Methods
Data Sources
The cohort was assembled using information from 4 provincial administrative databases that include all residents of Quebec, Canada. Those 4 databases were the physician billing claims database from the Quebec Health Insurance Board (Régie de l'Assurance-Maladie du Québec), the hospital admission and discharge database (MEDECHO), the Quebec Institute of Statistics database, and the Quebec Coroner Database. These databases include the place of residence, age, and sex of persons who use health care service, as well as information on their medical visits (date, diagnosis, specialty of doctor seen, and medical procedures), hospitalizations (diagnosis, date of admission, interventions, intensive care unit stay, and date of discharge), and date and cause of death. A material deprivation index (expressed in quintiles) based on an algorithm using the postal code of residence was provided by the Régie de l'Assurance-Maladie du Québec and was used to characterize SES.
Study Population and Design
We identified a cohort of 135,703 children aged 0–17 years who received medical services in the province of Quebec, Canada, during the 1987 calendar year. Four groups of children were defined based on the International Classification of Diseases (ICD), Ninth Revision diagnostic codes: 1) all children who received care for a TBI diagnosis (n = 7,894); 2) all children who had had a probable TBI (n = 47,537); 3) all children who had medical visits for fractures or dislocations of the extremities, labeled as musculoskeletal injuries (n = 24,841); and 4) a random sample of children who sought care for any other reason (n = 55,431, which is equal to the sum of groups 1 and 2).
For each child in the cohort, the Régie de l'Assurance-Maladie du Québec provided data on all medical services received until the end of the 2008 calendar year, for a maximum follow-up period of 21 years. The ethics review board of the Centre de Recherche Interdisciplinaire en Réadaptation du Montréal Métropolitain approved this study, and permission to access the data was obtained from the provincial commission on protection of personal information.
Study Variables
Outcome. The outcome variable was death by suicide as assessed using the coroner's database. In the coroner's database, suicide deaths were coded using an internal classification system based on the codes from the ICD, Ninth Revision, that had the prefix S until 2000, after which the ICD, Tenth Revision, was introduced.
Exposure Variables. The main exposure variable was having had a TBI, which was defined by the following diagnoses: concussion, intracranial hemorrhage, and/or cranial fracture (ICD, Ninth Revision, codes 800–804 and 851–854). A list of pertinent diagnostic and procedure codes has been published previously.
Subjects were classified as exposed at the time of the first occurrence of one of the above diagnoses recorded in the Régie de l'Assurance-Maladie du Québec, the hospital admission and discharge database, or both, during the 21-year follow-up period. Subjects did not necessarily sustain a TBI in 1987 (the year in which our groups were defined); they might have had 1 or more in subsequent years. To address this, we created time-dependent cumulative variables that were updated to include the individual number of TBIs as they occurred in children, adolescents, and adults during the follow-up. On the basis of these variables, we constructed time-dependent indicators of the presence or absence of TBIs during follow-up. We replicated the process for both probable TBIs and musculoskeletal injuries.
It is not uncommon for children to have multiple TBIs. To differentiate a visit for a new TBI from a follow-up visit for a previously sustained TBI, we considered 3 "clear zones": 15 days after the initial diagnosis and 90 and 180 days after that. These intervals represented the minimum number of days between 2 consecutive visits for TBIs that would indicate 2 different episodes. TBIs that occurred on the day of suicide were identified but not counted as new episodes because they most likely represented the method chosen to commit suicide.
Time-dependent indicators were further developed to denote whether injuries were sustained during childhood (<12 years of age), adolescence (12–17 years of age), or adulthood (≥18 years of age). This allowed us to address whether the risk of suicide varied with the age at which the injury was sustained. The severity of the TBI was determined using an algorithm that classified TBI in 3 categories: concussion, cranial fracture, and cerebral contusion/intracranial hemorrhage.
Confounders. The potential confounding variables that were included were sex, age in years at inclusion in the cohort, SES, probable TBI (based on both diagnostic and procedural codes, validated by Kostylova et al.), and musculoskeletal injuries, as well as mental disorders that were diagnosed before a TBI. Mental disorders that occurred before the TBI were examined by using an algorithm based on ICD, Ninth Revision, codes 291–318 and ICD, Tenth Revision, codes F00–F99. We only considered mental health disorders that were diagnosed before the first TBI because those that occur after are considered to be in the causal pathway.
Given the 21-year follow-up period, some subjects were followed until they barely reached adulthood (21 years of age), whereas others were followed until the age of 38 years. Suicide risk varies with age, resulting in potentially different risks of suicide in the cohort members. To verify whether the risk of suicide changed in relation to the age at study inclusion, we created 3 categories for age at inclusion in the cohort: younger than 6 years of age, 6–11 years of age, and 12 years of age or older.
SES was assessed using the deprivation index, a population-based proxy derived from postal codes and census tract data. Values for this index were missing for 952 subjects (0.70% of the cohort) who moved out of the province during follow-up; these people were excluded from the cohort. We dichotomized the SES variable by regrouping the first, second, and third quintiles into a higher SES category, whereas the 2 lower quintiles became the lower SES category. We expected the risk to differ across SES levels and therefore stratified the SES variable so that each SES category would have its own baseline risk.
We included other injuries, such as musculoskeletal injuries and probable TBI, in the model. A probable TBI was an injury that was not positively identified as a TBI but that leaves a strong suspicion that one might have occurred because of the diagnosis, medical procedures, or tests involved. For example, a dislocation of the jaw combined with a head or brain scan would be categorized as a probable TBI. Musculoskeletal injuries that were included were those that required medical attention, such as fractures, dislocations, and severe soft tissue injuries. The inclusion of these other types of injuries can help control for certain premorbid characteristics that might be common causes of both TBIs and suicide, such as an impulsive/aggressive personality or a genetic predisposition to injury.
Analysis
Descriptive statistics were measured to characterize our sample. We then used a Cox proportional hazards model with time-dependent covariates to analyze the survival time between injury and suicide. We used age as the time axis and defined time 0 as the age in years at enrollment in the study. Subjects who did not commit suicide were censored at the end of follow-up or at the time of death from other causes.
We modeled TBI as the main exposure variable in 4 different ways: 1) TBI at any age, 2) TBI by age group (children, adolescents, and adults), 3) TBI severity, and 4) repeated TBIs across age groups. Concussions and cranial fractures were grouped into a single low-severity category, and intracranial hemorrhages comprised the high-severity category.
For each model, we estimated the hazard ratio for suicide with 3 levels of adjustment: no adjustment, adjustment for demographic variables (age in years at injury, sex, and SES strata), and adjustment for demographic variables, other injuries (probable TBI, musculoskeletal injuries), and mental disorders diagnosed before the TBI. Finally, in order to determine whether the risk of suicide was higher for persons with a TBI and no mental health disorders, we conducted an analysis stratified by the presence of mental health disorders. The proportional hazards assumption was verified for each model.
Statistical analyses were conducted at the research data access center of the Institut de la Statistique du Québec. The statistical software used included SAS, version 9.2, and R, version 2.14.