Excess medical costs in patients with asthma and the role of comorbidity

Data sources

We retrieved data from January 1997 to December 2012 from the provincial health administrative databases of British Columbia (BC), a large province in Canada with 4.4 million residents as of 2011 [13]. These databases provide linked, individual-level information on demographics, vital statistics and healthcare encounters of all BC residents [14–18] (supplementary material S1). 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). Ethical approval was obtained from the University of British Columbia Human Ethics Board (H08–01287).

Study design and sample

This was a retrospective, propensity score-matched cohort study. The asthma cohort consisted of individuals newly identified with asthma. We first identified all individuals who satisfied a validated case definition of asthma [19], defined as the presence of, during any rolling 12-month period, one or more asthma-specific inpatient encounters (with the most responsible diagnosis being asthma), two or more asthma-related outpatient visits (in different dates), or three or more filled asthma-related prescriptions. Asthma-specific inpatient and outpatient encounters were determined based on International Classification of Diseases (ICD) codes (ICD-9: 493.x; ICD10: J45.x, J46.x). Asthma-related medications were pre-specified (supplementary material S2). The index date was defined as the date of the patient's first asthma-related inpatient or outpatient service based on their history of resource use, marking the beginning of follow-up. The cohort included individuals aged between 5 and 55 years at their index dates. To include only new patients, we restricted the sample to individuals who were registered with the healthcare system for at least 300 days per year (i.e. fully registered) for at least 5 years prior to the index date without having any asthma-related records. These individuals also had to have at least 1 year of follow-up data after the index date.

Because asthma medications significantly overlap with treatment for chronic obstructive pulmonary disease (COPD), to minimise the chance of including patients with COPD who received the same drugs, we excluded patients with any COPD-specific healthcare resource usage within 2 years of the asthma index date (COPD codes: ICD-9: 491.x, 492.x, 493.2x, 496.x; ICD-10: J43.x, J44.x). Patients with asthma can develop COPD [20], which is a legitimate comorbidity to consider in this study. Therefore, COPD-related records beyond the first 2 years after the asthma index date were included.

Propensity score matching

As the basis of our comparator group, we had access to a random sample of 100 000 BC residents who had no asthma-specific health resource use between January 1997 and December 2012. From this sample we created a comparison group. The BC Ministry of Health does not permit researchers to access the full provincial population due to privacy considerations.

We used propensity score matching to create a balanced 1:1 cohort of asthmatic and non-asthmatic subjects, as follows: First, we excluded non-asthmatic subjects with less than 5 years of coverage in the database, and defined the first day following the five continuous years of coverage for each remaining individual as their index date (to match the 5-year pre-index criterion applied to the asthma cohort). Comparison subjects also had to be between 5 and 55 years of age on their index date. We then merged the asthma and comparison cohorts, and fitted a logistic regression model predicting the probability that an individual had asthma. This model included age, sex, neighbourhood household income quintile, health services delivery area (all as observed in the index year) and baseline comorbidity status (observed in the 12 months before the index date). The latter was measured by the number of non-asthma-related hospitalisations, physician visits and medication dispensations, as well as a Charlson comorbidity index (CCI, excluding asthma from the score) [21]. Finally, we used the estimated propensity scores to perform matching based on nearest distance criterion, so that each member of the asthma cohort was matched to a member of the comparison cohort. Because the size of the pre-matching asthma cohort was larger than the comparison cohort (145 744 versus 50 110, see supplementary material S3), we adopted the matching-with-replacement strategy (in which an individual in the comparison cohort could be matched to more than one patient in the asthma cohort). This approach allowed us to efficiently use the comparison cohort subject and has proven to yield results comparable to the conventional matching-without-replacement method [22].

All study subjects were followed from the index date to date of death, loss to follow-up (i.e. end of registration with the healthcare system), or December 31, 2012, whichever came first. Supplementary material S4 provides a schematic presentation of the study design.

Outcomes

Our primary outcome was all-cause excess costs, defined as the difference in all-cause direct medical costs between the asthma and comparison cohorts. Costs were summed from three components, comprising inpatient encounters, outpatient encounters and filled prescriptions, and were converted to 2013 Canadian dollars ($) using the historical inflation rates [23]. Inpatient costs were calculated using the case mix methodology by multiplying the resource intensity weight (available for each hospitalisation record) with the average costs of hospitalisation in BC for the same fiscal year [24]. Outpatient and medication costs were directly available in the data. Costs of emergency department services were mostly captured by fee-for-service payments to physicians, with those leading to an inpatient episode captured by the corresponding inpatient records [25].

The secondary outcomes were condition-attributable excess costs, which were calculated by attributing resource use records to asthma, or to any of the 16 major comorbid areas as defined by the major disease categories in the ICD-10 system [26]. Resource uses attributable to the presence of symptoms categorised by the ICD-10 code range R00–R69 were assigned to related comorbid categories if the diagnosis code was grouped into that category in the Diagnosis-Related Groups based on General Equivalence Mappings [27]. ICD-9 codes, being used in the outpatient databases and up to 2002 in the inpatient databases, were converted to ICD-10 codes using validated cross-walk tables [28]. For medication records, we mapped the American Hospital Formulary Service (AHFS) major categories (available in the data) to ICD-10 categories [29] (supplementary material S5). Some records of resource use and medication dispensations could not be assigned to any disease or symptom categories (e.g. codes for injury, poisoning, burns and other external causes, laboratory tests, miscellaneous drugs, or when there were no ICD or AHFS codes). We grouped costs associated with such records into an “unattributable” category.

Statistical analysis

Descriptive analyses were performed using SAS 9.3 (SAS Institute Inc., Cary, NC, USA). Regression analyses were performed using Stata/IC (V.12.1. College Station, TX, USA). Differences in the distribution of matching variables between asthma and comparison cohorts were compared using the standardised difference, with a value below 0.20 indicating similarity between two groups [30].

The unit of analysis was person-year (PY). Follow-up time was divided into 12-month intervals starting from the index date, with the last interval potentially truncated. To avoid under-reporting due to individuals with temporary absences from the province, intervals with less than 300 days were removed unless death occurred. We used a multivariate general linear model with interval-specific costs as the dependent variable. Normal distribution and identity link were used, in line with expert recommendations on their robustness with large sample sizes [31]. Separate regression analyses were performed for the overall costs and by cost components. For all models, the independent variables included disease status (asthma=1, non-asthma=0), age group at the beginning of the interval (5–18 years, 19–45 years, 46–55 years), and the number of intervals since the index date. The models also included calendar year at the beginning of each interval.

Excess costs were estimated as the adjusted differences in the predicted costs between an asthmatic patient and the non-asthmatic match. Changes over age and time were captured by the regression coefficients for the three-way interaction between asthma status, age group and time from the index date. Because individuals in the comparison cohort could be matched to several individuals in the asthma cohort, our data were clustered around each comparator and all potential matches. Accordingly, we fitted a weighted regression as recommended [22], with comparison subjects weighted by the inverse of the frequency that they were matched to an asthmatic patient. We followed expert recommendation to use generalised estimating equations with the robust variance estimator to obtain valid inference for the nested clustered data [32].

Sensitivity analysis

To evaluate the robustness of our results against design features, we performed several sensitivity analyses. First, we repeated the cost analyses among individuals with short follow-up time (<12 months) who were excluded from the main analysis, adjusting for the length of follow-up in days. Second, to further minimise the risk of including patients with COPD, we estimated costs by removing all patient-years in which patients were over 45 years of age.