In this population-based study, we analyzed data for children who had physician visits for selected childhood infections from fiscal year (FY) 1996 (Apr. 1, 1996, to Mar. 31, 1997) to FY 2000. Two criteria for nonadherence to evidence-based antibiotic therapy were applied to the health care and prescription records of these children, and child, household and physician determinants of nonadherence were ascertained.
The study protocol was approved by the Human Research Ethics Board, University of Manitoba, and Manitoba's Health Information and Patient Confidentiality Committee. Data were obtained from 4 population-based electronic databases maintained by the Manitoba Health Services Insurance Plan (MHSIP), which provides health insurance for all Manitobans: registration files, physician reimbursement claims, hospital discharge abstracts and records of prescriptions dispensed. The MHSIP registry contains a record for every eligible individual who received insured health services, and each record includes birth date, sex and geographic location. Records of physician reimbursement for medical care are submitted under a fee-for-service arrangement and contain information on patient diagnosis according to the clinical modification of the International Classification of Diseases, 9th revision (ICD-9-CM). Discharge abstracts for hospital services include information on up to 16 ICD-9-CM diagnostic codes. Prescription records submitted by retail pharmacies contain information on date of dispensing, drug name and identification number, dosage form, quantity dispensed and type of drug insurance program. Household income, which is required to calculate the income-based deductible payment, is also captured in the prescription database. The reliability and validity of MHSIP's prescription and health care databases are high.24,25
Record linkages among databases were achieved through anonymized personal identifiers. Children were selected on the basis of age less than 19 years and the availability of data on household income.
Two measures of physicians' nonadherence to evidence-based antibiotic prescribing were evaluated: prescription of an antibiotic for a VRTI and prescription of a second-line antibiotic without prior use of a first-line agent. VRTIs were diagnoses identified from epidemiologic studies as the most likely to be of viral cause and included ICD-9-CM codes for bronchiolitis (466), bronchitis (490, 491), acute respiratory tract infections (464, 465) and the common cold (460). Acute episodes for VRTI were selected by identifying singular ambulatory physician visits for VRTI or clusters of visits, separated from each other by a period of 30 days. Antibiotic prescriptions dispensed within 7 days after a VRTI visit were defined as having been prescribed for the VRTI. When an acute episode consisted of multiple VRTI visits, the antibiotic prescription dispensed closest to the first visit was selected, and the 7-day criterion was applied to any visit within the cluster. To ensure that there were no competing conditions for which antibiotics were warranted, the VRTI visit was excluded from the analysis if there were physician visits for acute otitis media, pharyngitis, pneumonia, urinary tract infection or cellulitis between the VRTI visit and the dispensing of the antibiotic. Intervening physician visits for noninfectious diagnoses were allowed; however, if a hospital admission occurred between the VRTI visit and the dispensing of the antibiotic prescription, the VRTI visit was also excluded. These methods have been used by others to identify respiratory tract infections of viral origin.13,26,27
For the second criterion, the antibiotic dispensed closest to an acute episode of otitis media, pharyngitis, pneumonia, urinary tract infection, or cellulitis or impetigo (within 7 days after the episode) was selected. The definition of an acute episode was analogous to the VRTI definition. Children with more than 4 physician visits for otitis media per year (defined as having chronic otitis media) were excluded. A list of first-line antibiotics was obtained from the then-current published guidelines, which took into account usual patterns of resistance (see Appendix 1
All antibiotics that were not first-line therapy according to these guidelines were labelled as second-line agents.
Child, household and physician factors were selected on the basis of literature indicating an association with antibiotic use.18,19,20,21,22,23
Children were described by age, sex and household income. Physician characteristics included age, sex, location of training (in Canada or the United States v. elsewhere), specialty (general practitioner [GP], pediatrician or other specialist), period since licensure (less than 20 years v. 20 years or more), hospital affiliation (identified in the hospital database as the treating physician) and type of practice (solo v. group). A solo practitioner was defined on the basis of reimbursement claims received from one location with no reimbursement claims from other physicians being received from the same location.34
Physician visits and prescriptions were classified according to year and season (winter, spring, summer and fall). For the second criterion, antibiotic prescriptions were classified according to the drug insurance program that paid for the prescription: Pharmacare, Income Assistance, or self-payment or private insurance.
The odds ratio (OR) for receiving an antibiotic prescription for VRTI or for receiving a second-line antibiotic as initial therapy was determined according to child, household and physician factors. Recognizing the multilevel structure of the data, hierarchical linear modelling methods were pursued to determine the ORs.35
Child visits to a physician or prescriptions (level 1) were nested within physicians (level 2), and the most parsimonious model was selected at a significance level of p
< 0.05. Study measures were introduced into the models as continuous or binary variables. Information on household income was available in $10 000 ranges. A backward elimination process identified level 1 variables that met the significance level; these were placed into the model along with level 2 variables. The same backward elimination process was repeated to select significant level 2 variables. The final model was a population-average model that represented the average effects of each covariate across the level 2 units. Age, sex and household income were treated as random effects, and all other explanatory variables were fixed effects.