Social disparities in the use of colonoscopy by primary care physicians in Ontario

We used six data sources: (1) The Ontario Cancer Registry (OCR), a registry of all Ontario residents newly diagnosed with cancer or who died from cancer since 1964, estimated to be 95% complete [8]; (2) The Canadian Institute for Health Information-Discharge Abstract Database (CIHI-DAD), which contains information on all discharges from acute care facilities for residents of Ontario dating from 1988 including clinical information on diagnoses, procedures and discharge status; (3) The Ontario Health Insurance Plan (OHIP) database, which contains information on claims for physicians' services and all medical procedures to the Ontario Ministry of Health and Long-Term care made by fee-for-service physicians, community based laboratories and radiology facilities; (4) The Registered Persons Data Base (RPDB), which contains demographic information for all residents eligible for health care in Ontario; (5) The ICES Physician Database (IPDB), which contains information about physician demographics, specialty training and practice location in Ontario; and (6) The 2001 Canadian Census files, which contain aggregated data that describe general demographic information of the Canadian population at the census tract level.

We were interested in studying colonoscopy use among patients in whom colorectal cancer screening might be considered. Patients with a previous or new diagnosis of colorectal cancer, or those in whom cancer is strongly suspected, are very likely to undergo colonoscopy; variation between PCPs in the use of colonoscopy for these patients is likely to vary little. Colonoscopy performed for diagnosis of suspected colorectal cancer, known lower gastrointestinal disease or acute presentation is not highly discretionary. However, most colonoscopies are done in patients at low risk of having colorectal cancer.

Residents of Ontario aged 50-74 years who were eligible for OHIP benefits for each calendar year from 1996-2005, had no prior diagnosis of CRC, inflammatory bowel disease (IBD), Ulcerative colitis, Crohn's disease, or colonic polyps, and who received care from a PCP in Ontario, were considered as screen-eligible patients. Patients who had colon or rectal surgery at any time prior to January 1 of each calendar year, or who had a colonoscopy for any reason in the previous 4 years were excluded. Patients residing in regions where physicians do not bill directly for services were not included (approximately 4.5% of the Ontario population). The exclusion criteria and corresponding diagnosis and service codes are shown in Additional File 1: Appendix 1.

Each patient with two or more encounters with a general practitioner/family physician (GP/FP) in Ontario was assigned to a PCP based on health services received during each calendar year from 1996-2005. For each potential patient we identified all outpatient OHIP service codes claimed by a GP/FP and calculated the number of visits made to GP/FPs. For each calendar year, we assigned each patient to a single PCP. For those patients who had visits to more than one GP/FP over the year, we used the continuity of care (COC) measure [9, 10] to determine the PCP most responsible for providing continuous care.

We used the method for estimating COC described by Bice and Boxerman [11] which can be estimated for patients who have had at least two visits to a GP/FP over the year. All visits to either a usual provider or a referred provider are attributed to a single provider (the referring provider is considered to be the 'usual provider' for the purpose of estimating the COC measure). A score of one represents perfect COC.

Based on OHIP billing codes, we identified all colonoscopies (Z555) performed on patients linked to a PCP for each calendar year from 1996-2005. For individuals who had more than one colonoscopy during the study period, only the first colonoscopy was considered. We were also interested in differentiating colonoscopy performed for the investigation or work up of a known or highly probable CRC from those performed on a more discretionary basis, such as CRC screening. It was not possible to differentiate screening colonoscopies using our data sources. Therefore, we developed an algorithm to identify colonoscopies--in retrospect--that were very likely to have been performed for CRC screening, or for another relatively discretionary indication. A proportion of these colonoscopies would even considered "unnecessary" by some endoscopists. A discretionary colonoscopy was defined as a colonoscopy procedure performed on a patient of screen-eligible age, not performed during an inpatient stay, and not associated with diagnosis of CRC at the time of colonoscopy or within a 3 year period following the colonoscopy (Additional File 1: Appendix 2). Because most referrals to specialists in this time period originated from PCPs [12], we attributed performance of the colonoscopy to the patient's PCP although colonoscopy was in general performed by a specialist physician who had seen the patient at the request of a PCP.

We collected information from the IPDB on PCP characteristics such as age, sex, number of years in practice, education (Canadian training vs. International Medical Graduate), and rurality of the area of practice. Information on characteristics of patients treated by PCPs was derived both at the patient level (age, sex and comorbidity) and at the neighborhood level (income, rurality, level of education, employment, immigration status and language fluency). We determined co-morbid conditions for subjects using the Charlson comorbidity score [13, 14] based on hospital discharges up to four year prior to their entry into the study. Those with no hospital admissions were assigned a Charlson score of zero. Patient characteristics were aggregated at the PCP level to characterize the patient population for each PCP.

Patient neighborhood information at the level of the census dissemination area (DA, the smallest geographic area for which census data are made available by Statistics Canada) was used to specify covariates at an ecological level. Each patient's DA of residence was identified based on their postal code and then linked to the Statistics Canada postal code conversion file (PCCF). The neighborhood information was gathered from the 2001 Canadian census based on each patient's DA. For each patient we determined the percentage of patients who resided in the lowest income quintile (Q1) neighborhood. For each patient, we also determined the percentage of the population aged 20 and older with less than a high school diploma, the percentage that was a visible minority (persons other than Aboriginal peoples, who are non-white), the percentage of the population aged 25 and older that was unemployed, the percentage of the population living in a rural area and the percentage of the neighborhood population that did not speak an official language (English or French). Once these neighborhood-level variables were defined for each patient, they were then aggregated at the PCP level using a weighted average by taking the size of the DA into account to define the average characteristics of the patients within his or her practice.

The unit of analysis was the PCP, and all patient-level information was aggregated at the PCP level. PCPs' use of colonoscopy and discretionary colonoscopy was measured as a rate per 100 screen-eligible patients during each calendar year. To avoid unstable estimates based on small denominators, and to ensure we only included PCPs who routinely refer patients for colonoscopy, only PCPs who contributed more than two years of data and who were linked to at least 10 patients per year were included in analyses. We used negative binomial regression [15] to explore the relationship between the rate of colonoscopy and PCP characteristics. We modeled the number of colonoscopies for each PCP as a count, with the logarithm of the eligible population of each PCP as the offset variable, including other physician and patient-level characteristics as covariates. Since the data were right-skewed and the variance was greater than the mean, the negative binomial model provided an improved fit to the data and accounted for over-dispersion better than a Poisson regression model [16, 17]. The median rate of colonoscopy was calculated for each year stratified by different characteristics. We selected categories to make the number of subjects approximately equal. The categories were all defined a priori so as not to select optimal cut-points to maximize the study findings. Variation in colonoscopy use between PCPs was tested using F-test for equality of variance. In the multivariate analyses, a generalized estimating equations (GEE) approach [18] was used to account for repeated measures from each PCP. The trend over time was examined by modeling year as a continuous explanatory variable. Interactions between the trend over time and each of the patient characteristics were tested. A piecewise regression model was fitted to test for sudden changes in trend in any specific year. The only year with a significant instantaneous change was 1999. Due to the high degree of collinearity, the following variables were excluded in the final multivariate model: age of PCP, rural location of PCP practice and proportion of visible minorities. All tests were two-sided and all analyses were performed using SAS software system version 9.1. Adjusted rate ratios (RR) and 95% confidence intervals (CIs) are reported.

This study was approved by the research ethics board of Sunnybrook Health Science Centre, Toronto.