Incidence, distribution, seasonality, and demographic risk factors of Salmonella Enteritidis human infections in Ontario, Canada, 2007–2009

In Ontario, a confirmed case of salmonellosis is defined as the isolation of Salmonella (excluding Salmonella Typhi or Paratyphi) from an appropriate clinical sample (e.g., stool, urine, blood) with or without clinically compatible signs and symptoms [36]. Data pertaining to the S. Enteritidis cases’ age, sex, reporting PHU, and date of illness onset were acquired from the iPHIS database. The University of Guelph Ethics Review Board was consulted because our research involved human participants; however, ethics approval was not required because our data did not contain any personal or health information that could be linked back to the original identifiers. The data represent all cases of S. Enteritidis that were captured within the database between January 1, 2007 and December 31, 2009. Travel-related (i.e., those who had traveled outside of Canada within 3 days before the onset of illness) and outbreaks (two or more epidemiologically-linked cases) were included in the analysis because the study objectives were to describe the overall epidemiology of S. Enteritidis infections in Ontario.

The Census of Canada is administered every five years by Statistics Canada, to collect demographic and socioeconomic information on Canadians [37]. Estimates based on the 2006 Census of Population for each year, age category, sex, and PHU were obtained from Statistics Canada, Demography Division [38].

The distribution of values was examined, missing data and improbable values were identified, and the data were corrected wherever possible or eliminated from the analysis. Descriptive and statistical analyses were performed using Microsoft Excel 2000 (Microsoft Corporation, Redmond, WA, USA) and STATA Intercooled statistical software, version 10.1 (Stata Corporation, College Station, TX, USA).

Direct standardization [39‐41] was used with the PHU-, year-, age-, and sex-based population as the reference population to calculate annual age-and-sex-adjusted IRs, and annual and mean age-adjusted sex-specific IRs for S. Enteritidis cases in Ontario, and the mean age-and-sex-adjusted IR for each PHU.

To identify associations between S. Enteritidis IRs in Ontario and demographic and seasonal factors, a multivariable Poisson regression analysis was conducted. The dependent variable was the number of S. Enteritidis cases by year, season, age group, sex, and PHU (see Additional file 1 - Legend 1 for an example of the data structure). The categorical independent variables were year, season, age group, sex, and PHU. The variable PHU represented the 36 PHUs in Ontario. The PHU was included as a fixed effect because of the observed variability of the IRs across PHUs (Figure 1). The District of Algoma Health Unit, because it had the lowest IR, was used as the reference category to which the other PHUs were compared. The date of onset of illness reported by each S. Enteritidis case was used to assign the case to a particular year and season. When the date of onset was missing, the date when the sample was received by the laboratory or when the case was reported into the iPHIS database was used. Season was categorized as winter (December, January, and February), spring (March, April, and May), summer (June, July, and August), and fall (September, October, and November). The variable year was defined as a consecutive 12-month period from January 1^st to December 31^st; thus, there were three categories for year (2007, 2008, and 2009). The variable age included ten-year age categories, with the exception of children < 4 years of age and those 5–9 years of age, which were retained because of their biological importance [42, 43], and adults 60 years of age and older, which were pooled into one category because of the small number of cases in this age group. Pair-wise correlation coefficients using the Spearman’s rank test among all variables were examined. If the independent variables were highly correlated (Spearman’s rho > 0.70), variables with the smallest p-value were considered for the model building process. To address the differences in year-, age group-, sex-, and PHU-based population size estimates, we used the natural log-transformed population estimates as the offset, which accounted for the denominator when calculating incidence rate ratios (IRRs). An IRR was the IR in the category of interest compared to the IR in the reference category. Variables with a p-value equal to or less than 0.05 were considered significant and were kept in the model. Incidence rate ratios and their corresponding 95% confidence intervals were estimated. Interaction terms were created between each independent variable and tested for significance. If the interaction term was significant (p ≤ 0.05) it was retained in the final model. The model was evaluated by identifying influential observations (i.e. large values of Cook’s distance) and outliers (i.e. large values of Pearson, deviance, or Anscombe residuals) using residual plots. The overall fit of the model was assessed using Deviance and Pearson χ ² goodness-of-fit tests [44].

To account for lack of fit, a multi-level mixed-effects Poisson regression model was then constructed using the xtmepoisson command in STATA [45], which uses adaptive Gaussian quadrature to approximate the log likelihood. The model included the same dependent and independent (year, season, age group, and sex) variables as the first model with the exception that PHU was included as a random intercept instead of a fixed effect. The structure of the multi-level model included an offset representing the natural log-transformed year-, age group-, sex-, and PHU-based population size estimates. As part of assessing model fit, we examined the normality of the best linear unbiased predictors (BLUPs) [46]. Outlier and influential observations were assessed using residual plots. Bayesian information criterion (BIC) was used to compare the fit of the two models.

ArcGIS 10 (Environmental Systems Research Institute, Inc., Redlands, CA, USA) was used to create a choropleth map for mean age-and-sex-adjusted IRs across Ontario’s PHUs; Jenk's optimization classification method [47] was employed for defining the critical intervals. This method arranges data into classes based on their distribution by using an algorithm that reduces variance within groups and maximizes variance between groups.

The age of cases ranged from < 1 year to > 90 years. Children < 4 years of age and adults ≥ 60 years of age represented 13.1% and 13.6% of cases, respectively, while adults 20–29 years of age represented 16.8% of cases. Overall, 51.6% and 48.4% of cases were females and males, respectively.

The annual age-and-sex-adjusted IR per 100,000 person-years was 4.4 [95% CI 4.0-4.7] in 2007, and 5.2 [95% CI 4.8-5.6] in both 2008 and 2009 (Table 1). Over the study period, the annual age-adjusted sex-specific IR per 100,000 person-years ranged between 4.5 and 5.5 for females and between 4.2 and 5.2 for males (Table 1). The mean age-adjusted sex-specific IR per 100,000 person-years was 5.1 [95% CI 4.8-5.4] for females and 4.8 [95% CI 4.5-5.1] for males (Table 1).

Seasonal counts ranged from 135 to 187 cases in winter (mean over 3-year period = 160 cases), 155 to 189 in spring (mean = 173), 156 to 166 in summer (mean = 160), and 121 to 177 in fall (mean = 151) (Figure 2). The highest monthly count was 75 cases in March 2007 and the lowest was 28 cases in November 2007 (Figure 3).

The mean age-and-sex-adjusted IRs for the entire study period across Ontario’s PHUs ranged from 1.9 to 8.1 (Figure 1; Additional file 2 - Legend 2). Visually exploring the map, the highest IRs (> 6.0 per 100,000 person-years) were observed in three south-western PHUs (Halton Regional Health Unit, Huron County Health Unit, and Waterloo Health Unit), one northern PHU (Thunder Bay District Health Unit), and in the City of Toronto Health Unit.

The Deviance and Pearson χ ² goodness-of-fit test statistics for the Poisson model were 6,212.6 (P = 1.00) and 17,727.6 (P = 0.004), respectively. Several outlier and influential observations were identified; however, re-running the model without these observations did not change any of the coefficients. The BIC for the model was 9,906.4.

Because one of the two goodness-of-fit tests for the Poisson model indicated lack of fit, we used a multi-level model. No outlier or influential observations were identified for the upper level residuals of the multi-level model. The BLUPs for the PHU random intercept were normally distributed. The BIC for the multi-level model was 9,655.2, indicating a better fit.

The results of the multi-level model are shown in Table 2. Significantly higher IRRs of S. Enteritidis infections were reported in 2009 [IRR = 1.18, 95% CI 1.06-1.32] and 2008 [IRR = 1.17, 95% CI 1.05-1.31] compared to 2007. Compared to the fall season, a significantly higher IRR of S. Enteritidis infections was reported in the spring [IRR = 1.14, 95% CI 1.01-1.29]. Children 0–4 years of age (reference category), followed by children 5–9 years of age [IRR = 0.64, 95% CI 0.52-0.78] had the highest IRRs of infection. Adults ≥ 60 years of age and 40–49 years of age [IRR = 0.31, 95% CI 0.26-0.37] had the lowest IRRs of infection. No statistically significant difference in S. Enteritidis infection rates were detected between sexes.