Mpumalanga Province, which borders Mozambique and Swaziland, is a predominantly rural area with a population of approximately four million people. Malaria occurrence in this area is seasonal during the wet and humid months of October to April. Ninety-five percent of malaria cases are due to Plasmodium falciparum infection. This area has historically experienced malaria outbreaks and epidemics with relatively high mortality 2930. The main vector in this region is Anopheles arabiensis and vector control is predominantly by indoor residual spraying (IRS) with DDT 3132.

The seven rural towns that experience the highest malaria risk in Mpumalanga were included in this study (Figure 1)33, with malaria incidence ranging from 30 to 61 per 1,000 persons between 1997 and 2005.

Original town maps acquired in 2000 were upgraded from digitized aerial photographs produced in 2002 (Cadnet, Nelspruit, South Africa) and updates were converted into MapInfo version 6.4 files (MapInfo Corporation, New York, USA) with Geomedia software (Symmetry Systems Inc., New York, USA). Unique stand (household) numbers were allocated during the digitizing process. A stand may include one or more structures belonging to a family unit on a designated piece of land.

Malaria case investigators followed up all malaria cases reported at local health facilities for these towns as part of malaria control program activities for the 2002/2003, 2003/2004 and 2004/2005 seasons. A malaria case is defined as a person diagnosed either by positive rapid test or blood slide. An investigator would identify the probable source of infection by confirming the travel history of the patient recorded on the notification.

SaTScan™ software, version 5.1.3, using the Kulldorf method of retrospective space-time permutation and the Bernoulli purely spatial model 3435, was used by malaria control programme operational staff to detect malaria clusters in individual towns in the three seasons under investigation and over the combined time period. Individual malaria cases were used and recorded against source of infection households. Households who did not seek health care for malaria or tested negative during active case detection were used as controls for the Bernoulli method. More than one case can be reported per household. A household in these areas is typically defined as a family unit with a single land owner.

The circular scan statistic is isotopic with respect to the rotation of the geographical area http://www.satscan.org. This method has previously been validated for plotting and understanding local malaria time-space-clusters 2136373839.

Observed cases in a cluster were compared to the distribution of expected cases if spatial and temporal locations of all cases were independent. The model adjusts for entirely spatial or entirely temporal clusters. With spatial adjustment, time remained dormant and during temporal analysis seasons were considered. The distribution and statistical significance of the clusters were explored by means of Monte Carlo replication of data sets under the null hypothesis with replications greater than 999 to ensure adequate power for defining clusters http://www.satscan.org. Clusters were prioritized for public health action according to their statistical significance. Hard copy town maps showing cluster areas were distributed to malaria field staff investigating cases allowing the coordination of intervention efforts within communities.

Four hundred and twenty two malaria cases were notified during the three seasons from 341 households across the seven towns. Malaria incidence differed significantly between the towns during the three seasons (χ² = 6.442, p = 0.040) (Figure 2).

SaTScan analysis detected a number of clusters during the study period (Table 1). Albertsnek produced two space-clusters over the three-season period using the Bernoulli model, one in the northern (log likelihood ratio = 12.308, p = 0.003) and one in the south eastern part of town (log likelihood ratio = 12.187, p = 0.003) (Figure 3A). Three additional space-clusters were observed; in Goba (log likelihood ratio = 12.226, p = 0.001), Mbuzini (log likelihood ratio = 22.372, P = 0.001) and Thambokulu (Figure 4A) (log likelihood ratio = 22.372, p = 0.001). Only two towns, Albertsnek (test statistic = 5.548, p = 0.007) and Thambokulu (test statistic = 3.668, p = 0.004), had space-time clusters and both were during the 2004/2005 season (Figures 3B and 4B).

The third season for case cluster identification, 2004/2005, overlapped with the first season of implementation of an outbreak identification and response system in the Nkomazi area 29. Thambokulu and Albertsnek triggered outbreak declarations during this season (Table 1). Analysis demonstrated clusters in both towns during this season. Real time cluster detection, monitoring as cases occur, triggered at the same time as the binomial outbreak identification thresholds 29.

The physical plotting of cases by household during this season (2004/2005) as part of the cluster identification activities at district level and on hard copy maps provided a visual distribution of risk in the towns and assisted with the timely planning of public health activities especially logistics, including active case detection, early diagnosis and treatment of positive cases in the areas of the clusters, additional indoor residual spraying, focal larviciding and health promotion activities.