Health care costs
Direct health care costs, expressed in Canadian dollars, were estimated for each LTC resident from the date of LTC admission until death, discharge, or the end of the 365-day study observation period. The total accumulated cost for each LTC resident was the sum of costs for hospitalization (both inpatient and outpatient [i.e., day surgery]), general and specialist physician services, LTC residence, and dispensed prescription drugs. All costs were adjusted for inflation using 2015 dollars and the Saskatchewan health care component of the Consumer Price Index (CPI) from the Canadian Socio-Economic Information Management System (CANSIM [
26];). Costs were also cumulated for the 365-day period prior to LTC admission and were used to control for confounding in the statistical models.
We estimated the total cost using both micro-costing and macro-costing approaches [
27,
28]. The micro-costing approach, which combines appropriate unit costs with person-level utilization data (quantity x unit cost), was adopted to calculate the cost of physician services and prescription drugs. The macro-costing approach, commonly referred to as standard (or average) cost per service provided, was adopted to calculate the cost of inpatient hospitalizations, outpatient hospitalizations, and LTC stays.
The total hospital cost for each LTC resident was calculated by multiplying the cost of a standard hospital stay with its Resource Intensity Weight (RIW). The cost of a standard hospital stay for day surgery or inpatient hospital stay is the province-specific average direct cost per weighted case, which measures the province’s average total cost of treating an average acute inpatient for a given fiscal year [
29]. We used the same values for day surgeries and inpatient stays, which may have resulted in some overestimation of costs. Nonetheless, this should not be a major concern since day surgeries accounted for only 9% of the total number of hospitalization records in our data. The RIW measures the intensity of resource utilization associated with different diagnostic procedures, surgical procedures, and demographic characteristics of an individual [
30]. RIWs are assigned according to Case-Mix Groups (CMG), a patient classification algorithm developed by the Canadian Institute for Health Information, a national not-for-profit organization that provides information on the Canadian health care system [
31]. CMGs classify patients into homogeneous groups based on similar clinical and resource-utilization characteristics. It should be noted that our hospital cost estimates may underestimate the total cost of hospitalization as it excludes hospital operational and building capital costs (e.g., hospital administration, utilities, or other capital-related costs) as well as visits to ambulatory clinics such as emergency departments.
The cost of physician services for each LTC resident was calculated by summing the fees for all physician visits during the study period. These costs were calculated separately for general practitioners and specialist physicians. In the case of non-fee-for-service physician payments, shadow billing claims were used to compute physician costs using the equivalent fee-for-service value. Some non-fee-for-service physicians may not submit shadow-billed claims. However, existing research supports the validity and completeness of the physician claims data for capturing the vast majority of physician services [
25,
32].
The prescription drug cost for each LTC resident was the total cost for all dispensed outpatient prescriptions; this included drug materials, dispensing fee, and markup. Costs for drugs used in acute care facilities are captured in the total hospital cost.
The cost of LTC services for each LTC resident was determined by multiplying the annual per-diem cost by the number of days a member of the study cohort was an LTC resident. The annual per diem cost was calculated by dividing the same value of annual LTC provincial expenditures for all facilities by the total number of resident days. This information was obtained from the Ministry of Health Community Care Branch.
Covariates
The primary covariate of interest was pain group, which had two categories: CSP and NP/NDMP. This measure was derived from RAI-MDS assessment records, which captures information on pain frequency (i.e., no pain, pain less than daily, daily pain) and its intensity (i.e., mild pain, moderate pain, pain is horrible or excruciating) at the index date and quarterly thereafter. The RAI-MDS pain scale has been widely used for Canadian LTC residents [
33] and its validity has been established [
34]. Pain frequency and severity, stratified by sex is available in the
Additional file accompanying this article. Similar to Guliani et al. [
18], LTC residents with daily pain or less than daily pain that was of moderate or severe intensity were classified as having CSP. LTC residents with no pain or less than daily pain of mild intensity were classified as having NP/NDMP. Our approach to classification of patients in terms of their RAI-MDS pain status is consistent with quality indicator approaches used in Canada [
35].
We also included the RAI-MDS Cognitive Impairment Scale (CPS [
36]), Activities of Daily Living (ADL [
36]), and Changes in Health, End-Stage Disease and Signs and Symptoms (CHESS [
36]) as covariates in our regression analysis; these covariates measure residents’ physical and cognitive impairment. The CPS evaluates the level of cognitive impairment affecting a patient such as memory, decision-making ability, communication skills and eating impairments, and is scored on a seven-point scale with a minimum value of zero and a maximum value of six. Higher scores indicate more severe cognitive impairment [
36]. We used the ADL long-form scale to measure residents’ self-sufficiency. The scale is composed of seven ADL items; total scale scores range from zero to 28, with higher scores indicating less independence [
36,
37]. In our study, we categorized the ADL scores as 0–7, 8–14, 15–21 and 22–28. The CHESS scale is a measure of frailty and health stability designed to identify residents at risk of serious decline and takes on values from zero (no instability) to five (high instability). We have combined scores from three to five due to a small number of observations. A high CHESS score is shown to be a predictive of mortality in the LTC population [
38]. Further details on these measures are provided by the Canadian Institute of Health Information [
36].
Other covariates included residents’ demographic characteristics at the index date (age, sex, rural/urban residence), comorbidities, prior direct health care costs for the 365-days before the index date, and LTC facility characteristics. Age was categorized as 65–74, 75–79, 80–84, 85–89, 90–94, and 95+ years. A resident’s address from the PHRS at index date was used to classify the resident’s location as rural or urban. An address from one of the two health regions containing major urban centers was classified as urban, while an address from one of the remaining 11 health regions was classified as rural.
The Charlson comorbidity index (CCI [
39];) score was calculated for the 365-day period prior to the index date. The index is based on ICD–9 and ICD–10-CA diagnosis codes captured in hospital and physician data [
40,
41]. CCI scores were classified into four groups:, 0 (no comorbidities), 1, 2, and 3 or more comorbidities.
The natural log of total direct health care cost in the 365-day period prior to the index date was included as a covariate. The total health care cost was the sum of hospitalization, general practitioner visits, specialist visits, and prescription drug cost. These costs were adjusted for inflation using 2015 dollars and CPI for Saskatchewan health care component of CANSIM.
The Saskatchewan ISCH database classifies LTC facilities as integrated, special care homes, and hospital-based special care homes. Integrated LTC facilities provide the services of both an acute care facility and a special care home [
42]. Special care homes are publicly funded and provide LTC services to qualified residents who require care and supervision that cannot be provided in their own homes [
42]. Hospital-based special care homes serve individuals who are still in a hospital but do not require acute care services. The facility type variable allows us to capture any unknown and unmeasured variations in the services that facilities provide.
Statistical analysis
Frequencies, percentages, means, standard deviations (SD), and the interquartile range (IQR) were used to describe the cost data. Descriptive statistics were stratified by pain group at baseline (i.e., index date).
Generalized linear models (GLMs) with generalized estimating equations (GEEs) were used to test for differences in the marginal total direct health care costs between the CSP and NP/NDMP groups before and after adjusting for other covariates (i.e., unadjusted and adjusted models). GLM models are widely used for modelling health care costs [
43,
44]. GLM with GEEs allowed us to account for the repeated measurement of costs within patients over time and to make robust inferences by implicitly considering the correlation structure of the data. The assumption of a normal distribution for the outcome, which underlies ordinary least-squares regression, is violated for cost data because it has a skewed distribution. We initially analyzed the data using GLMs with gamma and inverse Gaussian distributions and used a likelihood-based goodness-of-fit statistic to assess which distribution provided the best fit for the data. Based on the assessment of model fit, we selected a gamma distribution and an identity link function.
Pain group was defined as a time-varying covariate; it was updated each time a resident’s pain status changed based on information recorded during the RAI-MDS assessments. We used a time-varying covariate because pain status is unlikely to remain constant over time.
We first fit an unadjusted model to the cost data; it contained pain group, time, and the pain group x time two-way interaction. The two-way interaction enabled us to test whether the association of pain group with cost varied over time. We then fit a fully adjusted model that incorporated all patient and facility characteristics in addition to pain group, time, and the pain group x time two-way interaction.
Two adjusted and unadjusted models were fit to the data: In Model 1, the outcome variable was total direct health care cost including the LTC component. In Model 2 the outcome variable was total direct health care cost excluding the LTC component. By removing the LTC facility costs, we were able to analyze the impact of pain on the remaining health care services cost.
An exchangeable correlation structure, which assumes constant correlation in successive periods for the repeated measurements was used. The model offset was the natural log of the number of person-days of follow up (the number of days in the LTC) which accounts for differences in exposure time (i.e., due to censoring). GLM model fit was evaluated using the ratio of scaled deviance to the model degrees of freedom without considering correlation; a value close to one indicates a well-fitting model. A penalized quasi-likelihood fit statistic was used to evaluate fit when the correlation structure was taken into account. Regression coefficient estimates and 95% confidence intervals (CIs) are reported. SAS v.9.4 [
45] was used to perform both the descriptive and inferential analyses.