^{1}

^{2}

^{*}

^{3}

^{4}

^{4}

^{5}

^{6}

^{4}

^{7}

^{3}

^{8}

^{9}

Conceived and designed the experiments: ACAC. Analyzed the data: ACAC. Wrote the paper: ACAC. Coordinated the surveys and organised data entry: EBO AFG. Provided detailed comments on the draft: EBO MS AL RD MT GC AFG AF SB. Managed field survey teams: MS AL RD GC. Provided leadership for the national control programme: MT AF. Involved in the 1984–1989 surveys: MT. Collated the 1984–1989 data and provided guidance and direction on the statistical analysis: SB.

The authors have declared that no competing interests exist.

We investigated changes in the spatial distribution of schistosomiasis in Mali following a decade of donor-funded control and a further 12 years without control.

National pre-intervention cross-sectional schistosomiasis surveys were conducted in Mali in 1984–1989 (in communities) and again in 2004–2006 (in schools). Bayesian geostatistical models were built separately for each time period and on the datasets combined across time periods. In the former, data from one period were used to predict prevalence of schistosome infections for the other period, and in the latter, the models were used to determine whether spatial autocorrelation and covariate effects were consistent across periods.

A decade after the apparently successful conclusion of a donor-funded schistosomiasis control programme from 1982–1992, national prevalence of schistosomiasis had rebounded to pre-intervention levels. Clusters of schistosome infections occurred in generally the same areas accross time periods, although the precise locations varied. To achieve long-term control, it is essential to plan for sustainability of ongoing interventions, including stengthening endemic country health systems.

Geostatistical maps are increasingly being used to plan neglected tropical disease control programmes. We investigated the spatial distribution of schistosomiasis in Mali prior to implementation of national donor-funded mass chemotherapy programmes using data from 1984–1989 and 2004–2006. The 2004–2006 dataset was collected after 10 years of schistosomiasis control followed by 12 years of no control. We found that national prevalence of

Mali was one of the first countries in sub-Saharan Africa to initiate a national schistosomiasis control programme. Control efforts started regionally in 1978 in Dogon Country (region of Mopti) after the construction of small dams for growing vegetables, and became a national programme in 1982. During the first 10 years, the programme was run by the Malian Ministry of Health in partnership with the World Health Organization and the German Technical Cooperation (Deutsche Gesellschaft für Technische Zusammenarbeit, GTZ)

GTZ support for the programme ceased in 1992, with the government taking over financial responsibility. However, lack of resources led to control activities being considerably reduced and the implications of this for infection levels were not assessed in the immediate post treatment period. From 1998, a new, decentralised control programme was approved by the Ministry of Health but, due to lacking continuous financial support from the government, many planned activities were not implemented. In 2004, a new initiative to recommence national control activities was established with support from the Schistosomiasis Control Initiative (SCI;

The potential of using risk mapping to describe the spatial patterns of infections is now well-established, and has been demonstrated for a range of diseases including malaria

Much of this work to date has used data from a single geographical area at a single point in time to develop predictions for similar locations. Preliminary work has investigated the spatial extent to which risk models can be reliably extrapolated

In this paper, we use unique data on schistosome infections, available from two nationwide surveys conducted in Mali, the first undertaken during the 1980s prior to the implementation of the GTZ-supported national control programme and the second between 2004–2006, 12 years after this programme had ceased and prior to implementation of the SCI-supported programme. We aim to determine whether the overall prevalence and spatial distribution of schistosomiasis in Mali is different in 2004–2006 compared to the 1980s and to determine whether the spatial distribution, including covariate relationships with environmental variables and parameters that describe the spatial dependence structure (i.e. clustering), have changed in Mali over the last two decades.

A nationwide survey was carried out between May 1984 and May 1989 prior to implementation of the GTZ-supported programme (see Traoré et al.

A more recent nationwide survey was conducted in 194 schools (including 15,051 school-aged children) between December 2004 and May 2006. Ethical approval for these surveys was obtained from St. Mary's Hospital Research Ethics Committee UK and the National Public Health Research Institute's (INRSP) scientific committee in Mali. All data collection activities were carefully explained to, and oral consent was obtained from traditional authorities in the village (the village head and the elders), the schoolmaster, the representative of the pupils' parents and the local health authorities. Child participants were given an explanation of the data collection activities and were free not to participate if they so chose. Written consent was not obtained and oral consent was not specifically documented because the survey was considered by the UK and Malian ethical committees as part of the monitoring and evaluation of routine health activities carried out by the Malian Ministry of Health's national schistosomiasis control programme.

Survey protocols (available on request) instructed survey teams to select 30 boys and 30 girls per school using systematic random sampling. Schools were selected to maximise geographical coverage of the study area; all parts of Mali excluding the northern desert and far eastern regions, where transmission is known not to occur

For the current study, numbers tested and positive (defined as one or more eggs for each species of schistosome) were calculated for each survey location. School or community-level raw prevalence was then plotted in the GIS. Electronic data for land surface temperature (LST) and normalised difference vegetation index (NDVI) were obtained from the National Oceanographic and Atmospheric Administration's (NOAA) Advanced Very High Radiometer (AVHRR; see Hay et al.

Multivariable logistic regression models were developed for each species of schistosome and each of the two survey periods in a frequentist statistical software package (Stata version 10.1, Stata corporation, College Station, TX). Prelimary results were similar for each species of schistosome and each study period. A quadratic association between LST and prevalence was assessed and was found to be significant and DPWB was also significantly and negatively associated with prevalence. NDVI was not found to be significantly associated with prevalence in the preliminary multivariable models and was excluded from further analysis. Therefore, it was decided to enter LST (in quadratic form) and DPWB as covariates into the final spatial models. Bayesian geostatistical models, developed in WinBUGS 1.4 (Medical Research Council, Cambridge, UK and Imperial College London, UK), were identically structured for each species of schistosome and each study period. Statistical notation is presented in

Three chains of the models were run consecutively. A burn-in of 1,000 iterations was allowed, followed by 10,000 iterations where values for the intercept and coefficients were stored. Diagnostic tests for convergence of the stored variables were undertaken, including visual examination of history and density plots of the three chains. Convergence was successfully achieved after 10,000 iterations in each model and the posterior distributions of model parameters were combined across the three chains and summarized using descriptive statistics. Geostatistical prediction across Mali was done in WinBUGS using the

To compare predictions accross time periods, the 1984–1989 model was used to predict infection prevalence at the 2004–2006 survey locations and

A stationary model is one where the parameters that define the spatial dependence structure are the same for the two time periods and a non-stationary model is one where the parameters are different (note we refer to stationarity across time periods, not different parts of the study area). Models were developed using the combined datasets, including with different intercepts for each time period and: 1) different coefficients, spatial dependence parameters and random effects (i.e. assuming separate sub-models for each time period); 2) the same coefficients but different spatial dependence parameters and random effects (i.e. allowing the sub-models to have common covariate effects); 3) the same coefficients and spatial dependence parameters but different random effects (i.e. allowing common covariate effects and stationary spatial dependence structures, but separate predicted risk surfaces); and 4) the same coefficients, spatial dependence parameters and random effects (i.e. a single model giving an overall predicted risk surface across the two time periods). Models 1 and 2 were non-stationary models and models 3 and 4 were stationary models. Statistical notation is presented in

The best-fitting model (of 1–4) was selected using the deviance information criterion (DIC). An additional comparison of the spatial distribution of schistosomiasis accross time periods was done by subtracting predicted prevalence from the best-fitting

The national prevalence of infection with

The Bayesian geostatistical models for each time period are presented in

Variable | ||||

1984–1989 | 2004–2006 | 1984–1989 | 2004–2006 | |

OR: DPWB | 0.50 (0.32,0.71) | 0.50 (0.25,0.91) | 0.59 (0.34, 0.94) | 0.49 (0.20,0.95) |

OR: LST | 1.41 (1.02,1.85) | 0.62 (0.33,1.02) | 0.48 (0.28, 0.84) | 0.40 (0.19,0.71) |

OR: LST^{2} |
0.96 (0.79,1.14) | 1.06 (0.81,1.36) | 0.88 (0.66, 1.13) | 1.02 (0.69,1.48) |

Intercept | −1.73 (−2.15,−1.35) | −1.42 (−2.33,0.23) | −5.40 (−6.29, −4.60) | −6.13 (−7.18,−5.25) |

Phi ( |
5.38 (3.69,7.60) | 1.68 (0.96,2.60) | 6.09 (2.94, 12.04) | 9.02 (2.01,54.25) |

Sill | 3.20 (2.43,4.26) | 8.24 (5.44,12.79) | 6.67 (4.55, 9.98) | 9.42 (5.66,15.79) |

Models developed on 1984–1989 and 2004–2006 data were generally able to discriminate infection prevalence for the other dataset to an acceptable level (

Observed prevalence threshold | ||||

Using 1984–1989 data to predict 2004–2006 status | Using 2004–2006 data to predict 1984–1989 status | Using 1984–1989 data to predict 2004–2006 status | Using 2004–2006 data to predict 1984–1989 status | |

≥50% | 0.70 (0.62, 0.78) | 0.73 (0.66, 0.79) | 0.81 (0.71, 0.92) | 0.93 (0.84, 1.00) |

≥20% | 0.73 (0.65, 0.80) | 0.72 (0.66, 0.78) | 0.78 (0.64, 0.91) | 0.86 (0.78, 0.95) |

≥10% | 0.78 (0.71, 0.84) | 0.74 (0.68, 0.80) | 0.82 (0.72, 0.91) | 0.79 (0.70, 0.87) |

>0% | 0.82 (0.73, 0.91) | 0.82 (0.70, 0.95) | 0.70 (0.62, 0.78) | 0.67 (0.60, 0.73) |

The deviance information criterion for models 1–4, for

Model | ||

1) Different coefficients and spatial structure | 2952.4 | 1352.2 |

2) Same coefficients, different spatial structure | 2947.5 | 1351.9 |

3) Same coefficients and spatial structure | 2949.9 | 1346.5 |

4) Data grouped, with single overall prediction | 2950.1 | 1347.6 |

Variable | Posterior mean (95% posterior interval) |

OR: DPWB | 0.51 (0.39, 0.67) |

OR: LST | 1.33 (1.02, 1.77) |

OR: LST^{2} |
0.95 (0.74, 1.12) |

Intercept: 1984–1989 | −1.72 (−2.11, −1.34) |

Intercept: 2004–2006 | −1.37 (−2.17, −0.71) |

Phi ( |
5.60 (3.59, 8.24) |

Phi ( |
6.82 (1.77, 45.75) |

Sill: 1984–1989 | 3.17 (2.42, 4.27) |

Sill: 2004–2006 | 6.35 (4.26, 9.70) |

Variable | Posterior mean (95% posterior interval) |

OR: DPWB | 0.57 (0.35, 0.82) |

OR: LST | 0.45 (0.31, 0.65) |

OR: LST^{2} |
0.92 (0.71, 1.15) |

Intercept: 1984–1989 | −5.39 (−5.99, −4.71) |

Intercept: 2004–2006 | −5.84 (−6.59, −5.18) |

Phi ( |
6.47 (3.28, 16.57) |

Sill | 7.15 (5.17, 9.86) |

Spatial predictions (showing the mean of the posterior distributions for predicted prevalence) based on the best model for each type of schistosome infection are presented in

Predictions are based on a non-stationary Bayesian geostatistical model.

Predictions are based on a stationary Bayesian geostatistical model.

Comparative maps show predicted prevalence in 1984–1989 subtracted from predicted prevalence in 2004–2006, using the best-fitting models (

Predictions for

Despite differences in survey design and study population between the time periods, this study demonstrated remarkable similarities in the spatial distribution of prevalence of infection with

Regarding the sampling strategies, the data were based on different sample locations, collected for different purposes and from different populations. The data from 1984–1989 were collected from the general population including adults, whilst the 2004–2006 data were from school-aged children. Age-stratified prevalence and intensity of

The 1984–1989 surveys had a less uniform geographical distribution than the 2004–2006 surveys, which is not surprising given that the 1984–1989 surveys were not explicitly designed with subsequent spatial analysis in mind, whereas uniform geographical coverage was an aim of the survey design for the 2004–2006 study to facilitate spatial analysis. Investigation of the impact of different sampling strategies on observed spatial correlation is an area of future research.

Factors potentially related to changing epidemiology include desertification, urban growth and rural-urban migration

In addition to the limitation of different survey designs between periods, we were not able to compare spatial variation in intensity of infection between time periods because location-specific mean egg counts were not available from the 1984–1989 surveys. Maps of intensity would be useful for determining any changes in transmission across the periods. Examination of a single urine slide or single stool sample as a diagnostic approach results in sub-optimal sensitivity and this will also have affected the accuracy of our maps. We also did not incorporate anisotropy (where the spatial correlation structure varies by direction) or non-stationary spatial variation between different parts of the country, within each time period; these are future potential refinements of the models. We should also point out that the model predictions are distributions and here we have only presented the posterior mean. Examination of the full posterior distribution of predicted prevalence enables assessment of uncertainties arising from sampling and measurement error (including in the model covariates). We have recently described how an understading of these uncertainties can assist decion making in schistosomiasis control programme planning

Our results show that, while there were differences in the raw data, the overall prevalence of neither

One of the most important conclusions arising from the current work is that it is essential to develop a sustainability strategy to ensure ongoing benefits from the current national control programme. Recognising this fact, SCI has developed a sustainability plan which is outlined in Fenwick et al.

The maps presented here can be used to target what are likely to be more limited national resources in the longer term to the highest-risk areas, where they will have the greatest impact on infection, morbidity, and (hopefully) transmission. The current move towards integration of control of neglected tropical diseases means that the government may have the opportunity to implement a cost effective control programme encompassing schistosomiasis, soil transmitted helminth infections, lymphatic filariasis, river blindness and trachoma. It is clear that a commitment from the Malian government and international donors for substantial resources is required long into the future, or alternative strategies need to be found, if control of schistosomiasis transmission in Mali is to be achieved.

STROBE Checklist

(0.08 MB DOC)

Statistical notation of Bayesian geostatistical models for prevalence of

(0.03 MB DOC)

Statistical notation of Bayesian geostatistical models of prevalence of

(0.07 MB DOC)

We thank the many field technicians, local administrators and teachers involved in the surveys, and the children who provided urine and stool samples in the 2004–2006 surveys. We also thank the Malian Ministry of Education, in particular Mrs Fatoumata Keita, for facilitating the surveys. Dr Adrian Barnett, Queensland Institute of Technology, assisted with mathematical notation of the models.