions, i.e.

remaining parameters. The full conditional distributions of σ2 and σ2k,

a
at location si and modelled the relation between the malaria risk and the posterior distribution of each parameter is conditional on the
environmental covariates Xi via the logistic regression logit(pi) = Xi T
ci

ß, where ß = (ß1,ß2,...,ßp)T are the regression coefficients. This model k=1,…, K are inverse gamma distributions and simulation from them
er
m

Table 3. Bivariate associations between malaria prevalence and environmental indicators arising from non-spatial and spatial logistic
models.
m

Environmental indicators Non-spatial models Spatial models

co

PE 95% CI PE 95% CI
LSTday
>300 K 0.59 0.55, 0.63 0.18 0.06, 0.53
on

>305 K 0.64 0.60, 0.68 0.14 0.03, 0.57

LSTnight 0.95 0.94, 0.96 1.00 0.99, 1.01
N

Altitude 1.001 0.999, 1.002

Land use
Forest 1.45 1.33, 1.58 0.12 0.02, 0.77
Build 0.70 0.68, 0.73 0.81 0.31, 2.00
Crop 10.65 8.84, 12.83 5.44 0.47, 86.53
Distance to water bodies (km)
1-2 1.04 0.97, 1.11 0.71 0.29, 1.78
2-3 1.25 1.18, 1.33 1.50 0.59, 3.71
3-4 2.11 1.97, 2.27 1.16 0.36, 2.98
>4 1.65 1.58, 1.72 1.37 0.63, 3.15
Vegetation
NDVI >0.35 1.53 1.48, 1.59 1.23 0.61, 2.64
EVI >0.20 1.21 1.17, 1.26
LAI 1.013 1.011, 1.014
Rainfall 1.00 1.00, 1.00 0.78 0.39, 1.28

Φ
Spatial correlation parameters

σ2
2303.0 136.8, 4486.0
0.88 0.56, 1.5
PE, posterior estimates; CI, confidence interval; LST, land surface temperature; NDVI, normalized difference vegetation index; EVI, enhanced vegetation index; LAI, leaf area index. Regression coefficients represent-
ing odds ratios estimated by the median of the posterior estimates of the corresponding coefficients.

[page 166] [Geospatial Health 2015; 10:306]

gh-2015_1.qxp_Hrev_master 29/10/15 07:54 Pagina 167

Article

is straightforward. The rest of the parameters do not have full condi- and Benue and in isolated parts of the northeastern and northwestern
tional distributions of known forms. We simulated from the non-stan- parts of the country. Isolated patches of low malaria prevalence were
dard distributions by employing a random walk Metropolis Algorithm scattered around the country with northern Nigeria having more such
(Tierney, 1994) having a normal proposal density with a mean equal to areas. The middle belt regions had general malaria prevalence of 40%
the estimate of the corresponding parameter from previous Gibbs iter- and above. This pattern of malaria distribution is shown in Figures 2
ation and a variance equal to a fixed number, iteratively adapted to and 3 with the standard deviation (SD) of the median malaria preva-
optimize the acceptance rates. We ran five chains with a burn-in of lence ranging from 0.1 to 0.32.
5000 iterations. Convergence was assessed by inspection of ergodic
averages of the selection model parameters. The analysis was imple-
mented using Fortan 95 (Compaq Visual Fortran Professional 6.6.0)
and standard numerical libraries supported by the Numerical
Algorithms Group (NAG) Ltd (http://www.nag.co.uk/).
The model was used to predict malaria risk throughout Nigeria. This
approach treats the malaria risk at a new location as random and cal-
culates its predictive posterior distribution, which not only provides a
single risk estimate, but also gives a whole range of likely values
together with their probabilities to be the true values at a specific loca-
tion. This makes it possible to estimate the prediction error, a substan-

ly
tial advantage over the classical Kriging methods. We estimated the
predictive posterior distributions at prediction locations by simulation.

on
se
Results
The univariate non-spatial analysis indicated that the environmental
lu
factors, NDVI, EVI, LAI, distance to water bodies, rainfall, LST for day
and night and land-use, were related to malaria prevalence. Malaria
a
prevalence in Nigeria, estimated at 46 locations, had a median of about
ci

45% with values ranging from 0 to over 70%. The results of the bivariate
er

associations between malaria prevalence and environmental indicators

arising from non-spatial and spatial logistic models are shown in Table
Figure 1. Malaria prevalence at the various survey locations in
m

2. Nigeria.
Estimates of the odds ratios (OR) indicate that in the non-spatial
m

analysis, all the environmental covariates were related to estimated

malaria prevalence after adjusting for other risk factors. All covariates
co

significant at the 5% significance level and having low area under the
curve (AIC) values were included in the spatial analysis. The relation
between malaria risk and rainfall was linear. The logarithmic transfor-
on

mation of NDVI described its relation with malaria risk the best.
Second order polynomial terms for minimum temperature, maximum
temperature and distance to water bodies gave the best associations
N

with the malaria prevalence. The spatial model suggests that only
LSTday and land use (forest) were related to malaria. Surprisingly, the
models estimated a positive relation with the distance to water bodies,
implying that the risk of malaria increased with the distance from per-
manent water bodies until about 4 km and then decreased. The func-
tion 3/rho*100 gives the minimum distance in km with a spatial corre-
lation less than 5%.
Figure 1 shows the malaria prevalence at the various survey loca-
tions in Nigeria. There were relatively fewer surveys from northern
Nigeria compared to the South. The southern part of the country also
generally had more areas with malaria prevalence above 70%. This
agrees with the generally held view considering that the southern part
has more water bodies, is heavily forested and has more rain than the
North.
The malaria risk maps (Figures 2 and 3) estimated from the spatial
model shows that malaria prevalence in Nigeria varied from less than
20% in certain areas to over 70% in others. The highest prevalence
rates of 70% and above were found in the Niger Delta states of Rivers
Figure 2. Median of the malaria prevalence across Nigeria.
and Bayelsa, the areas surrounding the confluence of the Rivers Niger

[Geospatial Health 2015; 10:306] [page 167]

gh-2015_1.qxp_Hrev_master 29/10/15 07:54 Pagina 168

Article

hardly any data and for which future surveys should be done. These
Discussion maps rely on predictions of risk at locations without observed preva-
lence data. The prevalence estimates from both the non-spatial and
Malaria is an environmental disease and environmental factors are spatial approaches have broad agreement, though the spatial estimates
good predictors of transmission, but the relationship between environ- tend to have larger standard errors. This may be due to the fact that in
mental factors, mosquito abundance and malaria prevalence is not lin- the non spatial analysis there is the assumption of independence
ear. This relationship can be established only by means of adequate, which was not upheld. There is still much to be done in the area of
spatial statistical models, which can be used for improving predictions malaria control and elimination in Nigeria. The data collection for this
of malaria transmission not only in space (for risk mapping) but also work was difficult, primarily due to the fact that malaria prevalence
in time (for developing early warning systems for malaria epidemics). data are few and old. The data that were used to produce the prevalence
The Bayesian model that was used has the advantage that it is age- maps were collected from 1983 to 2007 with about 60% of these surveys
adjusted and makes use of all the survey data available, so I did not done before 2000, which means that these maps may not accurately
have to discard any surveys because of inappropriate age groups. The capture the current malaria prevalence picture of Nigeria.
Bayesian approach also allows flexible model fitting and estimation The maps shown here (Figures 2 and 3) belong to a series of maps
and mapping of the prediction error. The prevalence map produced that have been produced showing malaria prevalence in Nigeria.
(Figures 2 and 3) broadly corresponds to the known distribution of Previous malaria maps (Kleinschmidt et al., 2000; Gemperli et al., 2003;
malaria in Nigeria and in particular indicate high transmission in the Gosoniu et al., 2006) concern all of West Africa and that of Gething et
areas around the main rivers. The result of the survey showed that al. (2011) for the whole world. A comparison of the estimated malaria
prevalence shown in the maps in this paper with those produced by

ly
NDVI, EVI, LAI, distance to water bodies, rainfall, LSTday, and land-use
were related to malaria prevalence with rainfall having a linear rela- Kleinschmidt et al. (2001a) for West Africa reveals similar patterns, but

on
tionship with malaria risk. The models, however, estimated a positive the predicted prevalence in the present maps show more regions with
relation with the distance to water bodies up to 4 km before decreasing. prevalence above 70% and below 30%. There is, however, agreement
This could be due to the fact that even if people like to live near perma- between the maps, with both showing high risk for malaria in the

se
nent water bodies, they do not usually live on the banks but rather region of central Nigeria. The maps shown here, however, are in dis-
about 500 m away. This implies that water collected in cans, open jars, agreement with the map produced by Gemperli et al. (2006b). Their
map shows only two regions of Nigeria with prevalence of malaria
pits and drainage canals – which serve as vector breeding grounds
lu
– will mostly be found at distances at least 500 m away from the banks. above 70%. These are in the far north in the area of Kano State and in
Thomas et al. (2013) found that 95% of female Anopheles gambiae the southwestern tip of the country in the area of Lago state. These dif-
a
moved within 1.67 km from the nearest breeding site, which means ferences may be due to the use of different dataset and also the fact
ci

that from about 500 meters from permanent water bodies, malaria
prevalence will increase for as far as the flight range of the female
er

anopheles mosquitoes before decreasing. Thomson et al. (1997);

Kleinschmidt et al. (2000); Carter et al. (2000); Gosoniu et al. (2006);
m

Gemperli et al. (2006) also found this relationship. Indeed,

m

Kleinschmidt et al. (2000), in their work on the analysis of malaria

prevalence in Mali, found that malaria risk was estimated to be reduced
co

to a level lower than that measured close to water only for distances of
more than 4 km away from the nearest permanent water body. While
vector abundance is supposed to be high closer to the breeding sites,
on

Carter et al. (2000) found a negative association between malaria

infection and vector abundance. They attributed this to the propensity
of people to use bednets (Thomson et al., 1997) or the stimulated devel-
N

opment of immunity during early childhood in high-risk areas (Thomas

and Lindsay, 2000). Most malaria surveys include people from areas of
several km2, so surveys close to water bodies may include some people
from the riverbank and some further away. It is therefore not obvious
what relationship to expect with respect to distance to water. The exact
relationship between proximity to rivers and malaria appears to be very
sensitive to which data points are included and to the details of the
model, especially when there are very few data points in the critical
areas. It may also be that the lack of adjustment for age in earlier mod-
els biased some of the covariate effects.
Empirical maps of malaria risk are important tools in malaria control
as they can guide interventions and help assess their effectiveness.
Maps are useful in malaria control by identifying areas according to
their degrees of risk, thereby guiding resource allocation. They are
easy to understand and can therefore be appreciated by people of dif-
ferent educational levels. With accurate maps the effectiveness of
malaria control measures can be evaluated after a few years, by com-
paring the baseline map with subsequent maps. By identifying survey Figure 3. The standard deviation from the median of malaria
points it was possible to show areas of the country, for which there are prevalence in Nigeria.

[page 168] [Geospatial Health 2015; 10:306]

gh-2015_1.qxp_Hrev_master 29/10/15 07:54 Pagina 169

Article

that their map was for malaria prevalence in children ten years old and 11:1032-46.
below. The Gemperli map also has 80% of the Nigerian population liv- Gemperli A, Vounatsou P, 2004. Fitting generalized linear mixed models
ing in places with malaria prevalence of 40% and below. The map pro- for point-referenced data. J Mod Appl Stat Methods 2:497-511.
duced here, on the other hand, shows that 50% of the Nigerian popula- Gemperli A, Vounatsou P, Sogoba N, Smith T, 2006b. Malaria mapping
tion live in places with malaria prevalence of 50% and above. The using transmission models: An application to survey data in Mali.
Gethin map, however, generally agrees with the map shown here as it Am J Epidemiol 163:289-97.
shows that virtually every region of Nigeria with a P. falciparum para- Gething PW, Patil AP, Smoth DL, Guerra CA, Elyazar IRF, Johnston GL,
site rate of at least 50% with the southern tip of the country having a Tatem AJ, Hay SI, 2011. A new world malaria map: Plasmodium
rate of 70% and above as shown here. Falciparum endemicity in 2010. Malaria J 10:378.
Gosoniu L, Vounatsou P, Sogoba N, Smith T, 2006. Bayesian modeling
of geostatistical malaria risk data. Geosp Health 1:127-39.
Guinovart C, Navia MM, Tanner M, Alonso PL, 2006. Malaria: burden of
Conclusions disease. Curr Mol Med 6:137-40.
Hay SI, Rogers DJ. Toomer JF, Snow RW, 2000. Annual Plasmodium fal-
A national risk map for Nigeria will allow planners to identify malar- ciparum entomological inoculation rates (EIR) across Africa: liter-
ia high-endemicity areas, where long term use of insecticide treated ature survey, internet access and review. T Roy Soc Trop Med H
nets (ITNs) would be useful. Two distinct areas that can be identified 94:113-27.
with the aid of these maps are epidemic-prone areas and malaria- Kitron U, 2000. Risk maps: transmission and byrden of vector-borne

ly
endemic areas. These areas require distinct intervention measures. diseases. Parasitol Today 16:324-5.
For the epidemic areas, surveillance, indoor spraying of insecticides Kleinschmidt I, Bagayoko M, Clarke GPY, Craig M, LeSueur DA, 2000. A

on
would be helpful in controlling malaria. In the endemic areas, wide- spatial statistical approach to malaria mapping. Int J Epidemiol
spread use of ITNs, behavioural changes, such as avoiding storing 29:355-61.
water in open cans outdoors and clearing of bushes around dwelling Kleinschmidt I, Omumbo J, Bret O, Van de Giesen N, Sogoba N, Mensah

se
places would help. Planners can also assess the possible health impacts NK, Windmeijar P, Moussa M, Teuscher T, 2001a. An emperical
of measures aimed at improving food security through the promotion lu malaria distribution map for west Africa. Trop Med Int Health
of large-scale irrigation and wetland management projects. Road con- 6:779-86.
struction companies should be requested to fill up construction ponds Kleinschmidt I, Sharp BL, Clarke GPY, Curtis B, Fraser C, 2001b. Use of
(burror pits) once they are no longer needed. Finally, the maps con- generalized linear mixed models in the spatial analysis of small-
a
structed will also guide public health researchers in identifying appro- area malaria incidence rates in KwaZulu, Natal, South Africa. Am
ci

priate study environments for intervention trials as well as assist with J Epidemiol 153:1213-21.
identification of populations potentially benefiting from new interven- Okonko IO, Soleye FA, Amusan TA, Ogun AA, Udeze AO, Nkang AO,
er

tions. Ejembi J, Faleye TOC, 2009. Prevalence of malaria Plasmodium in

Though the data used for this study came from an imperfectly sam- Abeokuta, Nigeria. Malays J Microbiol 5:113-8.
m

pled population of Nigeria, this study nevertheless is the first known Rogers DJ, Randolf SE, Snow RW, Hay SI, 2002. Satellite imagery in the
m

attempt to produce a malaria risk map for Nigeria based entirely on study and forecast of malaria. Nature 415:710-5.
malariometric data. It is anticipated that it will provide useful addition- Ribeiro JMC, Seulu F, Abose T, Kidane G, Teklehaimanot A, 1996.
co

al guidance to control programme managers, and that it can be refined Temporal and spatial distribution of anopheline mosquito in an
once sufficient additional data become available. Ethiopian village: implications for malaria control strategies. B
World Health Organ 74:299-305.
on

Snow EW, Omumbo JA, Lowe B, Molineaux CS, Obiero JO, Palmer A,
Weber MW, Pinder M, Nahlen B, Obonyo C, Newbold C, Gupta S,
References Marsh K, 1997. Relation between severe malaria morbidity in chil-
N

dren and level of Plasmodium falciparum transmission in Africa.

Berquist NR, 2001. Vector-borne diseases: new trends in data collection Lancet 349:1650-4.
and risk assessment. Acta Tropica 79:13-20. Thomas CJ, Cross DE, Bøgh C, 2013. Landscape movements of
Carter R, Mendis KN, Roberts D, 2000. Spatial targeting of interven- Anopheles gambiae Malaria vector mosquitoes in rural Gambia.
tions against malaria. B World Health Organ 78:1401-11. PLoS One 87:e68679.
Craig HM, Snow RW, Le Sueur D, 1999. A climate-based distribution Thomas CJ, Lindsay SW, 2000. Mapping and estimating the population
model of malaria transmission in sub-Saharan Africa. Parasitol at risk fromlymphatic filariasis in Africa. T Roy Soc Trop Med H
Today 15:105-11. 94:37-45.
Diggle PJ, Moyeed RA, Rowlinson B, Thomson M, 2002. Childhood Thomson MC, Connor SJ, 2000. Environmental information systems for
malaria in the Gambia: a case study in model-based geostatistics. the control of arthropod vectors of disease. Med Vet Entomol
J R Stat Soc 51:493-506. 14:227-44.
Gelfand AE, Smith AFM, 1990. Sampling-based approach to calculating Thomson MC, Connor SJ, Milligan P, Flasse SP, 1997. Mapping malaria
marginal densities. J Am Stat Assoc 85:398-409. risk in Africa: what can satellite data contribute. Parasitol Today
Gemperli A, 2003. Development of spatial statistical methods for mod- 13:313.
eling point-referenced spatial data in malaria epidemiology. PhD Tierney L, 1994. Markov chains for exploring posterior distributions.
Diss. University of Basel, Basel, Switzerland. Ann Stat 22:1701-28.
Gemperli A, Sogoba N, Fonjo E, Mabaso M, Bagayoko M, Briet O, USGS, 2014. Land processess distributed archive center (LP DAAC)
Anderegg D, Liebe J, Smith T, Vounatsou P, 2006a. Mapping malaria modis products. Available from: https://lpdaac.usgs.gov/dataset_
transission in West and Central Africa. Trop Med Int Health discovery/modis/modis_products_table/mcd15a3

[Geospatial Health 2015; 10:306] [page 169]

gh-2015_1.qxp_Hrev_master 29/10/15 07:54 Pagina 170

Article

Vaughan AM, Aly ASI, Kappe SHI, 2008. Malaria parasite pre-erythrocyt- map.wordpress.com/2010/04/01/world-health-organizations-health-
ic stage infection: gliding and hiding. Cell Host Microbe 4:209-18. mapper/
Weier J, Herring D, 2000. Measuring vegetation (NDVI & EVI) Nasa WHO, 2013. Expert Committee Report on Malaria. World Health
earth observatory. Available from: http://earthobservatory.nasa. Organization, Geneva, Switzerland.
gov/Features/MeasuringVegetation/measuring_vegetation_2.php WHO, 2014. Expert Committee Report on Malaria. World Health
Wijeyaratne P, 1999. Malaria prevention: lessons learned. Organization, Geneva, Switzerland.
Environmental Health Project Publ., Washington, DC, USA. World Resources Institute, 1995. African data sampler. World Resources
WHO, 2010. The HealthMapper database. Available from: https://health- Institute, Washington, DC, USA.

ly
on
se
a lu
ci
er
m
m
co
on
N

[page 170] [Geospatial Health 2015; 10:306]