A descriptive analysis of the spatio-temporal distribution of enteric diseases in New Brunswick, Canada

Laboratory confirmed case counts, as defined by New Brunswick provincial reportable disease system, for Campylobacter spp., Escherichia coli, Giardia spp., Salmonella spp. and Shigella spp. were obtained from the New Brunswick Department of Health and Wellness for the time period January 1, 1994 to December 31, 2002. Patient identification numbers had been removed from the data and a sequential identification number added to preserve patient confidentiality and to meet the University of Guelph’s Research Ethics Board Guidelines for the use of secondary data sources. The data provided contained information on the age, sex, reporting date, diagnosis date and postal code of the case. Due to privacy issues, full postal code data was unavailable after 2002, as such, data after this year was not included in the study.

A large number of cases were removed due to incomplete or missing postal codes which were needed to link case information to census subdivisions (CSD) or watershed (WS) identifiers for purposes of aggregation (Table 1). Other cases were removed because the provided postal code could not be mapped to known New Brunswick watershed or CSD areas (i.e. postal code information was complete but was geographically mapped outside of New Brunswick provincial boundaries). Other reasons why cases were removed included incomplete date information, invalid postal code and large discrepancies between admission date and reporting date. A summary of the number of cases used in the study can be seen in Table 1. The majority of the Escherichia coli cases were of the O157 serotype, with only 2.8 % (6/215) being non-O157. For the purposes of this study, where E. coli cases are referred to, they will be referred as E. coli thus encompassing both O157 and non-O157 strains.

To assess the temporal distribution of the 5 different pathogens, case counts, for each pathogen, were summed to create weekly and monthly case counts for the entire province. Week was defined as the week of the year (1–52) in which reporting occurred, or the month of the year (1–12) in which reporting occurred. These were used to estimate the incidence rate for each pathogen respectively using population estimates from the 1996 population census [26] in the denominator.

For spatial analysis, case counts were calculated for the two spatial scales (CSD and WS) that would be used in the spatial analysis. The postal code field contained in the reportable disease dataset obtained from New Brunswick Health and Wellness was used to link data from that dataset to CSD and WS identifiers (see "Population Data for Spatial Analysis" below) and population estimates. All cases in a given geographical area were summed over the duration of the study period. The case counts were then used to estimate incidence rates for each area using the population estimate for that area (see below) and expressed on a yearly basis (cases per person per year).

Monthly incidence rate time series were plotted for each reportable disease. A moving average for 3 month (seasonally) and 13 month (yearly) smoothed incidence rates were estimated. These estimates were plotted in conjunction with the raw monthly incidence rates.

Negative binomial regression was used to test the significance of long term trends by examining monthly incidence rates over the length of the study period [27]. A negative binomial distribution was used because in several cases disease incidence exhibited significant overdispersion, thus a Poisson distribution would not be appropriate. Month of study was used as an explanatory variable of monthly cases of disease to control for seasonal and holiday effects related to month. Population estimates of New Brunswick were obtained from the 1996 Canadian Census and were used as the exposure offset for modelling purposes.

Seasonal decomposition of monthly incidence rates was performed using the data driven LOESS smoothing method outlined in [28]. Monthly incidence was used for the seasonal decomposition to minimize noise present in the time series plots. The temporal scan statistic developed by [29] was used to examine the data for temporal clusters of disease. A purely temporal scan statistic gives an indication into the most likely temporal clusters of disease. Each year of the data series was examined for the occurrence of peak disease incidence. A Poisson distribution was used with a maximum temporal cluster size of 60 days. Kulldorff [30] recommends a maximum temporal cluster size that is 50 % of the length of the study period. That would represent a maximum temporal cluster size of 183 days as the test was conducted on each year of the data. A maximum temporal cluster size that was smaller than recommended value was used because the objective was to identify relatively short periods of time that may give an indication of seasonality of disease incidence. Data manipulation was carried out using Stata version 9.1 [31]. Graphing, moving average calculations and seasonal decomposition were performed in R [32]. Temporal clusters of disease were determined using the cluster detection software package SaTScan [30].

An enhanced postal code file was used to determine estimates of the population for each geographical area using population values for each postal code based on the 1996 census of population [33]. Postal codes were linked to CSD and WS information using a spatial join in ArcGIS [34]. A postal code was considered to associated with a CSD or WS if the center of the postal code inside the boundary of the CSD or WS. The population for each geographical region was estimated by summing the population estimates of all postal codes in the geographical area. Age and sex distribution was not available for each postal code, therefore population estimates for age and sex strata could not be determined. This prevented the estimation of age and sex adjusted rates for each CSD/WS.

Empirical Bayes (EB) estimates were calculated for incidence rates at both the CSD and WS geographical scales [35]. Mapping and estimation of EB smoothed incidence rates was performed using the spatial dependence (spedep) library [36] of the R software package [32].

Moran’s I was used to examine spatial clustering of the incidence rates. Specifically Moran’s I was applied to the EB smoothed incidences, because EB smoothing is a kind of spatial standardization with respect to varying sample sizes. Furthermore, the Queen contiguity structure was specified, which connects each region with all its direct neighbours [37].

The spatial scan statistic [29] was used to examine raw case counts for clusters of disease. Raw, rather than smoothed rates, were also used with the spatial scan statistic. Data were examined for high disease rate clusters using pure spatial clusters based on a Poisson model with a maximum geographical cluster size of 20 % of the population at risk. A maximal geographical cluster size of 20 % of the population was chosen as enteric diseases are rare in the population and it is unlikely that more than 20 % of the population would be at risk in a single outbreak. Since enteric disease outbreaks tend to be governed by localized risk factors and disease processes, only a fraction of the population should be affected. The optimal parameter setting for the maximum cluster size is often not clear [38]. Adjustments were made to allow for testing of significance of secondary clusters [39]. SaTScan software [30] was used to identify and test significance of spatial clusters. Statistical methodology for the spatial scan statistic has been described elsewhere [40]. Significant spatial clusters identified by the spatial scan statistic were mapped using ArcGIS [34].

Census subdivision shapefiles were obtained from Statistics Canada and watershed boundary shapefiles were obtained from Environment Canada. Choropleth maps of EB smoothed incidence rates were mapped using ArcGIS [34].

Summary statistics for weekly incidence rates are presented in Table 2 and indicate that Campylobacter was the most commonly reported enteric pathogen infection in New Brunswick based on the median and mean incidence. Salmonella and Giardia infections were second and third most commonly reported enteric pathogens, respectively. E. coli and Shigella infections had the lowest median incidence of reported cases. Reportable E. coli infections had a higher mean incidence than reported Shigella infections.

Time series plots of reportable Campylobacter, E. coli, Giardia, Salmonella and Shigella can be seen in Figure 1. Except for reportable Shigella infections, all time-series plots exhibited a marked seasonal component based on 3 month moving average values. Several peaks were seen throughout the year with the majority of the peaks occurring in winter and spring.

Negative binomial regression was performed for each of the disease outcomes to determine the significance of any secular trend. Table 2 reports the estimate of the slope parameter that represents the trend in the time series of reportable enteric pathogens. The incidence rate of Campylobacter, E. coli, Salmonella and Shigella infections reported in New Brunswick all exhibited a significant increasing secular trend (Table 2). The incidence of reportable Giardia infections did not exhibit a trend in time (Table 2).

Seasonal decomposition of reportable enteric disease rates are shown in Fig. 2. The incidence of Campylobacter infections peaked in early summer. The incidence of human E. coli infections showed a bi-modal seasonal pattern with the largest peaks occurring in the summer and a smaller peak occurring January. Several peaks were observed in the seasonal decomposition of Giardia disease rates. Two large peaks were observed in winter and spring. A smaller peak can be seen in late autumn. Seasonal decomposition of the incidence of Salmonella infections indicated a bi-modal seasonal pattern with the peak incidence occurring during the summer months with a peak about half the size occurring in spring. No discernable seasonal pattern could be extracted from the Shigella incidence.

Primary temporal clusters identified by the temporal scan statistic are presented in Table 3 for each year of the study period for each disease outcome. Six disease clusters were identified for the incidence of Campylobacter infections. All clusters occurred in the summer months (July, August, and September). There were five temporal clusters for the incidence of E. coli infections. Clusters occurred in late spring extending into the summer months. Three temporal clusters of the incidence for Giardia infections were identified and all occurred in spring. Five temporal clusters identified for the incidence of reportable Salmonella infections. The temporal distribution of these disease clusters was more varied than with the other pathogens. Two temporal clusters were identified in the fall (October, November, and December) and one in each of the winter, spring and summer. No temporal clusters were identified for the incidence of Shigella infections.

Choropleth maps of the EB smoothed yearly incidence of reportable Campylobacter, E. coli, Giardia, Salmonella and Shigella for CSD areas of New Brunswick can be seen in Fig. 3. Upon visual inspection, there appears to be some grouping of areas with high incidence, with the exception of Shigella infection rates. The spatial distribution of EB smoothed incidence rates of reported Campylobacter infections in CSD areas was highest in the middle of the province, with other potential areas of high incidence in the west and east of the province. The incidence of EB smoothed E. coli infections in New Brunswick CSD areas were evenly distributed throughout the province with a possible grouping of high incidence areas in the north-east corner of the province. Spatial distribution of the EB smoothed incidence rates of reported Giardia, Salmonella and Shigella infections at the CSD level were evenly distributed throughout the province.

Watershed based maps of the yearly incidence rates of reportable Campylobacter, E. coli, Giardia, Salmonella and Shigella are shown in Fig. 4. Higher incidences rates of reported Campylobacter infections can be seen in the north area of the province with a possible grouping of high incidence areas in the north-west. E. coli incidence rates were highest in the north, central and south-east areas of the province. EB smoothed Giardia rates were evenly distributed throughout the province with areas of higher incidence located in the north-west corner of the province. The spatial distribution of EB smoothed reported Salmonella incidence rates was evenly distributed through the province with the possible exception in the north of the province where areas of higher incidence were grouped. Visual inspection of EB smoothed Shigella incidence rates revealed no discernable spatial pattern.

Moran’s I and associated p-values for the incidence of enteric illness in New Brunswick are summarized in Table 4. Positive spatial autocorrelation was observed at both the CSD and watershed levels for the incidence of Campylobacter, Giardia and Salmonella infections. No evidence was found for autocorrelation within the incidence of Shigella and Escherichia coli infections at the CSD level. Moran’s I for E. coli at the watershed level was not significant indicating that there was no spatial autocorrelation present.

Maps showing the results of the spatial scan statistic can be seen in Fig. 5 for CSD level data and Fig. 6 for WS level data and Tables 5 and 6 summarize the findings for the CSD and WS levels, respectively. The relative risks represent the risk of disease inside the cluster (i.e. the exposed population) over the risk of disease outside the cluster (the unexposed population). Spatial clusters in the incidence of enteric diseases were observed for all pathogens being studied at both the CSD and WS geographical scales, with the exception of the incidence of Shigella infections where no significant spatial clusters were detected at either the CSD or WS level. Exact locations of spatial clusters of high incidence are shown in Figs. 5 and 6. Relative risk for spatial clusters of reported enteric pathogen infections at the CSD level ranged from 2.2 to 54.8, depending on the pathogen (Table 5); smaller relative risks were observed at the WS level with relative risks ranging from 1.8 to 34.2 (Table 6).

The spatial scan statistic identified that cases of enteric disease cluster. The general location and size of the reportable disease clusters detected by the spatial scan statistic did not differ greatly when comparing analyses from the CSD level (Fig. 5) to that of watershed level (Fig. 6). Previous studies have observed that the specification of a finer geographic resolution did not improve sensitivity of the scan test to detect variations in risk of disease (i.e. to detect disease clusters) [41]. The agreement between the results from both geographic scales strengthens the hypothesis that a possible common exposure contributes to the incidence of reportable enteric disease in New Brunswick; disagreement between the two geographical scales may be a result of the modifiable areal unit problem (MAUP), which is a consequence of aggregating data across different geographical scales [42, 43].

Aggregation of data to a higher geographical level provided a more stable estimate of disease rates as denominators were larger and the addition of a small number of cases (1–2 cases) did not greatly affect the estimated rate for that area. In several cases, the denominators for the CSD level data were very small or zero. This was less of a problem when data were expressed on the watershed level. Differences in the spatial pattern of disease incidence between CSD and WS levels were expected, due to the differences in population size between CSD and WS areas. The two levels of aggregation are also distinct in that CSD areas are political boundaries, whereas watershed boundaries follow natural geographical features. CSD areas are convenient from a resource allocation perspective, but underlying environmental exposure patterns typically do not follow such political boundaries. By examining disease incidence according to more naturally occurring boundaries, potential environmental exposures may be explored more appropriately.

Interpretation of EB smoothed maps was complicated by the nature of the CSD areas where rural areas tend to be represented by large geographical areas, which visually overwhelm urban areas that are represented by smaller geographical areas. In these cases, the larger areas provide a greater visual stimulus, more so than the smaller areas, thus attracting the eye. The use of choropleth maps in mapping health data provided a minimal level of information about the spatial distribution of disease incidence; their use is limited due to the problems outlined above.

There are a number of potential disease pathways that may contribute to the clustering of enteric illness incidence. Large municipalities in New Brunswick, as well as some incorporated areas, have municipal water supplies that draw water from surface and ground sources. These regions employ water treatment and distribution systems [44]. Rural areas, on the other hand, which account for about 40 % of the population (approximately 300,000 persons), primarily source their water from private wells [44]. These wells are prone to contamination and are not regularly tested [44]. Studies have identified living in rural areas as a risk factor for enteric illness [13, 15]. Small towns and rural areas may have lacked the necessary water treatment facilities to effectively treat water against contamination with enteric pathogens.

Livestock rearing and manure management practices may be contributing to the spatial distribution of enteric disease incidence. Studies in Ontario and Alberta have identified living in proximity to livestock and agricultural land treated with manure [15, 45‐47] as risk factors for infection. Not all pathogens examined in this study have previously demonstrated associations with agriculture. Odoi et al. [25] did not find an association between the incidence of reported Giardia infections and livestock density in Ontario. Only about 5 % of New Brunswick’s area is dedicated to agricultural land [48]. The effect of agriculture on enteric disease incidence in New Brunswick is not well understood. Further investigation is required, particularly in light of the fact that agricultural land use is limited and that agriculture alone does not explain the variation in enteric disease incidence [49].

Forestry is the primary industry of New Brunswick and approximately 80 % of New Brunswick’s land mass is productive forest [50]. Little, if anything, is known about the impact of forestry practices on the incidence of enteric illness. The harvesting of lumber has been shown to increase the rate of run-off from the land, increase stream flow, and increase the rate and timing of snow melt [51]. Increased stream flows have been associated with increased risk of flooding [51]. This has the potential to contaminate local surface water supplies, or contaminate improperly maintained or damaged wells.

Seasonal time series decomposition of Giardia incidence identified peaks that have previously been reported [13, 14], but also identified seasonal fluctuations in disease incidence that have not been reported before. In particular, the spring peak for Giardia did not follow the pattern observed in other Canadian studies of Giardia infections. Studies conducted in two Canadian provinces [13, 14, 52], Ontario and Alberta, and the United States [8] have reported peaks in Giardia incidence during the late summer to early fall. Seasonal Decomposition of Salmonella incidence and the temporal clusters of disease identified seasonal peaks that agree with the seasonal patterns reported by Guerin et al. [17], who reported peak disease incidence as occurring in July to September; although a report from Ontario indicated peak incidence in the spring [9] which was also observed in this study. The seasonality of other bacterial pathogens in the study is consistent with previously reported findings. Both the seasonal decomposition and the temporal scan statistic indicate that peak incidence occurred during the summer months. Previous studies have demonstrated peak incidence for E. coli and Campylobacter infections during the spring and summer months [15, 16, 19, 53‐55].

There are a number of potential risk factors that may play a role in disease incidence and explain the observed peaks in enteric disease incidence. Among these are natural weather patterns and agricultural activities. Weather has been identified as a factor in triggering waterborne disease outbreaks, primarily related to excess rainfall [3, 56‐58]. A study of waterborne outbreaks in Canada revealed that outbreaks peaked during the spring/summer and that these were primarily related to meteorological conditions or extreme weather events [59]. The peak incidence of Giardia and the smaller secondary peak incidence seen for Salmonella infections possibly correspond with peak run-off from spring melt. Atherholt et al. [60] observed increased level of Giardia cysts in river water following snow melt events. Spring snowmelt in New Brunswick peaks in mid-April to May and is responsible for the largest stream flow volumes [61]. This is in contrast to Ontario and Alberta, which get approximately half the amount of snow annually [62]. The timing of spring thaws for Ontario and Alberta also differs from New Brunswick with Ontario’s thaw occurring in March and Alberta having a much more variable winter with thaws throughout the season [62].

A number of studies have demonstrated the role of agriculture in enteric illness [15, 21, 45, 54]. Agriculture may impact disease rates through a variety of different routes. Direct contact with livestock or run-off from agricultural land has been identified as a risk factor for infections with Campylobacter, E. coli O157 and Salmonella [2, 54, 57, 63, 64].

Peaks in disease incidence could be related to the occurrence of outbreaks, as these were not adjusted for in the analysis. An examination of outbreaks during the study period (data not shown) identified several outbreaks that coincided with identified temporal clusters of disease. While this explains some of the temporal clustering of Salmonella cases, there did not appear to be any reported outbreaks in the literature that could explain the observed temporal clustering of Giardia cases.

The spring peaks observed in the incidence of Giardia and Salmonella infections in New Brunswick could be due to unmeasured factors. However, they may also be due to a different etiological pathway in New Brunswick compared to Ontario and Alberta. This study was unable to assess the effects of differences in population dynamics (e.g. age, sex, socio-economic factors), but the age and sex population distributions in New Brunswick are reflective of the national distribution [65]. This gives an indication that these unmeasured factors may be environmental in nature (e.g. geography or climate) as opposed to demographic factors [60].

Reportable enteric disease incidence rates appeared to be increasing in New Brunswick over the time period studied. Four of the five pathogens exhibited a linear increase in disease incidence over the study period (Table 1). The incidence rate of Giardia infections was the only disease that remained stable over the study period. This differs from trends seen at the National level [1]. Possible explanations for the increase in New Brunswick of incidence rates of Campylobacter, E. coli, Shigella and Salmonella include a true increase in the incidence of these infections over time or that there may also be an increase in the reporting of these diseases over the study period causing an apparent increase in disease incidence. The study period was also relatively short and this could have an impact on the temporal trends observed. An annual report by the New Brunswick Department of Health and Wellness, published in 2013 [66], indicated that Campylobacter and E. coli incidence rates have been stable since 2003; whereas reported Giardia, Salmonella and Shigella infections have increased slight in incidence. The differences observed between this study and current rates of disease may reflect difference in analytical techniques and temporal scales used to examine long term trends in the data.

Reportable disease data under represent the true number of cases in the population. It has been estimated for Ontario, Canada that for every reported case of enteric illness, 313 cases are not captured by the system [5]. This has the potential to introduce bias into the study if the population captured by the reporting system differs significantly from the New Brunswick population. It was not possible to remove outbreak-associated cases of enteric illness from the database, as information pertaining to whether a case arose due to an outbreak was not recorded in the database used. This could potentially impact the results of the spatial scan statistic since cases arising from an outbreak tend to cluster spatially. A large number of cases were removed from the data analyzed due to missing data or our inability to link cases to a spatial region, which may introduce bias if those cases are related to some spatial or temporal factor. Bias could be a problem if these cases differ significantly from the New Brunswick population that they represent. Lastly, it was not possible to distinguish between cases arising from foodborne, waterborne or person-to-person transmission. The etiology of disease incidence in the cases may be different and exhibit different spatial and temporal patterns.