Archived - Investigating the relationship between drinking water and gastroenteritis in Edmonton: 1993 - 1998

Table of contents:

• Authors
• Acknowledgements
• Executive Summary
• 1. Introduction
  • 1.1 Background
  • 1.2 Edmonton's Water Supply
    • 1.2 a) Water treatment plants
    • 1.2 b) Watershed characteristics
    • 1.2 c) Water purification procedures
  • 1.3 Study Objectives
• 2. Data Sources
  • 2.1 Water Quality Parameters
  • 2.2 Environmental Parameters
  • 2.3 Socio-Economic Data
  • 2.4 Health Outcome Data
    • 2.4 a) Canadian Institute for Health Information
    • 2.4 b) Alberta Health Care Insurance Payment Plan
  • 2.5 Assigning a Primary Water Source
• 3. Investigative Methods
  • 3.1 Descriptive Analyses
  • 3.2 Multivariate Logistic Regression
    • 3.2 a) Methodology overview
    • 3.2 b) Variables analysed
    • 3.2 c) Criteria for model-selection
  • 3.3 Time Series Analysis
    • 3.3 a) Methodology overview
    • 3.3 b) Variables analysed
    • 3.3 c) Criteria for model-selection
• 4. Results
  • 4.1 Descriptive Results
    • 4.1 a) Water quality parameters
    • 4.1 b) Environmental parameters
    • 4.1 c) Socio-economic data
    • 4.1 d) Health outcome data
  • 4.2 Statistical Analyses
    • 4.2 a) Multivariate logistic regression
    • 4.2 b) Time series analysis
• 5. Discussion and Conclusions
  • 5.1 Introduction
  • 5.2 Methodological Overview
    • 5.2 a) Multivariate logistic regression
    • 5.2 b) Time series analysis
  • 5.3 Discussion of Results
    • 5.3 a) Descriptive analyses
    • 5.3 b) Multivariate logi stic regression
    • 5.3 c) Time series analysis
  • 5.4 Conclusions
• 6. References

List of Figures

• Figure 1: Regional water supply in Edmonton
• Figure 2: Water service areas for the Rossdale and E.L. Smith plants, Edmonton
• Figure 3: Finished water quality descriptive results
• Figure 4: Raw water quality descriptive results
• Figure 5: Maximum and minimum daily temperatures
• Figure 6: Daily precipitation
• Figure 7: Comparing daily maximum temperature and daily mean raw water turbidity
• Figure 8: Comparing daily precipitation and daily mean raw water turbidity
• Figure 9: Distribution of average annual household incomes per postal code in Edmonton, 1995
• Figure 10: Data profiles
• Figure 11: Spatial distribution of postal codes within Edmonton and corresponding water service areas
• Figure 12: Spatial distribution of disease risk within the city of Edmonton between 1993-1998
• Figure 13: 3-D time series surface plots of disease risk within the Rossdale water service area between Feb. 10 '93 to Dec. 31 '98

List of Tables

• Table 1: Indicators of water quality received from EPCOR Water Services, Inc.
• Table 2: Criteria for selecting hospital admission gastroenteritis cases in the CIHI data source
• Table 3: Criteria for selecting hospital admission respiratory controls in the CIHI data source
• Table 4: Components of the analytic approaches utilised in the multi variatelogistic regression analyses
• Table 5: Independent variables analysed in the multivariate logistic regression analyses
• Table 6: Components of the analytic approaches utilised in the time series analysis
• Table 7: Independent variables analysed in the time series analysis
• Table 8: Summary statistics for indicators of raw and finished water quality
• Table 9: Summary statistics for environmental parameters
• Table 10: Multivariate logistic regression results: Comparisons among water service areas

Authors

• ^{1, 2}Gillian Lim, B.Sc., M.Sc.
• ¹Jeff Aramini, M.Sc., D.V.M., Ph.D
• ¹Manon Fleury, B.Sc.
• ¹Rita Ibarra, B.Sc.
• ¹Rob Meyers, B.Sc.

¹Division of Enteric, Foodborne and Waterborne Diseases,

Centre for Infectious Disease Prevention and Control, Population and Public Health Branch, Health Canada

²Corresponding Author: Gillian Lim

Division of Enteric, Foodborne and Waterborne Diseases CIDPC, PPHB, Health Canada
1 Stone Road West, 4^th Floor,
Guelph, Ontario,
Canada, N1G 4Y2
Tel: (519) 826-2262

Media Inquiries:

Paige Raymond Kovach
Health Canada Media Relations
(613) 957-1803

Acknowledgements

The authors of this report would first like to acknowledge the contributions of Dr. Les Gammie and Simon Thomas of EPCOR Water Services Inc. Their determination to provide us with a comprehensive database consisting of numerous water quality parameters made it possible to conduct an in-depth and detailed analysis. Furthermore, their expertise and advice on water quality and source water issues were invaluable. The authors also thank the Ministry of Health and Wellness, Alberta, for permitting us to access their healthcare administrative databases.

The authors also wish to recognise the contributions of Dr. Brian Allen, Dr. John Holt, and Mr. William Sears at the University of Guelph, for developing the time series methodology that was described in this report.

We would also like to acknowledge Dr. Paul Sockett (Population and Public Health Branch, Health Canada), and Mr. Will Robertson (Healthy Environments and Consumer Safety Branch, Health Canada) for providing input.

Finally, the authors would like to thank the members of the Enteric National Disease Surveillance (ENDS) Committee^{Footnote 1} for recognizing the need for this type of research, and for continuing to support waterborne disease-related research in Canada.

Executive Summary

The risk of microbial disease associated with drinking water is presently a priority concern among North American water jurisdictions. Numerous past outbreaks, together with recent studies suggesting that drinking water may be a substantial contributor to endemic (non-outbreak related) gastroenteritis, demonstrate the vulnerability of many North American cities to waterborne diseases. These findings have fuelled ongoing debates in Canada and the United States, and highlight the need for stricter water quality guidelines, changes in watershed management policies, and the need for additional water treatment.

EPCOR Water Services Inc. supplies water to the city of Edmonton and its surrounding region through two water treatment plants. The Rossdale plant has been upgraded several times since its initial construction during the 1940's, while the E.L. Smith treatment plant was built in 1976. Both plants draw raw water from the North Saskatchewan River, and utilise clarification, softening, recarbonation, filtration, and disinfection in the water treatment process. Downstream of the E.L. Smith plant and upstream of the Rossdale plant, are 85 stormsewers that discharge into the North Saskatchewan River. Covering 28,000 square kilometres, the North Saskatchewan River watershed has many potential points of pathogen introduction: recreational sites, treated wastewater discharge from upstream communities, private septic systems, and agricultural operations. An event of particular interest in this study, was the relocation of the Rossdale raw water intake site on December 10, 1997. The site was moved towards the centre of the river in an attempt to improve raw water quality. Also concurrent with this event was the introduction of particle counters at the Rossdale plant, which allowed for the optimisation of filter performance.

The primary objective of this investigation was to determine if endemic gastroenteritis among Edmonton residents was influenced by the municipal water supply between 1993 to 1998. Several analytic techniques were used to achieve the study objectives. Multi variate logistic regression using generalised linear (GLM) and generalised additive models (GAM) was used to determine if differences in risk for endemic gastroenteritis (as assessed by hospital admissions, emergency room visits, physician visits, and long-term care visits) existed among the water service areas (Rossdale vs. E.L. Smith), both prior and subsequent to the implementation of changes at the Rossdale plant on December 10, 1997 (relocation of the intake pipe and the introduction of particle counte rs).

A time series approach using generalised additive models (GAM) was used to investigate and quantify the temporal association between gastrointestinal-related health outcomes on a specific day (as assessed by hospital admissions, emergency room visits, physician visits, and long-term care visits) and environmental and water quality parameters (primarily finished water turbidity). Observed values for these parameters from 0 to 40 days prior to a health outcome event were assessed for their impact on the measured health outcomes. In addition to daily readings of finished water turbidity, the utility of particle counts data were assessed, together with raw water indicators including turbidity and coliform counts. Environmental parameters, including daily precipitation and maximum and minimum temperatures, were also investigated.

As anticipated, descriptive analyses and summary statistics of the raw water parameters indicated better quality of water entering the E.L. Smith plant in comparison to the Rossdale plant, both before and after December 10, 1997. These differences are most pronounced in daily total and faecal coliforms. Following December 10, 1997, the differences in raw water parameters between the plants are less prominent. Of primary interest for the time series analysis was the daily finished water turbidity. Prior to December 10, 1997, finished water turbidity values (mean and median) were slightly lower for the E.L. Smith plant. Following December 10, 1997, finished water turbidity values were equivalent between the two plants. Mean and median particle counts from the E.L. Smith plant were marginally lower compared to those of the Rossdale plant. However, particle count data were only available for 1998. Throughout the period of investigation (1993-1998), finished water turbidity levels for both plants were well within federal (Health Canada, 1996) and provincial (Alberta Environment Protection, 1997) drinking water turbidity guidelines. Results of the multivariate logistic model which fit a categorical variable for water source (GLM), suggest a slight decrease in the risk of gastroenteritis among residents supplied with Rossdale water following the changes that were implemented on December 10, 1997. The interpretation of the logistic regression results should, however, be made with caution as the magnitude of the effects were relatively small (odds ratios near one). Although it is plausible that this effect was due to the improvements at the Rossdale plant, other combinations of events could have resulted in a similar relationship. An obvious spatial pattern in gastroenteritis risk with respect to water plant service area was not identified using spatial regression models (GAM). This latter observation further underlies the tenuous nature of this causal hypothesis.

Given the descriptive findings, and the results of the multivariate logistic modelling, time series analysis was restricted to households supplied by the Rossdale water treatment plant prior to December 10, 1997. It was anticipated that if a temporal relationship between daily water quality and gastroenteritis existed, it would be most likely identified in this subset. Nevertheless, no significant lags were identified between daily finished water turbidity and gastroenteritis among Rossdale serviced residents (prior to December 10, 1997) using time series analysis. Lagged water quality parameters from 0 to 40 days prior to health outcome events were assessed for statistical significance and biological relevance using a time series approach. Finished water turbidity provided the best fit to the data. None of the raw water nor environmental parameters were adequate predictors of health in the final model. Similar to the spatial analysis, seasonal and temporal effects were adjusted for use in the final model. Despite limiting the analysis to the water supply posing the greatest potential risk, the time series analysis failed to identify temporal relationships between the finished water turbidity from the Rossdale plant prior to December 10, 1997 and endemic gastroenteritis.

In comparison to the results of the Vancouver time series study^{Footnote 2}, evidence in the present study suggests very little impact, if any, of drinking water on the level of endemic gastroenteritis among Edmonton residents. Despite markedly poorer raw water quality in Edmonton compared to Vancouver during the respective study periods, no significant temporal relationships were identified between water quality parameters and the risk of gastroenteritis. The lack of a temporal relationship supports the opinion that the high quality of Edmonton's drinking water adequately minimises the risk of waterborne endemic gastroenteritis. Furthermore, the lack of a significant relationship between raw water turbidity and faecal coliform counts and endemic gastroenteritis supports the effectiveness of the treatment processes utilised.

Under the multivariate logistic regression model, the suggestion of an overall decrease in risk of endemic gastr oenteritis among Rossdale serviced residents following improvements implemented at the plant on December 10, 1997 is consistent with the observed improvements in raw and finished water quality parameters. The lack of an obvious spatial relationship between endemic gastroenteritis and water service area (using spatial regression) together with the lack of a significant temporal relationship between endemic gastroenteritis and finished water turbidity (using time series analysis), suggest that this apparent decrease in risk is minor, and the relationship tenuous. Nevertheless, the lack of an obvious decrease in the risk of endemic gastroenteritis following the improvements at the Rossdale plant despite marked improvements in raw water quality, together with the lack of an obvious spatial relationship among the service areas, further supports the effectiveness of the operation and treatment processes utilised in Edmonton.

1: Introduction

1.1 Background

Contaminated municipal water supplies have been linked to numerous gastrointestinal outbreaks. Recent events in North America have highlighted the far-reaching impact of waterborne outbreaks on the community (Stirling et al, 2001; Grey Bruce Health Unit, 2000; MacKenzie et al, 1994). While outbreak investigations strive to estimate the health impact within a defined time interval, less is known about the effect of water quality on the endemic rate of illness within a community. Advances in analytical methodologies have enabled researchers to derive quantitative estimates of these impacts. By applying these techniques, several recent studies have identified links between water quality and endemic gastroenteritis (Aramini et al, 2000; Schwartz et al, 2000; Morris et al, 1998; Schwartz et al, 1997).

The current investigation is part of a series of national multi -centric studies being led by Health Canada to investigate risk factors and the burden of illness associated with drinking water. Since this investigation is an extension of the original study that was conducted in Vancouver (Aramini et al, 2000)^{Footnote 3}, a detailed account of the historic background and rationale for this study is not repeated in this report. A comprehensive review of common waterborne pathogens and their associated risk to public health is also provided in the Vancouver report.

The Vancouver study identified strong relationships between drinking water quality (measured by the level of turbidity in the water received at home) and gastroenteritis. At the time of the investigation (1993-1998), filtration procedures were not part of the water purification process in Vancouver. The primary objective of this study was to apply similar methodologies to determine if Edmonton's municipal water supply influenced endemic gastroenteritis in Edmonton, between 1993 to 1998.

1.2 Edmonton's Water Supply

EPCOR Water Services Inc., which is a subsidiary of EPCOR Utilities Inc., supplies water to the city of Edmonton and its surrounding region. An extensive pipe distribution system spanning approximately 3100 kilometres enables access to more than 40 communities. Twelve reservoirs interspersed throughout the city have a total holding capacity of 808 million litres, which is equivalent to a three-day reserve supply. The majority of consumers within the city of Edmonton receive water within two days of leaving the water treatment facilities (per. comm., EPCOR). This distribution system is depicted in Figure 1.

Figure 1: Regional water supply in Edmonton

During the year 2000, EPCOR distributed water to 658,000 Edmonton residents, and 187,000 persons in surrounding communities . With a diverse portfolio, only half of the water consumption comes from residential and multi-family dwellings. Other consumer types include commercial, industrial, wholesale and regional users.

1.2 a) Water treatment plants

Edmonton and many of the surrounding communities in the region are serviced by two water treatment plants located within the city: the Rossdale plant and the E.L. Smith plant. Using Geographic Information Systems application software, (Arc View 3.2, Environment Systems Research Institute, Inc.), water service areas for each plant were digitally mapped, and are shown in Figure 2. Regions where the water source could not be uniquely identified are shaded in grey.

The Rossdale plant consists of two treatment facilities - the first was built in 1947, and the second in 1956. Combined, they have the capacity to treat 275 million litres of water per day. The E.L. Smith plant was subsequently built in 1976 to meet the demands of an expanding population. With a treatment capacity of 240 million litres per day, this latter plant was built away from the downtown core, upstream of the Rossdale facilities. Both plants draw in water from the North Saskatchewan River, which is within the North Saskatchewan River Basin.

Figure 2: Water service areas for the Rossdale and E.L. Smith plants, Edmonton.

1.2 b) Watershed characteristics

The North Saskatchewan River originates from the foot of the Columbia Icefields in the Rocky Mountains south-west of Edmonton, and flows directly through Edmonton. En route, the river is fed by numerous rivers and creeks.

In contrast to the watersheds that serve as Vancouver's water source, the North Saskatchewan River Basin is not protected from human and agricultural influences. Covering 28,000 square kilometres, this watershed contains mountains, forests, several communities and farm lands.

The upper half of the watershed is mountainous and forested, with little human activity. In contrast, the lower watershed is flatter, more inhabited, a nd supports agricultural activities. The watershed contains a thriving livestock industry, with approximately 290,000 cattle. It also supports an extensive wildlife population that includes deer, elk, beaver, and moose. Approximately 76,000 people reside in the upstream basin in a number of small towns and hamlets. Four continuous discharge sewage treatment plants and 16 sewage treatment lagoons discharge waste to the North Saskatchewan River upstream of the city of Edmonton. The continuous discharge sewage treatment plants utilise secondary treatment to process waste materials. There is limited industry in the upstream basin apart from forestry, and oil and gas extraction. This latter activity includes many pipelines that collect at refineries in Edmonton (per. comm., EPCOR).

The soils in the watershed are primarily glacial clays, and result in very high turbidities during high river flow situations. The mean yearly flow in the North Saskatchewan River is just over 200 m³ per second, but peak flows can exceed 800 m ³ per second. The river has stable flows during the winter season under ice cover, but is susceptible to rapidly changing flows during the spring season when snowmelt and spring run-off situations occur. Heavy rainfalls during the summer also increase the river flow rate (per. comm., EPCOR).

1.2 c) Water purification procedures

The Rossdale and E.L. Smith plants utilise a multi -step approach to treat water drawn-in from the North Saskatchewan River. The process begins with intake pipes that draw-in water from the deepest part of the river. Screens at the end of these pipes prevent debris and fish from being drawn into the system. Coagulant chemicals are mixed with the raw water to achieve flocculation. During sedimentation, the aggregate particles settle to the bottom, and the clear water above is passed on to the next stage. Lime is then added, and the precipitated hardness is removed in a further sedmintation process. Carbon dioxide is added to reduce the pH after softening. Chlorine is added to the clarified water for disinfection, then ammonia is added to convert the remaining free chlorine to chloramine. Then the water is filtered through an almost 1-metre deep bed of fine anthracite coal and sand (E.L. Smith) or monomedia sand (Rossdale). Activated carbon is added during runoff periods to reduce the presence of organ ic material and to improve the taste and odour of the water. At the end of this process, the water is pumped out into the city to the customer, either directly, or through additional storage reservoirs located across the city.

1.3 Study Objectives

The primary objective of this investigation was to determine if endemic gastroenteritis within the city of Edmonton was influenced by the municipal water supply between January 1993 to December 1998. Using the level of turbidity in finished water as the primary measure of drinking water quality, a significant association was identified with endemic gastroenteritis in Vancouver. At the time of that investigation, filtration was not part of the water purification process in Vancouver. Other investigators have also found similar relationships using similar methodologies (Morris et al, 1998; Schwartz et al, 1997).

Geographic differences between the two Edmonton water treatment plants with respect to their point of access on the North Saskatchewan River warrant comparison in the levels of illness among their respectively serviced populations. The E.L. Smith plant is located upstream from the Rossdale plant, which in contrast, is situated in the downtown core (Figure 2). While the E.L. Smith plant is unaffected by city storm sewer runoff, the Rossdale plants are susceptible to runoff from 80 stormsewers upstream. In December of 1997, the Rossdale intake pipe was moved away from the shoreline towards the centre of the river in an attempt to minimise the impacts of stormsewer run-off on the quality of raw water. Concurrent with this event was the introduction of the use of particle counters at the plant, which allowed much better optimisation of the filters to minimise the effect of filter-to-waste periods, filter spikes, and particle increases at the end of filter runs.

Another objective of this investigation was to compare and contrast the utility of various water quality indicators for identifying potential relationships with gastroenteritis. In addition to finished water turbidity, particle counts in finished water were examined. Raw water indicators, such as turbidity and coliform counts, were also considered.

Finally, the effect of environmental parameters including temperature and precipitation were assessed. Heavy rainfall has been implicated as an inciting factor in several waterborne gastroenteritis outbreaks (Curriero et al, 2001; Grey Bruce Health Unit, 2000). In Edmonton, run - off potentially carries animal waste and irrigation chemicals into the North Saskatchewan River, which may then compromise the water purification process.

Achieving the objectives listed for this investigation allowed for an assessment of the nature of the relationship, if any, between Edmonton's municipal water supply and endemic gastroenteritis.

2: Data Sources

To assess the relationship between water quality and gastroenteritis, a comprehensive database was created from data that were provided by various agencies. Extensive retrospective water quality and environmental data were collected to assess their impact on community health. Data stemming from hospitalisation use and physician-billing claims were used to create appropriate health outcome groups. A description of these data are provided in the subsequent sections.

2.1 Water Quality Parameters

Water quality data from January 1, 1993 to June 30, 1999 were provided by EPCOR Water Services, Inc. A complete list of the types of data provided are listed in Table 1. With the exception of the particle count data, dates of availability for all the measured indicators are identical for both plants. However, data that were available for analysis differed for some parameters, depending on when automated data retrieval systems were implemented. Finished water turbidity was provided in the form of 5-minute readings for all on-line filters (nine in Rossdale, twelve in E.L. Smith). To facilitate analysis using these extensive data (288 readings per filter per day), these data were summarised into daily mean, median, and maximum values. Subsequent statistical tests were used to determine the optimal parameterisation for these data.

Compared to the Vancouver study, additional variables available for analysis in this investigation included finished water particle count data. The use of particle counts is a new approach to measuring water quality. Only recently implemented, particle counts are perceived to be a more accurate proxy measure of finished water quality than turbidity. Many of the raw water parameters listed in Table 1 were also unique to this investigation, and were not examined in the Vancouver study.

2.2 Environmental Parameters

Environmental parameters from January 1, 1993 to December 31, 1998 were provided by Environment Canada. Daily maximum and minimum temperatures in degrees Celsius, and daily precipitation measured in millimetres were used.

2.3 Socio-Economic Data

Average household income in 1995 was derived from the 1996 Census Data (Statistics Canada).

2.4 Health Outcome Data

Emergency visit data and hospitalisation admission data have been used previously to assess endemic gastroenteritis (Morris et al, 1998; Schawrtz et al, 1997). Many similarities exist between the health outcome data utilised in the Vancouver study and the current study. In the Vancouver study, a total of three different sources were used: hospitalisation admission data from the Canadian Institute for Health Information (CIHI); physician visit data from the Medical Services Plan in B.C.; and emergency room visit data from the British Columbia Childrens' Hospital. A comparison was then made among these datasets to determine consistency of results.

In the present study, two primary data sources were used: hospital admission data from CIHI; and physician-billing data from the Alberta Health Care Insurance Payment Plan. However, in contrast to the Vancouver study, it was possible to further differentiate the latter data source into three distinct categories: namely, billing that resulted from emergency room visits, physician-office visits, and long-long-term care facilities. Recognising that the various data-capture-sources would be more representative of certain populations, interpretations of the results were based on the findings from each of the four different data sources. The selection criteria for gastrointestinal cases and respiratory controls were based on the literature (Morris et al, 1998; Schwartz et al, 1987) and on the advice of general practitioners and a gastroenterologist.

2.4 a) Canadian Institute for Health Information

The Canadian Institute for Health Information receives data from approximately 85% of hospitals across Canada. Each record within this comprehensive database contains the specifics of each case admission, including the patient's date-of-birth, gender, postal code of residence, scrambled personal identifier, admission date, reason for admission, surgery type, diagnosis fields, and institute type. Only valid and complete records from January 1, 1993 to December 31, 1998 were further considered for the case and control selection process. These criteria selected patients that had a valid date-of-birth, admission date, and gender information. A postal code was also required to verify residence within Edmonton. Elective surgeries and elective admissions to acute care facilities were excluded from further investigation.

Gastrointestinal cases and respiratory controls were selected using a similar approach to that described in the Vancouver study. The 9^th revision of the International Classification of Diseases (ICD-9), published by the World Health Organisation, is designed for the classification of morbidity and mortality information for statistical purposes, and for the indexing of hospital records by disease and operations, for data storage and retrieval. Appropriate 3-digit and 4-digit ICD -9 codes were selected to designate admitted patients as gastroenteritis cases or respiratory controls.

Criteria for selecting gastroenteritis cases are outlined in Table 2. A gastroenteritis case was defined as any individual with a primary diagnosis^{Footnote 3} of one of the ICD-9 codes marked in column A, or with a primary diagnosis of one of the ICD-9 codes marked in column B and a secondary diagnosis of one of the ICD-9 codes marked in column A.

* Although labelled as "noninfectious", may in fact be an undiagnosed episode of infectious gastroenteritis.

Criteria for selecting respiratory controls are outlined in Table 3. A respiratory control was defined as any individual with a primary diagnosis^{Footnote 4} of one of the ICD-9 codes marked in column C and without an additional diagnosis of any of the ICD-9 codes marked in column D.

To prevent over-estimating the rate of illness and thus distorting the true relationship between gastroenteritis and water quality for both cases and controls, admissions with a similar diagnosis which occurred within 60 days of an initial visit were considered part of the same disease episode, and thus were not included in the analysis.

2.4 b) Alberta Health Care Insurance Payment Plan

The Alberta Health Care Insurance Payment Plan (AHCIPP) maintains a comprehensive database on all billing claims that originate from a variety of health care provider types. As with the CIHI database, each record within the AHCIPP database contains patient -specific information, including the age at time of treatment, gender, postal code of residence, scrambled personal identifier, diagnosis fields, and coded aggregate function centre type. This last field was used to differentiate between claims originating from various sources. Billing resulting from emergency room visits (EMRG), physician-office visits (PHYS), and long-term care centres ^{Footnote 5} (LTC) were analysed separately.

Within each of these three data sources, records from January 1, 1993 to December 31, 1998 with a valid age, gender, and postal code within Edmonton were selected. Since only 3-digit ICD-9 codes were available in the AHCIPP database, the criteria used to select cases and controls varied slightly from the CIHI database. For these data, a gastroenteritis case was defined as an individual who was diagnosed with an intestinal infectious disease (ICD-9 codes: 001 -009, refer to Tables 2 and 3) or other "noninfectious" gastroenteritis and colitis (558). A respiratory control was defined as an individual who was given a primary diagnosis of an acute respiratory infection (460-466), pneumonia (480-483, 485-486), or asthma (493), who did not also suffer from an intestinal infectious disease (001 to 009), disease of the digestive system (520 to 579), nor any of the other gastrointestinal-related illness described in Table 3 (027, 041, 070, 276, 487, 787, 789). All repeat visits within a 60-day interval were excluded.

2.5 Assigning a Primary Water Source

As in the Vancouver study, water received at home was chosen as the primary exposure variable of interest. Therefore, to identify potential associations between water quality and gastroenteritis, the same method of assigning cases and controls to their primary water source was used. The 6-digit postal codes provided in the health outcome data were linked to a digitised postal code file (Enhanced Postal Code File, Desktop Mapping Technologies, Inc.). This procedure designates a centroid point, along with geographical co-ordinates, for each postal code region. Applying GIS overlay techniques on the digitally-mapped water service areas shown in Figure 2, Edmonton residents were linked to a primary water source based on their location of residence.

3: Investigative Methods

3.1 Descriptive Analyses

Descriptive analyses were carried out initially to identify general trends and associations in both the health outcome and exposure data. Temporal trends in the health outcome data were investigated and a profile of each data source was created to determine the proportion of cases within each age group and for each water service area. The distributions of primary ICD-9 codes captured in the case definition were also determined. Preliminary statistics were determined for the water quality and environmental parameters. Comparisons were made between the two water treatment plants to determine if general differences in water quality existed.

3.2 Multivariate Logistic Regression

3.2 a) Methodology overview

The primary objective of this portion of the analysis was to compare the risk of gastroenteritis for residents within each of the water service areas in Edmonton. Results from the descriptive analyses demonstrated that some differences in water quality existed between the two water treatment plants. All individuals with a valid Edmonton postal code who could be linked to the Rossdale or E.L. Smith water service areas, as well as individuals residing in the mixed zone, were included in this portion of the analysis. Infants (< 2 years) were excluded from this investigation since it was hypothesised that their intake of tap water in the home was likely minimised due to neonatal feeding practices, and thus not compatible with the remaining study population with respect to exposure patterns.

A case-control study design was used for this analysis. The modeled binomial response was case status: gastroenteritis case or respiratory control. Respiratory controls were used to help control for the potential influences of environmental factors on health outcome. The effect of modeled variables on health outcome was given by odds ratios. Each health outcome data set for the multivariate logistic regression portion of the analysis contained binary health outcomes, and each data set was modeled independently. The components of these analyses are summarised in Table 4.

Generalised Linear Models (GLM)

A generalised linear model (GLM) with a binomial distribution and logit link, was the underlying statistical model used to analyse these data (McCullagh and Nelder, 1983). The resulting multivariate logistic modeling process was carried out using the GENMOD procedure in SAS, version 8.0. Separate models were evaluated for each of the four health outcome data sources.

Due to the enormity of the data originating from the physician-office billing data source, a random subset of the data was taken to facilitate analysis. A 1:3 ratio of cases to controls was formed by randomly selecting 20% of original cases. This randomised selection process was carried out using a random number function (RANUNI) in SAS, version 8.0.

Spatial Generalised Additive Models (GAM)

Spatial regression models were subsequently carried out to provide a visual representation of the geographic differences in risk of gastroenteritis within the city of Edmonton. A binomial generalised additive model (GAM) was utilised to analyse the same health outcome data (Hastie and Tibshirani, 1990 ). GAM models are an extension of GLM models, in that a flexible additive non-linear relationship may be modeled between the independent predictor (risk factor) and the response. The response is modeled as a sum of smooth functions in the predictors, where these functions are estimated using smoothers (a curve is fit to the data points locally, so that at any point, the curve depends only on the observations at that point, and some specified neighbouring points - this estimate of the response is referred to as a smooth, and procedures for producing such fits are called smoothers). Model fit is improved through the flexible parameterisation of variables in GAM models, and parameter estimates may be derived with greater accuracy.

GAM models were used for this spatial component of the analysis to accommodate for the inclusion of geographical co-ordinates in the modeling process. The location of cases and controls was determined by using their corresponding 6-digit residential postal code. Longitude and latitude coordinates assigned to the centroid point within each postal code area were incorporated into the loess smoothing function within the GAM model (Cleveland and Devlin, 1988). This method of fitting the location effect directly into the statistical model as a non-linear parameter has also been applied in other studies where the objective has been to detect spatial clusters (Preisler et al, 1997; Brillinger, 1994; Chambers and Hastie, 1992; Cook and Pocock, 1983). The gam function in S-PLUS 2000, release 2 (MathSoft, Inc.) was used to carry out this spatial component of the multivariate logistic regression analysis, as smoothing functions are easily accommodated in this analytical software package.

Similar to the GLM analysis, a random subset consisting of 20% of the original cases captured in the physician-office data source wa s created. Due to the computational intensity associated with fitting a loess smoother to the geographical co-ordinates, controls were selected to satisfy a 1:1 ratio of cases to control s using the same procedure described in the previous section.

3.2 b) Variables analysed

A list of variables investigated in both of the multivariate logistic regression processes is provided in Table 5.

Generalised Linear Models (GLM)

To estimate the risk of gastroenteritis for residents within each of the water service areas, a categorical variable was created to reflect the primary water source in the generalised linear modeling. Levels of this variable, SOURCE were:

• Mixed zone, before December 10, 1997
• Mixed zone, after December 10, 1997
• Rossdale water service area, before December 10, 1997
• Rossdale water service area, after December 10, 1997
• E.L. Smith water service area, before December 10, 1997
• E.L. Smith water service area, after December 10, 1997

The risks associated with each of the water service areas were distinguished between when the Rossdale intake pipe was moved, which coincided with the introduction of particle counters in this plant (December 10, 1997). This distinction was made since raw water quality from the Rossdale plant changed following these events. Statistical comparisons in risk among SOURCE levels were carried out using contrasts. Since residential location is a reflection of the quality of water received at home, none of the water quality nor environmental parameters associated with the time series analysis were included in these analyses.

Since SAS does not easily accommodate loess smoothing functions in regression models, a separate spline was created for each 2-month period, so that different slopes (risk estimates) could be determined for each time interval, and thus control for seasonal trends in viral gastroenteritis. This technique is used when it can be expected that the effects vary across different levels of a variable. Comparisons between models using splines in SAS and the loess smoothing function in S-PLUS to adjust for seasonal trends, produced similar results (data not provided).

Household income was used as an indicator of socio-economic status, which was hypothesised to influence the risk of illness. The average household income reported in 1995 for each enumeration area (Statistics Canada, 1996) was linked to each postal code, and subsequently to each case and control. This variable was included in all models.

Age was fit as a categorical variable in these models to adjust for potential differences in risk between age groups. Thus, an averaged estimate of risk for residents of different water service areas was obtained. Categories for age were identical to the levels of stratification used in the time series analysis: 2 to 18 years, greater than 18 to 65 years, and greater than 65 years.

Spatial Generalised Additive Models (GAM)

A loess smoother was fit to the longitude and latitude co-ordinates of the centroid points for each postal code. This independent (risk factor) variable was used to derive estimates of the risk of gastroenteritis for various locations within Edmonton. A categorical term, DEC1097 was also modeled to represent the risk prior to and after the relocation of the Rossdale intake pipe. The inclusion of these variables in the spatial model replaced the categorical term SOURCE used in the GLM modeling process.

Since S-PLUS easily accommodates smoothing functions, seasonal fluctuations in gastroenteritis were also modeled using the loess smoothing function. Average household income and age were also included in these spatial models.

3.2 c) Criteria for model-selection

For the GLM models, step-wise model selection based on the variables listed in Table 5 was conducted for each health outcome data set. The effect of each variable was evaluated by using the likelihood ratio test (Fahrmeir and Tutz, 1994). Criteria for inclusion in the final model was set at the 5% level of statistical significance. Model-fit was examined using Akaike's Information Criterion (Sakamoto et al, 1986). Variables in the spatial model (GAM) were chosen to reflect the variables identified in the final multivariate logistic regression model.

3.3 Time Series Analysis

3.3 a) Methodology overview

Similar to the Vancouver study, generalised additive models (GAM) were used to investigate temporal relationships between water quality and gastroenteritis. This approach has also been applied in other time series studies (Morris et al, 1998; Schwartz et al, 1997). The application of this methodology for time series investigations has been described in detail in the Vancouver study (Aramini et al, 2000). S-PLUS 2000, release 2 (MathSoft, Inc.) was used to carry out this portion of the analysis.

A loess smoothing function (Cleveland and Devlin, 1988) was used to describe the potentially non-linear relationship between the risk of gastroenteritis and turbidity, along with other water quality and environmental parameters described in sections 2.1 and 2.2. In addition, the smoothing function was also applied to fit a seasonal (long-term) parameter in an attempt to control for seasonal trends in gastroenteritis. A 220-day cycle, which was employed and described in the Vancouver study, was used to represent this seasonal trend.

To determ ine the influence of water quality from each plant on the corresponding serviced population, individuals from each of the four data sources were analysed separately, based on the primary water source. Individuals whose primary water source could not be uniquely identified (mixed zone - shaded grey area in Figure 2) were excluded from this analysis. By matching the date of health event to the date of the recorded water quality parameter, daily water quality values observed for each plant were then linked to the appropriate individuals.

As explained in the Vancouver study, both binomial (case-control) and poisson GAM models (Hastie and Tibshirani, 1990) were fit to the data. In the binomial models, the modeled outcome was case status: gastroenteritis case or respiratory control. The outcome for the poisson models was daily count of gastroenteritis cases. The effect of modeled variables on gastroenteritis was given by odds ratio and relative risks for the binomial and poisson models, respectively.

Since susceptibility to illness varies across different age groups, separate analyses were conducted for each of the following age groups: 2 to 18 years, greater than 18 to 65 years, and greater than 65 years. These age groups are identical to those used in the Vancouver study, with the exception that infants (< 2 years) were excluded from this investigation. Therefore, each health outcome data set used in this analysis was specific to each age group, water service area, and data source (refer to Table 6).

In the binomial analysis, a randomised subset of cases and controls were selected from the physician-office data, using a 1:3 ratio of cases to controls. This was the same subset described in the GLM modeling approach. Physician -office data were not reduced for the poisson analysis since gastrointestinal outcomes were collapsed into daily counts, and the resulting smaller data set was therefore easily processed by the software application.

3.3 b) Variables analysed

A list of the independent (risk factor) variables assessed in the time series analysis is provided in Table 7. Time series variables that were lagged are denoted with subscript i. A description of some of the key variables is provided below.

Separate analyses were conducted for each time series health outcome data set. For the water quality and environmental parameters, values from 0 to 40 days prior to the day of the health outcome event were modeled. This range of lagged values was selected to reflect multiples of the incubation periods commonly reported for prevalent waterborne pathogens. Therefore, for each time series health outcome data set and combination of variables analysed, 41 models were evaluated (one for each lag day).

Finished water turbidity (TB i) was the primary variable of interest for this analysis. However, the effects of other water quality parameters were also evaluated. Five-minute readings that were provided for some of the water quality parameters were summarised as daily observed values, including the mean, median, and maximum reading for that day. Statistical comparisons among models with the various parameterisations of these data, indicated that the daily mean provided the best fit to the data. Therefore, all 5-minute readings were summarised as daily means. Rare occurrences of missing values, which were occasionally the result of off-line filters, were replaced with the mean of adjacent observed values.

Environmental parameters were also examined. In addition to the effect of weather on raw water turbidity, precipitation was also hypothesised to influence the level of exposure at home. It was hypothesised that people are more likely to stay indoors as a result of inclement weather, facilitating person-to-person spread of infectious gastroenteritis. Therefore, precipitation and temperature lagged up to 40 days prior to the day of the health outcome event were similarly investigated.

Temporal confounders were included in the models to adjust for temporal variations, including the seasonal parameter. Day-of-week effects were significant in the Vancouver study and were similarly included in this study. Statutory holidays were also considered to determine if changes in the accessibility of medical services, as well as holiday behaviours, had an impact on gastroenteritis. To that end, separate categories were created for the following holidays and holiday weekends:

• Christmas and New Year's,
• Easter,
• Victoria Day,
• Canada Day,
• August Long Weekend,
• Labour Day,
• Thanksgiving, and
• Remembrance Day.

In addition, the week following each holiday event was also assessed (with the exception of Remembrance Day).

Finally, an autoregressive term was included in all of the models to adjust for the correlation arising from daily observations made of each health outcome event. As discussed in the Vancouver study, even in the absence of a relationship with water quality, the number of hospital admissions, physician visits, emergency room visits, and treatments provided in a long-term care facility on any given day can be expected to be related to the number the day before. Several possible reasons for this include persons sharing a common food or water source, and variable pathogen incubation times. The autoregressive term was expressed as a ratio of the number of cases to the number of controls, and was fit as a linear term in the model.

3.3 c) Criteria for model-selection

Based on the descriptive analyses and on the multivariate logistic regression results, quantitative time series analysis was conducted for the water service area that represented the highest potential risk of endemic waterborne gastroenteritis (Rossdale, prior to December 10, 1997). Similar to the multivariate logistic regression, the effect of each variable listed in Table 7 was determined by conducting deviance comparisons using the likelihood-ratio test, and comparing AIC values and parameter estimates. Significant lags were identified by comparing the change in deviance in models with and without the aggedl variable, thereby evaluating the overall effect of that lagged variable. However, multiple statistical comparisons increase the probability of erroneously detecting a statistically significant association, commonly referred to as Type I error (Steel et al, 1997). Therefore, when evaluating the effect of parameters up to 40 days prior to the health outcome event, only lags that were significant over two to three consecutive days were further assessed for consistency among the different age groups and health outcome data sources.

4: Results

4.1 Descriptive results

4.1 a) Water quality parameters

Summary statistics for the finished and raw water quality parameters of primary interest are provided in Table 8. Because of the changes that were implemented at the Rossdale plant on December 10, 1997 (movement of the intake pipe and the introduction of particle counters), results for this plant are further differentiated as prior to and following this event.

In general, raw water parameters indicate better quality of water entering the E.L. Smith plant in comparison to the Rossdale plant, both before and after December 10, 1997 (Table 8). The differences are most pronounced in total and faecal coliforms. Following December 10, 1997, the differences in raw water parameters between the plants are less prominent.

Of primary interest in the time series analysis was the daily finished water turbidity (TB_i). Prior to December 10, 1997, finished water turbidity values (mean and median) were slightly lower in the E.L. Smith plant (Table 8). Following December 10, 1997, summary statistics appeared to be equivalent. Limited data on particle counts in finished water were provided. Mean and median particle counts from the E.L. Smith plant are marginally lower. However, these data must be interpreted with caution, since they were only available in 1998.

Temporal trends for finished water parameters are further described in Figures 3a and 3b. These data have been presented using the loess smoother and a span appropriate to the data to reduce the 'noise' resulting from daily fluctuations in the data. These figures demonstrate that finished water turbidity values for the Rossdale experienced greater fluctuations (within 0.2 NTU) than those originating from the E.L. Smith plant. There is some suggestion, however, that following December 10, 1997, values became more stabilised.

For both plants, seasonal peaks in raw water turbidity during the summer months were

apparent, and were likely due to an increase in run-off from rainfall. Faecal coliforms also appeared to favour these temperate months, as their numbers increased during this time.

• ¹ Rossdale intake pipe relocated on December 10, 1997, at which time, particle counters were also introduced. Therefore, presented results are based on data up to and including December 10, 1997
• ² Presented results are based on data from December 11, 1997 and beyond
• ³ Data available March 13 - December 31, 1998
• ⁴ Data available January 1 - December 31, 1998
• ⁵ Data available January 1, 1994- December 31, 1998

Figure 3a: Finished Water Quality Descriptive Results (Jan. 1, 1993 to Dec. 31, 1998) Rossdale plant

(i) Daily mean finished water turbidity(NTU) (span 25 days) ^{Footnote 6}

(ii) Daily mean finished water particle counts (span 25 days)

(iii) Distribution of observed daily mean finished water turbidity values

(iv) Distribution of observed daily mean finished water particle counts

Figure 3b: Finished Water Quality Descriptive Results (Jan. 1, 1993 to Dec. 31, 1998) E.L. Smith plant

(i) Daily mean finished water turbidity (NTU) (span 25 days)

(ii) Daily mean finished water particle counts (span 25 days)

(iii) Distribution of observed daily mean finished water turbidity values

(iv) Distribution of observed daily mean finished water particle counts

Figure 4a: Raw Water Quality Descriptive Results (Jan. 1, 1993 to Dec. 31, 1998) Rossdale plant

(i) Daily mean raw water turbidity (span 60 days)^{Footnote 7}

(ii) Daily raw water faecal coliform counts (span 60 days)

(iii) Density plot of observed daily mean raw water turbidity values

(iv) Density plot of observed daily mean raw water faecal counts

Figure 4b: Raw Water Quality Descriptive Results (Jan. 1, 1993 to Dec. 31, 1998) E.L. Smith plant

(i) Daily mean raw water turbidity (span 60 days)

(ii) Daily raw water faecal coliform counts (span 60 days)

(iii) Density plot of observed daily mean raw water turbidity values

(iv) Density plot of observed daily mean raw water faecal counts

4.1 b) Environmental parameters

Summary statistics are provided for the environmental parameters in Table 9. Temporal trends for daily extreme temperatures and precipitation are provided in Figures 5 and 6, respecti vely. Strong seasonal patterns were apparent for these data, with peaks being observed during the spring and summer season.

Figures 7 and 8 visually describe the relationship between raw water turbidity and environmental factors. In general, increases in temperature and precipitation were accompanied by an increase in raw water turbidity.

Figure 5: Maximum and minimum daily temperatures (°C) (Jan. 1, 1993 to Dec. 31, 1998) (span=25 days) (^{Footnote 6})

Figure 6: Daily precipitation (mm) (Jan. 1, 1993 to Dec. 31, 1998) (span=25 days)

Figure 7: Comparing daily maximum temperature and daily mean raw water turbidity

Figure 8: Comparing daily precipitation and daily mean raw water turbidity

4.1 c) Socio-economic data

Average annual household incomes reported in 1995 are shown in Figure 9.

Figure 9: Distribution of average annual household incomes per postal code in Edmonton, 1995

4.1 d) Health outcome data

To ensure that all individuals could be linked to water quality values up to 40 days prior their date of illness, February 10^th 1993 was the first service date considered for analysis, and December 31, 1998 was the last. Detailed profiles of each data source are provided in Figures 10a to 10d. The majority of cases were captured by utilizing the physician-office data from the AHCIPP database. Although long-term care c entr es a lso pr ovid e regular treatment to individuals suffering from persistent mental illness and physical disabilities, the majority of patients (70%) captured in this database were greater than 65 years of age. In contrast, only 10% of selected cases in the emergency room and physician-visit billing data sources were greater than 65 years.

The distribution of cases by the most responsible ICD-9 code is also provided. For comparative purposes, the same ICD-9 codes are presented for each data source, and represent the majority of what was observed in the respective databases.

The seasonal distribution of cases on a bi-weekly basis was also apparent, with a higher incidence of cases observed during the spring and early summer months. More cases were also admitted to the hospital at the beginning rather than during the work week. In contrast, more cases sought emergency room services during the weekend.

Figure 10a: Data Profile (February 10, 1993 to December 31, 1998) : Hospital Admissions (CIHI data source)

(ii) Number of cases per 14-day interval (seasonal distribution)

(iii) Proportion of cases by most responsible ICD-9 code (iv) Proportion of cases by day of week

(iv) Proportion of cases by day of week

Figure 10b: Data Profile (February 10, 1993 to December 31, 1998) : Emergency room-related visits (EMRG data source)

(ii) Number of cases per 14-day interval (seasonal distribution)

(iii) Proportion of cases by most responsible ICD-9 code (iv) Proportion of cases by day of week

(iv) Proportion of cases by day of week

Figure 10c: Data Profile (February 10, 1993 to December 31, 1998) : Physician-office visits (PHYS data source)

(ii) Number of cases per 14-day interval (seasonal distribution)

(iii) Proportion of cases by most responsible ICD-9 code (iv) Proportion of cases by day of week

(iv) Proportion of cases by day of week

Figure 10d: Data Profile (February 10, 1993 to December 31, 1998) : Long-term care vists (LTC data source)

(i) Distribution of cases by age group and water service area

(ii) Number of cases per 14-day interval (seasonal distribution)

(iii) Proportion of cases by most responsible ICD-9 code (iv) Proportion of cases by day of week

(iv) Proportion of cases by day of week

4.2 Statistical Analyses

4.2 a) Multivariate logistic regression

As described in section 3.2, multivariate logistic regression analysis was conducted to determine if the risk of gastroenteritis varied for residents of different water service areas. The final model derived using the generalised linear model (GLM) was :

(refer to Table 5 for a complete description of variable names)

As time was both a component of the seasonal parameter and of the categorical source variable (SOURCE), the impact of water source was assessed by calculating the relative change after December 10, 1997 in the risk of gastroenteritis between Rossdale and E.L. Smith serviced residents. Presented in Table 10 are the results of this comparison, carried out using appropriate contrasts within the final logistic regression model for each health outcome data set. The odds ratios equate to the relative change in risk among service areas after Dec 10 '97. For example, the "risk" in hospitalisation for gastroenteritis among Rossdale serviced residents in comparison to E.L. Smith serviced residents fell by a magnitude of 0.18 times after December 10, 1997 (Table 10).

The results of Table 10 suggest a small lowering in risk of gastroenteritis among Rossdale residents (in comparison to E.L. Smith serviced residents) after Dec 10 '97. Statistically significant (P < 0.10) relative "risk" reductions were noted among hospitalisations and emergency room visits.

Based on the model above, the following spatial model was determined using generalised a dditive modeling (GAM):

(refer to Table 5 for a complete description of variable names)

Figure 11: shows the distribution of centroid points associated with all Edmonton postal codes and their assigned water source.

Figure 11: Spatial distribution of postal codes within Edmonton and corresponding water service areas.

Odds ratios corresponding to each longitude and latitude combination for each captured postal code were determined, both before and after December 10, 1997. For illustrative purposes, an odds ratio of 1.2 for each spatial location represents a 20% increase in the likelihood of illness associated with that centroid point, compared to the predicted probability associated with the mean location effect, after adjusting for other parameters in the model. Results for each data source are presented in Figures 12a to 12d. Spatially, there does not appear to be an obvious association between the risk of gastroenteritis and water supply area after controlling for other important risk factors.

Figure 12a: Spatial distribution of disease risks within the city of Edmonton between 1993-1998: CIHI hospitalisation data source, excluding infants (less than 2 year-olds)

(i) Prior to December 10, 1997

(ii) After December 10, 1997

Figure 12b: Spatial distribution of disease risks within the city of Edmonton between 1993-1998: Emergency room visit data source, excluding infants (less than 2 year-olds)

(i) Prior to December 10, 1997

(ii) After December 10, 1997

Figure 12c: Spatial distribution of disease risks within the city of Edmonton between 1993-1998: Physician-office billing data source (randomised subset), excluding infants (less than 2 year-olds)

(ii) Prior to December 10, 1997

(ii) After December 10, 1997

Figure 12d: Spatial distribution of disease risks within the city of Edmonton between 1993-1998: Long-term care data source, excluding infants (less than 2 year-olds)

(i) Prior to December 10, 1997

(ii) After December 10, 1997

4.2 b) Time series analysis

Based on the findings of the descriptive and multivariate logistic regression analyses, time series analysis was conducted only on individuals that resided within the Rossdale water service area. In addition, only water quality data prior to December 10, 1997 were analysed. Individuals supplied with Rossdale water prior to this date were more likely to be at the greatest potential risk of endemic waterborne gastroenteritis.

Using these data, final models for all time series health outcome data sets were derived using binomial and poisson analyses.

Poisson model:

Binomial model:

(refer to Table 7 for a complete description of variable names)

From these final models, no significant finished water turbidity lags were identified using the criteria outlined in section 3.3c. Similarly, none of the other measured water quality and environmental parameters, including raw water quality parameters and precipitation, were significantly associated with endemic gastroenteritis.

Although no relationship with finished water turbidity was identified in this time series analysis, 3-D surface plots were constructed to facilitate a comparison to what was observed in the Vancouver study. Figures 13a to 13d show the estimated relative rates (Poisson model) and odds ratios (Binomial model) in the final model for lagged daily mean finished water turbidity values. Interpretation for relative rates are similar to that described for odds ratios in the previous section. That is, a relative rate of 1.2 represents a 20% increase in the likelihood of gastroenteritis associated with that level of turbidity for that particular lag, compared to the predicted probability associated with the mean turbidity effect for that particular lag, after adjusting for other parameters in the model. Estimates for the binomial model using the physician-visit billing data are based on a randomised subset of the o riginal data (ref er to Sect ion 3.3a), since the software application that was used to analyse these data (S-PLUS 2000, release 2; Mathsoft, Inc.) could not manipulate the original extensive database. Only results from the greater than 65 year -old age group were presented for the long-term care data so urce since model estimates were highly unstable as a result of the limited data available for analysis in the other age groups.

Colour schemes in these figures are identical to those used in the Vancouver study. Results of point -wise comparisons for each lag and turbidity value were differentiated by the level of statistical significance. Red indicates that a point-wise relative rate or odds ratio was significantly greater than 1.0 (i.e. significantly greater than the average turbidity effect) at the 5% level of significance; yellow indicates significance at the 10% level. The turbidity scale was truncated to exclude days with turbidity values that were greater than the 95 ^th percentile and therefore observed less than 5% of the time. Model estimates for these values were extremely unstable due to the limited data available for analysis, and the correspondingly large standard errors did not provide reliable statistical results.

Although sporadic statistical associations were observed in these 3-D plots, the primary statistical criterion used for identifying significant lags was determined using the likelihood ratio test (Fahrmeir and Tutz, 1994). And, as already stated, no significant lags were identified using this criterion. Furthermore, no consistent patterns could be appreciated among the 3-D plots.

Figure 13a: 3-D time series surface plots of disease risk within the Rossdale water service area between Fe bruary 10, 1993 to December 10, 1997: CIHI data source

Figure 13b: 3-D times series surface plots of disease risk within the Rossdale water service area between February 10, 1003 to December 10, 1997: Emergency room billing data source.

Figure 13c: 3-D time series surface plots of disease risk within the Rossdale water service area between February 10, 1993 to December 10, 1997: Physician-office billing data source

Figure 13d: 3-D time series surface plots of disease risk within the Rossdale water service area between February 10, 1993 to December 10, 1997: Long-term care billing data source.

Poisson    Binomial

5: Discussion And Conclusions

5.1 Introduction

Waterborne outbreaks of gastrointestinal illness are well documented in North America. In Milwaukee, U.S.A. in 1993, an estimated 403,000 people were infected with Cryptosporidium through a contaminated water supply (MacKenzie et al., 1994). In Canada, between 1986 -1993, approximately 150 suspected infectious drinking water outbreaks were reported to Health Canada (Todd, 1991, 1994, 1996, 1997, 1998). In Walkerton, Ontario in May of 2000, seven deaths and approximately 2000 illnesses resulted from the contamination of municipal well water with E. coli O157:H7 and Campylobacter spp. (Grey Bruce Health Unit, 2000). In The Battlefords, Saskatchewan, approximately 6000 persons developed gastroenteritis associated with waterborne cryptosporidiosis in April of 2001 (Stirling et al, 2001). These past outbreaks, together with recent studies that suggest that drinking water may be a substantial contributor to endemic (non-outbreak related) gastroenteritis (Payment et al., 1997; Schwartz et al., 1997; Isaac-Renton et al., 1999), demonstrate the vulnerability of many North American cities to waterborne diseases and have fuelled ongoing debates in Canada and the United States concerning the need for stricter water quality guidelines, changes in watershed management policies, and the need for additional water treatment.

EPCOR Water Services Inc. (a subsidiary of EPCOR Utilities Inc.) supplies water to the city of Edmonton and its surrounding region. An extensive pipe distribution system spanning approximately 3100 kilometres enables access to more than 40 communities. Edmonton and the surrounding regions are serviced by two water treatment plants located within the city (Figure 1 and Figure 2): the Rossdale plant and the E.L. Smith plant. The E.L. Smith treatment plant was built in 1976, while the Rossdale plant was built in the 1940's. The Rossdale plant has been upgraded several times since it was first constructed. Of particular interest for this study, on December 10, 1997 the Rossdale raw water intake site was moved towards the centre of the river in an attempt to improve raw water quality. At the same time, the introduction of particle counters at the Rossdale plant allowed for much better optimisation of the filters to minimise the effect of filter-to-waste periods, filter spikes, and particle increases at the end of the filter runs. Downstream of the E.L. Smith plant and upstream of the Rossdale plant, there are 85 storm sewers that discharge into the North Saskatchewan. Both plants draw raw water from the North Saskatchewan River and use clarification, softening, recarbonation, filtration, and disinfection in the water treatment process.

The North Saskatchewan River originates at the foot of the Columbia Ice-fields in the Rocky Mountains south-west of Edmonton, and flows directly through Edmonton. En route, numerous rivers and creeks feed the river. In contrast to the watersheds that serve as Vancouver's water source, the North Saskatchewan River watershed is not protected. Covering 28,000 square kilometres, the watershed has many potential points of pathogen introduction: recreational sites, treated wastewater discharge from upstream communities, private septic systems, and agricultural operations.

The primary objective of this investigation was to determine if endemic gastroenteritis among Edmonton residents was influenced by the municipal water supply from January 1993 to December 1998. This was achieved by investigating the relationships between health outcomes associated with gastroenteritis (hospital admissions, emergency room visits, physician visits, and long-term care visits) and residential drinking water supply and finished water quality. The impact of the changes implemented at the Rossdale plant on December 10, 1997 was addressed in the investigation, together with the effects of environmental parameters such as precipitation. Another objective of this investigation was to compare and contrast the utility of various water quality indicators in identifying potential relationships with gastroenteritis. In addition to turbidity, particle counts in finished water were examined, together with raw water indicators such as turbidity and coliform counts.

5.2 Methodological Overview

Together with descriptive analyses, two main methodological app roaches were used in this study. A detailed discussion of the methodological approaches and data sources used in this study are provided in Section 3 of this report.

5.2 a) Multivariate logistic regression

Multivariate logistic regression was used to determine if differences in risk for endemic gastroenteritis existed among the water service areas (Rossdale vs. E.L. Smith), before and after the changes implemented at the Rossdale plant on December 10, 1997 (movement of the intake, and the introduction of particle counters). In the first of two approaches (GLM), cases (individuals with gastroenteritis related diagnoses) and controls (individuals with respiratory disease related diagnoses) were assigned to a water service area based on their place of residence. Separate models were evaluated for each outcome data source (hospitalisation, emergency room visit, physician visit, and long-term care visits), for all individuals greater than 2 years old. Age was fit as a categorical variable. In addition to water source and age, other variables included in the model were average household income and season. Household income was used as an indicator of socio-economic status and was assigned to individuals based on their enumeration area. Season was fit using splines.

In the second approach, multivariate logistic spatial modeling using a generalised additive model (spatial GAM) was carried -out to provide a visual representation of the geographic differences in endemic gastroenteritis risk among Edmonton residents. The objective of this analysis was to assess the spatial trend in risk as it related to water service areas. Modeling proceeded as above (multivariate logistic regression), but instead of fitting a water source categorical variable, latitude and longitude was fit using a loess smoothing function, and a variable corresponding to when the improvements took place at the Rossdale plant (December 10 , 1997). Season was also fit using a loess smoothing function.

5.2 b) Time series analysis

Similar to the Vancouver study (Aramini et al., 2000), a time series approach was used to investigate the temporal relationship between water quality and gastroenteritis. Using a generalised additive model (GAM), the objective of this analysis was to investigate and quantify the temporal associations between endemic gastroenteritis as measured by hospitalisations, emergency room, physician visits, and long-term care visits, and water quality parameters (primarily finished water turbidity) from 1993 to 1999. The modeling approach controlled for other significant and potentially confounding variables such as seasonal and long-term effects, and used nonparametric regression smoothers to assess the turbidity-gastroenteritis relationship. Two different, though related modeling approaches were used in this study: a Poisson regression approach and a Binomial (case-control) regression approach. For the Poisson modeling approach, the relationships between the daily numbers of health outcomes as measured by hospital admissions, emergency care visits, physician visits and long-term care visits and water quality from 0 to 40 days prior to health outcome event was investigated. For the Binomial (case-control) modeling approach, the relationship between case status (gastrointestinal vs. respiratory conditions) and water quality from 0 to 40 days prior to a health outcome event was investigated for each health outcome group. Each health outcome data set and age group (2-18 yrs, >18-65 yrs, >65 yrs) was modeled independently.

5.3 Discussion of Results

5.3 a) Descriptive analyses

Detailed descriptive summaries of the outcome data sets (hospitalisations, emergency room visits, physician office visits, and long-term care visits) and of the risk factor data sets (finished and raw water parameters, precipitation, average household income, etc) are provided in Section 4.1.

As anticipated, raw water parameters indicate better quality of water entering the E.L. Smith plant in comparison to the Rossdale plant, both before and after December 10, 1997. The differences are most pronounced in total and faecal coliforms. Following December 10, 1997, the differences in raw water parameters between the plants are less prominent.

Of primary interest in the time series analysis was the daily finished water turbidity. Prior to December 10, 1997, finished water turbidity values (mean and median) were slightly lower for the E.L. Smith plant. Following December 10, 1997, summary statistics appeared to be equivalent. Mean and median particle counts from the E.L. Smith plant were marginally lower compared to those of the Rossdale plant; however, particle count data must be interpreted with caution as these data were only available for 1998. Despite small improvements in finished water turbidity for the Rossdale plant after December 10, 1997, finished water turbidity values throughout the study for both plants were well within the Canadian drinking water turbidity guidelines ^{Footnote 8}. The differences in water quality parameters among the water supplies supported the efforts to investigate the influence of water source (E.L. Smith, Rossdale before December 10, 1997, Rossdale after December 10, 1997) on endemic gastroenteritis using multivariate logistic regression, and to limit the time series analysis to individuals presumably at the greatest risk of waterborne gastroenteritis (persons supplied by the Rossdale plant prior to Dec 10, 1997). It was anticipated that if a temporal relationship between daily water quality and gastroenteritis existed, it would be most likely identified in this later subset.

Also of interest was the prominent geographic disparity in average household income among Edmonton residents. As anticipated, household incomes were generally higher in the Southwest parts of the city and away from the city centre. As socio-economic factors are considered to be key determinants of health, the inclusion of this potential confounder in the multivariate logistic regression approach was supported.

5.3 b) Multivariate logistic regression

Multivariate logistic regression results are presented in detail in Section 4.2a.

After controlling for potential confounders (average household income, age, seasonal trends), results of the multivariate logistic model which fit a categorical variable for water source, suggest a small decrease in the risk of gastroenteritis among residents supplied with Rossdale water following the changes that were implemented on December 10, 1997. As many factors could potentially affect both the risks of gastroenteritis and respiratory diseases during the study period, comparisons between service areas were limited to within the same time frame (either before, or after December 10, 1997). For the same reason, interpretation of the results was based on the relative change in the relationship between Rossdale and E.L. Smith service areas following December 10, 1997, and not on the absolute value of the parameter estimates of the models. Analyses of the four health outcome data sets demonstrated a trend towards a decrease in the risk of gastroenteritis among Rossdale vs. E.L. Smith serviced residents after the improvements at the Rossdale plant.

The interpretation of the logistic regression results should be made with caution as the magnitudes of the effects were relatively small (odds ratios near one). Odds ratios between 0.5 and 2.0 are generally considered to be relatively susceptible to the influence of uncontrolled confounders. It is also important to note that the improvements at the Rossdale plant are only a potential cause of the associations observed. Results demonstrate a trend towards a decrease in the risk of gastroenteritis (in comparison to respiratory diseases) among Rossdale serviced residents after December 10, 1997 when compared to E.L. Smith serviced residents. Although it is plausible that this effect was due to the improvements at the Rossdale plant, other combinations of events could have resulted in a similar relationship. Factors (other than drinking water) that may have impacted on endemic gastroenteritis or respiratory diseases differentially across service areas could have potentially contributed to the observed relationship. The lack of an obvious spatial pattern in risk with respect to the service areas further underlies the tenuous nature of this causal hypothesis. Nevertheless, the associations observed are consistent with the improvements made at the Rossdale plant.

5.3 c) Time series analysis

Time series results are presented in detail in Section 4.2b.

Given the descriptive findings, and the results of the multivariate logistic modelling, time series analysis was restricted to households supplied by the Rossdale water treatment plant prior to December 10, 1997. Nevertheless, no significant lags were identified between daily finished water turbidity and gastroenteritis among Rossdale serviced residents, suggesting the absence of a significant association between finished water quality and endemic gastroenteritis. Lagged water quality parameters from 0 to 40 days prior to health outcome events were assessed for statistical significance and biological relevance using a time series approach. Similar to the spatial analysis, the final model was adjusted for seasonal and temporal effects. Despite limiting the analysis to the water supply posing the greatest potential risk, the Time Series analysis failed to identify temporal relationships between the finished water (daily mean turbidity, daily median turbidity, and daily maximum turbidity) from the Rossdale plant prior to December 10, 1997 and endemic gastroenteritis.

The criteria for identifying a significant lag were defined using both statistical and epidemiological principles. The sporadic statistically significant associations observed in the 3-D time series plots were subsequently assessed to acquire additional detail on the relationship between a specifically lagged turbidity measurement and gastroenteritis. The primary statistical criterion for identifying a significant lag was based on comparing the change in deviance in models with and without the lagged variable, using the likelihood-ratio test. This statistical comparison enabled the evaluation of the overall effect of the variable. However, by conducting multiple statistical tests for the 40 lags, the probability of erroneously detecting a statistically significant association was increased (Type I error). Therefore, only lags that were significant over two to three consecutive days were further evaluated for consistency among different age groups and data sources. Using the criteria outlined above (including consistency), no significant relationships were identified between lagged turbidity values and gastroenteritis.

The final models derived for the binomial and poisson models in the Edmonton time series analysis were similar to those found in the Vancouver study. However, compared to the results reported in the Vancouver time series study, the relationships (relative risks and odd ratios) between finished water turbidity and endemic gastroenteritis as assessed by hospital admissions, emergency room visits, and physician visits, were much smaller in magnitude in the Edmonton study. Furthermore, in the Edmonton study, only sporadic statistical associations were observed with turbidity, and co nsistency of the relationships was not maintained across the various health outcome data sources.

Finished water turbidity has been used as an indicator of water quality in other drinking water investigations (Schwartz et al, 2000; Morris et al, 1998; Schwartz et al, 1997). While other water quality parameters were available for analysis in this study, finished water turbidity was found to provide the best fit for the data, and provided the most stable model parameter estimates. It was this measured variable upon which temporal gastrointestinal relationships with water quality were determined. In contrast, the use of raw water quality indicators (turbidity and faecal coliform counts) and environmental parameters (precipitation and temperature) provided very unstable estimates, even after extensive smoothing of the data was carried out. The later observation demonstrates the lack of significant associations between the fluctuating raw water parameters and precipitation, and endemic gastroenteritis.

5.4 Conclusion

In comparison to the results of the Vancouver Time Series study, evidence in the present study suggests very little impact, if any, of drinking water on the level of endemic gastroenteritis among Edmonton residents. Despite markedly poorer raw water quality in Edmonton compared to Vancouver during the respective study periods, no significant temporal relationship was identified between finished water quality and the risk of gastroenteritis. The lack of a temporal relationship supports the opinion that the high quality of Edmonton's drinking water adequately minimises the risk of waterborne endemic gastroenteritis. Furthermore, the lack of a significant relationship between raw water turbidity and faecal coliform counts and endemic gastroenteritis supports the effectiveness of the treatment processes utilised.

The suggestion of an overall decrease in risk of endemic gastroenteritis among Rossdale serviced residents following improvements implemented at the plant on December 10, 1997 (using multivariate logistic regression, GLM), is consistent with the observed improvements in raw and finished water quality parameters. The lack of an obvious spatial relationship between endemic gastroenteritis and water service area (using spatial regression, GAM), together with the lack of a significant temporal relationship between endemic gastroenteritis with finished water turbidity (using time series analysis, GAM), suggest that this apparent decrease in risk is minor, and the relationship tenuous. Nevertheless, the lack of an obvious decrease in the risk of endemic gastroenteritis following the improvements at the Rossdale plant despite marked improvements in raw water quality, together with the lack of an obvious spatial relationship among the service areas, further supports the effectiveness of the operation and treatment processes utilised.

References

• Alberta Environment Protection. 1997. Standards and Guidelines for Municipal Waterworks, Wastewater and Storm Drainage Systems. http://www3.gov.ab.ca/env/protenf/publications/StormDrainageDec97.pdf
  
• Aramini, J., Wilson, J., Allen, B., Holt, J., Sears, W., McLean, M., Copes, R. 2000. Drinking water quality and health care utilization for gastrointestinal illness in Greater Vancouver. Health Canada.
  
• Brillinger, D.R. 1994. Examples of scientific problems and data analyses in demography, neurophysiology, and seismology. J. Comp. Graph. Stat 3:1 -22
  
• Chambers, J.M., Hastie, T.J. (eds). 1992. Statistical models in S. Wadsworth & Brooks, Pacific Grove, CA.
  
• Cleveland, W.S., and Devlin, S.J., (1988) Locally-weighted Regression: An Approach to Regression Analysis by Local Fitting. J. Am. Statist. Assoc., Vol. 83, pp 596-610.
  
• Cook, D.G., Pocock, S.J. 1983. Mutiple regression in geographical mortality studies, with allowance for spatially correlated errors. Biometrics 39:361-371.
  
• Curriero, F.C., Patz, J.A., Rose, J.B., Lele, S. 2001. The association between extreme precipitation and waterborne disease outbreaks in the United States, 1948-1994. Am. J. Pub. Health, 91:1194-1199.
  
• Fahrmeir, L., Tutz, G. 1994. Multivariate statistical modeling based on generalized linear models. Springer-Verlag, New York. p. 45
  
• Grey Bruce Health Unit. 2000. The Investigative Report Of The Walkerton Outbreak Of Waterborne Gastroenteritis.
  http: //www.publichealthgreybruce.on.ca/_private/Report/SPReport.htm
  
• Hastie, T.J., Tibshirani, R.J. 1990. Generalized additive models. Chapman and Hall, London, p. 335.
  
• Health Canada. 1996. Guidelines for Canadian drinking water quality - sixth edition. Canada: Supply and Services Canada.
  
• Isaac-Renton, J., J. Blatherwick, W.R. Bowie, et al. 1999. Epidemic and endemic seroprevalence of antibodies to Cryptosporidiu m and Giardia in residents of three communities with different drinking water supplies. American Journal of Tropical Medicine and Hygiene; 60:578-583.
  
• MacKenzie, W.R., Hoxie, N.J., Proctor, M.E., Gradus, M.S., Blair, K.A., Peterson, D.E., Mazmiercza k, J.J., Addiss, D.G., Fox, K.R., Rose, J.B., Davis, J.P. 1994. A massive outbreak in Milwaukee of Cryptosporidium infection transmitted through the public water supply. N. Engl. J. Med., 331:161-167.
  
• McCullagh, P., Nelder, J.A. 1983. Generalized Linear Models. Chapman and Hall, London, UK.
  
• Morris, R.D., Naumova, E.N., Griffiths, J.K. 1998. Did Milwaukee experience waterborne cryptosporidiosis before the large documented outbreak in 1993? J. Epidemiol., 9:264-270.
  
• Payment, P., J. Siemiatychi, L.Richardson, et al. 1997. A prospective epidemiologic study of gastrointestinal health effects due to the consumption of drinking water. International Journal of Environmental Health Research, 7: 5-31.
  
• Preisler, H.K., Rappaport, N.G., Wood, D.L. 1997. Forest Science. 43(1) 71-77. Regression methods for spatially correlated data: An examle using beetle attacks in a seed orchard.
  
• Schwartz, J, Levin, R., Goldstein, R. 2000. Drinking water turbidity and gastrointestinal illness in the elderly of Philadelphia. J. Epidemiol. Community Health, 54:45-51.
  
• Schwartz, J., Levin, R., Hodge, K. 1997. Drinking water turbidity and pediatric hospital use for gastrointestinal illness in Philadelphia. J. Epidemiol., 8:615-620.
  
• Sakamoto, Y., Ishiguro, M., and Kitagawa G. 1986. Akaike Information Criterion Statistics. D. Reidel Publishing Company.
  
• Steel, R.G.D., Torrie, J.H., Dickey, D.A. 1997.Principles and procedures of statistics: A biometrical approach, 3^rd edition. McGraw-Hill, Boston. p. 90.
  
• Stirling, R., Aramini, J., Ellis, A., Lim, G., Meyers, R., Fleury, M., Werker, D. North Battleford, Saskatchewan, Spring 2001: Waterborne cryptosporidiosis outbreak. Health Canada. http://www.health.gov.sk.ca/mc_dp_health_can_epi_report_NB.pdf
  
• Todd, E.C.D. 1991. Foodborne and waterborne disease in Canada: 1985-86. Polyscience Publications, p. 205.
  
• Todd, E.C.D. 1994. Foodborne and waterborne disease in Canada: 1987. Polyscience Publications, p. 196.
  
• Todd, E.C.D. 1996. Foodborne and waterborne disease in Canada: 1988-89. Polyscience Publications, p. 318.
  
• Todd, E.C.D. 1997. Foodborne and waterborne disease in Canada: 1990-91. Polyscie nce Publications, p. 286.
  
• Todd, E.C.D. 1998. Foodborne and waterborne disease in Canada: 1992-93. Polyscience Publications, p. 301.