Introduction

The individual biases held by medical providers may contribute to patient satisfaction following medical care, especially for patients from minority groups. Racial and ethnic health disparities are well documented in the United States [1–3] and there is substantial evidence that healthcare providers, like the general public, demonstrate a wide variety of implicit biases towards a wide range of marginalized groups [4,5]. Implicit biases refer to associations between a social category (e.g., Hispanic people) and a trait (e.g., “non-compliant”) or evaluation (e.g., “good”) that are considered more automatic than explicit, or self-reported, biases either because individuals are less aware that they hold these biases or because these biases inform behavior outside of conscious awareness [6]. Evidence suggests that providers associate individuals who are racial and ethnic minorities, people with obesity, members of the LGBTQIA+ community, people with disabilities, or people who have stigmatized health conditions with more negative concepts compared to individuals who are white, thin, straight, able bodied, and without stigmatizing health conditions [4,7–12]. The implicit prejudices and negative stereotypes that providers hold about patient groups may create barriers to providing equitable care.

Evidence tying provider biases to patient care varies based on the type of outcome considered. Most studies examining if provider implicit bias links directly to medical decision making have shown that, across a wide variety of health care providers (nurses, internal and family medicine physicians, surgeons, pain management specialists, etc.), the relationship between these biases and treatment recommendations is small and unreliable [13–20]. However, research examining the relationship between provider bias and patient perceptions of care suggests that physician bias may be more reliably linked with interpersonal aspects of patient care. Black patients whose doctors demonstrate greater anti-Black implicit prejudice report less positive attitudes toward their healthcare providers [21,22], less confidence in completing treatment recommendations, perceive fulfilling treatment recommendations as more difficult, and remember less about their encounter with more biased physicians [22]. Additionally, physicians who hold stronger implicit non-compliance stereotypes towards Black people have patients who feel less involved in their medical care and trust their physicians less [21]. One pilot study suggests that physician implicit bias correlates with broader aspects of patients’ well-being [23]. In a study examining the relationship between physician implicit bias and Black spinal cord injury patients, greater implicit racial bias was associated with higher rates of patient depression, lower life satisfaction, and less ability to maintain social relationships. Whereas racial concordance between physicians and patients can attenuate the relationship [21–23], these findings suggest physicians communicate biases to patients that are reflected in their perceptions of medical care and, potentially, to patient well-being. Nevertheless, an important limitation to the previous work is that it has focused largely on the relationship between provider bias and outcomes for Black patients. It is important examine if such patterns extend to other racial and ethnic patient groups, such as Hispanics, who are the largest ethnic minority group in the US [24].

Another potential limitation is that each of the above studies presents the relationship between physician implicit bias and patient perceptions of care as a simple bi-variate relationship. One study, drawing on aversive racism theory [25,26], examined the extent to which the effects of implicit bias on Black patient perceptions were moderated by explicit bias [27]. Overall, patients with more implicitly biased physicians were less likely to view themselves and their physician as a team. Additionally, physician explicit bias moderated the effect of implicit bias on patients’ overall feelings about their medical care. Consistent with aversive racism theory, Black patients had more negative perceptions of their care when physicians were high in implicit but low in explicit bias compared to any other combination of implicit and explicit bias. These findings suggest that implicit bias may not always operate directly on patient perceptions, but may be selectively communicated depending on other interpersonal factors, such as the language that providers use during a clinical interaction.

Although providers appear to communicate bias to patients non-verbally, relatively little is understood about whether or how physicians’ verbal behaviors factor into these interactions. Word choice is often intentional and can communicate meaning through the semantic content of words speakers decide to use. However, there is also evidence that word choice (e.g., pronouns, verbs, and prepositions) can communicate information about the speaker’s motivations, intentions, and their identity [28–30]. For example, participants asked to write an essay about a time they felt more powerful were more likely to speak about work-related topics [31]. Other studies find that using nouns compared to adjectives to describe others (e.g., “Jane is a liar” vs. “Jane was dishonest”) is associated with more essentialist beliefs [32], and that greater use of first person singular pronouns (e.g., “I”, “me”,”my”) is associated with greater attention to the self and stronger perceptions of narcissism [29,33]. Thus, physicians who implicitly associate Hispanic patients with negative traits and emotions may also view themselves as more powerful, hold more essentialist beliefs about their patients [34], or engage in more self-focused attention during interactions with Hispanic patients. To the extent that a physician’s subtle word choice unintentionally communicates these internal states during clinical interactions with Hispanic patients, they may negatively influence the patients perceptions of the care they receive.

Additionally, research suggests that context, intention, and longevity of relationship can all inform the ways in which language is interpreted [30,35–38]. For example, language style matching or the extent to which two individuals align in their use of function words is often associated with positive interpersonal outcomes such as greater liking and relationship satisfaction [30,37]. However, in the early stages of transactional relationships, greater language style matching is viewed as a marker of deception [38]. Thus, the same language patterns can be perceived differently depending on context. In a medical visit, physician non-verbal expressions of bias may create contexts in which word use is interpreted differently by their patients than when unbiased physicians use the same language.

To date, only three studies have examined the relationship between physician implicit bias and verbal behavior [21,39,40]. Greater physician implicit racial bias is associated with both how physicians use language [21,41] and with physician word choice [39,42]. Hagiwara and colleagues (2017) measured the implicit anti-Black prejudice of 18 non-Black primary care physicians and audio recorded their interactions with 117 Black patients. Physician word use was quantified using the Linguistic Inquiry and Word Count (LIWC) software [43]. They found that physicians with greater anti-Black implicit bias used more first-person plural pronouns and anxiety-related words compared to their less biased colleagues. The finding related to first-person plural pronouns may seem counterintuitive. Although first-person plural pronouns can be used to a shared group, it may also an indicator of interpersonal coldness in some contexts (Zimmerman et al., 2013). Similarly, Wolsiefer and colleagues (2023) examined physician implicit biases towards Hispanic patients in a larger sample including both Hispanic and non-Hispanic physicians (N = 59 residents, N = 291 patients). Their findings suggest that physicians with greater implicit anti-Hispanic bias used more interrogatives and work-related language, but fewer prepositions and words related to the future, relativity, and time (these effects did not differ by ethnicity of the physician). These differences in findings may result from sampling from different populations of physicians (e.g., residents vs. more established physicians, physicians in the U.S. Midwest vs. Southwest, etc.) or from a focus on biases towards two different target groups (e.g., Black vs. Hispanic patients). Notably, all of these studies examine whether physician word use is directly related to physician bias but did not examine how verbal behavior and physician bias predict patient perceptions of care.

The goal of the present study is to examine the relationship between physician implicit bias, verbal behavior and patient perceptions of their care by conducting a secondary analysis of an existing dataset of approximately 300 doctor-Hispanic patient clinical interactions [40]. The present analysis extends the findings of Wolsiefer et al. (2023) by examining the relationship between physician implicit bias and patient satisfaction, and how this relationship may vary as a function of the words that physicians with different levels of implicit bias use when communicating with Hispanic patients. A secondary analysis of this large-scale study has the potential to provide further insight into the correlates of physician bias with actual patient interactions.

Method

Results

Implicit bias and patient satisfaction

As reported in Wolsiefer et al. (2023), average scores for implicit anti-Hispanic stereotypes and prejudice were both positive and significantly greater than 0, indicating that resident physicians associated Hispanic individuals with both negative evaluations and non-compliance stereotypes (compared to White individuals), stereotypes: M = 0.26, SD = 40, t(52) = 4.73, p < .001, Cohen’s D = 0.65; prejudice M = 0.37, SD = 42, t(52) = 6.41, p < .001, Cohen’s D = 0.88. Scores on the measure of implicit prejudice and implicit stereotyping were highly correlated, r(51) = 0.64, p < 0.001, 95% CI (0.44, 0.77), so we averaged them to create one overall implicit bias score for each resident physician [40].

Given the racial diversity of physicians in this sample, it was possible that physicians differed in implicit bias based on their own racial/ethnic identity. A previous analysis of these data did not find evidence of race differences in physician implicit bias (40). Additionally, we tested whether race moderated the linear or quadratic effects of implicit bias on patient satisfaction. There was no evidence that this was the case (race X IAT linear: b = −.35, t(283) = −0.94, p = 0.35, 95% CI [−1.07, −.37]; race X IAT quadratic: b = 0.99 t(281) = 1.17, p = 0.24, 95% CI [−0.64, 2.62]. Since physician bias did not differ as a function of race/ethnicity, we did not include physician race/ethnicity as a factor in subsequent analyses.

To examine whether physician implicit bias was related to patient satisfaction, we estimated a linear mixed effects model. This model regressed patient satisfaction on physician implicit bias and estimated a random intercept to account for non-independence since patients were nested within residents. Unlike in previous research [21], there was no evidence of a linear relationship between resident bias and patient satisfaction, b = −0.22, t(285) = −1.23, p = 0.22, 95% CI (−0.56, 0.13). However, we reasoned that it is possible that patients may only be sensitive to their physician’s implicit biases at relatively high levels, which would correspond to a non-linear effect. Thus, in a subsequent exploratory analysis, we tested a non-linear relationship between physician implicit bias and patient satisfaction and found a significant quadratic effect, b = −1.01, t(284) = −2.93, p = 0.004, 95% CI (−1.69, −.34). This effect remained significant whether or not Hispanic identifying physicians were excluded from the analysis. To interpret this quadratic effect, we examined the simple effect of implicit bias on patient satisfaction at one standard deviation below the mean IAT score, the mean, and one standard deviation above the mean IAT score. High levels of physician implicit bias were significantly related to patient satisfaction, b = −1.04, t(284) = −3.15, p = 0.002, 95% CI (−1.68, −0.39), = 6.25. However, at low and mean levels of implicit bias, there was no evidence of a relationship between implicit bias and patient satisfaction, low: b = 0.47, t(284) = 1.60, p = 0.11, 95% CI [−0.10, 1.04], = 6.46; mean: b = −0.28, t(284) = −1.59, p = 0.11, 95% CI [−0.63, 0.06], = 6.49. This suggests that, among physicians with higher levels of implicit bias, implicit bias was associated with lower levels of patient satisfaction (see Fig 1). Whether physicians self-reported any implicit bias training in the past did not relate to patient satisfaction or moderate the effect of implicit bias (linear or quadratic effects) on patient satisfaction, all ps > 0.22.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Quadratic relationship between physician implicit bias and patient satisfaction.

Note. Figure represents relationship between physician implicit bias and patient satisfaction at −1 SD, mean and +1SD levels of bias.

https://doi.org/10.1371/journal.pone.0338699.g001

Resident language

To examine whether physician language played a role in whether physician implicit bias was communicated to patients, we conducted 3 additional analyses. For each of the 3 sets of models below, we began by examining language variables that previous literature has deemed relevant for physician-patient interactions, namely: 1^st person singular pronouns (e.g., “I”, “me”), anxiety words (e.g., “worried”, “fearful”, interrogatives (e.g., “how”, “who”), work-related language (e.g., “job”, “xerox”), prepositions (e.g., “to”, “with”), future words (e.g., “may”, “will”), relativity words (e.g., “area”,”bend”), time-related words (e.g., “end”,”until”), and language style matching. After running analyses on these variables, we conducted exploratory analyses using every variable provided by the LIWC program.

In the first analysis, we examined quadratic effects of physician implicit bias on the physician word use found in previous research. To test these relationships, we estimated a series of mixed effects models that regressed each of these LIWC language variables on the linear and quadratic effects of resident implicit bias and allowed the intercept to vary by resident. There was no evidence of a quadratic relationship between implicit bias and any of the LIWC language categories even before performing any corrections for multiple significance tests, all ps > .06. See Table B in S1 Table for a summary of these quadratic effects.

Second, we examined the relationship between resident physician language and patient satisfaction by estimating linear mixed effects models that regressed patient satisfaction on each LIWC language variable (in separate models) and allowed the intercepts to vary by resident physician, focusing first on language variables that have previously demonstrated a relationship with physician implicit bias. This yielded one significant linear relationship—patients whose physicians used more work-related language reported greater satisfaction, b = 0.31, t(285) = 1.98, 95% CI: [0.00, 0.61]. This effect did not persist after corrections for multiple significance tests or when Hispanic identifying physicians were excluded from the analysis.

In an exploratory analysis, we estimated the same models for the remaining LIWC language variables (see Table C in S1 Table). Several additional relationships between resident language and patient satisfaction emerged; however, given the large number (75) of hypotheses tested, we controlled for the false discovery rate [52]. This approach is an alternative to family wise error corrections such as Bonferroni corrections and results in higher statistical power. After these corrections, only the effect of use of articles on patient satisfaction remained significant, b = 0.21, t(285) = 3.04, 95% CI: [0.09, 0.34]. Patients whose physicians used more articles during their visits reported higher satisfaction. This effect remained significant whether or not Hispanic identifying physicians were included in the analysis. All relationships that reached statistical significance without correcting for Type 1 error inflation are described in S1 Appendix. Since these relationships may be the result of Type 1 errors they should be interpreted with caution. However, given the uniqueness of this dataset, we include these supplemental results for the reference of future researchers examining similar relationships.

Third, the quadratic relationship between implicit bias and patient satisfaction suggests that implicit bias relates to patient satisfaction primarily when physicians express higher levels of bias. Thus, we examined whether physician word choice might be related to patient satisfaction differently for physicians with differing levels of implicit bias. To test these relationships, we regressed patient satisfaction on physician bias, each language variable (separately), and their interaction. Once again, we began by examining language variables supported by previous literature. This initial approach resulted in two significant interaction effects: physician implicit bias interacted with language style matching, b = −5.71, t(283) = −2.34, 95% CI: [−10.47, −0.94], and with physician use of 1^st person singular pronouns, b = −0.46, t(283) = −2.56, 95% CI: [−0.81, −0.11]. These effects remained significant whether or not Hispanic identifying physicians were included in the analysis.

To examine these moderating effects, we calculated the simple effects of language at low (−1 SD from the mean), mean, and high (+1 SD from the mean) levels of physician implicit bias. The results of these simple effects tests can be found in Table 1. Both of these language variables demonstrated a negative and significant relationship with patient satisfaction only for physicians high in implicit bias. That is, for physicians high in implicit anti-Hispanic bias, greater language style matching and greater use of 1^st person singular pronouns was associated with lower levels of patient satisfaction.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Simple effects for language X implicit bias interactions.

https://doi.org/10.1371/journal.pone.0338699.t001

Again, we ran an exploratory analysis that examined parallel models for the remaining LIWC language variables. These effects are reported in Table D in S1 Table. Several relationships met the uncorrected standard for statistical significance and are described in S1 Appendix. After controlling for the false discovery rate, no interaction effect met the corrected criterion for statistical significance.

Discussion

This study yields several novel findings regarding the interplay between physician implicit bias, word use, and patient satisfaction. First, the relationship between physician implicit bias and patient satisfaction was not linear as reported in previous research [21,53]; implicit bias was only significantly (and negatively) related to patient satisfaction when it was high compared to moderate or low. Whereas this is consistent with the assumption that physicians with high implicit bias communicate their negative attitudes and stereotypes through word choice, this quadratic effect could also indicate that Hispanic patients expect some level of prejudice and are, thus only impacted when physicians display high levels of prejudice [54]. Alternatively, this quadratic effect may suggest that physicians with lower and moderate levels of implicit bias are better able or more willing to control these biases. Indeed, in this study, physicians with higher levels of implicit bias tended to display lower levels of external motivation to control prejudice, suggesting that higher bias physicians may be less motivated to appear non-biased due to normative concerns.

Other possibilities for the difference in findings could be the target groups studied in the respective studies. Whereas Cooper and colleagues (2012) and Gonzalez and colleagues (2024) examined the relationship between anti-Black implicit bias and patient (or standardized patient) perceptions of care, the current research examined bias towards Hispanic patients. It is possible that bias towards Hispanic patients is different in nature or operates differently than specific anti-Black stereotypes about healthcare. That is, although previous research suggests that non-compliance stereotypes exist for both Black and Hispanic patients, it may be that other stereotypes are more relevant for predicting physician behaviors towards Hispanic patients; whereas non-compliance stereotypes are the most relevant for predicting behavior towards Black patients. Differences in the operationalization of implicit bias may have also led to the difference in findings. Cooper and colleagues (2012) also measured both implicit prejudice and implicit non-compliance stereotypes, but treated these measures as separate linear predictors of patient perceptions of care. Given the strong correlation between the implicit prejudice and stereotype measures in the current study, we aggregated these scores to form a single implicit bias measure when predicting patient satisfaction.

Finally, we note the ease with which people can correctly identify the race or ethnicity of a patient may play a role in how implicit bias impacts patient care. Research indicates that people are less accurate at correctly categorizing Hispanic compared to Black faces. In a clinical context, unless their ethnic identity is present on a chart or in the EHR, physicians, on average, may be less accurate at visually identifying Hispanic compared to Black patients. This may result in less consistent activation of prejudice and non-compliance stereotypes when physicians interact with Hispanic compared to Black patients. Further, more racially prejudiced individuals and those higher in race essentialism are more careful to correctly categorize others based on race/ethnicity [34,55,56]. Thus, it may also be that higher bias physicians are more likely to accurately identify the ethnicity of Hispanic patients and more susceptible to activation of negative associations than physicians with lower or moderate levels of anti-Hispanic implicit bias.

Our second set of findings demonstrated two relationships between physician language and patient satisfaction. Across levels of physician implicit bias, greater use of work related language and greater use of articles were associated with higher patient satisfaction, and these relationships did not depend on the physician’s level of implicit bias. Some previous research on work-related language has characterized use of work-related language as an indicator of focus on the task at hand [57,58] while other research shows that people who are made to feel powerful use more work-related language [31]. These findings present two potential explanations for the findings on work-related language and patient satisfaction. First, it may be that patients view physicians who use more work-related language as more task focused and respond favorably to this focused approach to patient care. Alternatively, patients view physicians who use more work-related language as being more powerful and, counterintuitively, prefer physicians who take a more hierarchical approach to patient care. Additional work should attempt to replicate this finding and examine these possible explanations.

There is very little research on the use of articles in spoken language so our explanation for this effect remains speculative. However, there is some evidence that when nouns are used to describe a person’s behavior rather than adjectives, they link that noun more strongly to the individual’s identity [59,60]. Since articles precede nouns, it may be that greater use of articles is indicative of physicians using more identity-focused language with their patients (e.g., “You’re not an exerciser” vs. “You haven’t been exercising much.”). It is unclear why such language use would be associated with positive patient impressions of care. Future work should examine the context in which articles are used during patient visits and explore additional explanations for this finding.

Our third set of findings found that two language variables interacted with physician implicit bias to predict patient satisfaction. For physicians high in implicit bias, greater use of first person singular pronouns and language style matching were associated with less patient satisfaction. Previous research has also found that the use of first-person singular pronouns relates to implicit bias and lower patient-centered care [61], possibly because the use of first-person pronouns signal that the physician is more focused on themselves than the patient. However, previous research on language style matching showed positive associations with positive interpersonal experiences [30,37]. Whereas our findings may seem at odds with previous literature, additional research suggests that impressions of language style matching may be complex. First, individuals with less power are more likely to match their language to high power individuals than the reverse and greater language style matching by low-power individuals is associated with more positive impressions [35]. Thus, the present findings might suggest that, when interacting with high bias physicians, Hispanic patients change their patterns of speech to match their physicians in an attempt to increase rapport, but resent the need to change their own behavior to seek better medical care. Alternatively, if physicians are changing their language to match that of their patients, patients may view this with distrust. In transactional relationships, individuals who engage in greater language style matching are viewed as more dishonest [42]. In the context of a medical visit, it may be that language style matching from a high-bias physician is seen as a cue for dishonesty or mistrust by Hispanic patients.

As a whole, these findings suggest that language may be one route by which physicians express implicit bias to their Hispanic patients, but this depends on the level of bias of their physician. Previous work suggests that implicit racial bias is associated with non-verbal and paraverbal behaviors [21,61], and the current findings suggest that this may primarily be the case for physicians at higher levels of implicit bias. However, the analysis of language by implicit bias interactions suggests that bias may be important for doctor-physician interactions in specific contexts. These moderated effects may arise for a few reasons. First, language use is contextual [62] and the language may be differentially predictive of interaction patterns in different situations. For example, verbal productivity scores in public, day-to-day contexts was predictive of extraversion, but the same verbal productivity scores generated from private speech were not, speaking to how communication contexts can modulate the psychological meaning of language markers [63]. Thus, the same words may be used in different contexts by high vs. low bias physicians and therefore may have differential implications for the overall evaluations of patients. Second, patient impressions are likely determined by multiple factors. For example, prosody impacts the interpretation of language, especially when language is positive [64]. Thus, seemingly innocuous word choice when viewed in the context of a transcript may communicate disdain or unfriendliness if communicated with a sarcastic tone. Alternatively, previous research suggests that non-verbal friendliness, which was also associated with implicit anti-Black bias, was positively associated with a Black interaction partner’s impression of a White participant [65]. Thus, it may be that the unique combination of both what a physician said and how they said it contributes to patients’ satisfaction with their outpatient visit.

Finally, the current study may have allowed for detection of these more nuanced effects due to its unique study design, in which several patient visits were nested under individual resident physicians. This approach, in contrast to approaches that measure only one patient visit per physician, may allow for more stable estimates of relationships between physician implicit bias and aspects of (statistically noisy) patient interactions. Future research should explore these possible mechanisms by examining the contexts in which physicians of different levels of bias are using different types of language and should explore whether verbal and non-verbal behaviors interact to influence patient outcomes.

Supporting information