To our knowledge, this is the first study using health centre catchment areas as spatial units for the spatio-temporal analysis of MM over a whole sub-Saharan country. The study's first finding was the more frequent detection of spatial clusters within nine southern districts, mainly on the southern border with Nigeria. Second, clusters most often encompassed only a few HCCAs within a district, without expanding to the entire district. In addition, no consistent annual spatio-temporal pattern for cluster emergence and epidemic spread could be observed, thus precluding the capacity to predict where the next epidemic would break out, and what geographical direction it would follow. These findings rely on laboratory-based data and have important public health implications as discussed hereafter.
The first asset of this study was the quality of the microbiological data. We used laboratory-confirmed N. meningitidis
cases data, coming from a surveillance system managed by CERMES and DSSRE throughout the country. Most other spatio-temporal studies on meningitis epidemics in sub-Saharan Africa 
are based instead on suspected cases reported in the framework of the national surveillance systems. In our dataset, none of the three typical bacterial aetiologies (N. meningitidis
, S. pneumoniae
and H. influenzae
) could be identified in almost 60% of the CSF analysed by CERMES over the study period (see ). Relying only on suspected cases would therefore introduce a large number of misclassified cases. However, our system may suffer from underreporting from areas where performing a lumbar puncture and shipping the samples to CERMES may represent logistical difficulties. Further analyses (not shown here) have documented that indeed the districts the most remote from CERMES (in Maradi and Zinder regions) were sending less CSF specimens than the closer ones, for a similar number of suspected cases notified to DSSRE. However, the proportion of negative cases among the received CSF specimens was fairly similar among the healthcare centres (outside the capital Niamey). This suggested that the decision to take or not a CSF sample from a patient based on clinical criteria had no significant spatial variability. Moreover, our cluster analyses enabled us to detect the importance of remote regions in the epidemic dynamics according to the recurrent clusters identified there. Like in many other settings, the surveillance system may not cover the entire population of Niger affected by meningitis. However, we can reasonably assume that most meningitis cases, because of their severity, end up reaching the healthcare centres, with or without prior self-treatment or consultation of a tradi-practitioner. Moreover, free healthcare offered to all people suffering from meningitis in Niger probably reduces social and spatial disparities in care-seeking behaviours. Thus, for all the reasons above, we are confident that the surveillance system is representative enough and that underreporting did not substantially affect the validity of our results, which are more likely to reflect the dynamism peculiar to meningitis than the spatial disparities in the surveillance system efficiency. Incidence estimates were based on the 2001 census and constant population growth rates were applied. We could not take into account possible variations of population growth rate over time and space, due to the difficulty in quantifying population migrations.
The second asset of this study was the use of HCCAs as spatial units for the spatio-temporal analysis of MM. They represent a more accurate spatial unit of analysis than the district level on which reactive vaccination strategies and spatio-temporal studies are usually based 
. Analysing data at the HCCA level has greater relevance for understanding the epidemic dynamics, for making decisions in response to starting epidemics and for assessing control strategies.
Indeed, this study has shown that clusters most often included only a few HCCAs within a district. This finding, previously suggested by 
, is important for understanding meningitis epidemics and should encourage surveillance at the health centre level. Clusters occurred in different HCCAs within the same districts in consecutive years, demonstrating strong intra-district heterogeneity and year-to-year variability of the affected HCCAs. This could result from outbreaks limited to HCCAs without exceeding the threshold at the district level: the district is not vaccinated and may be affected by a large outbreak the following year. Besides, waiting for the threshold to be reached at the district level to initiate reactive vaccination may incur unnecessary delays: we showed that a decision based on threshold estimated at the health centre level might lead to earlier detection of outbreaks, so more reactive and possibly more cost-effective vaccination strategies. Thus, adding HCCA-level surveillance to the current district-level surveillance would improve the timeliness of epidemic detection.
With the introduction of a new meningococcal A conjugate vaccine (MenAfriVac™) in the meningitis belt over the next few years, the use of the health centre catchment areas as spatial units can also help to monitor more accurately the vaccine supply at a finer spatial scale, saving doses that could be given inadequately, and to evaluate its impact and protective efficacy in the population (herd immunity) at the same level. Although this vaccine brings new hope to the control of meningitis epidemics, reactive vaccination with polysaccharide vaccines and research to improve control strategies will still be needed in the coming years, since it will take several years to immunize against the A the vulnerable population across the belt and since other serogroups like W135 may replace meningitis A as the dominant serogroup 
. New decision criteria will have to be found for reactive vaccination. With the additional use of a finer spatial scale like the HCCAs, an interesting strategy would be real-time cluster detection, with prospective space-time scan statistic 
or other existing methods 
In the context of a resource-limited country, this study can also assist public health authorities in their decision-making regarding resource allocation. The spatial clusters detected in our study were located in different HCCAs from year to year, but nine of the 42 districts were more recurrently affected by clustering of MM cases. Thus, these findings provide approaches to better adjust allocation of resources, including a ready supply of antibiotics and rapid diagnostic tests 
, as well as additional health care personnel. In order to reduce the reaction time of the vaccination, one may consider allocating vaccines to these districts' hospitals prior to the meningitis season, provided the cold chain can be maintained. Given cost and organizational constraints, further cost-effectiveness and feasibility analyses are needed to evaluate this strategy, before any policy recommendation.
Clusters were more often found in nine districts, including five bordering Nigeria within a 500 km distance between Doutchi and Aguie, most likely because of intense mobility of border populations 
. However, no consistent annual spatio-temporal pattern could be found over the study period; hence, no spread in a systematic geographical direction from a fixed source could be identified. This is contrary to a study carried out in Mali, which highlighted a potential south-north spread, with Bamako and Mopti as probable sources 
. Instead, our results suggest the emergence of scattered sources, likely from a pool of carriers when conditions are favorable to the occurrence of the invasive disease. Favorable conditions may include climatic conditions occurring during the dry season (low absolute humidity and dust-laden Harmattan wind), which would damage the nasopharyngeal mucous membrane and increase the risk of bloodstream invasion by a colonizing meningococcus 
. In this study, we observed that the latest spatio-temporal clusters during the epidemic season were often the northernmost ones, which could be correlated with the northward advance of the Intertropical Front preceding the arrival of rains from the south, thus raising relative humidity. However, climatic factors do not entirely explain these spatio-temporal epidemic patterns. As suggested by Mueller's hypothetical explanatory model 
, their role may be limited to the hyperendemic increase during the dry season, while transition from a hyperendemic state to highly localized epidemics may be due to increased transmission, possibly caused by viral respiratory co-infections. Moreover, in equivalent climatic conditions, an area in which the proportion of susceptible individuals is higher due to waning immunity (acquired by infection or vaccination) would be more prone to outbreaks 
. Recently, Irving et al 
suggested that population immunity may be a key factor in causing the unusual epidemiology of meningitis in the Belt. Although density and distance to primary roads were not individually correlated with MM incidence at the HCCA level, other socio-demographic factors (poverty, overcrowded housing, migrations, markets…) may also have an influence on local transmission of the bacteria and carriage and contribute to the risk of micro-epidemics of co-infections 
. Of note, one spatio-temporal cluster of four adult cases was detected in February 2009 in Bilma district (see ), in the oasis town of Dirkou, located on an important south-north route of trans-Saharan trade and transit migration. Meningococcal strain variations most likely play a role in the occurrence of epidemic waves 
. In this study, the spatio-temporal distribution of all N. meningitidis
cases was analysed irrespective of the serogroups. A subsequent analysis will differentiate serogroups of meningococci as their spatio-temporal patterns may significantly vary 
. Further etiologic studies are needed to explore causality of the spatio-temporal patterns highlighted in this paper.
Finally, our findings provide an evidence-based approach to reflect on public health policies and indicate a promising strategy to improve prevention and control of meningitis in sub-Saharan Africa. They can serve as an example for other meningitis belt countries, illustrating what finer scale surveillance and spatial analyses can offer for prevention and control of meningitis. Research efforts should now focus on investigating the role of dust, socio-demographic factors, co-infections and vaccination strategies on cluster occurrence at the HCCA level, and on developing an operational decision support tool to respond better to meningitis outbreaks with the introduction of the new conjugate vaccine.