describes the individual- and area-level characteristics of our study population. Our study sample included 3,292,605 individuals living in 49 non-rural LHAs. The sample age (40.2) and sex (51% female) distribution was equal to census figures for the province. Residents in our sample had diagnoses in their administrative health records that indicated an average of 3.2 ADGs in 2006. The LHAs included in this study varied moderately (coefficient of variation, CV>0.20) in terms of population health (PYLL), average income, and primary health care supply; LHAs varied considerably (CV>1.00) in terms of percentages of Chinese and South Asian populations. Of the drug types studied, antihypertensives (15%), opioids (12%), and antidepressants (10%) were most commonly used.
Characteristics of the study population.
Determinants of utilization
Likely because the correlation of errors within clusters was low, results of the GEE models on a 2% sample of the data and logistic models were virtually identical – all adjusted odds ratios were equal to the third or fourth decimal and no tests of statistical significance changed 
. lists results from the full sample logistic regressions for prescription drug purchases by therapeutic category. All five of the models containing only individual-level variables suggest that individual-level health needs, predisposing factors and enabling factors are significant for explaining variations in prescription drug use (p<0.001). Moreover, individual-level health needs and predisposing factors had expected impacts on the likelihood of category-specific prescription purchases. There was a u-shaped relationship between income and the likelihood of purchasing drugs from each class studied except for opioids (for which there was a negative income gradient).
Adjusted odds ratios for the likelihood of purchasing one or more prescription from specific therapeutic categories in 2006, non-rural local health areas of British Columbia.
Adjusted odds ratios on individual-level variables did not change significantly when area-level variables were added to the model, and the models containing both individual- and area-level variables were better fit (LR test, p<0.001), though the change in the predictive power of the model was very modest (small changes in C-statistics). Area-level variables had different impacts on the likelihood of purchasing drugs of different types. The level of population health needs (PYLL) was positively associated with the likelihood of purchasing antidepressants and opioid drugs and negatively associated with the likelihood of purchasing statins. Higher concentrations of ethnic minorities in LHAs were generally associated with a lower likelihood of prescription purchases, but the results varied by drug category. For example, the share of the local population that identified as Chinese was negatively associated with the likelihood of purchasing antidepressants and opioid drugs but not significantly associated with the likelihood of purchasing antihypertensives, statins, or acid-reducing drugs. Area-level supply of primary care physicians was not significantly associated with the likelihood of purchasing any of the five types of medicine studied.
Impact on measures of regional variations
lists summary statistics describing the distribution across LHAs of prevalence rates for prescription drug purchases from the five therapeutic categories. The table summarizes variations in crude rates of medicine use, as well as standardized rates based on predictions from the logistic regression with adjustments for individual-level determinants and from the logistic regression with adjustments for individual- and area-level determinants. The magnitude of variation in crude rates of prescription purchases across regions was comparable for all five drug classes studied. The extent to which measured regional variation was attenuated by the addition of individual- and area-level predictors of drug use differed by specific type of prescription drug.
Summary statistics for regional variations in rates of purchasing one or more prescription from specific therapeutic categories in 2006, non-rural local health areas of British Columbia.
The addition of individual-level variables to create adjusted measures of prevalence reduced measures of regional variation in the purchase of each type of prescription drug; however, measured variation fell most notably for antihypertensives and statins when individual-level factors were accounted for. The CVs for these categories changed from 0.18 to 0.07 and 0.20 to 0.11, respectively. The addition of individual-level variables had the least effect on measures of regional variation in the purchase of antidepressants and opioid drugs. In contrast, while the addition of area-level variables to the adjustment model reduced measured variations for all drug types, the effects of area-level variables were greatest for measured variation in the use of antidepressants and opioid drugs - the CVs for these categories changed from 0.17 to 0.07 and 0.16 to 0.07, respectively, with the addition of area-level variables.