Purpose: Patients with advanced cancer experience problems in pain management, understanding their prognosis, and consideration of their preferences in care. The level of such problems is considered as a function of communication and shared decision-making. The existing literature on understanding this function has focused on casual studies using linear regression analysis. Typical questions include whether race and patient activation have strong impacts on physician behaviors. In this study, we explored the applicability of clustering analysis, an alternative descriptive statistical analysis method, to predicting physician behaviors, but with an emphasis on identifying non-standardization of the behaviors.
Method: We analyzed the contextual data extracted from 39 audio-recordings of physician visits with unannounced standardized patients (SPs), including visits to 19 oncologists and 20 primary care physicians. SP portrayals were calibrated through hours of extensive training to achieve role accuracy in terms of both verbal content and emotional value of the role. The SPs were well-received by physicians and their staff. Decision-making sections of the audio-recordings were extracted and coded for topic and behavior by supervised undergraduate students. For our analysis, we aggregate conversation topics to form a list of 10 distinct categories. For each topical category, we encoded its appearance in each communication sequence (or case) in three distinctive manners: occurrence, i.e., binary indicator on whether the topic appears in the sequence; frequency, i.e., the number of times the topic appears in the sequence; and the length of the subsequence on the topic contained in the sequence. With each coding style, we constructed a case by category matrix. We then performed clustering analysis to identify outlier physicians, using COBWEB, which is a hierarchical conceptual clustering approach.
Results: We present a summary of the ten topic categories in terms of the three analyses. For the counting style of index occurrence, we identified 3 outlier physicians. For the case of frequency, we identified 1 outlier. For the case of the length, we identified 2 outliers. Given different counting styles, we obtained distinct outliers.
Conclusions: We assessed the applicability of clustering analysis to the context of identifying non-standard physician and patient behaviors in their shared decision-making. Additional work needs to be conducted to compare the effectiveness of various clustering methods and compare the clustering methods with regression approaches.
