Spatial statistical analysis, modelling and mapping of malaria in Africa

Estimates of the disease burden due to malaria in Africa show that the toll it is
exacting in terms of loss of life, episodes of serious illness, and impediment to
economic development is enormous. In many areas the situation has become worse
due to failing drugs, failing insecticides, failing health systems, large scale population
movements and possibly due to co-infection with HIV. On the other hand, recent
studies have shown that widespread use of insecticide treated bed nets has the
potential for making substantial inroads into this disease burden, particularly in areas
of high endemicity. Recording the geographical distribution of any major disease forms an important basis
for locating appropriate interventions for its control and a means to monitoring their
effectiveness. It also provides a possibility for identifying ecological factors with
which the disease may be associated. The objective of this thesis was to produce evidence-based maps of malaria
prevalence and incidence by means of spatial statistical modelling; to evaluate and
advance the application of methodology in the analysis of spatially correlated disease
data; and to undertake detailed analysis of malaria incidence for one particular area in
order to establish underlying patterns of malaria risk over space and time and in
relation to population, climatic and environmental factors. Altogether six individual
studies were carried out, which modelled malaria distribution at three different levels
of scale. These levels and their locations, were: regional level in sub-Saharan West
Africa, country level in Mali and district level in Ubombo and Ngwavuma in
KwaZulu Natal, South Africa. In the case of the regional and country maps, the
malariometric measure was parasite prevalence in children, obtained from the MARA
database. In the case of the district-level analysis, routinely recorded small area
malaria incidence data were used, which were obtained from the provincial malaria
control programme. Three of the studies modelled malaria distribution over space and
time. There are well-documented difficulties with the mapping of raw disease rates, since
such maps will be dominated by sampling variability and analyses based on them will
be flawed due to the lack of independence in the rates. Spatial statistical methods can
be used to overcome these difficulties, but these have rarely been applied in the
context of malaria distribution modelling. In this thesis two such approaches were
employed: 1) classical geo-statistical methods, based on variograms and generalised
linear mixed models, and 2) autoregressive models in a Bayesian context using
Markov Chain Monte Carlo (MCMC) methods. Some minor adaptations of the
methods have been suggested. The main findings of the studies carried out in this thesis were: Both classical geostatistical and autoregressive MCMC methods are feasible
for modelling malaria distribution and advantages and limitations of each
method have to be weighed up in a particular context. The development of
extensions to the MCMC spatial modelling approach to cater for point
referenced (as opposed to areal) spatial data will make this method more
generally applicable. The ability to adequately reflect the effects of random
errors comprehensively in the resulting map estimates is an important
advantage of the Bayesian modelling approach. It is feasible to produce evidence-based maps of transmission intensity, which
are a refinement of expert opinion maps, from parasite ratio surveys. Malariometric measures of transmission intensity (and their proxies) are often
highly correlated in space as well as in time and this must be taken into
account in any modelling, particularly at the short range scales. Due to strong spatial heterogeneity it is difficult to model malaria transmission
intensity without leaving considerable unexplained, residual variation, which
may be spatially correlated. It is therefore unsatisfactory to map model
predictions directly. One method of overcoming this problem is to produce a
map of kriged (interpolated) model residuals, and to add these to model
predictions which can then be mapped. In large heterogeneous regions, models
should be derived within ecological zones, and special smoothing methods should be employed in boundary areas between these zones, rather than
attempting to derive a single unified distribution model for the whole region. Spatial variation in malaria transmission intensity is significantly associated
with basic climatic factors in areas of endemic stable malaria and in areas of
epidemic unstable malaria, but the relationship is usually not straightforward.
However, an association between temporal variation in malaria transmission
intensity and variation in weather, whilst plausible, could not be proven in the
data that were analysed. Sharp increases in malaria caseloads in Kwa Zulu Natal appear to originate
mainly from areas of previously low incidence, whilst high incidence areas
have partly stabilized. This suggests a geographical expansion of malarious
areas, and the acquisition of clinical tolerance to disease in some individuals in
high incidence areas. The finding that adults in high transmission sub-regions
of the province experience lower incidence rates than teenagers, supports the
hypothesis of clinical immunity to infection in these relatively high incidence
areas. Children under five in the same area, experience the lowest incidence
rates compared to other age groups, possibly as a result of being more
adequately protected by vector control measures than older children and
adults. In areas of unstable fluctuating malaria transmission intensity, incidence in
individual localities is highly correlated to incidence at the same locality in
previous seasons. One of the maps (West Africa) that were produced in this thesis has already been put
to use in malaria control. The findings relating to Kwa Zulu Natal will be presented
directly to the provincial malaria control programme. Two of the six studies have
been published, three have been submitted for publication and one is being prepared
for submission, to ensure widespread dissemination of the findings. A number of future research questions arise out of this work. These are, amongst
others: Methodological development of Bayesian spatial modelling software,
particularly to accommodate point referenced spatial data. Further analysis using the MARA database to produce endemicity maps of
other regions in Africa. Prospective studies should be undertaken to assess the relationship between
malaria and weather changes in epidemic prone areas, with a view to further
exploring the feasibility of epidemic forecasting systems. Further investigation of factors that influence the acquisition of clinical
immunity in adults in areas of moderate transmission intensity; investigation
whether this is confirmed in similar areas elsewhere (e.g. Namibia, Botswana),
and whether it is supported by age specific differences in case-fatality rates.