To guide physicians in treating the disease, Osteoporosis Canada has developed and circulated the 2002 Clinical Practice Guidelines for the Management of the disease [15
]. Nonetheless, even with guidelines, it has been demonstrated that a care gap exist in managing patients in everyday clinical settings [16
There are several potential reasons as to why osteoporosis is under-treated. For instance, family physicians may have a tendency to overlook the impact of osteoporosis given that individuals with osteoporosis are no symptoms (other than fracture), and may instead focus on other conditions they consider more severe. Physicians may lack confidence regarding the management of osteoporosis and thus are not able to provided appropriate treatment strategies, they may not be aware or may not understand the Osteoporosis Canada's management guidelines for osteoporosis that outline treatment options, or they may be discouraged by low patient adherence to their medications possibly due to poor patient education and as a consequence prescribe less often.
Given this care gap, and the availability of effective osteoporosis therapies, our study was designed to improve patient care. Following the educational interventions that included the distribution and discussion of educational materials related to the 2002 OC guidelines, the evaluation of physician profiles, an educational workshop, and facilitators led small group discussions that identify barriers to the management of osteoporosis and strategies to improve patient care, family physicians demonstrated greater odds of administering osteoporosis therapy appropriately over a two year period. The appropriate use of therapy was observed in both high and low risk patients, which revealed that physicians did not just prescribe more therapy in general, but reduced their osteoporosis prescriptions in low risk patients.
There were differences in the use of appropriate therapy between the two low risk groups (appropriate therapy increase significantly in the low risk group with no fractures and remained unchanged in the low risk group). This finding is interesting and may indicate that physicians are more willing to treat patients with fractures regardless of their bone density measurements (i.e. normal or no bone density measurement) in spite of the guidelines recommendations.
To compare the results produced by the generalized estimating equations analyses (our primary analysis), a random coefficient (mixed effects model) and standard logistic regression analyses were conducted. For the generalized estimating equations analysis, a correction is made for the within-physician correlations between the repeated measurements by assuming a working correlation structure. For the current study, the correlation between individual physicians and their patients were 0.11, 0.10, and 0.09 for the high risk, the low risk, and the low risk group without fractures, respectively. The generalized estimating equations method is robust to an assigned correlation structure; such that if a correlation structure is misspecify the analysis will still obtain consistent parameter estimates [21
]. The random coefficient analysis, deals with within-physician correlations by allowing different coefficients to be random. The logistic regression analysis does not take into account the correlation and treats the data as independent [21
Both the logistic and random coefficient analysis confirmed our finding with the generalized estimating equations analysis. However, conclusions drawn from the results of these methods should be viewed in the context of the models. The main concern with ordinary logistic regression is that the analysis does not take into account the within subject correlation of the physicians resulting from repeated measures. As a consequence, while the parameter estimates from the logistic regression analyses are similar to the other models, the standard errors of the estimates are generally smaller resulting in smaller confidence intervals and p-values. As a consequence, logistic analysis revealed more significant results for both the low risk and low risk without fracture groups as compared with the generalized estimating equations and random coefficient analyses.
Furthermore, it is important to realize that the regression coefficients calculated with generalized estimating equations method are population averaged. The regression coefficients calculated with random coefficient analysis can be seen as subject specific [21
]. As a result, these two model types have different targets of inference. Given that we performed a trial concerned with changes in the mean responses over time in the study population and we were not interested in individual physician change over time, the generalized estimating equations technique is a more appropriate analysis as compared with the random coefficient method.
There are other methods available for analyzing clustered data including unweighted linear regression analysis or robust standard errors [26
]. In general, analyses may be performed using individual-level data or aggregated at the cluster level. Analyses at the aggregated level are conducted using data derived from summary statistics for each cluster.
Given the large number of methods available and the fact that results may differ among techniques, it is important that investigators specify the primary data analysis in advance. However, while simulation studies show that generalizing estimating equations have the greatest statistical power among several commonly used methods for analyzing binary clustered data, there are no guidelines that evaluated what method is the most appropriate; therefore, individual investigators will have to make judgments as to which analysis to conduct [26
This national study has several important strengths. For example, the project selected a large number of family physicians from across the country who examined over 19000 patients' charts from their own practices, which will improve the generalizability of our findings. Furthermore, the patient chart audits were chosen at random and did not depend on physicians self-reports, which may mirror physicians' attitudes about their practice but not reflect the true practice. Because single-component interventions have not consistently been found to transform behavior [28
], the QC project combined various methodologies into one multifaceted intervention, which may be more useful [31
]. In addition, physicians' behavior change observed in our study was long lasting, given that the project was conducted over a two year period.
Nonetheless, our study has some limitations. All patients examined in the project were postmenopausal women and as a consequence, the relationship between appropriate management over time may be different in premenopausal women or male patients. Furthermore, given that physician recruitment was based on the clinician's interest in women's health and osteoporosis, these members may have more knowledge in treating the disease prior to enrollment. Also, because only physicians who were interested and willing to commit to the second year of the project completed the final follow-up data collection, a lower sample of physicians reduce the power of the study to detect change over that time period. Moreover, a randomized control trial of physicians is necessary to determine the precise mechanism that the educational intervention had on physicians' behavior. Finally, it is important to consider that Osteoporosis Canada practice guidelines were developed and distributed to provide clinicians with a summary of the best evidence from clinical trials to help physicians make health care decisions regarding osteoporosis; however, clinical judgment and the patient's preference, will determine if therapy is initiated. As such, 100% compliance to the guidelines is not reasonable.