Notes were documented as close to verbatim as possible. We used the audio-recordings to check notes and verify quotations and subsequently erased them. We coded the notes with a coding scheme that mirrored the domains in the discussion guide, and used an iterative consensus process to develop the final coding scheme.16 In place of a measure of inter-coder reliability, we used established methods for addressing differences in coding by reconciling them through discussion and consensus.17, 18, 19 One investigator (Mynti Hossain) completed coding of all discussion notes using qualitative research software (NVivo9.0, developed by QSR International). A second investigator (Dana Petersen or Joseph Zickafoose) reviewed and approved all final coding.
To distill findings, we used pattern recognition analysis to identify similarities and differences within domain categories and then examined patterns and associations.20 Research team members developed summaries for selected topics. Each summary included quotes and estimates of the frequency of commentaries by domain and theme, and was reviewed by another team member. The full team discussed themes to clarify, confirm, refine, or elaborate them and consider implications. We used the following categories to describe the relative frequency of commentaries related to our findings: With respect to the number of physicians, "few" indicates <3, "some" indicates 4-6, "many" indicates 7-9, "half or more" reflects 10-14, and "most" reflects 15-20.