All Ontario residents (13.6 million people in 2013) are covered by the Ontario Health Insurance Plan (OHIP) for medically necessary services. Our cohort population for this study consisted of a 10% random sample of the adult population who had a valid health care card at the index date (April 1, 2012). Individuals were excluded from the study sample if they incurred zero primary care costs (meaning that they did not use physician services) or if they died during the study period (April 1 2012 to March 21st 2013).
Data source and variables
Encrypted data for this cross-sectional study were obtained from administrative databases at the Institute for Clinical Evaluative Sciences (ICES). The study received ethics approval from the Research Ethics Boards of the University of Toronto and the Sunnybrook Health Sciences Centre. The sample was extracted from the adult population in the Ontario Registered Persons Database (RPDB) which contains basic demographic information on individuals. Using unique patient identifiers, the ICES Key Numbers (IKN), patients’ enrolment data at the beginning of the study period from the Client Agency Program Enrolment (CAPE) database were used to link patients to their primary care physicians and to the corresponding payment model to which the physician belonged. Patients who were not formally enrolled with a physician were considered as a separate group – the not enrolled. The IKN were also used to retrieve each individual’s health services utilization. All health care services paid for by the MOHLTC were included: physician services were extracted from the OHIP database; general hospital services from the Canadian Institute for Health Information Discharge Abstract Database, the Same Day Surgery Database, and the National Ambulatory Care Reporting System; drugs for adults on social assistance and for people aged 65 and over from the Ontario Drug Benefit Plan database; home care services from the Home Care database; lab and non-physician billings from the OHIP. Utilization from long term episodes of care were combined to include residents of nursing homes and of specialized hospitals with data from the Complex Continuing Care database, the Ontario Mental Health Reporting System database and the Long Term Care database.
The two outcome variables were primary care costs (PCC) and total health care costs (THCC).
Costs were calculated for each individual in the study population based on the individual’s utilization of health care services during the study period and the prices of services, as paid by the MOHLTC. Algorithms for the prices of health care services have been previously developed by a team of Ontario researchers and implemented at ICES [
16]. Primary care costs were composed of the prices and quantity of each service billed by primary care physicians to OHIP as well as capitation costs, and shadow billing costs. The capitation rates depended on the primary care model that the physician belonged to.
Total health care costs were calculated based on the utilization of health care services that each individual made. They included services paid for on a fee schedule as well as institutional care. Costs for institutional care were adjusted for the resource intensity weight of the care setting. In the case of long term care, costs were based on a per diem fee paid by the MOHLTC.
The costs associated with the payments for the establishment and operations of FHTs, as well as performance payments to primary care physicians were not available for inclusion at the individual patient level.
The independent variables of interest were the primary care models that physicians can belong to: CCM, FHG, FHN, FHO, FHT-FHN, FHT-FHO, using the FFS as the reference group.
Since patients may be differentially distributed across models [
17], the statistical models controlled for patient characteristics that could affect health care utilization and costs. Explanatory variables included patient’s age (continuous variable), sex (dichotomous variable), and socio-economic status, using the income quintile as a proxy (based on postal codes with the lowest quintile (1) as the reference group), the Adjusted Clinical Group (ACG®) weight, and the Rurality Index of Ontario (RIO) score for the practice location of the primary care physician.
The ACG® system is a measure of case-mix developed by the Johns Hopkins University to reflect patients’ health care needs based on the combination of their diagnostics, age and sex [
18]. The ACG® weight ranges from 0.000 to 4.666, with a higher weight representing a poorer health status. There are only a few scores over 1.000, and these reflect a high level of complexity with a combination of: at least 2 major adjusted diagnostic groups (ADGs); at least 6 other ADG combinations; and an age over 34. The algorithm accounts for the mix of diagnosis over a defined period of time and across health care settings. The ACG® has been tested and it was validated as a predictor of utilization and mortality in Canada [
19‐
21]. A recent review of various morbidity measures suggests that the ACG® system is the strongest in predicting health care utilization [
22]. The RIO is a continuous variable that takes a value between 0 and 100, with lower values indicating an urban location and it adjusts for the geographic location of the primary care practice. The RIO is a measure that was developed by the Ontario Medical Association for Ontario communities [
23]. The RIO includes the following 10 variables: travel time to nearest basic referral center, travel time to nearest advanced referral center, community population, number of active general practitioners (GP), population-to-GP ratio, presence of a hospital, availability of ambulance services, social indicators, weather conditions, and selected services.
Statistical Analysis
There are multiple approaches to the analysis of cost data, with the most common being the OLS with a log transformation and the generalized linear model [GLM] [
24‐
28]. Although the OLS with log transformation is widely known and used, it does not eliminate heteroscedasticity and the retransformation could lead to bias. The GLM is a preferred approach in the presence of heteroscedasticity in a statistical model with multiple covariates. The GLM was selected for this study after the Breusch-Pagan test found heteroscedasticity and because of other advantages of the GLM. The GLM takes into account heteroscedasticity, and does not require retransformation so as to express estimates in dollar amounts and accommodates skewness, which is typical of health care cost data [
29‐
31]. This method does require specifying a distribution for the mean-variance relationship and a link function. Gaussian, Poisson, Gamma and inverse Gaussian distributions for the mean-variance relationship were tested with the modified Park test [
24,
30]. The results reported here are based on the GLM with an identity link and a Gaussian family based on the results from the Park test.
Two regressions were conducted: the first regression modelled patient primary care costs, and dummy variables for the models; the second analysis examined total health care costs as a function of the primary care model of the primary care physician and also of the other explanatory variables identified.
Regression models for both primary care and total health care costs were defined as:
$$ {\mathrm{Cost}}_{\mathrm{i}}={\upbeta}_0+{\upbeta}_1\ {CCM}_{\mathrm{i}}+{\upbeta}_2\ {FHG}_{\mathrm{i}}+{\upbeta}_3\ {FHN}_{\mathrm{i}}+{\upbeta}_4{FHO}_{\mathrm{i}}+{\upbeta}_5\ FHT-{FHN}_{\mathrm{i}}+{\upbeta}_6\ FHT-{FHO}_{\mathrm{i}}+{\upbeta}_7{RIO}_{\mathrm{i}}+{\upbeta}_8\ {\mathrm{ACGweight}}_{\mathrm{i}}+{\upbeta}_9\ {\mathrm{age}}_{\mathrm{i}}+{\upbeta}_{10}\ {\mathrm{sex}}_{\mathrm{i}}+{\upbeta}_7\ \mathrm{income}\_{\mathrm{quintile}}_{\mathrm{i}}+{\upvarepsilon}_{\mathrm{i}} $$
Where: Costi is either the primary care or the total health care cost of the services for patient i for a 12 month period; β0 is the intercept; CCMi, FHGi, FHNi, FHOi, FHT-FHNi,and FHT-FHOi are dichotomous variables for the primary care models, using FFS as the reference group; RIOi is the value of the RIO of the practice the patient i belongs to; agei, sexi, and income_quintilei are the adjustments for the patient’s age, sex, and neighborhood income quintile; ɛi is the error term for patient i.