Data sources

Healthcare resource use data from January 1996 to December 2016 of residents of British Columbia (BC), a Canadian province of 4.4 million (as of 2016, 13% of Canada’s population),14 were obtained through the provincial health administrative databases (covering >98% of legal residents). These databases provide linked, individual-level information on demographics, vital statistics and healthcare encounters of all BC residents15–19 (Details are provided in online appendix 1). Previous analyses have showed a very low prevalence of missing data, under-reporting and misclassification in these databases.20 21 For example, in the current study cohort, only 5% of cost data was missing in the pharmacy dispensation database while the outpatient services database did not have any missing cost data. All inferences, opinions and conclusions drawn in this study are those of the authors and do not reflect the opinions or policies of the Data Steward(s). Ethics approval was obtained from the University of British Columbia Human Ethics Board (H17-00938).

Study design and sample

This was a retrospective matched cohort study. First, we identified individuals who satisfied a validated case definition of asthma (sensitivity=0.632, specificity=0.997).22 According to this algorithm, asthma was defined as the occurrence of one or more asthma-related inpatient admissions or two or more asthma-related outpatient physician visits (on different dates) or three or more filled asthma-related medications (on different dates) during any 12 month period. Asthma-related inpatient and outpatient encounters were defined as those in which asthma was the most responsible diagnosis (International Classification of Diseases codes, ninth edition (ICD-9): 493.xx or tenth revision (ICD-10): J45.xx, J46.xx). For asthma-related medications, we used a prespecified list (see online appendix 2 for the list of medications). For each patient identified, we defined the date of the first asthma diagnosis as the index date, marking the start of follow-up. To study the evolution of multimorbidity costs over time, we included only new patients who had been in the databases for at least 3 years prior to their first asthma diagnosis. To further reduce the risk of including patients with COPD who might have received the same medications as those with severe asthma, we limited our sample to patients aged between 5 and 55 years at the time of asthma diagnosis and excluded patients who incurred any COPD-specific hospitalisation or three or more COPD-specific physician visits in the first 2 years following the index date for asthma (COPD codes: ICD-9: 491.x, 492.x, 493.2x, 496.x; ICD-10: J43.x, J44.x). Once the patients were included, we followed them over time regardless of age.

Of note, although we removed individuals with a COPD diagnosis in the first 2 years after the index date, COPD is considered a legitimate comorbid condition of asthma.23 As such, beyond the first 2 years, we still considered COPD-related healthcare use in cost calculations.

Second, we divided the asthma cohort into two groups according to disease severity: the severe asthma group experienced an incident year of severe asthma (without having asthma before) and the ‘non-severe asthma’ group experienced only mild and/or moderate asthma throughout the study period. Severity was calculated for each complete 12-month interval from the index date based on a validated algorithm,24 defining severe asthma as the use of high doses of ICS in beclomethasone-chlorofluorocarbon equivalent (average daily dose of at least 1000 µg/day) or reliever medications (more than 10 doses of short-acting β2 agonists per week) while still experiencing moderate-to-severe exacerbations (online appendix 3 provides the details of severity classification). Specifically using Canadian administrative health data, this severity algorithm was based on the Canadian Asthma Consensus Guidelines and was validated against well-established indicators of asthma severity such as pulmonary function measurements.25 The definition of severity in the Canadian guideline is concordant with the definition proposed by the joint European Respiratory Society/American Thoracic Society Task Force on severe asthma26 and has been used in multiple peer-reviewed publications.27 28

As the basis for comparison, we created a third group from a random sample of BC residents who did not satisfy the asthma case definition during the same study period (‘no asthma’). We excluded non-asthmatic subjects with less than 3 years of coverage in the database to match a similar inclusion criterion imposed on the asthma groups. For each included subject in this group, we assigned an index date corresponding to the data of a randomly chosen healthcare encounter which occurred after being in the data for 3 years. Similar to the asthma cohorts, we excluded patients with COPD from the non-asthma cohort.

All study participants were followed from the index date to date of death, loss-to-follow-up (ie, end of registration with the healthcare system) or 31 December 2016, whichever came first.

Matching

Using a pragmatic common-referent pair matching approach,29 we first estimated the propensity scores by fitting two separate logistic regression models that predicted the probability of an individual having severe asthma (common referent group) relative to having non-severe asthma or no asthma, adjusted for the index year, age, sex, area of residence (ie, urban versus rural) and neighbourhood household income quintiles (all observed in the index year). Next, we matched each patient with severe asthma to a patient with non-severe asthma and an individual without asthma, with respect to age (±5 years), sex, index year (±2 years) and propensity score (±0.2). The histogram of propensity scores across three matched cohorts is provided in online appendix figure 1. The remaining imbalances in these covariates across the three groups, as well as pre-existing comorbidity status in the 12 month prior to the diagnosis of asthma (Charlson comorbidity index, number of hospitalisations, physician visits and dispensations of medications), were additionally adjusted for in the regression analysis model.

Outcome variables

The primary outcome was incremental costs of severe asthma, defined as the difference in mean total healthcare costs between severe asthma and non-asthma and between severe and non-severe asthma groups. A secondary outcome was condition-attributable incremental costs, defined as additional costs that were attributable to either asthma or identifiable comorbid conditions. Of note, in line with the definition set forth by the UK Medical Research Council,6 comorbidity was defined as the coexisting condition, rather than the secondary diagnosis of asthma. All costs were adjusted to 2016 Canadian dollars using historical inflation rates from the Consumer Price Index of BC,30 in which CAD$1=US$0.75=₤0.56=€0.68.31 Total costs were summed from three components: hospitalisations, outpatient services and filled prescriptions. We used case mix methodology to calculate hospitalisation costs, which involved multiplying the resource intensity weight for each inpatient admission with the average costs of hospitalisation in BC for the same fiscal year, as described previously.32 Costs of outpatient services and prescription medications were directly available in the data. Of note, most emergency department encounters (78%) were already captured by fee-for-service payments to healthcare practitioners within the outpatient encounter database (eg, billing codes with emergency visits/consultation and 1-day onsite hospital visits), and the rest were captured within hospitalisation costs.33

Based on the most responsible diagnosis code recorded on the billing claim of each healthcare encounter, we separated multimorbidity-attributable costs to either asthma or to one of the 16 major disease categories of comorbidities. These categories were defined according to the highest level classification (the first letter) in the ICD-10 codes. As for the healthcare encounters that were associated with the presence of symptoms (ICD-10 code range R00-R69), we assigned their costs to the corresponding ICD-10 comorbid areas based on the Diagnosis-Related Groups (part of the ICD system). ICD-9 codes, being used in the entire outpatient databases and up to 2002 in the inpatient databases, were converted to ICD-10 codes using validated cross-walk algorithm.34 Medication records were assigned into corresponding comorbidity categories by mapping the American Hospital Formulary Service (AHFS) major categories (available in the data) to ICD-10 categories35 (online appendix 4). Inpatient, outpatient or medication records that could not be assigned to any multimorbidity category (eg, injury, poisoning, burns and other external causes, lab tests, miscellaneous drugs) or when they had undefined ICD or AHFS codes, were grouped into an ‘unattributable costs’ category.

Statistical analysis

The unit of observation was each 12 month period following the index date. To avoid biased estimates of incremental costs due to temporary absence of individuals from the Province, periods with fewer than 300 days of registration were removed unless death occurred within the period.

Descriptive analyses were performed on the distribution of matching variables and additional selected confounders at baseline and the prevalence of comorbidity diagnoses during follow-up period across study cohorts. In the adjusted analyses, we used generalised linear models to estimate the differences in healthcare costs between patients with severe asthma versus non-asthma matches and between severe asthma versus non-severe asthma patients. The exposure variable was the cohort membership (no asthma, non-severe asthma, severe asthma). Estimates were reported as averaged incremental costs per person-year. To account for the potential residual imbalances in matching, we further adjusted the model for age, sex, rural residence and neighbourhood household income, all measured at the index year. To control for the effect of pre-existing conditions (prior to asthma incidence), we further adjusted for the number of all-cause hospitalisations, physician visits, medication dispensations and Charlson Comorbidity Index, all observed in the 12 months before the index date. Normal distribution and identity link were used in the estimation. The assumption of the normality of the mean is justified in this context given its robustness with large sample sizes,36 and was used in previous asthma cost studies.37 38 Generalised estimating equations, with an independent working correlation matrix and a robust sandwich estimator, were applied to obtain robust inference for the clustered time series data (given the clustering of observation units within individuals). Separate regression analyses were performed for the all-cause and condition-attributable costs and by cost components. All analyses were performed using SAS 9.3 (SAS Institute, Cary, North Carolina, USA).