The bias-reduction intervention opened with a didactic on health disparities, stereotypes, microaggressions, interracial provider-patient interactions and racism. Then, a guided, interracial eye-contact mindfulness exercise was performed to increase providers' awareness and acceptance of subtle bias that occurs in interracial interactions. Then, in small, mixed-race groups, participants practiced the above mindfulness skills while reciprocally sharing and responding with empathy to each other's personal life histories and personal narratives of loss and/or betrayal. The intervention ended with explicit practice component, involving practice and feedback.

The control condition was a waitlist condition. Doctors were given workshop materials after the study ended.

A training for doctors

We used the Roter Interactional Analysis System (RIAS), a turn-by-turn coding system in which each provider and patient utterance is coded into one of 40 categories of speech, which can be combined to create various summary scores. RIAS codes have demonstrated validity with a variety of samples and medical contexts. For the current study, two trained RIAS coders provided ratings, and both coders coded a subset of tapes (10%) for reliability (r= 0.92). Because standardized patients in our study were following a protocol and not responding naturally, we analyzed only provider codes, specifically the Emotional Rapport Building summary score which is a composite of doctor codes including emotional statements, legitimizing statements, concern statements, partnership statements, self-disclosure statements, and reassurance statements.

We modified observer-based responsiveness coding systems used in previous research that operationalized positive responsiveness in dyads as concrete instances of understanding, validation, and caring on a 0 to 3 scale. We expanded the previous 0 code, which originally combined both no responses (e.g., simply ignoring the challenge) and invalidating responses (e.g., defensive or microaggressive responses), into a negative responsiveness dimension and provided anchors and examples for each score, producing a Likert scale from - 3 to + 3, as shown in Table 3. Four coders were trained to high reliability.

Single-item question analyzed by responsiveness coders, asking "did the doctor say anything offensive?"