Population and sample
We linked Canadian Community Health Survey (CCHS) 2005–2017 to National Ambulatory Care Reporting System (NACRS) 2003–2017 to examine the association between food insecurity status and injury-related ED visits.
CCHS is an annual cross-sectional survey administered to roughly 65,000 households in Canada. One member 12 years or older is randomly selected per household to answer the survey. The responses generalize to 98% of the non-institutionalized population in the country. Questions on food insecurity have been formally incorporated in the survey since 2005 though certain jurisdictions chose not to answer them when given the option.
NACRS is the largest database of administrative records on ED visits in Canada, containing 64% of the ED cases nationwide. Ontario since 2002, Alberta since April 2010, and Yukon territory since 2015 have been mandated to report all ED records to NACRS; only a limited share of ED records were reported in other jurisdictions and years (e.g. 2% in Manitoba), where the reporting mandate was either partial or absent . NACRS provides case-level information on timing and cause of each ED visit, which is assigned a main cause and up to nine joint causes, all coded in International Statistical Classification of Diseases and Related Health Problems, Tenth Revision, Canada (ICD-10-CA).
We limited our sample to Ontario and Alberta given their full coverage by NACRS. The two provinces combined represented roughly half of the country’s population. We linked all respondents from Ontario in CCHS 2005–2017 to NACRS 2003–2017 through unique person identifiers; likewise, respondents from Alberta interviewed in April/2012–2017 were linked to NACRS from April/2010–2017. Such linkages ensured taking full advantage of the available data to build the outcomes (injury-related ED visits in the past year) and a key covariate (ED visits in the year before). We excluded the other jurisdictions and years due to potential sampling bias; Yukon was excluded due to its lack of food insecurity measurement in CCHS 2005 and 2013–2016. Of the 241,500 survey respondents from Ontario and Alberta, we excluded 1800 individuals with invalid food insecurity status and all 27,400 Ontario respondents in CCHS 2015–2016 when the province opted out of food insecurity measurement. The final sample consisted of 212,300 individuals 12 years or older, of whom 20,900 had injury-related ED visits in the past year.
Our primary outcome of interest was a count variable for the number of injury-related ED visits during the 12 months preceding the interview. We identified causes of injury mainly using the ICD-10-CA codes for secondary and tertiary causes of visit because main (primary) cause almost always described diagnosis (e.g. fracture) versus underlying causes (Table S1). As a sensitivity check, we constructed a narrower definition for injury, considering only the secondary cause. We separated intentional injuries from non-intentional injuries and examined two types of intentional injuries (violence, self-harm) and nine types of non-intentional injuries (fall, medical complications related to surgery and non-surgical interventions, being struck by objects, overexertion, transport incident, skin piercing, animal bite or sting, poisoning, and other miscellaneous). For transport incidents, medical complications, fall, and struck-by injuries, we further investigated subcategories under each. Given the demographic heterogeneity in injury risk [2, 3], we also stratified analyses of all-cause injury by sex and age.
Our key independent variable was food insecurity status in the past 12 months. This is a four-level variable built from the 18-item questionnaire in CCHS, developed by United States Department of Agriculture and adapted by Health Canada . Ten questions asked about adults’ access to food in the past year while eight questions concerned food access among children below 18 years if there were any in the household. Based on the number of affirmative answers, a household was categorized as either food secure or marginally, moderately, or severely food insecure (Table S2).
We adjusted for factors that may confound the relationship between food insecurity and injury-related ED visits, including respondent’s sex (male, female), age at interview (years), race-ethnicity (white, Black, Indigenous, others), immigrant status (Canadian-born, immigrant), tobacco smoking status (never, former, current), past-year alcohol consumption (none, up to once a week, more than once a week), and number of ED visits in the year before (count, as a proxy for baseline health and injury risk). We also controlled for household characteristics including income categories (in $20,000 intervals), highest education (high school incomplete, high school graduate, some college, college degree), housing tenure (renter, homeowner), household type (couple with children, couple without children, lone parents, others), province (Ontario, Alberta), and survey cycle. We selected the confounding variables a priori based on their established associations with injury [10, 14, 15, 21, 24, 26, 28, 30, 31]. Missing values in covariates were dummy coded and kept in the analyses.
We used chi-squared tests and t-tests to compare sample characteristics by injury status; we also used t-tests and trends analyses to compare crude rates of injuries across food insecurity levels. We fitted Poisson models on the all-cause and cause-specific injury outcomes, adjusting for confounding factors. Models on all-cause injury were further stratified by sex and age. We conducted sensitivity analyses on all-cause injury, expanding the sample to include jurisdiction-years with partial ED records, using a narrower definition of injury, and applying person weights, respectively. Given that most people from our sample did not visit ED for injury in the past year, we also experimented with standard and zero-inflated negative binomial models, respectively, to compare their fitness to data to the Poisson model’s. We showed two-sided robust confidence intervals. Results were considered statistically significant at p < 0.05. Analyses were conducted unweighted in Stata SE 15.1. Numbers of observations were rounded to protect respondents’ identity. Ethics approval was obtained from the Health Sciences Research Ethics Board at the University of Toronto.