Introduction The most up-to-date global estimates of childhood anaemia indicate that 293.1 million children aged 0.05), and for children aged 1–2 y (87 g/l) than for children aged 2+ y (99 g/l) (p 0.8 (Table 3). The risk of anaemia in children aged 1–4 y was consistently high across the entire study area, with maximal prevalence (>95%) in a large focus straddling the borders of Burkina Faso and Mali (Figure 4). Smaller sized foci of high prevalence of anaemia were also predicted for southern areas of Mali, central areas of Burkina Faso, northern areas in Ghana, and areas adjacent to Volta Lake in Ghana. Phi (ϕ) indicates the rate of decay of spatial autocorrelation and varied from 13.68 in model 3 to 14.80 in model 5. Therefore, after accounting for the effect of covariates in model 6, the radii of the foci were approximately 23 km (note, ϕ is measured in decimal degrees and 3/ϕ determines the cluster size; one decimal degree is approximately 111 km at the equator). 10.1371/journal.pmed.1000438.g004 Figure 4 Predictive geographical risk of anaemia in children aged 1–4 y, based on a model-based geostatistical Bernoulli model. 10.1371/journal.pmed.1000438.t002 Table 2 Associations with anaemia risk, based on model-based geostatistical Bernoulli models. Variable Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Male (versus Female) 0.09 (−0.15, 0.31) 0.09 (−0.15, 0.32) 0.09 (−0.15, 0.32) 0.09 (−0.14, 0.32) 0.10 (−0.14, 0.32) 0.09 (−0.14, 0.231) Number of members in householda −0.04 (−0.08, 0.01) −0.04 (−0.08, 0.007) −0.04 (−0.08, 0.006) −0.04 (−0.08, 0.01) −0.04 (−0.08, 0.007) −0.04 (−0.08, 0.007) Age in monthsa −0.03 (−0.05, −0.01) −0.03 (−0.05, −0.01) −0.03 (−0.05, −0.01) −0.03 (−0.05, −0.01) −0.03(−0.05, −0.01) −0.03 (−0.05, −0.01) Rural (versus urban) 0.69 (0.27, 1.13) 0.68 (0.26, 1.11) 0.69 (0.24, 1.13) 0.67 (0.24, 1.11) 0.67 (0.25, 1.10) 0.69 (0.27, 1.11) CIAF Groups Y and F (versus CIAF Group A) −0.19 (−0.79, 0.43) −0.19 (−0.76, 0.43) −0.17 (−0.76, 0.45) −0.20 (−0.78, 0.41) −0.19 (−0.75, 0.38) −0.02 (−0.77, 0.41) CIAF Group C (versus CIAF Group A) 0.11 (−0.44, 0.64) 0.10 (−0.41, 0.63) 0.13 (−0.42, 0.66) 0.10 (−0.44, 0.63) 0.11 (−0.41, 0.59) 0.10 (−0.42, 0.60) CIAF Group D (versus CIAF Group A) 0.67 (0.13, 1.19) 0.66 (0.15, 1.19) 0.66 (0.12, 1.19) 0.66 (0.13, 1.18) 0.66 (0.16, 1.14) 0.66 (0.15, 1.16) PfPR2–10 a 0.37 (0.17, 0.57) 0.32 (0.11, 0.53) 0.30 (0.04, 0.51) 0.36 (0.17, 0.56) 0.28 (0.03, 0.49) 0.29 (0.04, 0.50) Prevalence of S. haematobium monoinfectionsa 0.05 (−0.13, 0.25) Prevalence of hookworm monoinfectionsa 0.11 (−0.09, 0.33) S. haematobium/hookworm coinfectiona 0.23 (−0.05, 0.53) Prevalence of S. haematobium a 0.05 (−0.13, 0.25) 0.11 (−0.09, 0.32) Prevalence of hookworma 0.22 (0.03, 0.47) 0.35 (0.11, 0.67) Interaction: prevalence of S. haematobium × prevalence of hookworma 0.28 (0.04, 0.57) Intercept 1.61 (0.95, 2.37) 1.60 (0.94, 2.31) 1.59 (0.88, 2.25) 1.62 (0.91, 2.23) 1.64 (1.02, 2.28) 1.63 (0.97, 2.29) ϕ (rate of decay of spatial correlation) 14.24 (3.14, 19.73) 13.80 (2.80, 19.73) 13.68 (1.86, 19.68) 15.38 (3.23, 19.71) 14.80 (2.80, 19.71) 14.25 (2.01, 19.63) σ2 (variance of spatial random effect) 1.45 (0.97, 1.95) 1.45 (0.95, 1.97) 1.36 (0.91, 1.98) 1.32 (0.01, 1.94) 1.40 (0.91, 1.98) 1.39 (0.92, 1.97) DIC 2,584.4 2,583.5 2,587.8 2,586.5 2,583.7 2,572.5 Association values are given as posterior mean (95% Bayesian credible interval). a Variables were standardised to have mean = 0 and SD = 1. 10.1371/journal.pmed.1000438.t003 Table 3 Summary of validation statistics for predictive models of anaemia prevalence and haemoglobin concentration in Burkina Faso, Ghana, and Mali. Validation Measure Prevalence of Anaemia Haemoglobin Concentration Area under the ROC curve (95% CI) 0.82 (0.75, 0.88) 0.77 (0.69, 0.83) Mean errora 0.03 (4.88) −7.99 (9.36) Mean absolute errora 0.12 (18.57) 10.96 (12.83) Correlation 0.79 0.82 a The observed values were compared to the mean of the posterior distribution of the each predicted value of prevalence of anaemia and Hb. The estimates in parenthesis are the percentage of the overall mean attributed to the error estimate. CI, confidence interval; ROC, receiver operating characteristic. Predicted Mean Haemoglobin Concentration All individual-level variables except number of members in household, age in months, and single anthropometric failures (CIAF Groups Y and F) were significantly associated with mean Hb in all models tested (Table 4). While rural residences and two or more anthropometric failures were significantly and negatively associated with the mean Hb, there was a significant positive association with mean Hb in male children. As with the models of risk of anaemia, Pf PR2–10 was significantly associated with mean Hb in all models tested. At the 5% level neither S. haematobium nor hookworm infection was significantly associated with mean Hb; however, S. haematobium/hookworm coinfections, hookworm monoinfections, and hookworm prevalence of infection were negatively associated with mean Hb. Estimates presented in Figure 5 are the mean posterior predicted mean Hb values from model 6 (the model that yielded the lowest DIC); this model was able to predict Hb greater than 90 g/l with an AUC >0.7 (Table 3). Figure 5 shows overlapping similarities to the map showing the predicted risk of anaemia (Figure 4) in that areas where Hb was predicted to be lowest ( 95%). These results suggest that resources for the treatment of moderate to severe anaemia, such as iron supplementation, deworming, and blood for emergency transfusion, should be prioritised towards populations located in the clusters of high anaemia risk identified in this study. This study reveals that malnutrition plays a central role in preschool anaemia burden in West Africa. The model including malnutrition, PfPR2–10, and helminth coinfection (Model 6) indicates that almost 40% of anaemia cases in preschool children in 2011 would have been averted by improving the nutritional status of children. Socio-economic status is a well-known risk factor for anaemia and infection at small spatial scales [47], and our results show that rural households are at significantly increased risk of anaemia compared to urban households. The same model also underlines the role of malaria infection in preschool children anaemia burden in the West African region in that the proportion of anaemia attributable to malaria was approximately 15%. These results are supported by earlier findings in Kenya using individual-level data (14% for infected preschool-age children and 7% for the whole population) [17]. The risk of anaemia attributable to hookworm infection (4.2%) is comparable to that estimated for S. haematobium (3.7%) and is significantly increased in hookworm/S. haematobium coinfections. This is consistent with evidence suggesting that morbidity associated with these infections is more pronounced in individuals with multiple infections [21]. Hookworm and S. haematobium infections have the smallest attributable risks both because the relative risk for these factors is modest and more importantly because the frequency of their mean prevalence in the population is low compared to malnutrition and malaria. Nevertheless, these results suggest that hookworm and S. haematobium infections are also important in the aetiology of anaemia in preschool children in West Africa, and deworming should be included in programmes aimed at controlling anaemia in this age group. We calculated that a total of 6.7 million children aged 1–4 y in Burkina Faso, Ghana, and Mali are anaemic. Our regional- and country-level estimates of number of children with anaemia are in line with estimates recently put forward by the World Health Organization in the three study countries [1]. In that regard, our study generated an important cartographic resource, providing important new information about sub-national priority areas for targeting anaemia control in the region and the quantity of resources needed in those areas (Figure 6). Using Predictive Parasite Infection Maps to Model Anaemia Important uncertainties should be noted from the anaemia DHS datasets and the prediction surfaces for parasite infection used in our models, which are likely to be propagated through the modelling framework. The outcome input data from the DHS surveys (anaemia and Hb) were collected in different years (2003 for Burkina Faso and Ghana and 2006 for Mali), and the covariate input prediction surfaces for parasite infection (malaria and helminth predictive surfaces) were for 2007. In order to assess relationships between anaemia indicators and potential contributors, we assumed that there was no contraction in anaemia cases in the three countries between the year anaemia data was collected and 2007. Although this temporal disparity may not be so problematic in the case of the DHS data for Mali, it may be problematic for Burkina Faso and Ghana; an overestimation of effects in those countries could be observed particularly in areas where the effects of intervention efforts to control anaemia were substantial. However, the degree to which the observed relationships are obscured by past spatially variable intervention efforts is not quantified in the literature. A rigorous assessment of the uncertainty associated with the mapped outputs of the input African malaria map was undertaken by [36]. This assessment provides great confidence about the input surface for the countries in our study in that the probability of correct endemicity class prediction was highest in West Africa. In this region, uncertainty was most important in small areas in southwest Ghana and northwest Mali. These latter estimates adjust for population density (using the population-weighted index in uncertainty) and reflect the co-occurrence of both low density of PfPR2–10 surveys and large populations in these regions. Despite the fact that point predictions generated by the malaria model are reasonably accurate, the model was shown to underestimate the probability of PfPR2–10 taking low values. This means that in low endemicity areas the PfPR2–10 may be overestimated [36]. However, our study is located in countries where malaria endemicity is high, and therefore we do not expect this suboptimal performance to significantly affect the point values of malaria endemicity used in our models. Similarly, the results of uncertainty assessment for the helminth infection covariate surfaces give us great confidence about their use in our models. The predictive ability of endemicity class membership (<50% for S. haematobium infection, 10% for hookworm infection, and 5% for S. haematobium/hookworm coinfection) was moderately good, with all AUC values above 0.7 [37]. Finally, by using existing continental-level and other mapped layers as proxies of parasite infection, we have adopted an ecological approach to modelling anaemia prevalence and Hb. This approach was chosen because comparable individual-level infection data were not available for the study area. Instead, the mean prevalence of parasite infection was used as a proxy for the true infection status of preschool children included in the analysis. This approach provides a somewhat imprecise measurement of exposure to P. falciparum and helminth infection and therefore may result in regression dilution bias arising from imprecise exposure measurement, which is most likely to lead to underestimation of the observed effects of parasite infections [48]. Although the observed relationships are biologically plausible, in the absence of individually collected data it is not possible to know to what extent the magnitude of relationships represent an artefact introduced by ecological fallacy. Using Population Attributable Fractions to Determine the Role of Competing Factors in Anaemia We used PAFs to represent the fraction of the total anaemia risk in the population that would not have occurred if the effect associated with the contributor of interest were absent while distributions of other contributors in the population remained unchanged [48],[49]. The PAF estimates attributable outcome and not necessarily preventable outcome numbers, as it may not be possible to remove the risk factor from the population altogether. Hence the numbers may overestimate achievable impact and are therefore measures of potential impact. An alternative statistic could have been used, namely, the population impact of eliminating a risk factor (the potential number of disease events prevented in a population over the next t years by eliminating a risk factor) [50]. PAF estimation is of public health significance when the risk factors being investigated are clearly the most proximal in the causal pathway and when there is consensus that the exposure is amenable to intervention [38],[51]. The nutritional factors and infections included in our anaemia model are well known to be causally related to anaemia, but as outlined above, these do not represent the complete multifactorial nature of anaemia. Haemoglobinopathies and thalassemias are importance inherited haematological conditions, particularly in the population of West Africa [52], but predictive surfaces for the sickle cell trait have only recently become available [53]. This study adopted an ecological approach to anaemia modelling in that the true infectious status of children is assigned by spatially overlaying available mapped parasite endemicity surfaces. In doing so, the estimated relative risks for these factors are prone to regression dilution bias, which may contribute to more conservative PAF estimates. In the absence of comparable individual-level data, the practical and logical limitations of including surrogate factors in PAF estimation are not trivial to assess, but our results are consistent to the only study available using individual-level data [17]. Another issue related to the interpretation and public health relevance of a PAF concerns specification of the exposure group [51]. For PAF estimation we have retained the continuous nature of the parasite surfaces to enable spatial prediction across all the areas and to avoid arbitrary categorisation of parasite endemicity surfaces, which could yield reference levels with few or no observations, resulting in PAF estimates with low power. We calculated the PAF for the mean of each parasite surface in the region, which corresponds to the fraction of total anaemia risk in the population that would have been reduced had the children been living in areas where the mean prevalence of the risk factors was very close to zero. Full consideration of continuous covariates is theoretically possible and is a matter of statistical modelling, and PAF estimates (model-based) have been developed for continuous exposures [54]. Our PAF estimation may be extended in future work to estimate a more general measure than PAF, namely, the generalised impact fraction (the fraction reduction of anaemia risk that would result from changing the current distribution of the contributing factors to some modifiable distribution) [55]. However, to set the level of reduction of the risk factor would require evidence of the effectiveness of malnutrition and parasite interventions, which is not objectively available. Accuracy of the PAF estimates also depends on the representativeness of the input data from the population of interest and the completeness of the multivariable model. The DHS anaemia data are to the best of our knowledge the most complete and representative anaemia data available in the public domain. The anaemia data were collected using standardised methods and quality control protocols (see http://www.measuredhs.com/start.cfm). The input data used to produce smooth maps of malaria included 3,384 geo-positioned records where parasite rates had been diagnosed either using microscopy (2,764 [81.7%]) or rapid diagnostic tests (n = 587 [17.3%]) [36]. The schistosomiasis and hookworm data were obtained in nationally representative surveys using Kato Katz and urine filtration methods [41],[42]. In PAF estimation the multivariable model needs to be as complete as possible; if one or several factors act as true confounders of the association between exposure and disease, then the crude PAF estimates are in general biased and there is a need for adjustment when estimating the PAFs [55]. Regression models allow one to take into account adjustment factors as well as interaction of exposure with some or all adjustment factors [54]. We are confident in our statistical control of confounding by adjusting our analysis for age, sex, and socio-economic factors; we also considered interactions between proximal parasite infections. However, even if one uses adjusted estimates of the relative risk, PAF estimates can be biased in the presence of unaccounted confounding factors, and overestimation of PAFs can occur [49],[56]. Malaria endemicity values may be confounded by the presence of bed net usage, which in turn is known to be influenced by socio-economic status. We found collinearity between bed net usage and socio-economic indicators in the DHS data, which provided statistical support for the inclusion of socio-economic indicators only. Furthermore, these indicators are also related to a broader group of distal factors contributing indirectly to anaemia (e.g., water, sanitation, and deworming). The order of a variable in the causal pathway and the way it is entered in a multivariable model influence its PAF estimation [57]. The impact of different combinations of proximal infection contributors on the observed relationships with anaemia indicators was assessed by building different models (Tables 2 and 4). In so doing, we noticed the effect of variable order on the resulting coefficients, and PAF estimation was conducted based on the model with best statistical support for model complexity and fit to the data. Furthermore, indirect effects can be noticed when more distal factors impact proximal risk factors by increasing their rate or prevalence. Some of the anthropometric failures used in our models as proxies of malnutrition, and stunting, in particular, can be the result of an indirect effect of both parasite infections and malnutrition, but collinearity between these factors was not identified at variable screening. Finally, the PAFs refer not to the general population but rather to the study population in West Africa. The results generated from an adjusted PAF model for a specific population may not fit settings in other populations [49]. PAFs in other populations may differ because of varying prevalence of risk factors and the impact of additional socio-demographic factors that were not included in the original sample [56]. Accuracy of Geostatistical Anaemia Modelling and Potential Refinements The frequency distributions for the predicted anaemia and Hb surfaces cover substantially smaller ranges of values than those of the DHS input data. The resulting anaemia and Hb predictive surfaces are certainly smoother than the raw data from which they are predicted because the MBG modelling approach makes predictions at unsampled locations using linear associations between covariates and the DHS survey data. This smoothing effect (or interpolation) has important repercussions on the models' ability to accurately predict anaemia endemicity over very short distances. The models performed satisfactorily when predicting point values and endemicity classes of anaemia indicators. However, certain aspects of the uncertainty statistics are suboptimal in that the anaemia risk model tends to overestimate prevalence by 5% and the Hb model tends to underestimate Hb by 10 g/l. Nevertheless, despite the different sources of uncertainty that are embedded in the MBG modelling approach, the resulting predictive maps represent an important evidence base for operational managers of anaemia control in the region. The computational demands of the MBG modelling approach restricted the range of modelling procedures we could utilise to improve the predictive ability of the anaemia and Hb models. A number of potential improvements to the geostatistical approach could be employed in the following ways. First, future iterations of these maps should consider the incorporation of other covariates, particularly the assessment of the additional influence of inherited blood disorders (haemoglobinopathies and thalassemias) once these become available. Second, our approach could be updated once the existing mapped surfaces have been revised with the inclusion of diagnostic uncertainty into their modelling frameworks. This is particularly important for schistosomiasis in low transmission settings [58]. Third, prediction surface uncertainty around the predicted mean of infection covariates could be incorporated in the modelling framework by modelling the distribution of probable values using a beta distribution parameterized by the predicted posterior mean and the posterior standard deviation for each parasitological survey location. Fourth, the inclusion of spatial variation of spatial dependency in anaemia risk (non-stationarity) could be another possible refinement but was considered computationally infeasible. Future iterations of the present models could incorporate non-stationarity: models could assume separate regional fixed coefficients and include a series of random coefficient models incorporating different correlation structures. Fifth, the 5×5 km resolution may not have been sufficiently precise to classify exposures, and a reduced resolution could have been chosen at the expense of computational run time. For example, an urban-rural map of 5×5 km resolution may not be sufficiently precise to define clusters as rural or urban, since settlements may vary in size across the study area. Finally, infections considered here are known to cause multiple competing morbidities, and the methods presented here could be extended to investigate spatial heterogeneity of co-morbidities attributable to malaria and helminth infections. This would involve applying a multinomial analogue of the present model. Although analysis of the spatial variation in other childhood morbidity indicators, such as stunting, fever, pneumonia, and diarrhoea, has been attempted at the national scale in Malawi [59],[60] and Burkina Faso [61], and at the continental scale in the case of paediatric fevers associated to malaria infection [62], none of these studies have investigated the differential role of malnutrition and parasite infection metrics in prevalence of co-morbidities at a regional or continental scale. Conclusions The combination of anaemia and mean Hb predictive maps has allowed the identification of communities in West Africa where preschool-age children are at increased risk of morbidity. The use of anaemia maps as an alternative to aggregated country-level estimates has important practical implications for targeted control in the region and could contribute to the efficient allocation of nutrient supplementation programmes and delivery of fortified foods as well as the planning and evaluation of resource needs for geographical delivery of transfusion services for severe anaemia cases. This study shows that existing continental-level disease and other mapped layers can be used to predict anaemia risk. The development of maps indicating the geographical risk profile of anaemia controlling for malnutrition and major infections would allow assessment of the risk of anaemia due to different causes, which would in turn constitute an important evidence base to work out the best balance between interventions. In the future, these maps could be updated in subsequent methodological iterations to incorporate further modelling refinements. Supporting Information Text S1 Supplementary technical information tables. (0.11 MB DOC) Click here for additional data file.
