BMC Infectious Diseases BioMed Central Research article

Background The outbreak of the pandemic flu, Influenza A H1N1 (Swine Flu) in early 2009, provided a major challenge to health services around the world. Previous pandemics have led to stockpiling of goods, the victimisation of particular population groups, and the cancellation of travel and the boycotting of particular foods (e.g. pork). We examined initial behavioural and attitudinal responses towards Influenza A, H1N1 ("Swine flu") in the six days following the WHO pandemic alert level 5, and regional differences in these responses. Methods 328 respondents completed a cross-sectional Internet or paper-based questionnaire study in Malaysia (N = 180) or Europe (N = 148). Measures assessed changes in transport usage, purchase of preparatory goods for a pandemic, perceived risk groups, indicators of anxiety, assessed estimated mortality rates for seasonal flu, effectiveness of seasonal flu vaccination, and changes in pork consumption Results 26% of the respondents were 'very concerned' about being a flu victim (42% Malaysians, 5% Europeans, p < .001). 36% reported reduced public transport use (48% Malaysia, 22% Europe, p < .001), 39% flight cancellations (56% Malaysia, 17% Europe, p < .001). 8% had purchased preparatory materials (e.g. face masks: 8% Malaysia, 7% Europe), 41% Malaysia (15% Europe) intended to do so (p < .001). 63% of Europeans, 19% of Malaysians had discussed the pandemic with friends (p < .001). Groups seen as at 'high risk' of infection included the immune compromised (mentioned by 87% respondents), pig farmers (70%), elderly (57%), prostitutes/highly sexually active (53%), and the homeless (53%). In data collected only in Europe, 64% greatly underestimated the mortality rates of seasonal flu, 26% believed seasonal flu vaccination gave protection against swine flu. 7% had reduced/stopped eating pork. 3% had purchased anti-viral drugs for use at home, while 32% intended to do so if the pandemic worsened. Conclusion Initial responses to Influenza A show large regional differences in anxiety, with Malaysians more anxious and more likely to reduce travel and to buy masks and food. Discussions with family and friends may reinforce existing anxiety levels. Particular groups (homosexuals, prostitutes, the homeless) are perceived as at greater risk, potentially leading to increased prejudice during a pandemic. Europeans underestimated mortality of seasonal flu, and require more information about the protection given by seasonal flu inoculation.


Background
As most endemic countries begin to re-focus their malaria control goals, including in some cases a target of elimination [1], contemporary maps that reliably define subnational variation in disease risk are required to inform priority setting and the selection of appropriate suites of intervention. Recent efforts at developing empirical global maps of Plasmodium falciparum risk herald a new era of using maps to define regional populations at risk of malaria to guide the future global malaria control agenda [2]. However, the applicability of malaria risk mapping to make predictions at spatial scales and time points necessary for effective health service planning and review depends largely on the amount and resolution of information available. For example, it is recognized that continental risk maps may not provide sufficient detail and precision for national and sub-national level control priority setting [2].
Kenya is one of very few countries that have a plethora of malaria risk data, spanning over 30 years. The earliest attempts to describe the spatial distribution of malaria risk in Kenya were based on expert opinion of malaria seasons and climate [3]. Between 1998 and 2005 several attempts were made by our group to model the predictive accuracy of this historical map [4,5] or use historical parasite prevalence data and remotely sensed proximates of climate to predict risk using sub-optimal spatial methods that were unable to define fully the uncertainty in the modeled maps [6,7]. Here we present a more robust Bayesian model-based geo-statistical spatial-temporal method to predict the risk of malaria in Kenya in 2009 using the largest assembled contemporary empirical evidence for any country in Africa. As a new phase of malaria control in Kenya begins, the implications of the resulting malaria risk map for decision makers and the prospects for the future of malaria control nationwide are discussed.

P. falciparum parasite rate as a marker of risk
There are many measures of the intensity of malaria transmission [8,9]. Direct measures of transmission intensity applicable for malaria modeling include the entomological inoculation rate (EIR) and the basic reproductive number (R o ). EIR is the number of parasite-specific infectious bites received by a person per unit time and R o is the average number of secondary infections resulting from one infected individual being introduced into a nonimmune host population. These indices are rarely measured, however, thus limiting their utility for spatial modeling [10]. An alternative measure of P. falciparum malaria risk is the parasite rate (PfPR), which is the proportion of a random sample of population with malaria parasites in their peripheral blood, used frequently to define transmission intensity since the 1950's [11] and has a predictable mathematical relationship to the rarely sampled measures of EIR and R o [12][13][14]. The PfPR has therefore become the benchmark indicator by which malaria risk is modeled and mapped in Africa [2,6,7,[15][16][17].

Data identification
PfPR survey data were identified using basic search principles and inclusion criteria described elsewhere [18] with two notable exceptions: firstly survey data were included if surveys were undertaken from 1 st January 1975, because of the rich Ministry of Health survey data between 1975 and 1984 [19]; and secondly no restriction was placed on sample size for inclusion in the spatial modeling [2].
Data searches included online searches for peer-reviewed publications using PubMed [20] and African e-repositories [21]; manual searches of monthly returns archived from over 40 field stations maintained by the Ministry of Health's Division of Vector Borne Diseases; reviews of master's and doctoral thesis titles and abstracts from the Universities of Nairobi and Jomo Kenyatta; access to household survey data supported as part of national [22] or sub-national sample surveys on malaria or nutrition [23,24]; and an extensive correspondence and data sharing exercise with the prolific malaria research community in Kenya or those working in the country but based overseas [25]. Data searches began in 2005 and were completed with final reviews of published reports and correspondence with national research groups on 31 st March 2009. All data were entered into a customized Microsoft Access (Microsoft 2007) database to include information on survey location, survey timing (month and year), age ranges of the sampled population, sample size, numbers reported positive for P. falciparum infection and the methods of parasite detection [18].

Pre-processing of PfPR survey data
Geo-location techniques A series of independent databases of cities, towns and villages developed since 2004 with longitude and latitude coordinates from Global Positioning System (GPS) recordings are available in Kenya. These include a national schools database developed through a mapping project in 2008 by the Ministry of Education [26]; a database of settlements connected to the classified motorable road network compiled as part of a road mapping project by the Ministry of Roads and Public Works [27]; and a variety of smaller databases developed as part of research projects or development programmes. In addition, a database of villages digitised from topographical maps in 2002 was obtained from the International Livestock Research Institute. These databases were first used to geoposition survey locations with priority given to the GPS sources. Where survey locations could not be geo-positioned from any of these national databases, digital data-bases such Microsoft Encarta (Microsoft 2007), Alexandria Digital Library [28] and Falling Rain Genomics Inc. Global Gazetteer [29] were used. A database of enumeration areas for the 1999 census obtained from the Kenya National Bureau of Statistics was used as a final source if survey data could not be positioned using the other sources. Survey locations were classified as points if they could be positioned to an area ≤10 km 2 ; wide area (>10 km 2 to <25 km 2 ); or polygon (≥25 km 2 ) [18].

Age standardization of PfPR
Under stable endemic transmission PfPR is age-dependent and rises during early childhood, peaks in older children and falls through adolescence and adulthood, the rate of decline a consequence of development of anti-parasitic immunity [14]. PfPR surveys, however, are often reported for a variety of age-ranges. The assembled PfPR data were therefore standardized to the classical age-range of 2 to less than 10 years using an algorithm based on catalytic conversion models first adapted for malaria by Pull and Grab [30] and modified by Smith et al., [14]. This agestandardized parasite rate, henceforth referred to as PfPR 2-10 , was computed for each survey report [2].
The relationships of the covariates in their continuous and categorical forms were first visually examined against PfPR 2-10 data using scatter and box plots. These were used to aggregate the covariates into suitable categories that corresponded to biologically appropriate definitions, previous applications of remotely sensed variables and retention of effective sample sizes (see Additional File 1). A univariate non-spatial binomial logistic regression model was then implemented for each covariate with PfPR 2-10 as the dependent variable in Stata/SE Version 10 (Stata Corporation, College Station, TX, USA). The results of the univariate analyses were used to determine the relative strength of each candidate covariate as a predictor of PfPR 2-10 and identify those which qualified for inclusion in the Bayesian geostatistical model. First, where there was more than one plausible way of categorizing a covariate, the size of the odds ratio, the Wald's p-value and the value of Akaike Information Criterion (AIC), a measure of the goodness of fit of an estimated statistical model [35], were used to determine which approach resulted in categories with the strongest association with PfPR 2-10 [SI 1]. Once the best categorizations were determined, a collinearity test of all the covariates was undertaken and if a pair had a correlation coefficient > 0.9 [43], the variable with the highest value of AIC was dropped from subsequent analysis. The selected covariates were then analysed in a binomial multivariate logistic regression with PfPR 2-10 as the dependent variable. Using backwards variable elimination, covariates with Wald's P > 0.2 were removed stepwise until a fully reduced model was achieved.

Bayesian space-time models
Using the Kenya PfPR 2-10 data and the selected covariates, a spatial-temporal Bayesian generalized linear geostatistical model [2] was implemented to predict a malaria map of Kenya for 2009. Bayesian geostatistical models provide the ability to predict values of a spatially continuous event at unsampled locations using combinations of the sampled data in space and time, and importantly allow for calculation of robust uncertainty estimates around model predictions [2,43,44]. The underlying assumption of the Kenya PfPR 2-10 model was that the probability of prevalence at any survey location was the product of two factors. First, a continuous function of the time and location of the survey, modified by a set of covariates, and modelled as a transformation of a space-time Gaussian random field. Second, a factor depending on the age range of individuals sampled in each survey. The distribution of the second factor [2] was based on the procedure described by Smith et al. [14]. The Bayesian spatial-temporal model was implemented in two parts starting with an inference stage in which a Markov Chain Monte Carlo (MCMC) algorithm was used to generate samples from the joint posterior distribution of the parameter set and the space-time random field at the data locations. This was followed by a prediction stage in which samples were generated from the posterior distribution of PfPR 2-10 at each prediction location on a 1 × 1 km grid. Details of the spatial-temporal Bayesian geostatistical models are presented in Additional File 2.

Model validation and measures of uncertainty Selection of model validation test data
To ensure that the validation data were spatially representative of the whole country, a spatial declustering algorithm [2] was implemented. This algorithm defined Thiessen polygons whose boundaries enclosed the area that was closest to each point relative to all other points around each survey location. A 10% sample of the larger Kenya PfPR 2-10 dataset was then drawn randomly. Each data point had a probability of selection proportional to the area of its Thiessen polygon so that data located in densely surveyed regions had a lower probability of selection than those in sparsely surveyed regions [45]. The Bayesian spatio-temporal geostatistical model was then implemented in full using the remaining 90% of data.

Computing model accuracy and uncertainty
A series of validation statistics were computed by comparing the predicted PfPR 2-10 values to actual PfPR 2-10 observed at the validation locations. The validation statistics were: the linear correlation coefficient; mean error (ME); and mean absolute error (MAE) is a measure of the bias of predictions (the overall tendency to over or under predict). Finally, the probability of membership of a survey location to its assigned endemicity class (see next section) was computed as a measure of uncertainty. These probabilities, ranging from 1 (no uncertainty in class membership) to 0.14 (membership equally likely to all classes) were computed from the posterior distributions resulting from the Bayesian geostatistical model as explained in detail in Additional File 2.

Malaria risk classifications and estimations of populations exposed to risk
Seven endemicity classes of PfPR 2-10 were selected: <0.1%; ≥0.1% and < 1%; ≥1% and <5%; ≥5% and <10%; ≥10% and <20%; ≥20% and <40%; ≥40%. These classes were selected as they can be used to compute approximates of the traditional measures of endemicity [11], are congruent with recommendations for the selection of suites of vector control and the timelines to effective transmission control [9,46,47] and allow for interpretation of lower risk categories where the predominant spatial risks are not among the higher endemicity classes. The probability of membership to each endemicity class was estimated from the posterior probability distributions of PfPR 2-10 for each pixel generated by the Bayesian geostatistical model, as described in Additional File 2.
A high-resolution (100 × 100 m) population distribution map of Kenya [48] was used to compute the number of people in each of the malaria endemicity classes. This map was constructed from a combination of satellite imagery and land cover maps which were used to develop models that identified the location of settlements [48,49]. The modelled settlements map was then used to redistribute census population counts within the small enumeration area polygons. The resulting high-resolution map represented estimated population distribution in Kenya for the year 2000. This raster population surface was then projected to 2009 using provincial inter-censal growth rates from the 1999 national census [50]. The raster malaria endemicity map was then overlaid on the projected population map and the number of people in each endemicity class, overall and by province, was extracted using ArcGIS 9.2 Spatial Analyst tool.

Assembled data
A total of 2,756 PfPR random sample surveys were assembled for the period 1975-2009. Of these, 74 survey locations were excluded from analysis because they were polygons (n = 30); were not positioned (n = 41); or were missing survey month (n = 3). Of the remaining 2,682 data points (Table 1)

Testing of climate and ecological covariates
The univariate analysis showed that all the biologically selected categorical covariates were statistically significant predictors of differences in PfPR 2-10 (see Additional File 1 and Table 2). There was reduced risk of infection in areas that were: urban compared to rural; of minimum average annual temperature of <16°C compared to ≥16°C; of maximum average annual temperatures of <25°C or >30°C compared to between 25°C -30°C; of zero or 1-3 sets of three continuous months of precipitation >60 mm in an average year compared to corresponding precipitation patterns that occurred >3 sets in an average year; where EVI was ≤0.3 compared to >0.3; where the survey was located at an altitude of ≤500 m or >1500 m compared to between >500-1500 m above sea level; and were at a distance to main water bodies of >12 km relative to ≤12 km (see Additional File 1 and Table 2).
In the multivariate regression, however, only the classifications of urban-rural; maximum temperature; precipitation; EVI and distance to main water bodies were included ( Table 2). Altitude and minimum temperature were excluded from the multivariate analysis because they were highly correlated with each other (R 2 = 0.97) and with maximum temperature (R 2 > 0.70) and both had comparatively higher AIC values [SI 1]. From the multivariate analysis the risk of malaria parasite infection was lower in locations that were: urban compared to rural (odds ratio, 95% CI: 0.50, 0.36-0.70, p < 0.001); of maximum temperatures <25°C (0.25, 0.12-0.52, p < 0.001) or >30°C (0.61, 0.44-085, p = 0.003) compared to between 25°C-30°C; of zero (0.53, 0.36-0.83, p = 0.005) or 1-3 (0.63, 0.46-0.85, p = 0.003) sets of three continuous months of precipitation >60 mm in average year compared to >3 sets; and at distance to water bodies of >12 km (0.62, 0.49-0.77, p < 0.001) relative to ≤12 km (Table 2). Although there was a reduced risk of infection prevalence at EVI of ≤0.3 (0.77, 0.57-1.06) compared to >0.3, this was not sta-tistically significant (p = 0.114). This, however, did not preclude the inclusion of EVI in the final model set as it still met the inclusion criteria with a P value < 0.2 and the AIC value of the multivariate model was lower with it compared to without.

Bayesian predicted risk projected to 2009
The 2009 map of the predicted posterior mean distribution of PfPR 2-10 is shown in Figure 3a. The predicted malaria endemicity class map is shown in Figure 3b and indicates that the majority of the country's surface area falls into endemicity classes of <5% PfPR 2-10 . The lowest endemicity class (< 0.1% PfPR 2-10 ) covers most of Nairobi and Central provinces and some parts of the Eastern and  Rift Valley provinces (Figure 3b). The endemicity class of between 0.1 and 1% covers most of the North Eastern, Eastern, Rift Valley and Coast provinces. High transmission areas (endemicity class ≥40% PfPR 2-10 ) were predicted mainly in small parts of Nyanza province along the shores of Lake Victoria and cover <2% of the total area of Kenya.

Model validation
The mean error in the prediction of PfPR 2-10 for 2009 revealed low overall bias with a slight tendency to underestimate predictions by -0.15% (Table 3). The mean absolute error also showed a relatively moderate model precision with low average error of predictions of 0.38%. The correlation between the actual and predicted values for the hold-out set was 0.81 indicating a strong linear agreement ( Figure 4). In assessing the endemicity classes, the overall probabilities of membership of the predicted class were all greater than the chance assignment value of 0.14 and in most of the country was greater than 0.45 ( Figure  5).

Population at risk in 2009
Of the estimated 40 (Table 4).

Discussion
We have assembled over 2,600 independent, empirical survey estimates of P. falciparum infection prevalence in Kenya and used these data to generate a contemporary map of infection prevalence at a 1 × 1 km resolution for the year 2009 using space-time geostatistical models within a Bayesian framework. The modeled distribution had a high predictive accuracy as shown by the low values of ME and MAE and high correlation between predicted and observed PfPR 2-10 ( Table 3 & Figure 4). The probability of endemicity class membership were also generally moderate to high across the country with the exception of small pockets of the low population-density areas of the northern districts where there was relatively sparse distribution of input data in time and space ( Figure 5). This mapped distribution of malaria risk represents the most accurate depiction of parasite exposure described for Kenya since efforts to map risk began in the 1950's [2,3,5,6]. More importantly it represents a distribution of risk in 2009 serving as a contemporary basis upon which to design the future of malaria control in Kenya.
The use of a carefully selected suite of covariates to inform the prediction of risk is a departure from the current Malaria Atlas Project approach, with the exception of the use of urban-rural classification [2], but consistent with other approaches to modeling malaria distributions [5,6,[15][16][17]35]. In fact several discrete categories of the covariates were as different in infection prevalence as the differences described for urban versus rural surveys. We elected not to include a mask of zero or unstable transmission based upon temperature and aridity as developed previously by Guerra et al, [40]. Rather we have assumed that these climatic drivers of transmission would be captured within the model and have chosen to bin the risk classes within much smaller PfPR 2-10 ranges at the lowest end of the transmission spectrum. The lowest risk class encompasses predicted PfPR 2-10 between 0 and < 0.1% and covers approximately 91,000 km 2 within the Nairobi and Central provinces and small parts of Eastern and Rift Valley (Figure 3). Defining absolute zero transmission is conceptually difficult and practically impossible to measure empirically using PfPR 2-10 and we therefore feel that the more conservative and inclusive approach used here allows for the possibility of transmission until proven otherwise.
Sample semi-variograms of PfPR 2-10 dataset (n = 2,682) indi-cating the presence of spatial autocorrelation in the PfPR 2-10 data up to lags of 1 decimal degree or the equivalent of ~111 km at the equator  What is striking about the contemporary 2009 distribution of malaria infection risk is the enormity of Kenya's land surface under very low intensity transmission. Over 94% of Kenya's surface area is predicted to be exposed to a PfPR 2-10 of less than 5% and is home to 86% of Kenya's projected population in 2009 (Table 3). Approximately 66% of the 2009 population live in areas where infection prevalence is less than 1%, including a large majority where risks are hard to detect empirically (PfPR 2-10 < 0.1%) ( Table 3). Conversely areas of high transmission, as defined by a PfPR 2-10 of ≥40%, representing areas expected to be intractable to immediate reductions in parasite prevalence with scaled-up use of insecticide treated nets [47] are located in the strip of land along the shores of Lake Victoria (Figure 3). In 2009 only 11% of Kenya's population was exposed to this highest transmission intensity class (Table 3). Historically holo-to-hyperendemic transmission (≥50% PfPR 2-10 ) was thought to exist across much larger reaches of the Kenyan coast, around Lake Victoria and along the Tana River [3,51]. In the present modeled iteration of PfPR 2-10 holo-endemic transmission (>75% PfPR 2-10 ) no longer exists and hyper-endemic transmission is constrained to small pockets within the darkest red class shown in Figure 3. Although this study doesn't present change of infection risk over time, it seems plausible that across much of Kenya the extent and intensity of P. falciparum transmission has undergone a recent decline with increasing spatial areas and populations becoming exposed to lower and lower risks of parasite exposure. This Spatial distribution of P. falciparum malaria in Kenya at 1×1 km spatial resolution Figure 3 Spatial distribution of P. falciparum malaria in Kenya at 1×1 km spatial resolution. a) continuous posterior mean PfPR 2-10 prediction; b) endemicity classes: PfPR 2-10 < 0.1%; ≥0.1 and < 1%; ≥1 and <5%; ≥5 and <10%; ≥10 and <20%; ≥20 and <40%; ≥40%.  has implications for a changing clinical epidemiology in areas undergoing transition, with older children becoming increasingly at risk of severe clinical outcomes [52][53][54][55] but more importantly as communities transition to very low levels of parasite exposure overall malaria morbidity and mortality will decline substantially [52,56].
Although the model is characterized by generally low uncertainties, the pockets of greatest predicted uncertainty are located in the northern districts of Turkana, Marsabit and Moyale ( Figure 5). Surprisingly, pockets of risk >10% PfPR 2-10 were observed in these hot and generally arid parts of the country traditionally regarded to be of unstable low risk. These areas, which generally have low population densities and have traditionally not been targeted for scaling of malaria preventive interventions, exhibit highly focal transmission close to water features, such as the Turkwell, Tana and Kerio rivers and were referred to in historical maps as 'malarious near water' [3]. Because of their presumed low risk, few empirical studies of malaria have been undertaken in these areas. The malaria situation among these poor, pastoralist communities remains ill-defined. In addition there are some important methodological constraints to defining risk in areas of very low transmission and new approaches to micro-geographic Bayesian modeling of risk based upon a presence/absence criterion may be required to improve risk mapping in these areas where the majority risk is negligible, seasonal and exceptionally heterogeneous, associated with the presence of water features.
Further improvements in malaria risk mapping using PfPR 2-10 might be achieved if the prediction models were corrected for whether microscopy or RDT was used to examine parasitaemia given the varying sensitivities and specificities of the two methods [57]. In this study, however, this was not possible because information on the type of RDT used and the quality of microscopy was lacking for most surveys. In future, it may be feasible to develop universal models that correct for sensitivity/specificity differentials of the methods used to test for infection, preferably based on large-scale population surveys which have used both RDT and microscopy for the same individuals with the appropriate quality assurance and external validity.
The prospects for Kenya to transition the majority of its population living in high transmission areas in the next Spatial distribution of probability of membership of P. falciparum malaria endemicity class in Kenya at 1 × 1 km spatial resolution Figure 5 Spatial distribution of probability of membership of P. falciparum malaria endemicity class in Kenya at 1 × 1 km spatial resolution. Given that there are seven endemicity classes, the lowest probability of class assignment is 0.14. Any value above 0.14 is better than a chance allocation to the endemicity class. Lines shown on the map represent the contours of the different endemicity classes shown in Figure 3.
10 years to areas of low (PfPR 2-10 <5%) or very low endemicity (PfPR 2-10 <1%) look promising. It is however important to emphasize the control implications of this low stable endemic control. There appear to be areas along the Kenyan coast that currently experience risks associated with a PfPR 2-10 <5% and are likely to have transitioned to this state from meso-hyperendemic conditions. If this has been achieved through the scaled-up use of insecticide treated nets (ITN) then universal coverage must be maintained as withdrawal of ITN would result in a devastating rebound where vectors persist but functional immunity has been modified among the human host population. Conversely in areas that have historically had low or very low transmission, for example in semi-arid areas, the adoption of ITN may not be the most cost-effective strategy. As such all areas of similar contemporary risk may not be equivalent in terms of strategic control. One therefore must interpret contemporary distributions of risk for control planning in concert with the potential vulnerabilities of transmission based upon vector distributions or historical descriptions of risk. For Kenya it is also important to recognize that there are vast areas where infection risks are low and have historically been low because of their ecological niches (arid, urban or at high elevation). While these communities enjoy a low risk of infection, risks are not absent and thus cost-efficient suites of interventions must be tailored to meet their needs. This poses a challenge where universal coverage of ITN and presumptive fever treatment with Artemisinin based combination therapy remain the single bedrock of most national malaria control strategies across Africa.

Conclusion
There remains some debate over the feasibility of malaria elimination in Africa [58][59][60][61]. Kenya is an example where infection prevalence is low across large parts of the country. However moderate-to-high risks remain in well defined areas, some of which share borders with neigh-  Mean error is a measure of the bias of predictions (the overall tendency to over or under predict) whilst mean absolute error is a measure of overall precision (the average magnitude of error in individual predictions).