Excess economic burden of comorbidities in COPD: a 15-year population-based study

Data source

We used the provincial health administrative databases of British Columbia, Canada (with 4.4 million residents as of 2011 [12]) between January 1, 1996 and December 31, 2012 (University of British Columbia Human Ethics Certificate H13-00684). These databases provided linked, individual-level information on healthcare encounters of all legal British Columbia residents, including inpatient and outpatient services, dispensed prescription medications, community care services, demographics, vital statistics, and registration status (details of these databases are provided in supplementary appendix S1) [13–18]. 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).

Study design and sample

This study has been registered in the European Network of Centres for Pharmacoepidemiology and Pharmacovigilance (EUPAS18242). We used a matched longitudinal study design by creating a cohort of patients with newly diagnosed COPD, and comparing their health services and medication utilisation to a comparable cohort of non-COPD individuals over the same period. A schematic presentation of the study design is provided in supplementary appendix S2.

The COPD cohort included individuals who satisfied a validated case definition of COPD [19], defined as the presence of one or more hospitalisations or two or more outpatient visits on different dates, with COPD as the most responsible diagnosis (International Classification of Diseases (ICD)-9: 491.xx, 492.xx, 493.2x, 496.xx; ICD-10: J43.xx, J44.xx) during any rolling 12-month period. This case definition, validated against chart reviews, has high sensitivity and specificity [19], and has been used in previous studies [20]. Once identified, we defined the index date as the next day after the patient first received a diagnosis of COPD (the choice of the “next” day was to avoid bias due to guaranteed resource use on the index date for the COPD cohort, which is not necessarily the case for the comparison group). The index date marked the beginning of the follow-up period. To ensure that we identified incident cases, we included only those patients who had been fully registered in the databases for at least 3 consecutive years prior to the index date. Patients also had to be registered for at least 1 year following the index date. To further improve the specificity of the COPD case definition, we restricted our analysis to patients who were aged ≥45 years on their index date.

The non-COPD cohort was obtained as a random sample from the general population who had never been diagnosed with COPD and were matched to COPD patients with regard to sex, birth year, health service delivery area and neighbourhood income quintiles. Unlike in COPD patients for whom the index date was a specified date, the index date for the non-COPD cohort represented a random date during the study period. To maximise the possibility that COPD patients would find a match in the comparison group, we randomly assigned three index dates from the COPD cohort to each non-COPD person so that each potential control subject could have three chances to be matched to COPD patients. To create a cohort comparable to the COPD cohort, we excluded the index dates on which the individual in the comparison cohort was not fully registered in the databases for at least 3 years or was aged <45 years.

We then performed a multivariable logistic regression to calculate a propensity score, which was the probability that an individual belonged to the COPD cohort at baseline. Variables used for propensity score calculation included sociodemographic indicators as measured in the index year (age, sex, neighbourhood household income, health services delivery area), comorbidity status as measured in the 12 months prior to the index date (Charlson comorbidity index (CCI, excluding COPD), number of non-COPD hospitalisations, outpatient visits and medications), and the calendar year and season of the index date. Using a nearest-neighbourhood matching algorithm, we matched each person in the COPD cohort to a distinct person in the non-COPD cohort, with differences in their propensity score values of ≤0.01.

Outcome variables

All costs were inflated to 2015 Canadian dollars using historical inflation rates that were obtained from the Consumer Price Index for British Columbia [21] and were converted to 2015 Euros (CAD1.000=EUR0.706) [22]. The total medical costs were derived from four components: costs of inpatient encounters, outpatient encounters, filled prescriptions and use of community services (the latter included care at long-term care facilities, assisted living facilities, family care homes and group homes, adult day-care programmes, home care, and home supports). Using a case-mix methodology, inpatient costs were calculated by multiplying the resource intensity weight of each hospitalisation episode with the mean costs of hospitalisation in British Columbia during that fiscal year [23]. Costs of outpatient services, medications and community care were directly summed from the data. Emergency department visits were determined from the billing codes identifying emergency consultations and 1-day onsite hospital visits, and their costs were either captured by fee-for-service payments to physicians or by the corresponding inpatient records for those leading to an inpatient episode (22% of all visits) [24].

All-cause medical costs were divided into costs attributable to COPD or comorbid conditions. COPD-attributable costs were calculated by attributing resource use records to a primary diagnosis of COPD. Comorbidity-attributable costs were calculated in the same fashion, with resource usage attributed to any of the 16 major comorbid areas in the ICD-10 system [25]. Some resource use records were associated with diagnoses of symptoms (ICD-10 code range R00–R69), which were also mapped to related comorbid categories. ICD-9 codes were used in the outpatient databases throughout the study period and in the inpatient databases up to 2002 (after which ICD-10 codes were used). We thus applied validated crosswalk tables to convert them into ICD-10 codes [26]. Medication records were also mapped to ICD-10 categories from the American Hospital Formulary Service (AHFS) major categories (supplementary appendix S3) [27]. For resource use records and medications that could not be associated with any disease or specified symptoms as described above (e.g. codes for injury, poisoning, burns and other external causes, and lab tests), as well as for miscellaneous ICD or AHFS codes, we grouped their related costs into a “nonattributable” category. Of note, community care costs were also grouped under the nonattributable category because the billing records were not associated with any ICD code.

Statistical analysis

All analyses were performed using SAS version 9.4 (SAS Institute, Cary, NC, USA). We compared the distribution of matching variables between COPD and non-COPD cohorts by visualising the distributions and by calculating the standardised difference for each variable (similarity defined as difference ≤0.10) [28].

Excess costs were estimated as the adjusted difference in predicted costs of a COPD patient and their matched non-COPD subject. We estimated excess costs across conditions (COPD and 16 comorbid categories), as well as across specific cost components (hospitalisation, outpatient services, medications and community care). The unit of observation was person-year starting from the index date. To avoid underreporting bias when individuals were temporarily absent from the province (e.g. travelling outside of the province), we excluded person-years with <300 days of follow-up (unless death occurred in the year). Differences in excess costs across baseline age groups (45–54, 55–64 and ≥65 years), sex and baseline comorbidity burden (none, CCI score 0; mild, CCI score 1, moderate, CCI score 2; high, CCI score ≥3) were examined using the regression coefficients of interaction terms between COPD and these three indicators. Since the first year of costs might be higher than subsequent years due to the acute treatment of COPD, we included a binary indicator for the first year.

To accommodate zero-inflated and highly right-skewed cost data, we used a two-part generalised linear model. The first part was a logistic regression model to calculate the probability of incurring any costs. The second part was a gamma regression model with a logarithmic link to estimate the value of costs among patient-years with nonzero costs, which has been shown to perform well in the estimation of population-averaged healthcare costs [29]. The overall per-person-year costs were estimated as the probability of incurring any costs from the first part of model multiplied by the predicted costs for nonzero costs from the second part. Generalised Estimating Equations were applied to obtain valid inference for the clustered data around each matched pair. The adjusted effects of baseline risk factors on excess costs were estimated by combining the two parts using G-computation [30]. Inference (calculating p-values and confidence intervals) was performed with 50 rounds of bootstrapping. A detailed description of the estimation is provided in supplementary appendix S4.