Health Reports
Linkage of the nationally representative Canadian Community Health Survey – Nutrition 2004 to routinely collected mortality records

by Mahsa Jessri, Deirdre Hennessy, Anan Bader Eddeen, Carol Bennett, Didier Garriguet, Claudia Sanmartin and Douglas Manuel

DOI: https://www.doi.org/10.25318/82-003-x202200900002-eng

Abstract

The Canadian Community Health Survey (CCHS) – Nutrition 2004 (n=35,107; interview dates from January 2004 to January 2005) linked to the Canadian Vital Statistics – Death Database (CVSD) (2011) represents a novel linkage of a population-based, nationally representative nutrition survey with routinely collected mortality records (including date and cause of death). The linkage was done through individual tax data in Canada, and contains longitudinal records for 29,897 Canadians aged 0 years and older—1,753 of whom died—in the 10 provinces of Canada. The median follow-up time was 7.49 years, with 102,953 person-years among males and 114,876 person-years among females (unweighted), and included a special sampling survey weight (for linked data) to account for those who did not agree to share and link their information. The CCHS – Nutrition 2004 linked to CVSD has been used to evaluate associations between lifestyle and sociodemographic characteristics and mortality. Using these data, statistical methods have been developed and tested to control random and systematic measurement errors when evaluating the relationship between different dietary exposures (evaluated using repeated 24-hour dietary recalls) and health outcomes. The linked data are available through Statistics Canada’s Research Data Centres.

Keywords

Statistics Canada’s Data Linkage Program, National Survey Data, Vital Statistics, Universal Healthcare System

Authors

Masha Jessri (mahsa.jessri@ubc.ca) is with the Food, Nutrition and Health Program and the Centre for Health Services and Policy Research, at the University of British Columbia. Masha Jessri is with the Clinical Epidemiology Program at the Ottawa Hospital Research Institute, the School of Epidemiology and Public Health at the University of Ottawa, and presently holds the Canada Research Chair in the Food, Nutrition and Health Program, Faculty of Land and Food Systems at the University of British Columbia. Mahsa Jessri, Deirdre Hennessy, Didier Garriguet, Claudia Sanmartin, and Douglas Manuel are with the Health Analysis Division at Statistics Canada, Ottawa, Ontario. Anan Bader Eddeen, Carol Bennett, and Douglas Manuel are with the Institute for Clinical Evaluative Sciences in Ottawa, Ontario. Carol Bennett and Douglas Manuel are also with the Clinical Epidemiology Program at the Ottawa Hospital Research Institute, and Douglas Manuel is also with the Department of Family Medicine at the University of Ottawa, and the C.T. Lamont Primary Health Care Research Centre Program at the Bruyère Research Institute in Ottawa, Ontario, Canada.

Introduction

National health surveys and vital statistics registries are the cornerstones of surveillance, monitoring and policy development in most developed countries.^Note 1 National health surveys collect detailed information on a broad range of health behaviours, health statuses and sociodemographic characteristics from large, representative samples of populations, with the aim of monitoring risk factors and health status at the national level and to inform policy.^Note 2 However, these surveys are cross-sectional, and therefore no information on participants’ health status is collected beyond the survey date (via self-report). By contrast, vital statistics registries provide objective measures of mortality at the national population level, despite being limited in providing any information beyond basic demographic characteristics.^{Note 1 Note 3 Note 4} Increasingly, vital statistics are being linked to national health surveys, ^{Note 5 Note 6 Note 7 Note 8 Note 9 Note 10 Note 11 Note 12 Note 13 Note 14} thus providing a unique and prospective component to these nationally representative surveys for in-depth assessment of mortality determinants^{Note 8 Note 9 Note 10 Note 11 Note 12} including socioeconomic and lifestyle behaviours, such as dietary intake, smoking and physical activity.^{Note 15 Note 16 Note 17 Note 18 Note 19 Note 20 Note 21} Particularly, national nutrition surveys are not routinely linked with health administrative databases—this results in a lack of evidence on the relationship between health and unhealthy diets at the population level.

Aims

This cohort profile describes the linkage of the Canadian Community Health Survey (CCHS) – Nutrition (2004) to the Canadian Vital Statistics – Death Database (CVSD) (2011). Furthermore, it demonstrates how these data can be used in terms of potential analysis strategies and their strengths and limitations—considerations that could apply to any national nutrition survey worldwide.

In addition, a summary of findings on the associations between all-cause mortality and selected sociodemographic characteristics and lifestyle behaviours are presented. This resource provides the first nationally representative linked nutrition survey in Canada, allowing for a generalizable evaluation of the potential link between dietary intake and mortality.

Data resource description

Canadian Community Health Survey – Nutrition

Statistics Canada has collected cross-sectional, nationally representative CCHS data since 2000 with the objective of providing timely information on health status and determinants, as well as health system utilization. The CCHS is funded through the partnership of Statistics Canada, the Public Health Agency of Canada, Health Canada and the Canadian Institutes for Health Information. Since the inception of the CCHS, it was determined that a focused survey cycle (longer version) dedicated to collecting reliable and timely information on dietary intakes was required to complement the information on other health determinants. The CCHS –Nutrition survey was conducted in 2004 to provide detailed information on usual dietary intakes (in terms of food groups, nutrients, nutritional supplements and eating patterns) and measure the prevalence of household food security.^{Note 22 Note 23 Note 24} The CCHS – Nutrition 2004 (January 2004 to January 2005) was a complex, multistage cross-sectional survey that included 35,107 non-institutionalized Canadians aged 0 years and older from 10 provinces (response rate=76.5%, representing 98% of residents in the provinces).^{Note 22 Note 23 Note 24} Data on sociodemographic characteristics, lifestyle behaviours and selected health conditions were collected using interviewer-administered questionnaires, and weight and height were directly measured by trained interviewers. A modified version of the United States Department of Agriculture (USDA) automated multiple-pass method (AMPM) was used to collect two standardized 24-hour dietary recalls,^{Note 25 Note 26} with the second recalls collected approximately 3 to 10 days after the first interview on a different day of the week, for about 30% of the sample. The second dietary recall can be used to assess the day-to-day variation in dietary intakes for the entire sample—including those with only one day of recalls—when using the National Cancer Institute (NCI) method.^{Note 27 Note 28} The nutrient composition of foods was analyzed using Health Canada’s Canadian Nutrient File.^Note 29 Listing below presents the CCHS – Nutrition 2004 components:

• 24-hr recall (all ages)
• Administration (data sharing) (all ages)
• Alcohol (age ≥ 12 years)
• Children's physical activity (age 6 to 11 years)
• Chronic conditions (all ages)
• Food security (all households)
• Fruit and vegetable consumption (age ≥ 6 months)
• General health (age ≥ 12 years)
• Household and education (all ages)
• Income (all ages)
• Labour force participation (age 15 to 75 years)
• Measured height and weight (age ≥ 2 years)
• Physical activities (age ≥ 12 years)
• Sedentary activities (age 12 to 17 years)
• Self-reported height and weight (subset age ≥ 18 years)
• Smoking (age ≥ 12 years)
• Sociodemographic characteristics (all ages)
• Vitamin and mineral supplements (all ages)
• Vitamin and mineral supplement details (all ages)
• Women's health (age ≥ 9 years)

More details on the survey are available elsewhere.^Note 22

The Canadian Vital Statistics – Death Database (formerly the Canadian Mortality Database)

The CVSD is a census of all deaths registered in Canada since 1950. Information on deaths is collected by Statistics Canada from provincial and territorial vital statistics registries. The CVSD includes demographic (name, date of birth, date of death, postal code at the time of death) and medical (cause of death) information. The cause of death is coded using the World Health Organization’s International Classification of Diseases, version 10. Deaths occurring between January 1, 2004, and December 31, 2011, were eligible for linkage to the CCHS – Nutrition survey (Table 1). More information regarding the CVSD is available elsewhere.^Note 30

Data collected

Linkage of the CCHS – Nutrition 2004 was conducted as part of a more comprehensive linkage of the CCHS national health surveys to mortality data.^Note 1 Data were collected under the authority of the Statistics Act and were analyzed and accessed at Statistics Canada. Of the 35,107 participants of the CCHS – Nutrition 2004, 89.33% (n=29,897) consented to share their survey information with federal and provincial ministries of health and to link their responses to administrative databases. Special sampling survey weights were created for the linked CCHS – Nutrition 2004 data to adjust for those who did not agree to share and link their information.

The linkage was conducted based on Fellegi-Sunter’s theory of probabilistic record linkage using G-Link, the generalized record linkage system developed by Statistics Canada to support large-scale record linkages.^{Note 31 Note 32} The following unique identifiers were used in the linkage process: given name, last name, date of birth, sex and postal code. The Historical Tax Summary File (HTSF) was used to assist in record linkage through the provision of additional linkage information (given name, last name, postal code) and in the manual resolution of doubtful links. The HTSF is a compilation of tax return files and represents unique individuals for whom a tax declaration was created in a given year. Internal and external validation was conducted to evaluate the accuracy of the linkage process and the fitness of the data for use in the analysis. Global false positive and false negative rates for the more comprehensive linked data were estimated at 0.04% and 2.43%, respectively, using a weighting function. Further information regarding the linkage process has been described previously.^Note 1

Data linkage was approved by Statistics Canada’s Executive Management Board.^Note 33 The use of the data was governed by the Directive on Record Linkage.^Note 34 Participants’ privacy during record linkage and use of the linked files were ensured by Statistics Canada. Access to the unique identifying information (e.g., names) was limited to employees directly involved in linking the databases, and these individuals did not access health information. After record linkage, all identifying information was removed (de-identified), and an analytical file was created for subsequent use and analysis.

The record linkage was conducted by Statistics Canada with funding support from Health Canada, the Ottawa Hospital Research Institute and the Canadian Institutes for Health Research.

Quality assessment

Overall, 5.9% of CCHS – Nutrition respondents (n=1,753) eligible for linkage died and were linked to a mortality record. Table 1 presents the crude unadjusted mortality rates by follow-up year in the CCHS – Nutrition 2004 data linked with the CVSD. The median follow-up time was 7.49 years, with 102,953 person-years for males and 114,876 person-years for females (unweighted) as of December 2011. Linkage rates were compared across respondents’ characteristics (Table 2) to assess whether the pattern of linkage rates reflected the differential mortality risk among groups at a higher risk.^{Note 35 Note 36} Generally, the linkage rates (weighted) were higher for older adults 75 years and older (39.94% vs. 0.09% for children 12 years and younger) and males (5.06% vs. 3.85% for females), reflecting the expected higher mortality risks in these groups. Participants in Alberta, British Columbia and Ontario were less likely to link to the CVSD than those in other provinces, while participants in Nova Scotia were more likely to be linked. The overall pattern of linkage was in line with the general CCHS linkages conducted by Statistics Canada.^Note 1 Further external validation was conducted to compare the mortality patterns in the linked data with official mortality statistics.^Note 37 The distribution of death across several variables for the linked CCHS – Nutrition data is presented in Table 3. The age distribution in the linked CCHS – Nutrition data showed an increased occurrence of deaths with age, with about 51.08% of these deaths concentrated in those aged 75 years and older (weighted estimate). A similar pattern was observed in the CVSD dataset (60.7%), considering that deaths among the institutionalized population are not captured in the linked CCHS – Nutrition 2004 data since institutional residents were not part of the CCHS survey. Geographic distributions of deaths in the linked CCHS were also similar to those derived from the CVSD, with the majority of deaths occurring in the largest provinces (Ontario and Quebec). As expected, the weighted distributions of deaths resembled those in the CVSD more closely than the unweighted distributions. The distribution of deaths by sex and cause were also similar to those in the CVSD dataset.

Figure 1 and Table 4 presents the age-standardized mortality rates (ASMRs) derived from the linked CCHS – Nutrition 2004 file compared with the official mortality rates. Annual adjusted mortality rates are from the Canadian population in 2001. As expected, annual ASMRs in the linked CCHS – Nutrition 2004 file were lower than those derived from official mortality databases reported by Statistics Canada because linkage does not capture 100% of the population. People living on First Nations reserves, people living on Crown lands, institutional residents and full-time Canadian Forces members were excluded from the CCHS sampling. There is also a risk of healthy respondent bias in which those who respond to surveys tend to be healthier than non-respondents. The selection bias could affect the mortality rate in the years in this cohort but would likely be attenuated over time. The ASMRs were lowest in the first year of follow-up for the CCHS – Nutrition 2004, and generally increased in the following years.

Data resource use

This new data resource has several potential applications, and can be used to evaluate associations between lifestyle and sociodemographic characteristics and mortality. An example of such an application is presented below. The analyses presented in the following section pertain to respondents aged 20 or older from the CCHS – Nutrition 2004 linked to CVSD data (unweighted person-years: 117,323) since the mortality rate in those aged younger than 20 is low; including those aged younger than 20 years means there is not sufficient power for the analysis.

Example: Application of the CCHS – Nutrition 2004 linked to CVSD data

Here a brief example of how the data can be used in the evaluation of the relationship between dietary factors (i.e., sodium, saturated fats and added sugars intakes) and mortality is presented.

Data: This analysis included 16,438 participants aged 20 years and older. For analyses of nutrients (sodium, saturated fats and added sugars intakes), 24-hour recalls were used. Invalid 24-hour dietary recalls were removed, as well as records for participants who were pregnant or breastfeeding, leaving us with 16,212 (n=4,901 of whom had a second recall) complete dietary recall data for nutrient-specific analyses (median follow-up: 7.48 years). Outliers were identified and winsorized through observing the model residuals after Box-Cox transformation of exposure (diet) variable, and all analyses accounted for age as well as the nuisance effects including the indicator of sequence of diet recall (first vs. second) and the day of the week when the 24-hour dietary recall was collected (weekend vs. weekday). The distribution of usual intakes for sodium density (grams per 1000 kcal)^{Note 38 Note 39} and percentage of energy from saturated fat and added sugars^Note 40 were estimated using the NCI method, which can be used to account for day-to-day variation in intakes when more than one day of intake is available for some of the respondents.^{Note 41 Note 42} Since added sugar values are not included in the CCHS food composition database, an estimation method proposed by Brisbois et al. and previously tested by our team^{Note 43 Note 44 Note 45 Note 46 Note 47} was used to derive approximate estimates of added sugars.

Variables: Mortality, dietary variables (i.e., sodium, saturated fats and added sugars intakes), and covariates (i.e., age).

Analysis: In the first step of the NCI analysis, univariate likelihood-based models were used to estimate the distribution of intake (simulated population) and to generate intake percentiles. In the next step, regression calibration models were fitted for estimating the association of nutrients of public health concern with all-cause mortality. Age-adjusted Cox proportional hazard ratios (HRs) (95% confidence intervals [CIs]) were calculated for all-cause mortality, with the usual intake of nutrients entered as continuous in all models. All 95% CIs for HRs were estimated from coefficient of variations after 500 bootstrap runs. The misreporting status was used to address systematic measurement errors in dietary intakes. Outliers in terms of dietary intakes were identified and replaced with the next closest plausible value to address random measurement errors.

Results: Table 5 presents the crude weighted mean of the estimated usual intakes of sodium density, percentage of energy from added sugars and saturated fat. Generally, weighted and age-adjusted HRs were revealed increased mortality with each 100-unit increase in the usual intake of sodium density in males (HR: 1.06 95% CIs: 0.95 to 1.17), even though it did not reach statistical significance. Each 5% increase in usual intake distribution of added sugars was associated with a significant 1.32 (1.1 to 1.57) times higher all-cause mortality risk in females, but not in males. As expected, the association of the sodium density, percentage of energy from saturated fats and added sugars with all-cause mortality were weaker when using single 24-hour dietary recalls. Adjustment for usual intake distribution of energy intakes using bivariate likelihood-based models did not change the magnitude and direction of results significantly.

Future work: Under the same funding umbrella used to create this database, a more in-depth and completely adjusted analysis of nutrients (i.e., sodium, added sugars, saturated fats) and dietary patterns in relation to mortality has been completed. These data are published elsewhere.^Note 48

Considerations: These analyses are not fully adjusted and are for presentation of proper data use only, and therefore should not be interpreted as complete adjusted analyses. Reliance of CCHS – Nutrition 2004 linked to Canadian Vital Statistics – Death Database on self-reported dietary data from 24-hour recalls poses challenges and limitations when evaluating long-term outcomes (e.g., mortality). The use of appropriate methodology for handling random and systematic measurement errors is essential to avoid misleading and conflicting results. Systematic errors can cause the estimate to be biased in any direction.^Note 49 Implementing these methodologies requires specialized training. Overall, the analyses of these data require careful planning and attention to the presence of measurement errors, especially when evaluating long-term outcomes.

Strengths and weaknesses

The newly created CCHS – Nutrition linked to CVSD has several strengths. First, it is a population-based nationally representative nutrition survey linked with mortality, which provides a rich dataset to examine the relationship between dietary intake and health outcomes. It is also straightforward to update the dataset with additional follow-up years as the data becomes available at Statistics Canada. Longitudinal studies provide a powerful resource to investigate the association of dietary intakes with health outcomes. Most longitudinal studies in the field of nutritional science have been cohort studies using food frequency questionnaires for collecting dietary data; these cohorts have a less complex sampling framework because they are not designed to be representative of the population. While the emergence of record linkage techniques has enabled us to link detailed survey information to outcomes, the analyses of nutritional data in this context warrant careful planning and attention to the presence of random and systematic errors.^Note 50 It is necessary to use specific methodologies, such as those described in this paper, to ensure robust estimates, especially when evaluating long-term outcomes (e.g., mortality). There are also some provinces (e.g., Alberta) that were less likely to have CCHS – Nutrition 2004 respondents linked with the CVSD. Although it is unclear why this may be, this discrepancy should be a caution when using the data.

With regard to the example presented, there are limitations and challenges to using nationally representative linked nutrition surveys with 24-hour recalls as the main method of data collection (e.g., CCHS – Nutrition 2004 linked with CVSD). These challenges and limitations include the time-consuming, computationally intensive analyses and the need for proper adjustment for measurement errors in analyses to avoid misleading and conflicting results. In the presence of random error, systematic error and lower power, potential limitations also include attenuated estimates, especially in the use of repeated dietary recalls. Although within-individual random errors do not always attenuate estimates.^{Note 51 Note 52} Specifically, there are time-varying covariates for which the link with mortality may be weaker as time progresses away from the baseline observation in 2004.

Data resource access

The linked data are available through Statistics Canada’s Research Data Centres.

Funding

Authors acknowledge funding from Statistics Canada, Health Canada and the Canadian Institutes of Health Research (CIHR) for the linkage project. This study was supported by a grant from Statistics Canada.

M.J. was supported by the CIHR Banting Postdoctoral Fellowships program, a CIHR Fellowship and Banting Foundation Award. 

The record linkage was conducted by Statistics Canada with funding support from Health Canada, the Institute for Clinical Evaluative Sciences and the Ottawa Hospital Research Institute.

Acknowledgement

The authors would like to thank the linkage team with the Health Analysis Division, Statistics Canada.

Conflict of interest

The authors declare no conflict of interest.