INTRODUCTION
Plague, a severe bacterial disease caused by Yersinia pestis, has a worldwide distribution.1 In recent decades, the majority of human cases have been reported from eastern or southern Africa and Madagascar.2–4 The West Nile region of Uganda, located in the northwest of the country, represents the current primary epidemiologic focus in that country. 5,6 However, because of limited capacity for diagnostic laboratory testing in that region, few clinically diagnosed cases are confirmed and the true burden of disease is yet to be determined.
Yersinia pestis is maintained primarily in zoonotic cycles by rodents and their fleas. Although the bacterium can be transmitted through direct contact or inhalation of infectious respiratory droplets, humans are most often infected with Y. pestis via flea bites during epizootic periods when rodent hosts perish from infection and infectious fleas are forced to parasitize alternative hosts, including humans.1 Plague infections have a short incubation period (typically 2–6 days for the bubonic form of the disease) followed by acute onset of illness. If antibiotic treatment is delayed or inadequate, the disease is often fatal.2 Outcome of infection is improved by early diagnosis followed by appropriate antibiotic therapy. 1,7–14
Raising awareness among health care providers, environmental health specialists, and the public in areas where transmission to humans is most likely to occur could aid in prevention and control of plague in Africa. In the United States, fine-scale spatial risk models have been constructed to identify areas posing the greatest threat to humans within key plague foci such as the Four Corners Area. 15,16 Although a plague model has been constructed for the African continent, 17 its resolution is very coarse and public health utility limited. Fine-scale models developed in African plague foci are needed to define high risk areas and are in line with the World Health Organization’s (WHO) recommendations for implementation of public health interventions, including predictive surveillance and preventative control for plague. 18,19
In the study, we sought to 1) describe the spatial distribution of suspect and probable human plague cases in the West Nile region, 2) identify remotely sensed ecologic correlates of parish-level disease incidence, and 3) incorporate these variables into a predictive model that defines areas of plague risk in the region. This is useful for defining existing epidemiologic foci, and for identifying areas where enhanced surveillance and preventative measures may be implemented to decrease the burden of plague.
MATERIALS AND METHODS
Study area.
Two districts, Arua and Nebbi, located in the West Nile region of Uganda were included in the study (Figure 1). These districts, each comprised of three counties, encompass an elevation gradient from 626 to 1,668 m with low-lying areas in the northeast rising to higher elevations across the Rift Valley escarpment, which runs from north to south roughly bisecting the districts. The eastern parts of Arua and Nebbi districts, which include portions of Padyere and Madi-Okollo counties, are characterized by sandy soil and low rainfall, whereas the western highlands of Okoro and Vurra counties contain lush vegetation, fertile soil, and numerous rivers and tributaries. The annual rainfall average in this region is 1,250 mm with March–June experiencing less reliable rainfall and heavier and more reliable precipitation occurring from late August through November (http://www.nebbi.go.ug/; http://www.arua.go.ug/).20,21 The 2002 Uganda Population and Housing Census reported a population of 833,928 in Arua district and 435,360 in Nebbi district (Uganda Bureau of Statistics, 2002).
Epidemiologic data.
An epidemiologic database of reported human plague cases within Nebbi and Arua districts was developed based on review of health records from a total of 31 health facilities (27 district clinics and 4 hospitals) (Figure 1). The year 1999 was selected as the starting year because a uniform recording system was initiated that year and required clinics to record all consults and the presumptive diagnoses. All cases for which plague was listed as the primary clinical diagnosis were extracted from clinic and hospital log books. Information was available for age, sex, place of residence, date of clinic visit, concurrent diagnoses, and treatment. Additionally, Health Management Information System (HMIS) forms from the district Ministry of Health (MOH) offices were reviewed and information pertaining to plague cases such as onset of illness and outcome, including date of death, was extracted and cross-referenced with the information obtained in the health clinics. Finally, beginning in 2000 in Okoro County, a separate standardized reporting form was used that included more detailed information about the cases, including environmental data (e.g., reported rodent die-offs) for the site of residence.
Standard criteria for diagnosis of plague in Uganda are sudden onset of fever, chills, malaise, headache or prostration accompanied by either painful regional lymphadenopathy (bubonic), hematemesis or hematochezia (septicemic), or coughs with hemoptysis (pneumonic). For the purpose of this study, a suspect plague case was defined as a patient who was seen in a health facility where he/she was diagnosed and treated for plague. A probable plague case was a patient seen at the referral hospital in the region where he/she was diagnosed and treated for plague and there was additional environmental data suggesting ongoing plague activity, such as a rat die-off in the area where the patient likely contracted the infection. The ratio of suspect cases by parish relative to total suspect cases was calculated; the same calculation was completed for probable cases by parish. The proportion probable cases was subtracted from the proportion suspect cases and a Wilcoxon signed-rank test applied to the difference setting the mean equal to zero to identify whether the proportional representation of suspect cases versus probable cases was similar.
Cases were excluded from the analysis if 1) the village of residence was located in the Democratic Republic of Congo (DRC), or 2) the parish of residence was not listed. Although village of residence was known for the majority of cases, many village locations have not been georeferenced and, therefore, could not be used to create spatial risk models. Parishes, which represent the next smallest spatial unit and have been geo-referenced and linked with population data, therefore serve as the spatial unit for analyses. Cumulative plague incidence rates for 1999–2007 by parish were calculated based on population data acquired from the 2002 Uganda Population and Housing Census (Uganda Bureau of Statistics, 2002).
Environmental data.
Geographic Information System(GIS)-based data used in the spatial modeling included: 1) administrative boundaries within Uganda including district, county, and parish (International Livestock Research Institute [ILRI], 2006); 2) 90-m digital elevation model (DEM) (Shuttle Radar Topography Mission [SRTM] Elevation Data Set, 2008; accessed August 2008 at http://seamless.usgs.gov/); 3) Multipurpose Landcover database, 1 km resolution (Africover; Geographic Information Support Team [GIST], 2003); and 4) Landsat 7 Enhanced Thematic Mapper Plus [ETM+] images dated January 1, 2007 (Row/Path: 58/172) and January 10, 2007 (Row/ Path: 58/173). The Landsat imagery is remotely sensed (RS) data and was collected by the Landsat 7 satellite. This satellite scans the entire earth’s surface and collects 8 bands or channels of reflected energy that may be used to discriminate between earth surface materials as these materials (e.g., soil versus vegetation) have unique amounts of emitted and reflected radiation, which vary by wavelength and may be captured by satellite sensors. 22 The Landsat images were acquired through a cooperative agreement with the National Geospatial-Intelligence Agency [NGA]. Images were captured during clear atmospheric conditions and radiometric and geometric distortions of the imagery had been corrected. Landsat images from 1 and 10 January 2007 were mosaiced together based on spatial reference to cover the entire study area, which included two different paths that are covered on different dates. To test whether raster values from each separate image were similar within the overlapping region, 30 points were randomly selected from each of the individual Landsat images within the overlap area. There was no difference for values from the two images at the randomly selected points (Wilcoxon signed-rank test, df = 29, P = 1.0), which confirmed that the mosaicing process had been successful.
Landsat ETM + band 6 (low gain and high gain bands) was converted from digital number (DN) to absolute radiance and then to surface temperature (°C) using the raster calculator in ArcGIS 9.3 (ESRI, Redlands, CA). 23–27 Additionally, several dimensionless measures indicating landscape characteristics were calculated based on manipulation of Landsat bands using the bandmath function in ENVI 4.5. 22,28,29 These included the Normalized Difference Vegetation Index (NDVI), wetness, brightness, and greenness indexes. The NDVI (based on Landsat band 3 and band 4) is directly related to the photosynthetic capacity of plant canopies and therefore may be used to detect coverage of green vegetation. A higher NDVI pixel value indicates abundance of green biomass. Wetness, greenness, and brightness indices are compositive values calculated through tasseled cap transformation, which weights the sums of separate Landsat bands resulting in relative measures of soil brightness, greenness of the vegetation, and wetness of the land cover. 22,29 For example, in the greenness image, a higher pixel value indicates greater biomass in that area; a higher pixel value in the wetness image indicates greater moisture status; a high value for brightness is indicative of bare soil.
Because of the paucity of classified vegetation and soil layers with fine spatial resolution for this area, it was necessary to use pure band values as proxies for vegetation and soil variability. Individual Landsat bands measure distinct wavelengths within the electromagnetic spectrum that may be used to discriminate between different landscape variables, however it is difficult to make conclusions about specific landscape characteristics based on individual band values compared with indices. 22 Minimum, maximum, and average values for all indices, individual Landsat bands (1–5 and 7, 30 m × 30 m spatial resolution; band 6, thermal infrared band, 60 m × 60 m resolution; band 8, panchromatic band, 15 m × 15 m), and DEM data were extracted for each parish using the zonal statistics function of the ArcGIS 9.3 Spatial Analyst; all of these variables were continuous.
In addition, a dichotomous variable for mean elevation by parish was created based on an elevation cut-off of 1,300 m. This cut-off value was selected because qualitative observations indicated plague to be present in parishes with mean elevations above the cut-off, but absent in parishes at lower elevations. Finally, a variety field was calculated for the Multipurpose Landcover database (Africover) indicating the number of separate land classes within each parish. All layers were projected to World Geodetic System (WGS) 1984 projection.
Multivariate logistic regression model.
We applied a multivariate logistic regression model to determine the probability of plague case occurrence by parish. Covariates included in the logistic regression model were chosen via forward stepwise regression (probability to enter of 0.25) and were restricted to variables significantly associated with plague incidence in univariate tests of association (Wilcoxon test, P < 0.05) but not strongly correlated with each other (Spearman’s rank correlation, ρs < 0.8). A goodness of fit test was applied to determine whether the model covariates adequately explained the distribution of plague occurrence. Receiver operating characteristic curves (ROCs) assessed the overall discrimination of the model based on the area under the curve (AUC), which was also used to determine the optimal probability cut-off to characterize each parish as having elevated or low risk for plague thereby optimizing sensitivity and specificity simultaneously.
The logistic model is described by the following equation:
where P is the probability of a parish being classified as having elevated risk, β0 is the intercept and β1…βk represent the coefficients associated with each independent variable x1…xk The probability that a particular parish is classified as having elevated or low risk for plague cases to occur can be derived from Equation (1) using the following expression:
Each parish was assigned a single probability value. The optimal probability cut-off value was chosen by maximizing sensitivity and specificity simultaneously using ROC curves. Parishes with values above the probability threshold value were classified as having elevated risk, whereas all others were considered low risk.
Clinic buffer.
Minimal access to public transportation makes access to health care facilities difficult in parts of the study area and may result in under-reporting of plague cases from some parishes. We therefore only included parishes with centroids within 15.4 km of the nearest designated health clinic in the logistic model. The 15.4 km “clinic buffer” distance was chosen as it represented the longest travel distance from a reported plague case to a health clinic within the 1999–2007 study periods. The clinic buffer excluded Ayivu county in Arua district and Jonam county in Nebbi district thereby decreasing the study area to include parishes within two counties (Vurra and Madi-Okollo) in Arua district and two counties (Padyere and Okoro) in Nebbi District.
Multivariate linear regression model.
To predict parish-scale incidence of human plague from environmental GIS/RS-based data, multivariate linear regression models were constructed using data from the parishes classified by the logistic regression model as areas with elevated plague risk. 30 Candidate models were identified using forward stepwise regression (probability to enter of 0.25). Predictive variables were restricted to those significantly associated with plague incidence in univariate tests of association (Wilcoxon test; P < 0.05) but not strongly correlated with each other (Spearman’s rank correlation; ρs< 0.8). Models with the lowest Akaike information criterion (AIC) were considered the most parsimonious models, but models within two AIC units of the minimum AIC value were considered competing. 31 The best model had the lowest AIC and provided the most robust validation. A Moran’s I statistic was calculated using ArcGIS 9.3 to ensure that the residuals of the selected final model were not spatially autocorrelated. The predictive capability of the linear regression model was evaluated by regressing actual reported incidence on predicted incidence. The predictive equation was extrapolated to all parishes included in the study area and not excluded by the clinic buffer. If a parish was predicted to have an incidence less than zero, it was assigned an incidence equal to zero. Statistical analyses were carried out using the JMP statistical package, 32 and results were considered significant when P < 0.05.
RESULTS
Summary of epidemiologic data.
A total of 2,011 human plague cases reported from January 1999 to December 2007 were included in the epidemiologic database. Of these, 152 cases were excluded from the analysis because either the village of residence was located in the DRC or parish was not listed. The remaining 1,859 cases were comprised of 76 probable cases and 1,783 suspect cases. From 1999 to 2007, the annual mean number of reported cases was 199 (range 11–445) with the greatest number of cases reported in 2001. Onset of plague symptoms followed a seasonal pattern with cases increasing in September, peaking in November, and decreasing in December and January; few cases were reported in the remaining months. A further description of epidemiologic features and clinical presentation of plague cases will be described in more detail in a separate publication.
Qualitative analysis of the data revealed that suspect and probable cases had similar spatial trends such that case numbers were higher in the western reaches of Nebbi and Arua districts, especially at elevations above 1,300 m. Calculation of the mean difference between suspect and probable cases across all parishes indicated that the proportional representation of suspect cases versus probable cases was similar (Wilcoxon signed rank test: df = 25, P = 1.0) and justified the use of both probable and suspect cases in the analysis.
Logistic model for elevated or low risk of plague.
A multivariate logistic regression model based on GIS/RS-derived environmental data was developed to predict parishes with elevated versus low plague risk (Table 1). The model revealed that elevated parish-level plague risk was positively associated with mean parish elevation being > 1,300 m and with maximum brightness and average wetness, but negatively associated with average greenness. A lack of fit test indicated that the most parsimonious model included sufficient numbers of covariates with appropriate functional relationships (χ2 = 56.27, df = 111, P > 1.0) and a whole model test denoted good overall fit (χ2 = 96.69, df = 4, P < 0.001). Accuracy of the best model, based on the area under the ROC curve, was 0.96 indicating that randomly selected presence and absence pairs would be correctly ordered by their probability scores 96% of the time. Probability of plague case occurrence was calculated based on the model described, and dichotomized into “elevated risk” or “low risk” categories based on a cutoff probability value (P = 0.55), which was derived from the ROC curve and simultaneously optimized the sensitivity and specificity of the model.
The model predicted 46 (40%) of included parishes to report plague cases. These parishes were located in the western region of the study area, along the DRC border (Figure 2). Nearly all parishes predicted to have elevated plague risk were located above the 1,300 m elevation cut-off. Evaluation of the model against actual case reports by parish revealed a sensitivity of 93% (i.e., 93% of the parishes with plague reported were correctly classified by the model as having elevated plague risk). Model specificity was 92% (i.e., 92% of parishes where plague was not reported were correctly identified as having low plague risk). The model predicted plague cases in 6 parishes where no cases were actually reported thereby producing a positive predictive value of 87%. Three parishes below the 1,300 m elevation cut-off actually reported plague cases, but were misclassified by the logistic model, and were predicted to pose a low risk decreasing the negative predictive value to 96% (Table 2).
Linear regression model for plague incidence.
The linear regression model was developed based on 49 parishes either predicted by the logistic regression model to have plague cases present (N = 46), or misclassified by that model as low risk, but where cases actually had been reported (N = 3). The best linear regression model indicated that within these parishes, plague incidence continued to be positively associated with elevation above 1,300 m as seen in the logistic model. Positive associations were also observed between plague incidence and the parish level average value of band 3, minimum elevation values in parishes that were above the 1,300 m (e.g., low areas within a high plateau), surface temperature, and the variety of land-cover classes (Africover) (Table 3). Plague incidence was negatively associated with average values of band 7 and minimum brightness (Table 3). Combined, these variables explained 68% of the total variation in parish-level incidence (F7,41 = 12.58, r2 = 0.68, P < 0.001). Parameter estimates for covariates of the selected model are shown in Table 3. The model predicted cases to occur within parishes throughout the model build area, including parishes in northern Vurra County and eastern Okoro County where elevations were lower than 1,300 m.
Two parishes in northern Vurra County with very high plague incidences (Ayavu parish, incidence = 64.7 cases per 1,000 population; Ozoo parish, incidence = 54.1 cases per 1,000 population) prevented the model residuals from achieving a normal distribution. Once these parishes were removed from the analysis, the residuals of the model were normally distributed (Shapiro–Wilk test, W = 0.98, P = 0.66). The residuals from the entire model development area (including Ayavu and Ozoo parishes) were not spatially autocorrelated (Moran I = −0.08, Z[I] = −0.92). Regressing actual incidence on predicted incidence showed that 74% of the variation in actual incidence was explained by the predictive model (F1,47 = 130.42, r2 = 0.74, P < 0.001) (Figure 3).
Extrapolation of the linear regression model.
The linear regression model explained approximately 50% of the total variation in parish-level plague incidence when extrapolated east to all parishes within the 4 county study area and in the clinic buffer (F1,114 = 114.05, r2 = 0.50, P < 0.001). The most noteworthy areas of discordance between the predictive model and actual reported incidence were 1) those in which very high incidence was expected but no cases were reported or 2) areas where actual incidence was much higher than predicted. Specifically, three parishes in northern Vurra County did not report any cases from 1999 to 2007, whereas the model predicted incidences ranging from 26.0 to 48.1 cases per 1,000 population (Figure 3). Two of these parishes were noted as outliers in the analysis of residuals. Conversely, three parishes located in central and southern Vurra County were predicted to report cumulative incidences of 29.0 to 42.5 cases per 1,000 population (1999–2007), but case reports were higher than predicted (42.6–64.7 cases per 1,000 population) (Figure 3).
DISCUSSION
Previous studies have identified the West Nile region of Uganda as an epidemiologic focus for plague. 5,6,33 However, within this region the spatial distribution of case occurrence and the ecologic factors associated with elevated plague risk are poorly defined. We have addressed this by 1) mapping parish-level incidence of plague in Arua and Nebbi districts from 1999–2007 and 2) developing statistical models to predict plague incidence based on elevation and remotely-sensed (RS) environmental variables. The study demonstrated that plague incidence can be modeled at the parish-level scale based on environmental variables and identified several parishes where plague may be under-reported and enhanced plague surveillance is necessary.
Similar to previously published models of plague risk for the United States 15,34,35 and descriptive historical observations from Africa, 20,21,36,37 elevation was an important predictor of risk within the study area in Uganda’s West Nile region. Parishes located in western counties and situated above the Rift Valley escarpment (average parish elevation > 1,300 m) reported a higher incidence of plague than counties located below the escarpment. Consistently recorded meteorologic data are lacking from this area but the higher elevation sites are perceived to receive some of the highest rainfall in the West Nile region and experience lower temperatures than neighboring low-lying areas, which report average monthly temperatures of 22 to 35°C. 21 This combination of elevated rainfall and moderate temperature is consistent with conditions observed in other plague-endemic regions of Africa, such as Kenya, Tanzania, Democratic Republic of Congo, Madagascar, and southern Africa, as observed by Davis 37,38 upon evaluation of the relationship between seasonal occurrence of plague and meteorologic variables in localities throughout Africa. Initiation of plague activity in Africa was associated with increased precipitation and temperatures above 15°C, whereas case reports decreased during drier periods and when average temperatures exceeded 27°C, an observation similar to ones made earlier in Vietnam. 39–41 Likewise, in North America occurrence of plague epidemics and epizootics have previously been associated with temperature and rainfall. 35,37,42–44 The reasons for why occurrence of plague cases is related to temperature in Africa are unknown and merit further studies into both the ecology of plague in rodent reservoirs, flea vectors, and Y. pestis transmission dynamics, and potential alterations in human behavior that might increase the risk of Y.pestis transmission. For instance, it has been proposed that lower temperatures in this region cause individuals to sleep in the cooking hut near the fire to stay warm, which will increase their risk for plague by bringing them into contact with rodent fleas responsible for plague transmission. 21 Alternatively, temperature and relative humidity could influence host-seeking behavior of fleas.
Accuracy of the models to predict case occurrence or incidence of reported plague cases was improved by incorporating RS environmental indicators (e.g., temperature, brightness, land cover) together with elevation. However, because these variables were aggregated by parish to match the finest spatial resolution that was georeferenced for plague cases, the biologic meaning of these indicators needs to be interpreted with care. Several variables included in the models are indicative of differences in soil or vegetation types (e.g., Landsat bands 3 and 7, brightness, greenness, and wetness). 22,29 It is difficult to interpret the significance of individual Landsat bands that are included in the models as in what environmental characteristics a high or low band value indicates. Although these Landsat bands may not provide specific information on the environmental characteristics correlated with plague incidence, the Landsat satellites provide global coverage and provide opportunities to extrapolate the model to areas where surveillance information is lacking (e.g., the DRC located west of the study area). Further classification of vegetation and soil types in the West Nile region is required to identify precise environmental correlates of risk and develop hypotheses regarding biologic linkages between the RS indicators and risk of Y. pestis transmission from rodent-flea cycles to humans.
Plague incidence was also positively associated with the number of land cover types occurring within parishes. Others have suggested that habitat heterogeneity or fragmentation may be important for inter-epizootic maintenance of Y. pestis.1,45–49 For example, sufficiently large plague-susceptible small mammal populations or communities may be separated into distinct subpopulations or metapopulations by landscape features that restrict movement and thus either allow some subpopulations to be unaffected by plague epizootics or at least slow epizootic spread enough for some populations to recover before being affected by the next epizootic. Thus, habitat heterogeneity could increase the probability of local persistence of plague foci. This underscores the need for studies on rodent and fleas faunas, rodent movement patterns, and the effect of agricultural practices within clearly defined habitat types in plague-endemic areas in Africa.
Although 74% of the variation in parish level incidence was explained by the model, there were some notable outliers that affected model performance. We believe the most critical areas of discordance between actual and predicted plague incidence are those parishes that did not report cases, but where the model expects cases to occur. This includes Eruba, Kuluva, and Ezuku parishes in northern Vurra County, which were expected to report incidences of 25 to 50 cases per 1,000 population but no cases were observed between 1999 and 2007 (Figure 3). Possible reasons for the overprediction include underreporting of cases during the study period because of unstable political conditions along the DRC border. In addition, model development was based on 9 years of epidemiologic surveillance. However, plague is characterized by long periods of quiescence followed by rapidly spreading epizootics. 1,50 Therefore, it is possible that some parishes that did not report cases, but were identified as ecologically conducive for plague activity, could have been sampled during a locally quiescent period. For example, the model predicted an incidence of 18 cases per 1,000 population for Central Parish in southeastern Okoro County, yet no cases were reported from 1999 to 2007. However, a retrospective record review from 1986 to 1998 indicated hundreds of plague cases were reported to Paidha Clinic located within Central Parish. Similarly, during 2008 a patient with plague was reported from Ezuku parish in northern Vurra County where no cases were reported from 1999 to 2007. Together, these observations suggest that the areas identified by the model as high risk are ecologically conducive for plague activity but outbreaks may occur more sporadically. Additionally, ecologic similarities between other parishes where plague was reported and the outliers in northern Vurra County lead to the speculation that the overprediction of the model in these parishes may be a result of reporting anomalies. These parishes in northern Vurra County are therefore prime candidates for prospective studies and enhanced plague surveillance.
There were several parishes in southern and central Vurra County where the model expected fewer cases than actually occurred. Our model development was based on clinically diagnosed plague cases because standard laboratory confirmation of cases was not a common practice in this area. 51,52 It is possible that during outbreaks the index of suspicion for plague was elevated and this may have resulted in over-reporting of plague cases in some parishes. Additionally, the under-prediction of the model may be attributed to over-reporting of plague cases from health clinics bordering the DRC resulting from patients crossing the border to seek care in Uganda. Although we attempted to account for this by excluding patients who reported that their village of residence was in DRC, village of residence was self-reported and anecdotal evidence suggests that some patients from DRC may inaccurately state that their village of residence is in Uganda.
Plague prevention and control recommendations typically include implementation of vector control, reduction of rodent abundance through elimination of harborage or food sources in and around homes, 12,53–55 or rodent control measures when appropriate, 54 maintenance of surveillance activities that can aid in predicting where future cases are likely to occur, 19 and education campaigns aimed at raising awareness of the disease among health care workers and the public. 56 The spatial resolution of the model is useful for identifying parishes at elevated risk for plague, which may aid in identifying clinics that are most likely to see plague patients and suggest areas where surveillance should be enhanced. This information can be used to target educational campaigns and increase awareness at the parish level. However, because risk within individual parishes is considered homogeneous by the model, it is not sufficient for identifying villages that are most at risk and therefore is of limited use for spatially targeting vector control activities. Future development of spatial risk models based on exposure site locations for plague cases (and negative control locations) would provide fine resolution outputs that may be used to target vector control activities and could serve as a tool for further targeting enhanced surveillance activities. Such information could also explain parish-level model errors. For example, if villages at highest risk within a parish are a long distance to a health care clinic, cases may be under-reported. Identifying these villages could aid in mobilizing resources to these under-served areas at elevated risk. In addition, fine-scale spatial risk modeling could aid in refining the understanding of the ecologic factors that are predictive of risk. This information would be useful for extrapolating the risk assessment model to areas where consistent enhanced epidemiologic surveillance activities are lacking and where the burden of disease remains unclear. For example, models from the West Nile region in Uganda very likely are applicable to areas with similar elevations and ecologic characteristics in the northeastern part of the DRC directly across the border from the study area.
Parameter estimates for the multivariate logistic regression model predicting elevated vs. low risk of human plague in portions of Arua and Nebbi districts, Uganda
Evaluation of the logistic regression model for parishes classified as having elevated or low plague risk compared with reported plague cases
Multivariate linear regression models for incidence of human plague including model parameters*
Reported cumulative incidence per 1,000 population of human plague (1999–2007) by parish in Arua and Nebbi districts. This figure appears in color at www.ajtmh.org.
Citation: The American Journal of Tropical Medicine and Hygiene Am J Trop Med Hyg 80, 6; 10.4269/ajtmh.2009.80.1014
Reported cumulative incidence per 1,000 population of human plague (1999–2007) by parish in Arua and Nebbi districts. This figure appears in color at www.ajtmh.org.
Citation: The American Journal of Tropical Medicine and Hygiene Am J Trop Med Hyg 80, 6; 10.4269/ajtmh.2009.80.1014
Reported cumulative incidence per 1,000 population of human plague (1999–2007) by parish in Arua and Nebbi districts. This figure appears in color at www.ajtmh.org.
Citation: The American Journal of Tropical Medicine and Hygiene Am J Trop Med Hyg 80, 6; 10.4269/ajtmh.2009.80.1014
Parishes reporting plague cases compared with the prediction from the multivariate logistic regression model for elevated vs. low risk of plague. This figure appears in color at www.ajtmh.org.
Citation: The American Journal of Tropical Medicine and Hygiene Am J Trop Med Hyg 80, 6; 10.4269/ajtmh.2009.80.1014
Parishes reporting plague cases compared with the prediction from the multivariate logistic regression model for elevated vs. low risk of plague. This figure appears in color at www.ajtmh.org.
Citation: The American Journal of Tropical Medicine and Hygiene Am J Trop Med Hyg 80, 6; 10.4269/ajtmh.2009.80.1014
Parishes reporting plague cases compared with the prediction from the multivariate logistic regression model for elevated vs. low risk of plague. This figure appears in color at www.ajtmh.org.
Citation: The American Journal of Tropical Medicine and Hygiene Am J Trop Med Hyg 80, 6; 10.4269/ajtmh.2009.80.1014
Multivariate linear regression model predicting plague incidence per 1,000 population. Inset graph identifies parishes in Vurra County where the model over (diamond) or under (triangle) predicts incidence. This figure appears in color at www.ajtmh.org.
Citation: The American Journal of Tropical Medicine and Hygiene Am J Trop Med Hyg 80, 6; 10.4269/ajtmh.2009.80.1014
Multivariate linear regression model predicting plague incidence per 1,000 population. Inset graph identifies parishes in Vurra County where the model over (diamond) or under (triangle) predicts incidence. This figure appears in color at www.ajtmh.org.
Citation: The American Journal of Tropical Medicine and Hygiene Am J Trop Med Hyg 80, 6; 10.4269/ajtmh.2009.80.1014
Multivariate linear regression model predicting plague incidence per 1,000 population. Inset graph identifies parishes in Vurra County where the model over (diamond) or under (triangle) predicts incidence. This figure appears in color at www.ajtmh.org.
Citation: The American Journal of Tropical Medicine and Hygiene Am J Trop Med Hyg 80, 6; 10.4269/ajtmh.2009.80.1014
Address correspondence to Anna M. Winters, Division of Vector-Borne Infectious Diseases, Centers for Disease Control and Prevention, 3150 Rampart Rd., Fort Collins, CO 80522. E-mail: AWinters1@cdc.gov
Authors’ addresses: Anna M. Winters, Division of Vector-Borne Infectious Diseases, Centers for Disease Control and Prevention, 3150 Rampart Rd. Fort Collins, CO 80522 and Department of Microbiology, Immunology and Pathology, Colorado State University, Fort Collins, CO 80523, E-mail: AWinters1@cdc.gov. J. Erin Staples, Paul S. Mead, Kevin Griffith, Russell E. Enscore, Kenneth L. Gage, and Rebecca J. Eisen, Division of Vector-Borne Infectious Diseases, Centers for Disease Control and Prevention, 3150 Rampart Rd. Fort Collins, CO 80522. Asaph Ogen-Odoi, deceased, formerly of the Uganda Virus Research Institute, Entebbe, Uganda. Nicholas Owor and Nackson Babi, Uganda Virus Research Institute, Entebbe, Uganda. Lars Eisen, Department of Microbiology, Immunology and Pathology, Colorado State University, Fort Collins, CO 80523.
Acknowledgments: We thank Chris Sexton for assisting in the extraction and compilation of epidemiologic data for the area; Patrick Kariuki and Wm. Steve Helterbrand for geographic information system data for the study area; Emily Zielinski-Gutierrez, Heidi E. Brown, and Ingrid Weber for helpful discussions.
REFERENCES
- 1↑
Gage KL, Kosoy MY, 2005. Natural history of plague: perspectives from more than a century of research. Annu Rev Entomol 50 :505–528.
- 2↑
Dennis DT, Gage KL, Plague. Cohen J, Powderly WG, eds. Infectious Diseases. Volume 2. London: Mosby, 2003; 1641–1648.
- 3
WHO, 2004. Human plague in 2002 and 2003. Wkly Epidemiol Rec 79 :301–308.
- 5↑
Kilonzo BS, 1999. Plague epidemiology and control in eastern and southern Africa during the period 1978 to 1997. Cent Afr J Med 45 :70–76.
- 7↑
Aikimbajev A, Meka-Mechenko T, Temiralieva G, Bekenov J, Sagiyev Z, Kaljan K, Mukhambetova AK, 2003. Plague in Kazakhstan at the present time. Przegl Epidemiol 57 :593–598.
- 8
Hull HF, Montes JM, Mann JM, 1987. Septicemic plague in New Mexico. J Infect Dis 155 :113–118.
- 9
Mann JM, Schmid GP, Stoesz PA, Skinner MD, Kaufmann AF, 1982. Peripatetic plague. JAMA 247 :47–48.
- 10
MMWR, 1994. Human plague—United States, 1993–1994. MMWR 43 :242–246.
- 11
MMWR, 2002. Imported plague: New York City, 2002. MMWR 52 :725–728.
- 13
Crook LD, Tempest B, 1992. Plague. A clinical review of 27 cases. Arch Intern Med 152 :1253–1256.
- 14↑
Pollitzer R, 1954. Plague. World Health Organization Monograph Series No. 22. Geneva, Switzerland: World Health Organization.
- 15↑
Eisen RJ, Reynolds PJ, Ettestad P, Brown T, Enscore RE, Biggerstaff BJ, Cheek J, Bueno R, Targhetta J, Montenieri JA, Gage KL, 2007. Residence-linked human plague in New Mexico: a habitat-suitability model. Am J Trop Med Hyg 77 :121–125.
- 16↑
Nakazawa Y, Williams R, Peterson AT, Mead P, Staples E, Gage KL, 2007. Climate change effects on plague and tularemia in the United States. Vector Borne Zoonotic Dis 7 :529–540.
- 17↑
Neerinckx SB, Peterson AT, Gulinck H, Deckers J, Leirs H, 2008. Geographic distribution and ecological niche of plague in sub-Saharan Africa. Int J Health Geogr 7 :54.
- 19↑
Gage KL, 1999. Plague surveillance. Plague Manual: Epidemiology, Distribution, Surveillance and Control. Geneva: WHO, 135–165.
- 20↑
Hopkins GHE, 1949. Report on Rats, Fleas and Plague in Uganda. Nairobi, Kenya: East African Standard, Ltd., 52.
- 21↑
Orochi-Orach, 2002. Plague Outbreaks: The Gender and Age Perspective in Okoro County, Nebbi District, Uganda. Nebbe, Uganda: Agency for Accelerated Regional Development.
- 22↑
Jensen J, 2007. Remote Sensing of the Environment: An Earth Resource Perspective. Second edition. Upper Saddle River: Prentice Hall.
- 23↑
Chander G, Markham B, 2003. Revised Landsat-5 TM radiometric calibration procedures and postcalibration dynamic ranges. IEEE Trans Geosci Rem Sens 41 :2674–2677.
- 24
Chen Y, Guo D, Zhou T, 1999. Noise removal of thermal image and the determination of the temperatures. Remote Sensing Information 11 :7–9.
- 25
Han-qiu X, Ben-qing C, 2004. Remote sensing of the urban heat island and its changes in Xiamen City of SE China. J Environ Sci (China) 16 :276–281.
- 26
Landsat 7 Science Data Users Handbook. Available at: http://landsathandbook.gsfc.nasa.gov/handbook/handbook_toc.html. Accessed December 3, 2008.
- 28↑
Getting Started with ENVI—version 4.5. Available at: http://www.ittvis.com/ProductServices/ENVI/ProductDocumentation.aspx. Accessed December 3, 2008.
- 29↑
Jensen J, 1995. Introductory Digital Image Processing: A Remote Sensing Perspective. Second edition. Upper Saddle River: Prentice Hall.
- 30↑
Gelman A, Hill J, 2007. Data Analysis Using Regression and Multilevel/Hierarchical Models. New York: Cambridge University Press.
- 31↑
Burnham K, Anderson D, 2002. Model Selection and Multimodel Inference: A Practical Information-Theoretic Approach. New York: Springer.
- 32↑
Sall J, Creighton L, Lehman A, 2005. JMP Start Statistics. Third edition. Belmont: Brooks/Cole, 584.
- 33↑
Eisen RJ, Borchert JN, Holmes JL, Amatre G, Van Wyk K, Enscore RE, Babi N, Atiku LA, Wilder AP, Vetter SM, Bearden SW, Montenieri JA, Gage KL, 2008. Early-phase transmission of Yersinia pestis by cat fleas (Ctenocephalides felis) and their potential role as vectors in a plague-endemic region of Uganda. Am J Trop Med Hyg 78 :949–956.
- 34↑
Eisen RJ, Enscore RE, Biggerstaff BJ, Reynolds PJ, Ettestad P, Brown T, Pape J, Tanda D, Levy CE, Engelthaler DM, Cheek J, Bueno R, Targhetta J, Montenieri JA, Gage KL, 2007. Human plague in the southwestern United States 1957–2004: spatial models of elevated risk of human exposure to Yersinia pestis. J Med Entomol 44 :530–537.
- 35↑
Eisen RJ, Glass GE, Eisen L, Cheek J, Enscore RE, Ettestad P, Gage KL, 2007. A spatial model of shared risk for plague and hantavirus pulmonary syndrome in the southwestern United States. Am J Trop Med Hyg 77 :999–1004.
- 36↑
Velimirovic B, Zikmund V, Herman J, 1968. Plague in the Lake Edwards focus: the Democratic Republic of Congo, 1960–1966. Z Tropenmed Parasitol 20 :373–387.
- 37↑
Davis DH, 1949. Current methods of controlling rodents and fleas in the campaign against bubonic plague and murine typhus. J R Sanit Inst 69 :170–175.
- 38↑
Davis DH, 1953. Plague in Africa from 1935 to 1949; a survey of wild rodents in African territories. Bull World Health Organ 9 :665–700.
- 39↑
Cavanaugh DC, Marshall JD Jr, 1972. The influence of climate on the seasonal prevalence of plague in the Republic of Vietnam. J Wildl Dis 8 :85–94.
- 40
Cavanaugh DC, Dangerfield HG, Hunter DH, Joy RJ, Marshall JD Jr, Quy DV, Vivona S, Winter PE, 1968. Some observations on the current plague outbreak in the Republic of Vietnam. Am J Public Health Nations Health 58 :742–752.
- 41↑
Marshall JD, Ouy DV, Gibson FL, Dung TC, Cavanaugh DC, 1967. Ecology of plague in Vietnam: commensal rodents and their fleas. Mil Med 132 :896–903.
- 42↑
Enscore RE, Biggerstaff BJ, Brown TL, Fulgham RE, Reynolds PJ, Engelthaler DM, Levy CE, Parmenter RR, Montenieri JA, Cheek JE, Grinnell RK, Ettestad PJ, Gage KL, 2002. Modeling relationships between climate and the frequency of human plague cases in the southwestern United States, 1960–1997. Am J Trop Med Hyg 66 :186–196.
- 43
Parmenter RR, Yadav EP, Parmenter CA, Ettestad P, Gage KL, 1999. Incidence of plague associated with increased winter-spring precipitation in New Mexico. Am J Trop Med Hyg 61 :814–821.
- 45↑
Eisen RJ, Gage KL, 2009. Adaptive strategies of Yersinia pestis to persist during inter-epizootic and epizootic periods. Vet Res 40 :1.
- 46
Keeling MJ, Gilligan CA, 2000. Metapopulation dynamics of bubonic plague. Nature 407 :903–906.
- 47
Snall T, O’Hara RB, Ray C, Collinge SK, 2008. Climate-driven spatial dynamics of plague among prairie dog colonies. Am Nat 171 :238–248.
- 48
Collinge SK, Johnson WC, Ray C, Matchett R, Grensten J, Cully JF, Gage KL, Kosoy MY, Loye JE, Martin AP, 2005. Landscape structure and plague occurrence in black-tailed prairie dogs on grasslands of the western USA. Landscape Ecol 20 :941–955.
- 49↑
Davis S, Leirs H, Viljugrein H, Stenseth NC, De Bruyn L, Klassovskiy N, Ageyev V, Begon M, 2007. Empirical assessment of a threshold model for sylvatic plague. J R Soc Interface 4 :649–657.
- 50↑
Barnes AM, Maupin GO, 1982. Observations on the biting of humans by Euhoplopsyllus-Glacialis-Affinis (Siphonaptera, Pulicidae) and a review of is plague-transmission potential. J Med Entomol 19 :748–749.
- 51↑
Poland JD, 1999. Diagnosis and clinic manifestations. Plague Manual: Epidemiology, Distribution, Surveillance and Control. Geneva: World Health Organization.
- 52↑
Chu MC, 2000. Laboratory Manual of Plague Diagnostics. Geneva: Centers for Disease Control and Prevention and World Health Organization, 129.
- 53↑
Akiev AK, 1982. Epidemiology and incidence of plague in the world, 1958–79. Bull World Health Organ 60 :165–169.
- 54↑
Gratz NG, 1999. Control of plague transmission. Plague Manual: Epidemiology, Distribution, Surveillance and Control. Geneva: World Health Organization, 97–134.
- 55↑
Poland JD, Barnes AM, 1979. Plague. Steele JH, ed. CRC Handbook Series in Zoonoses. Section A: Bacterial, Rickettsial and Mycotic Diseases. Volume I. Boca Raton: CRC Press Inc., 515–559.
- 56↑
Gage KL, 1999. National health services in prevention and control. Plague Manual: Epidemiology, Distribution, Surveillance and Control. Geneva: WHO, 167–171.