Study setting

British Columbia is Canada’s westernmost and third most populous province (~4.7 million circa 2014). The annual incidence of foodborne infections in British Columbia is comparable to annual incidences in Canada, the USA, Australia and Western Europe.2 36 37

All British Columbia residents (defined as citizens or permanent residents of Canada who are physically present in British Columbia for at least 6 months in a calendar year), their dependents and certain other individuals (eg, some holders of study or work permits) are covered by the province’s health insurance programme.38 Enrolment is mandatory, and this programme covers nearly all the British Columbia population (with the exception of members of the Canadian military, Royal Canadian Mounted Police and some First Nations individuals covered by federal insurance programmes). Physician visits, laboratory tests, hospitalisations, outpatient hospital services and drugs for certain populations are among the publicly funded benefits. The administrative datasets that contain these healthcare use data, along with vital statistics (VS) (eg, births, deaths), demographic data and other datasets, are held by the British Columbia Ministry of Health and are accessible to researchers via Population Data BC,39 a central repository and ‘multiuniversity, data and education resource’ that ‘support[s] research access to individual-level, deidentified longitudinal data on British Columbia’s 4.7 million residents’.40

The British Columbia Public Health Act mandates that reportable diseases, including several foodborne infections,41 be reported by health professionals and laboratories to the local and provincial public health authorities, and these data are managed provincially by the British Columbia Centre for Disease Control (BCCDC). These data are housed within Panorama, the provincial public health database of reportable diseases. In the Panorama database, as well as the administrative health and VS databases, individuals are recorded by their unique Personal Health Number (PHN), allowing information from these data sources to be linked by individual.

Study design, population and timeframe

This is a retrospective cohort study of the population of British Columbia, with additional descriptive, cost and population attributable risk analyses. Because this is a dynamic population in which exposure status of individuals changes over time, our study design assesses the risk and effect of exposure in terms of person-time. Thus, rather than using fixed cohorts of exposed versus unexposed individuals, we will track individuals over time and assign their person-time at risk to either ‘unexposed person-time’ (eg, prior to foodborne infection) or ‘exposed person-time’ (eg, after foodborne infection), as described further below.

The study population is all individuals in British Columbia registered with the provincial health insurance programme at any point during the study period, that is, all individuals with the following from 2005 to 2014 inclusive: one or more record in one or more of the Medical Services Plan (MSP), Discharge Abstracts Database (DAD), VS Deaths or PharmaNet; or record of coverage under the provincial insurance programme within the Consolidation File database (see table 1). The 10-year study period is 1 January 2005 to 31 December 2014, inclusive, with additional 2 year wash-in (1 January 2003 to 31 December 2004) and wash-out (1 January 2015 to 31 December 2016) periods. During these periods, we will identify occurrences of foodborne infections, sequelae and death. The 10-year study period was selected to more than encompass timeframes for initial sequelae development and ensuing healthcare use currently reflected in the literature (ie, days to years), although there is some evidence that sequelae can develop over longer timeframes (eg, over decades).42

We assume that enrolment in the provincial health insurance programme (ie, entry into the study population) and reasons for exit from the cohort (eg, moving away from British Columbia) are not related to the exposures nor the outcomes of interest.

Data sources and linkage

The study will use individually linked, longitudinal data from eight databases to investigate both acute and longer-term health outcomes following foodborne infection (table 1). In totality, these data contain information on 14 reportable foodborne infections, physician and hospital visits, prescription medications, VS and various demographic descriptors, for the British Columbia population across the study period. All data will be stored and analysed within Population Data BC’s virtual Secure Research Environment.

Population Data BC will link the eight databases directly by individual using PHN, with additional identifiers (eg, name, age, sex) used to validate linkages and link records probabilistically when PHNs are missing. Population Data BC’s detailed linkage process is fully described elsewhere.43 Note that because the Statistics Canada Income Bands database contains area-level data (whereas the other seven databases contain individual-level data), these data are first linked to individuals (and their PHNs) using their six-digit postal code. Once the linkage is complete, each individual is then assigned a unique study identifier. All individually linked, deidentified databases are provided by Population Data BC within their Secure Research Environment, a centralised online platform, accessible via virtual private network within Canada, for accessing and analysing research data, with security standards that meet Data Steward requirements.

Measuring exposure and outcomes

The exposures of interest are infections with the 14 foodborne pathogens. These 14 infections were selected because they (1) are considered a priority in terms of prevention potential and health impacts and (2) capture nearly all reportable foodborne infections in British Columbia. Note that brucellosis and paralytic shellfish poisoning were also reportable foodborne infections in British Columbia during the study period. However, since brucellosis is very rare and nearly always travel-related, and paralytic shellfish poisoning (also rare) is syndromic and diagnosis is uncertain, these two were not considered for inclusion in this study.

Individuals will be considered exposed when and if they have a laboratory-confirmed and provincially reported case of any of the following, recorded in the Panorama database during the study period: Clostridium botulinum, Campylobacter, Cryptosporidium, Cyclospora, Giardia, hepatitis A virus, Listeria, Salmonella spp (non-typhoidal, Typhi, Paratyphi), STEC, Shigella, Vibrio parahaemolyticus and Yersinia (excluding pestis). Case definitions for each of these infections are specified by the BCCDC.44 Individuals without a reported foodborne infection, but who have International Classification of Disease (ICD) codes either for one of our infections (eg, A02.0, ‘Salmonella enteritis’) or for non-specific gastroenteritis (eg, A08.4, ‘viral intestinal infection, unspecified’45) within the MSP and DAD databases will be considered potentially exposed. We will describe these individuals as a separate group in our descriptive, economic and population attributable fraction analyses, but will remove them from the main analyses of sequelae risk. We will, however, estimate the risk of sequelae among those who are potentially exposed as a secondary analysis.

It is possible for individuals to have more than one reported foodborne infection during the study period, either as a simultaneously occurring coinfection or as two or more distinct events. For these individuals, we will treat this as a complex exposure problem; sequelae will be associated with the most plausible explanatory infection, considering biology and timing, and we will adjust for the presence of concurrent foodborne infections if applicable.46 47

Our primary outcomes of interest are those sequelae for which the link to a given foodborne infection is either established or is possible (table 2). We selected sequelae (1) with evidence of an association with any of the 14 individual foodborne infections (eg, 55)48 and (2) that occur via direct effects of pathogens or their toxin, or via autoimmune or chronic inflammatory processes that can be triggered by the infection. We will classify individuals as having the sequelae via administrative case definitions, using ICD diagnostic codes within the MSP, DAD and VS data, with the exception of stillbirths which will be determined using recorded events in the VS-Stillbirths database. The ICD codes in the MSP data are generally considered accurate to the third digit.49 Although ICD codes in the VS and DAD data are ordered by most probable diagnosis, we will consider all codes, regardless of order.

Validation of the ICD codes is currently in progress, via a literature review to identify administrative case definitions that have been validated in the Canadian context, medical expert consultation and for those sequelae without a relevant validated definition, a targeted chart review in British Columbia. In addition to the ICD codes, we may also use PharmaNet data to improve sequelae classification if needed by identifying pharmaceuticals given to patients (eg, tumour necrosis factor inhibitor use as an indication of reactive arthritis; intravenous immune globulin use for GBS).

Some of our sequelae of interest are lifelong (eg, Graves’ disease), and some are transient in that complete recovery is possible (eg, GBS, stillbirth). For lifelong sequelae, we will consider the individual as having the sequela on the earliest date they meet the administrative case definition for that sequela (with subsequent records considered as a continuation of the original event). For sequelae from which recovery and subsequent return to being at risk is possible, we will consider the individual as first having the sequela on the earliest date they meet the administrative case definition for that sequela; we will then apply a postsequela recovery time to determine the date on which the individual can be considered to be at risk for a new, subsequent occurrence of that sequela.

Individuals may develop more than one sequela during the study period, either because they develop multiple different sequelae (eg, HUS and GBS) or because they develop multiple occurrences of a single sequela from which complete recovery is possible (eg, GBS, stillbirth). In all instances, the occurrence of multiple sequelae will be recorded and described. When individuals develop multiple different sequelae during the study period (eg, HUS and GBS), we will treat these as distinct outcomes in our risk estimates. When individuals develop multiple occurrences of the same sequela, we will treat these as distinct outcomes but account for recurrent events.47

Individuals with foodborne infections who develop a sequela listed in table 2, but for which there is no current evidence of an established or possible link to the specific pathogen (eg, Campylobacter and stillbirth), will be excluded from our estimates of sequelae risk (but included in sensitivity analyses).

For all 14 infections, the secondary outcome of interest is death which will be classified using recorded events in the VS-Mortality database. Finally, we are also interested in (1) the acute illnesses related to these infections (regardless of whether the individual develops sequelae or dies) and (2) additional outcomes following acute sequelae (eg, end-stage kidney disease and kidney transplant following HUS); these will be included only in our descriptive and economic analyses.

Measuring time at risk

For all individuals, time at risk for sequelae (figure 1) will be measured from the start of their entry into the study, which we define as the earliest registration date in the provincial health insurance programme (recorded in the Consolidation File). Individuals with foodborne infections may contribute to both the exposed time at risk (during the postinfection ‘at-risk’ period) and the unexposed time at risk (prior to, and after, the postinfection ‘at-risk’ period), while individuals without foodborne infections will only contribute to the unexposed time at risk. Time at risk for a specific sequela will be measured in days, from the date of entry into the study, until the development of that sequela, death, loss to follow-up or the end of the study. We define loss to follow-up as the last date of coverage in the provincial health insurance plan, calculated using the start day registered in the most recent year plus the total days registered in that year.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Study follow-up period and time at risk for development of Sequela X (solid lines: unexposed time at risk; dashed lines: exposed time at risk. *For sequelae where recovery and return to being at risk is possible: this date + recovery time for Sequela X = date of return to being at risk for a subsequent occurrence of sequela X. B.C., British Columbia, Canada; DAD, Discharge Abstracts Database; MSP, Medical Services Plan; VS, Vital Statistics.

During the unexposed time at risk, we will treat all individuals as having the potential to develop any of the sequelae (with the exception of neonatal listeriosis and stillbirth, for which only those who are pregnant are at risk). For those who develop a foodborne infection, unexposed time at risk will end on the onset date of the infection. Infection onset date will be determined using the onset date reported in Panorama, and where this is missing, the date that the infection was reported minus the number of days between onset to reporting (eg, estimated using the Panorama data or from the literature).50

Exposed time at risk will be measured starting from the infection onset date, plus any additional induction periods (specific to each sequela and currently being determined via literature review and medical expert consultation). The end of the exposed time at risk period is currently being determined via literature review and medical expert consultation. During the exposed time at risk, individuals will be classified as having a sequela specific to their infection (table 2) on the date within the ‘at-risk’ period on which they meet the administrative case definition for that sequela (eg, the date of the physician visit or hospitalisation). After the postinfection at-risk period ends, individuals will revert to contributing to the unexposed time at risk.

These data are subject to censoring and truncation. Individuals will be censored for the sequela in the event of: death, last date of coverage in the provincial health insurance plan (ie, loss to follow-up, calculated as above) or the end of the study period, whichever comes first.51 52 In our descriptive and economic analyses, we will include all related healthcare use and prescription medication costs over the course of the infection and sequela(e), and in our estimates of mortality we will include any deaths recorded during the study period, following the sequela. We will determine whether healthcare use is related to infection and sequelae using ICD diagnosis codes, and we will determine whether prescription medication use is related via medical expert consultation.

Analysis plan

Data will be analysed and results reported following the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) and RECORD (REporting of studies Conducted using Observational Routinely-collected Data) guidelines.53 54 Analyses will be performed using SAS V.9.4 (SAS Institute, Cary, North Carolina, USA) and R.55 The nature and extent of missing data will be described. If imputation is used to complete missing data, specific methods and assumptions will be reported. We will emphasise estimation over tests of statistical significance by reporting relative measures of effect along with associated 95% CIs.

The datasets in table 1 contain the variable ‘sex’ (identified via government records) that denotes whether individuals are ‘male’ or ‘female’, thereby capturing a composite of sex and gender. To reflect for potential sex and gender differences, we will report and interpret findings stratified by this variable, in addition to overall findings.

Objective 1: To determine the risk of developing sequelae following foodborne infection, we will estimate hazard ratios using Cox regression models56 that adjust for confounders and comorbidities (see below), along with the possible effect-modifying role of age, sex, comorbidity and medication use (eg, antibiotics). We will include comorbidities as a composite score, using the revised Charlson Comorbidity Index and its associated coding algorithms.57–60 Following foodborne infection, we will compare the cumulative risk of first diagnosis of each infection-specific sequela using life table and Kaplan-Meier approaches.56 61 In the event an individual dies, we will use competing risk analysis.62 63 For those who experience more than one foodborne infection across the study period, we will explore the impacts of having multiple infections on the risk of sequelae and mortality. We will also estimate the likelihood of dying following foodborne infection using the same methods described above.

Objective 2: To describe the epidemiology and clinical progression across the range of outcomes, including acute illness (typically diarrhoea and other gastrointestinal symptoms), sequelae and death, for each foodborne infection we will calculate incidence rates; demographic, geographic and temporal distributions; timing and progression of outcomes; and case fatality rates, for both the acute stage, and sequelae associated to the foodborne infection.

Objective 3: To quantify the direct healthcare costs due to these infections and their various outcomes, we will determine health service use (ie, patterns of use by type, frequency, timing of physician visits and hospitalisations), for both the acute foodborne infection and any sequelae. We will estimate direct healthcare costs of out-of-hospital physician visits, hospitalisations and prescription medications. Out-of-hospital costs will be estimated using the MSP variables ‘Fee Item’ and ‘Paid Service’ and fee rates from the British Columbia fee schedule.64 Costs of inpatient and day-case hospitalisations will be calculated using established case-mix methodology (ie, using the ‘Resource Intensity Weight’ of each hospitalisation),65 and the British Columbia Ministry of Health unit costs for hospital stays.66 Total prescription medication costs will be calculated using the drug cost claimed by the pharmacist which includes the ingredient cost, professional dispensing fee and other special service fees (if applicable). Because these costs are captured directly in the PharmaNet data, they will be tallied directly. We will also apply these methods to determine the direct costs per sequelae, regardless of exposure. Costs will be adjusted for inflation using the Canadian Consumer Price Index.67 Results will be reported to allow comparability with other estimates (eg, 2010/2011 Canadian dollars and US dollars).

Objective 4: To determine the risk of sequelae in the population attributable to foodborne infections, we will calculate population attributable fractions using standard formulae.68 We will also describe the proportion of cases of each sequela with specific foodborne infections. Here, the total number of cases of each sequela occurring in British Columbia during the study period will be the denominator (eg, total number of cases of acute kidney injury), and the numerators will be the numbers of cases of each sequela occurring in those with specific foodborne infections (eg, total number of cases of hepatitis A virus, and of STEC, with acute kidney injury). We will also describe the proportion of individuals with sequela who do not have a foodborne infection, but who do have an ICD code for prior gastroenteritis, and use this to estimate the additional proportions of sequelae that may have an unidentified foodborne infection. We will calculate population attributable fractions and proportions for both established and possible sequelae, but clearly distinguish between the two when reporting findings.

Potential confounders and their adjustment

We will use propensity score matching and inclusion of potential confounders as covariates in our analyses, as our primary methods to adjust for confounding.69 The databases in table 1 include direct measurements of important known, strong confounders (eg, age, sex), as well as other potential factors (eg, use of protein pump inhibitors and antibiotics, disease severity). We will consider the following variables as potential confounders: age, sex, local health area, income band/area income (as a proxy for socioeconomic status), month/year, seasonality, immune status (eg, indication of immunosuppressant drugs, presence of conditions like cancer, pregnancy), use of medications like antibiotics and Charlson Comorbidity Index. Because none of our data sources include ethnicity nor race, we are unable to adjust, or conduct subanalyses, for these factors. Although ethnicity may impact healthcare-seeking behaviours, we anticipate these impacts will apply equally regardless of exposure, and thus we expect any bias in our findings to be negligible. Nevertheless, to assess residual confounding, we will conduct sensitivity analyses70 71 and perform indirect adjustments.72

Planned sensitivity analyses and study limitations

We are planning several sensitivity analyses to explore assumptions, methodological decisions, limitations in the data and robustness of results. We will explore the impact of propensity score matching on our sequelae and mortality risk estimates by also using (1) the whole unexposed population and (2) a random sample of unexposed individuals (matched on time), instead of propensity score-matched individuals. We may also explore additional matching and control strategies (eg, matching on age and sex). We will also explore the impacts of including individuals with foodborne infections who develop a sequela for which there is no current evidence of an established or possible link to the specific pathogen (eg, Campylobacter and stillbirth), in our estimates of sequelae risk. Our primary analyses will use a composite of all infections (ie, a report of any of the 14 foodborne infections) and their various sequelae. We will also analyse and present results for each of the individual foodborne infections and for each of the sequela.

A main recognised limitation of reportable disease data, such as the Panorama data in this study, is the under-ascertainment of foodborne infections. Here, this limitation means that individuals with foodborne infections who do not seek care, do not get tested, or who test negative will be misclassified as unexposed. We will assess the impacts of such potential misclassification via sensitivity analyses that illustrate how our findings could be impacted by different misclassification rates, using estimates of misclassification from the literature2 73 and from our data (eg, individuals with non-specific gastroenteritis). An additional limitation is that if sequelae develop over longer timeframes than our 10-year study (eg, over decades),42 our study cannot assess this scenario.