Geographic variation in the costs of medical care for people living with HIV in British Columbia, Canada

We conducted a retrospective cohort study using linked, provincial health administrative databases and disease registries, including the BC Centre for Excellence in HIV/AIDS drug treatment program and virology registries (antiretroviral dispensations, plasma viral load (pVL) and CD4 tests) [20], the BC Centre for Disease Control HIV testing database (HIV diagnosis/risk group classification) [21], the Medical Services Plan (MSP) database (physician billing records) [22], the Discharge Abstract database (hospitalizations) [23], the BC PharmaNet database (non-antiretroviral drug dispensations) [24], and the BC Vital statistics database (deaths) [25]. Our data set comprised all HIV positive individuals identified within these databases, observed for at least two quarters between January 1, 2010 and March 31, 2014. Details regarding databases and cohort construction have been published elsewhere [17]. Our data included quarterly observations from HIV diagnosis to death, administrative loss to follow-up (no record of health service use for at least 18 months), or censorship as of March 31, 2014. We chose quarterly observations according to CD4 and pVL monitoring guidelines from the International Antiviral Society–USA (IAS-USA) [26]. The health care setting featured government-funded, single-payer care, covering inpatient and outpatient care, selected prescription drugs, laboratory monitoring, and antiretroviral medications.

Our primary outcome variable was the quarterly cost of medical care; calculated as the sum of inpatient, outpatient (physician fee-for-service billing claims), prescription drug costs (ART and non-ART), and diagnostic testing costs for each individual in our cohort. We estimated medical costs for PLHIV including the costs of ART medication, excluding ART medication, as well as costs of ART medication only for those receiving ART. We presented each of these analyses separately, as differential rates of ART uptake across HSDAs, combined with the high relative costs of antiretroviral medication, could potentially obscure differences in health care costs that indicated divergence from clinical practice standards or higher intensity of service provision among PLHIV. The methods for deriving these costs have been described previously [27, 28]. We adjusted all costs to 2018 Canadian Dollars (CAD), using the Canadian Consumer Price Index.

To support our primary analysis, and to provide further context as to which health care components showed greater regional differences among PLHIV, we estimated costs and utilization rates for hospitalizations, physician billing records and non-ART prescription drugs, respectively. We defined utilization outcomes as hospitalization days (days with an inpatient record), physician billing days (number of days in which an individual had at least one physician billing contact) - chosen to reflect the frequency of contacts rather than the number of individual MSP billing line items - and drug days prescribed (the total of all prescription days in a given quarter).

Our primary independent variables were a set of quarterly indicator variables for an individual’s primary HSDA. We assigned one representative HSDA to each person in each quarter, depending on where an individual accessed health care services. We assigned time-varying indicators of HSDA due to high rates of intra-provincial migration for treatment among PLHIV [29, 30]. When an individual accessed health services in more than one HSDA in a quarter, we selected the HSDA in which an individual recorded the highest frequency of health care contacts as the representative HSDA.

We adjusted for a range of clinical factors, including measures of HIV disease progression (CD4 < 200 cells/μL; 200–499 cells/μL; ≥500 cells/μL; unmeasured) and ART status (ART-naïve, on-ART, and off-ART post-initiation (ART-dropout)). Additionally, we included the area under the log₁₀ plasma viral load curve (AUC pVL) for 12-months prior to the start of the quarter (< 2.7; 2.7–2.99; 3–3.49; ≥3.5; unmeasured) to capture individuals’ cumulative viral load over the previous 12-months. CD4 counts were taken from the most recent test prior to the start of the quarter, with previous test results carried forward for any missing observations, otherwise CD4 count was classified as unmeasured. AUC pVL measurements were also carried forward until the next non-missing observation, or the end of follow-up, otherwise AUC pVL was classified as unmeasured.

Demographic control variables included: age, sex, calendar year, transmission risk group, including: men who have sex with men (MSM); people who inject drugs (PWID); MSM who inject drugs (MWID); heterosexual/other (HET), year of diagnosis, grouped as: pre-1996; 1997–99; 2000–03; 2004–07; 2008–14, and two separate measures of medical comorbidity, the drug prescription-based Chronic Disease Score (CDS) [31], and the Charlson Comorbidity Index (CCI) [32], based on hospital records, calculated as a moving average for the year prior to quarter start date. Finally, to account for individuals potentially re-locating for specialized, high-cost health care services, we included a variable indicating if an individual had moved in the previous 12 months, based on changes to annually updated MSP-registered billing addresses.

To estimate quarterly health care costs, we fit a two-part generalized linear model (GLM) with logit link and binomial distribution for the probability of an individual incurring any costs, and a log link with gamma distribution for observations with non-zero costs [33]. We estimated costs per person-quarter by HSDA, holding all other covariates fixed at their overall sample means.

We estimated utilization rates for each component, using a two-part generalized linear model (GLM) with logit link and binomial distribution for the probability of an individual having any utilization, and a log link and negative binomial distribution for observations with non-zero utilization.

We repeated our primary analysis using representative HSDAs assigned by highest cost, rather than highest frequency of visits, in a given quarter. We conducted additional sensitivity analysis removing costs for services received through the Provincial Health Services Authority (which provides specialized health services to the entire province), which were otherwise assigned to an individual’s representative HSDA, to determine if estimates for some HSDAs were disproportionately affected.

We created our analytical sample in SAS 9.4, and conducted statistical analysis in Stata 14.1.

This analysis had several limitations. First, HSDA assignment was subject to some degree of misclassification due to the possibility of individuals receiving care in multiple HSDAs in a given quarter or missing information on HSDA, however, our results were robust to alternative HSDA assignment methods in sensitivity analysis (Additional file 1: Table S4). Second, PHSA costs were assigned to an individual’s representative HSDA, and could potentially have affected some regions disproportionately, however, our HSDA-specific cost estimates were robust to the removal of these costs (Additional file 1: Table S4). Furthermore, due to a small number of observations, results from the East Kootenay HSDA should be interpreted with caution. Finally, as with all non-experimental studies, we cannot rule out the possibility that our coefficient estimates may have been subject to some degree of bias from unmeasured confounding factors [49]. It is likely that some degree of the variation in regional costs could be attributable to incomplete case-mix adjustment. While our study included detailed information on HIV treatment and disease progression, it is possible that the prevalence and severity of comorbid conditions were not fully accounted for. Finally, our goal was to estimate variation among HSDAs in costs and utilization rates, and we did not examine the association between health care costs and quality of care, clinical outcomes or long-term spending by region. Further investigation is needed into the association between utilization rates of different health care components, particularly in rural and remote regions, to determine if lack of access, socioeconomic factors or other barriers to care are leading to higher hospitalization utilization and costs, or other adverse outcomes.