Documenting Race and Gender Biases in Pain Assessment and a Novel Intervention Designed to Reduce Biases

Abstract

Disparities in pain care are well-documented such that women and people of color have their pain undertreated and underestimated compared to men and White people. One of the contributors of the undertreatment of pain for people of color and women may be the inaccurate assessment of pain. Understanding the pain assessment process is an important step in evaluating the magnitude of and intervening on pain disparities in care. In the current work, we focus on documenting intersectional race and gender biases in pain assessment and present the results of a novel intervention for reducing these biases. Across 3 studies (N = 532) and a mini meta-analysis using real videotaped people in pain as stimuli, we demonstrate that observers disproportionately underestimated women of color’s pain compared to all other groups (men of color, White women, and White men). In study 3 (N = 232), we show that a novel intervention focused on behavioral skill building (ie, practice and immediate feedback) significantly reduced observers’ pain assessment biases toward marginalized groups compared to all other types of trainings (raising awareness of societal biases, raising awareness of self-biases, and a control condition). While it is an open question as to how long this type of intervention lasts, behavioral skills building around assessing marginalized people’s pain more accurately is a promising training tool for health care professionals.

Perspective:

This article demonstrates the underestimation of pain among people of color and women. We also found support that a novel intervention reduced observers’ pain assessment biases toward marginalized groups. This could be used in medical education or clinical care to reduce intersectional pain care disparities.

Women and people of color’s pain is systematically underestimated, inaccurately recognized, and mistreated.^1–6 Yet, little research has examined differences in the assessment of pain intensity using stimuli of real people experiencing pain or has tested interventions directed at closing the disparity gap in pain assessment. The current research establishes that inequities in pain assessment exist among laypeople and tests the effectiveness of three interventions to reduce the underestimation of pain for women and people of color.

Disparities in Pain Care

People of color are rated as being in less severe pain, are less likely to receive comprehensive diagnostic and treatment approaches for pain, and wait longer for treatment than White patients,^4–19 despite reporting higher levels of pain and lower pain tolerance than White patients.^20,21 Women also have an inequitable pain burden compared to men, experiencing heightened chronic and acute pain (ie, pain sensitivity),²² yet are more likely to have their pain underestimated,^23–25 undertreated,^7,26,27 less attended to,²⁸ and referred for psychological treatment rather than recommended analgesics.^24,29

Limited Intersectional Approaches to Pain Assessment and Interventions

While previous work has documented the existence of these biases, common experimental limitations restrict the ecological validity of bias assessments (eg, relying on vignettes or static photographs). Much of this past work has focused primarily on race or gender biases. Yet, the impact of living at the intersection of multiple minoritized identities is multiplicative, rather than additive, as multiple systems of oppression shape people’s lives based on the intersecting axes of oppression (eg, race, gender, gender, sexuality, and ability).³⁰ Therefore, to better understand and address disparities in pain assessment, an intersectional approach is essential.³¹

The primary goal of empirically documenting intersectional pain assessment biases is to reduce these biases. Yet, we are unaware of any interventions specifically focused on reducing biases in the assessment of others’ pain, revealing a critical gap in health care training and education. It is possible that providing observers an opportunity to practice making pain assessments and receive immediate feedback (ie, behavioral skills training) might be one effective method to improve pain assessment.³² If so, these results could provide support for an important training tool to address provider pain assessment bias.

Overview

Across three studies, observers viewed videos of individuals (ie, targets), who varied in race and gender, experiencing real pain, and were asked to infer how much pain each target was experiencing on the same scale targets had self-reported. We hypothesized that 1) observers would underestimate targets of color and women targets’ pain more than White targets’ or men targets’ pain, 2) target race and target gender would interact such that the underestimation of women’s pain compared to men’s pain would be greater among people of color than White people and that the underestimation of people of color’s pain compared to White targets’ pain would be greater among women than men. We did not anticipate that observer race or gender would moderate these effects based on past research suggesting bias exists across observers (eg, face memory,³³ emotion recognition,³⁴ lie detection,³⁵ pain deception³⁶).

In study 3, we tested a novel intervention to reduce observer bias in the assessment of pain for marginalized targets. We hypothesized that behavioral skill building would reduce observers’ biases in assessing pain, especially for women and people of color, compared to two building awareness trainings or a control condition.

Method

Study 1

Undergraduate students (ie, observers) viewed videos of targets who varied in race and gender experiencing real pain and inferred their pain intensity. We compared assessments of pain to actual self-reported pain provided by targets at the time of the pain task.

Participants (Observers)

Observers were N = 250 (191 women, 59 men) undergraduate students recruited using a convenience sample through a campus-wide email listserv sent to all undergraduate students from Massachusetts College of Pharmacy and Health Sciences for a 60-minute online study in which they received a $20 Amazon.com gift card for participation. This study was approved by the Massachusetts College of Pharmacy and Health Sciences’ Institutional Review Board and informed consent was obtained from each subject.

Observers ranged in age from 18 to 40 years old (M = 20.61, SD = 2.57). Additionally, 169 were White, 40 were Asian or Asian American, 15 were Black or African American, 1 identified as American Indian or Alaska Native, 1 identified as Hawaiian or Other Pacific Islander, 7 identified as Hispanic/Latine, 15 identified as “something else,” and 2 did not respond. There were no attention checks employed in study 1.

A sensitivity analysis for a mixed-measures F-test using G*Power³⁷ with the following input parameters indicated that our sample was sufficiently powered to detect small to medium effect sizes (f² = .05): “as in SPSS” option, α (2-sided) = .05, power = .80, number of groups = 4, number of measurements = 2 (largest numerator df + 1), non-sphericity correction = 1.

Pain Assessment Task

Observers’ pain assessment bias was measured with a perception task that contained 40, 10-second videos of real people (ie, targets) varying in race and gender undergoing a standardized laboratory pain task (ie, the tourniquet procedure). The tourniquet procedure involves wearing an inflated blood pressure cuff around the upper arm while performing several handgrip exercises standardized to the participants’ handgrip strength and standardized in terms of timing. Targets made pain ratings every 30 seconds on a numeric rating scale from 0 (no pain at all) to 10 (the most pain imaginable). Targets could end the procedure at any time. The entire procedure was video recorded. Ten-second clips leading up to targets’ self-reported pain were randomly selected for inclusion in the Pain Assessment Task so there would be variability in pain experienced and expressed. (More details about the methodology can be found in a previously published article.)³⁸ Only 1 clip was selected from each target. Race, gender, and pain ratings of targets are in Table 1. Researchers interested in accessing the videos may contact the first author.

Table 1.

Study 1, 2, and 3 Target Race and Gender and Self-Reported Pain Rating

	TARGET RACE AND GENDER	TARGET SELF-REPORTED PAIN RATING BY RACE/GENDER M (SD)
Study 1 N = 40	5 men of color	6.30 (2.59)
	4 Asian men
	1 Hispanic man
	13 White men	4.08 (1.85)
	8 women of color	6.81 (1.96)
	5 Asian women
	2 Black women
	1 Hispanic/Latine woman
	14 White women	5.50 (2.35)
Study 2 N = 35	8 Men of color	6.38 (2.66)
	3 Asian men
	3 Black men
	2 Hispanic/Latine men
	9 White men	4.81 (2.56)
	8 women of color	8.00 (1.51)
	1 Asian woman
	3 Black women
	4 Hispanic/Latine women
	10 White women	6.55 (2.09)
Study 3 N = 40	10 men of color	5.50 (2.80)
	4 Asian men
	4 Black men
	2 Hispanic/Latine men
	10 White men	3.3 (2.11)
	10 women of color	6.10 (2.33)
	2 Asian women
	4 Black women
	4 Hispanic/Latine women
	10 White women	6.60 (1.71)

Abbreviation: SD, standard deviation.

Observers were asked to make the same pain rating on a 0 (no pain at all) to 10 (the most intense pain imaginable) scale as the original targets in the videos. We took the difference between observers’ ratings of each target’s pain and the target’s own pain rating to create a pain assessment bias score, which served as our dependent variable across all analyses. Negative scores thus reflect observers’ tendency to underestimate pain, while positive scores represent observers’ tendency to overestimate pain. Scores closer to 0 reflect no bias in pain assessment. Capturing pain assessment bias in this way translates to treatment outcomes as it is likely that targets who consistently have their pain underestimated are rarely receiving enough treatment or timely treatment, while those who have their pain overestimated may be receiving inappropriate amounts of treatment. After observers completed ratings of all 40 videotaped targets’ pain, they completed demographic information about their own age, race, ethnicity, and gender.

Analytic Plan

A 2 (target race: White vs people of color) × 2 (target gender: men vs women) × 2 (observer race: White vs people of color) × 2 (observer gender: men vs women) mixed-measures analysis of variance (ANOVA) was conducted to determine whether observers’ pain assessment bias was affected by targets’ race and gender, which were entered as within-subject variables, as well as observers’ own race and gender, which were entered as between-subjects variables. Least significant difference (LSD) corrections for multiple comparisons were employed for all pairwise comparisons. Following recommendations from Cumming and Finch,³⁹ we determined that any difference between a set of pairwise comparisons was significant at P < .05 if the point estimate for 1 pairwise comparison was not captured in the confidence interval for the other pairwise comparison. All data analyses were performed in SPSS version 25. Alpha was set at .05 for all statistical tests and effect sizes are reported using either partial eta-squared statistic derived from a multivariate analysis of variance using Pillai’s trace to describe the effect size of the difference between 2 estimated marginal means or Cohen’s d. We interpret partial eta-squared values of .14 to be large, .06 to be medium, and .01 to be small and Cohen’s d values of .80 to be large, .50 to be medium, and .20 to be small.⁴⁰

Results

Study 1

Results revealed a large main effect of target race (F [1, 243] = 1,949.50, P < .001, η_p² = .89), where targets of color’s pain (M = −2.45, (standard deviation) SD = 1.67, 95% (confidence interval) CI: −2.66, −2.24) were significantly underestimated in comparison to White targets’ pain (M = −.75, SD = 1.62, 95% CI: −.96, −.55). There was also a large main effect of target gender (F[1, 243] = 576.80, P < .001, η_p² = .70), where women targets’ pain (M = −2.14, SD = 1.67, 95% CI: −2.35, −1.93) was significantly underestimated in comparison to men targets’ pain (M = −1.07, SD = 1.63, 95% CI: −1.27, −.86). There were no significant main effects of observer race (F[1, 243] = .13, P = .717, η_p² = .00) or observer gender (F [1, 243] = .00, P = .976, η_p² = .00) on observers’ pain assessment bias.

As hypothesized, the main effects of target race and gender were qualified by a medium and significant target race by target gender interaction, F(1, 243) = 19.26, P < .001, η_p² = .07. Following recommendations, we determined that any difference between a set of pairwise comparisons was significant at P < .05 if the point estimate for 1 pairwise comparison was not captured in the confidence interval for the other pairwise comparison.³⁹ Examining the effect of target gender within target race, pairwise comparisons did reveal that the underestimation of women targets’ pain compared to men targets’ pain was significantly greater among targets of color (M_difference = −1.27, 95% CI: −1.41, −1.13; P < .001, η_p² = .57) than White targets (M_difference = −.87, 95% CI: −.98, −.76; P < .001, η_p² = .50; see Fig 1). The other pairwise comparisons of the effect of target race within target gender also revealed the underestimation of targets of color’s pain compared to White targets’ pain was significantly greater among women targets (M_difference = −1.90, 95% CI: −2.01, −1.80; P < .001, η_p² = .84) than men targets (M_difference = −1.50, 95% CI: −1.63, −1.37; P < .001, η_p² = .68) (Table 2).

Figure 1.

Boxplot of observers’ pain assessment bias across studies 1, 2, and 3 preintervention by target race and gender.

Table 2.

Study 1, 2, and 3 (Preintervention) Pain Assessment Bias Means, Standard Deviations, 95% Confidence Intervals, and Pairwise Comparison Results From 2 (Target Race) × 2 (Target Gender) × 2 (Observer Race) × 2 (Observer Gender) ANOVA

	Pain assessment bias by target race and gender
	White men	White women	Men of color	Women of color
Study #	M (SD) [95% Cl]	M (SD) [95% Cl]	M (SD) [95% Cl]	M (SD) [95% Cl]
Study 1	−.32 (1.59) [−.52, −.12]	−1.19 (1.77) [−1.41, −.97]	−1.82 (1.84) [−2.05, −1.59]	−3.09 (1.67) [−3.30, −2.88]
Study 2	−1.26 (1.47) [−1.53, −.98]	−1.57 (1.80) [−1.91, −1.22]	−3.67 (1.53) [−3.96, −3.38]	−4.64 (1.75) [−4.97, −4.31]
Study 3†	−.78 (1.30) [−.97, −.58]	−3.92 (1.34) [−4.13, −3.72]	−2.93 (1.21) [−3.11, −2.74]	−2.97 (1.26) [−3.12, −2.77]

	Pairwise comparisons
	White men versus White women		Men of color versus women of color		White men versus men of color		White women versus women of color		White men versus women of color		White women versus men of color
Study #	Mdifference (95% CI)	ηp2	Mdifference (95% CI)	ηp2	Mdifference (95% CI)	ηp2	Mdifference (95% CI)	ηp2	Mdifference (95% CI)	ηp2	Mdifference (95% CI)	ηp2
Study 1	.87* (.76, .98)	.50	1.27* (1.13, 1.41)	.57	1.50* (1.37, 1.63)	.68	1.90* (1.80, 2.01)	.84	2.77* (2.66, 2.89)	.90	.63* (.51, .75)	.32
Study 2	.32 (−.11, .74)	.02	.97* (.54, 1.40)	.15	2.42* (2.19, 2.64)	.81	3.07* (2.87, 3.26)	.90	3.38* (2.94, 3.83)	.69	2.10* (1.68, 2.51)	.49
Study 3	3.20* (2.96, 3.44)	.81	.11 (−.17, .39)	.00	2.12* (1.96, 2.28)	.80	−.97* (−1.19, −.75)	.31	2.23* (1.96, 2.50)	.62	−1.08* (−1.34, −.82)	.29

Abbreviations: CI, confidence interval; SD, standard deviation.

^*P < .001.

^†Descriptive statistics do not include observers randomly assigned to the practice and feedback condition.

Although not hypothesized, we also observed a small, but significant interaction between target gender and observer gender (F[1, 243] = 4.28, P = .040, η_p² = .02), where men observers were significantly more likely to underestimate women targets’ pain than men targets’ pain (M_difference = −1.16, 95% CI: −1.32, −1.01; P < .001, η_p² = .68), compared to women observers (M_difference = −.98, 95% CI: −1.06, −.89; P < .001, η_p² = .48). There were no other significant (P < .05) interactions.

Discussion

Study 1

In line with hypotheses, observers underestimated women’s pain and people of color’s pain significantly more than men’s pain and White people’s pain. However, these main effects do not illustrate biases associated with multiple marginalized identities. That is, women of color’s pain was underestimated on average by more than 3 points on the 0 to 10 pain scale, whereas White women’s pain was underestimated on average by 1 point. Men of color’s pain was underestimated by nearly 2 points, while White men had their pain underestimated by < 1 point. Comparing this nearly 0 value for White men to the underestimation of women of color’s pain (M = −3.06) likely explains much of the disparate pain treatment and care in medical settings as observers in study 1 were attending a College of Pharmacy and Health Science where most students were training to become clinicians. Thus, these biases are held early in education, likely predate clinical training, and likely persist throughout clinical care and practice.

Method

Study 2

The purpose of study 2 was to replicate study 1’s findings and test the pervasiveness of these pain assessment biases with a different sample of undergraduate student observers. In addition, we ensured the target stimuli were equally balanced by race and gender.

Participants (Observers)

Observers were N = 110 (43 men, 66 women, 1 gender-diverse individual) undergraduate students who were recruited through convenience sampling from Northeastern University for a 60-minute in person study in which they received partial course credit for participation. This study was approved by Northeastern University’s Institutional Review Board and informed consent was obtained from each subject.

Observers were recruited from the Department of Psychology participant pool in which students enrolled in Introductory Psychology (and other select courses) were asked to participate in several research studies during the semester. Observers ranged in age from 18 to 23 years old (M = 19.42, SD = 1.36). Sixty-eight observers were White, 20 were Asian or Asian American, 6 were Black or African American, 2 were American Indian or Alaska Native, and 9 were Hispanic or Latine. A sensitivity analysis for a mixed-measures F-test using G*Power indicated that our sample was sufficiently powered to detect small to medium effect sizes (f² = .11).³⁷ There were no attention checks employed in study 2.

Pain Assessment Task

Observers’ pain assessment bias was again measured with a similar pain assessment task as in study 1. The new pain assessment task contained 35, 10-second videos of targets undergoing a standardized laboratory pain task (ie, the tourniquet procedure; for more details about the methodology, see Pain Assessment Task and a previously published article).³⁸ Race, gender, and pain ratings of targets are reported in Table 1. Observers were asked to make the same pain rating on a 0 (no pain at all) to 10 (the most intense pain imaginable) scale as the original targets in the videotapes had done so we could directly compare the assessment of pain. As before, we calculated observers’ pain assessment bias score by taking the difference between observer’s ratings of each target’s pain and targets’ ratings of their own pain to serve as our dependent variable. After observers completed ratings of the videotaped targets’ pain, they completed demographic information about their own age, race, ethnicity, and gender.

Analytic Plan

As in study 1, we ran a 2 (target race: White vs people of color) × 2 (target gender: men vs women) × 2 (observer race: White vs people of color) × 2 (observer gender: men vs women) mixed-measures ANOVA to determine whether observers’ pain assessment bias was affected by target gender and target race, which were entered as within-subject variables, and observer gender and observer race, which were entered as between-subject variables. LSD corrections for multiple comparisons were employed for all pairwise comparisons. Following recommendations from Cumming and Finch,³⁹ we determined that any difference between a set of pairwise comparisons was significant at P < .05 if the point estimate for 1 pairwise comparison was not captured in the confidence interval for the other pairwise comparison. All data analyses were performed in SPSS version 25. Alpha was set at .05 for all statistical tests and effect sizes are reported using either partial eta-squared statistic derived from a multivariate analysis of variance using Pillai’s trace to describe the effect size of the difference between 2 estimated marginal means or Cohen’s d. We interpret partial eta-squared values of .14 to be large, .06 to be medium, and .01 to be small and Cohen’s d values of .80 to be large, .50 to be medium, and .20 to be small.⁴⁰

Results

Study 2

Consistent with study 1, there was a large significant effect of target race (F[1, 105] = 1,196.43, P < .001, η_p² = .92), where observers underestimated targets of color’s pain (M = −4.15, SD = 1.22, 95% CI: −4.38, −3.93) significantly more than White targets’ pain (M = −1.41, SD = 1.20, 95% CI: −1.64, −1.18). Additionally, there was a medium and significant main effect of target gender (F[1, 105] = 10.12, P = .002, η_p² = .09), where observers underestimated women targets’ pain (M = −3.10, SD = 1.71, 95% CI: −3.43, −2.78) significantly more than men targets’ pain (M = −2.46, SD = 1.37, 95% CI: −2.72, −2.20). We also observed a small and significant main effect of observer gender on pain assessment bias scores (F[1, 105] = 4.22, P = .042, η_p² = .04), where men observers underestimated targets’ pain (M = −3.01, SD = 1.08, 95% CI: −3.33, −2.68) to a greater degree than women observers (M = −2.56, SD = 1.14, 95% CI: −2.84, −2.29). However, there was not a significant main effect of observer race on pain assessment bias nor were there any significant interactions with observer characteristics.

As in study 1, the target main effects were qualified by a large and significant target race by target gender interaction, F(1, 105) = 20.11, P < .001, η_p² = .16. Pairwise comparisons of the effect of target gender within target race revealed a similar pattern in study 1. Specifically, following the same recommendations as study 1,³⁹ evidence supported that the underestimation of women’s pain compared to men’s pain was significantly greater among targets of color (M_difference = −.97, 95% CI: −1.40, −.54; P < .001, η_p² = .16) than White targets (M_difference = −.32, 95% CI: −.74, −.11; P = .140, η_p² = .02; see Fig 1). The other pairwise comparisons of the effect of target race within target gender also revealed that the underestimation of pain among targets of color compared to White targets was significantly greater among women targets (M_difference = −3.07, 95% CI: −3.26, −2.87; P < .001, η_p² = .90) compared to men targets (M_difference = −2.42, 95% CI: −2.64, −2.19; P < .001, η_p² = .81) (Table 2).

Discussion

Study 2

In 2 separate studies, we demonstrated that undergraduate students hold biases in pain assessment such that women have their pain underestimated significantly more than men, people of color have their pain underestimated significantly more than White people, and women of color’s pain is underestimated in a magnitude greater than what would be expected by considering either of their marginalized identities independently. In fact, women of color’s pain was underestimated by about 3 to 5 points on a 0 to 10 pain scale. This work is consistent with past research on pain assessment biases and disparities in care as well as intersectionality theory that posits that woman of color experience unique biases due to their multiple and intersecting marginalized identities in society and health care. While the main effects demonstrate gender and race biases, it is the interaction of target race and gender that allows for a more nuanced perspective on bias.

In addition, the only observer factor predicting pain assessment bias in study 2 was observer gender, such that men underestimated pain significantly more than women regardless of target race or gender. This effect was small in magnitude and only occurred in 1 study, so we caution any overinterpretation of this observer gender effect and urge future researchers to continue examining observer factors including race and gender as we did not purposefully sample for these types of analyses.

Overview

Study 3

Both study 1 and study 2 demonstrated that observers, regardless of their own race and gender, make biased judgments of others’ pain, especially for women, people of color, and, more specifically, women of color. This is consistent with the literature in medical settings that highlights women and people of color receive inadequate pain treatment and care.⁴¹ The purpose of study 3, therefore, was to test 3 different interventions aimed at reducing pain assessment bias that disproportionately impacts women, people of color, and, in particular, women of color.

Bias reduction trainings that have been developed generally fall within 3 categories. The first is Allport’s contact hypothesis⁴² which rests on the assumption that providing opportunities for positive and cooperative contact between members of groups previously hostile to one another reduces prejudice. The second is increasing one’s awareness of biases. Awareness training focuses on getting observers to be more aware of general cultural assumptions, values, and biases.^43,44 Finally, behavioral training, also known as skill-building training, educates observers on monitoring one’s own actions and appropriate responses in assessing and responding to intergroup interactions.³² In the current study, we focus on testing the latter 2 types of trainings, awareness training, and behavioral (skill building) training as Allport’s contact hypothesis⁴² has resulted in mixed evidence in terms of bias reduction.⁴⁵

A meta-analysis across 40 years of research on bias reduction trainings tested, across almost 30,000 participants, the effectiveness of awareness alone, behavioral training alone, and awareness and behavioral training combined.³² The type of training was a significant moderator such that the effect size for behavioral training alone (g = .46) was larger than awareness alone (g = .31) and similar in magnitude to trainings that combined awareness and behavioral training (g = .46). However, researchers rarely tested behavioral trainings alone (k = 11 compared to k = 121 for awareness alone and k = 118 for combined awareness and behavioral trainings). It is important to test the effects of behavioral training alone to understand the effects of individual components in reducing bias before combining many factors into a training.

Previous research has demonstrated that one form of behavioral training, practice and immediate feedback, increases accuracy in assessing emotion, thoughts, and feelings in others, compared to practice alone or raising awareness.^46,47 However, this type of behavioral training has never been tested to reduce biased judgments in pain assessment. We believe this type of behavioral training may be particularly effective at reducing biased pain assessments because it gives observers an opportunity to practice and get immediate feedback in a private setting where they can unlearn gender and race stereotypes with real nonverbal expressions of pain. For example, they could unlearn stereotypes that Black people feel less pain than White people^10,48–51 or that women exaggerate their pain expression compared to men²⁵ by getting feedback about their assessments in real time.

In study 3, we tested the effectiveness of behavioral training, awareness of self-biases, and awareness of others’ (ie, societal) biases against a control condition that did not receive any training.

Method

Study 3

Participants (Observers).

Observers were N = 232 (125 men, 104 women, 2 gender-diverse individuals, 1 person did not respond) undergraduate students who were recruited using a convenience sample from the University of Maine’s Department of Psychology participant pool in which students enrolled in introductory psychology (and other select courses) are asked to participate in several research studies during the semester. Those who participated in this 60-minute in person study received partial course credit for participation. This study was approved by the University of Maine’s Institutional Review Board and informed consent was obtained from each subject.

Three observers were excluded for failing an attention check question, leaving a final sample of N = 229. Observers ranged in age from 18 to 26 years old (M = 18.77, SD = 1.20) and were predominantly White (N = 201), 4 were Asian or Asian American, 6 were Black or African American, 2 were American Indian or Alaska Native, 3 responded with “other,” and 13 observers did not respond. Additionally, 11 observers identified their ethnicity as “Hispanic or Latine.” A sensitivity analysis for a mixed-measures F-test using G*Power³⁷ indicated that our sample was sufficiently powered to detect small to medium effect sizes (f² = .08) in our mixed-measures ANOVA.

Pain Assessment Task.

Observers watched 40, 10-second videos of targets undergoing a standardized laboratory pain task (ie, the tourniquet procedure; for more details about the methodology, see Pain Assessment Task and in a previously published article).³⁸ The pain task contained an equal number of unique men and women who varied in race (ie, White vs people of color) experiencing pain. Race, gender, and pain ratings of targets can be found in Table 1. Observers were asked to make the same pain rating on a 0 (no pain at all) to 10 (the most intense pain imaginable) scale as the original targets in the videotapes had done, so we could directly compare the assessment of pain. The pain test was split into 2 sets for this study. Observers were trained on one half of the pain tests (20 targets) and tested on the other half of the pain test (20 targets). The pain video sets that were used for the training versus the test trials were counterbalanced. After observers completed ratings of the videotaped targets’ pain, they completed demographic information about their own age, race, ethnicity, and gender.

Description of Training Conditions.

Observers were randomly assigned to 1 of 4 training conditions: 1) practice alone condition (control; N = 58), 2) raising awareness about societal bias (N = 53), 3) raising awareness about one’s self-bias (N = 58), and 4) behavioral training: practice and immediate feedback (N = 60). In all conditions, observers were allowed to watch each clip only once.Practice alone condition (control)

In this condition, no feedback on performance was provided during or after taking the practice portion.Raising awareness of societal bias

In this condition, after watching and making pain ratings of each clip in the first half of the task, observers read an article about general societal race and gender biases in pain assessment and took a short multiple-choice quiz about the article to ensure they read and understood the article (Supplementary Appendix A). After reading the article and completing the multiple-choice quiz, observers were told the following: “Healthcare providers aren’t the only ones who have a tendency to underestimate women and people of color’s pain; research suggests that lots of people have these tendencies too. In a few moments, we will have you complete a second pain assessment task, keep these tendencies in mind as you perceive pain to try to avoid relying on them. Your goal is to perceive pain as accurately as possible and free from these biases.” Observers then completed their post-intervention pain assessment test without any further information.Raising awareness of self-bias

In this condition, after watching and making pain ratings of each clip in the first half of the task, observers were given feedback about their own tendencies to over or underestimate pain across the clips, specifically when assessing women’s pain compared to men’s pain and people of color’s pain compared to White people’s pain. That is, after viewing and making ratings about the first set of videos, observers were given their mean pain assessment bias scores by race and gender. These bias scores were computed by taking the mean difference score of each observer’s ratings of pain with each respective target’s actual self-reported pain rating across targets who shared the same gender or race so that each observer received an overall bias score for men, women, people of color, and White people’s pain. In this condition, the observers were given the following information: “Sometimes we have tendencies to over or underestimate pain based on physical or social characteristics of others. For example, women’s pain tends to go underestimated compared to men’s pain. Research suggests that the best way to reduce these biases is to be made aware of them. In a minute, I’m going to share with you some information about your own tendencies to over or underestimate others’ pain, specifically comparing judgments of women’s pain to men’s pain, people of color’s pain to White people’s pain. On the paper in front of you are your tendencies in pain assessment. Please read these carefully, and if you have any questions about what they mean, please ask. In a few moments, we will have you complete a second pain assessment task, keep these tendencies in mind as you perceive pain to try to avoid relying on them. Your goal is to perceive pain as accurately as possible and free from these biases.” Observers then completed their post-intervention pain assessment task without any further information.Behavioral training: practice with immediate feedback

In this condition, during the first half of the task, after observers entered their answer for each clip in the practice portion (ie, their assessed pain intensity for each target), the correct answer was displayed on the screen (ie, what each target actually reported at the time in terms of their pain intensity). Observers then completed their test version without any further information or feedback.

Analytic Plan.

To replicate baseline pain assessment bias from studies 1 and 2, we first conducted the same 2 (target race: White vs people of color) × 2 (target gender: men vs women) × 2 (observer race: White vs people of color) × 2 (observer gender: men vs women) mixed-measures ANOVA on pain assessment bias from the pre-intervention. Only observers assigned to the practice and feedback training condition were excluded from this analysis as these observers were exposed to an intervention in their pre-intervention, and therefore their scores are not representative of a true baseline. However, all other conditions received their intervention after their preintervention and before their post-intervention pain assessment test and were thus included (N = 169).

A mini meta-analysis⁵² was performed to examine the magnitude of pain assessment biases combining data from studies 1, 2, and 3 for each target group of interest (White men, White women, men of color, and women of color). To determine the magnitude of pain assessment bias for each target group across these 3 studies, we used the effect size (ie, Cohen’s d) from a series of one-sample t-tests against 0. This approach captured the magnitude of under or overestimation of pain as a larger negative effect signifies more underestimation of pain and a larger positive effect signifies more overestimation of pain while an effect closer to 0 signifies less bias and a more accurate assessment of pain.

We then conducted a 2 (target race: White vs people of color) × 2 (target gender: men vs women) × 4 (training condition: behavioral training, awareness of self-biases, awareness of societal biases, control) mixed-measures ANOVA to investigate the impact of training condition on post-intervention pain assessment biases. Observer gender and observer race were not included in this final model for 3 reasons: 1) We did not have reason to believe our experimental interventions would be more or less effective for specific observer groups, 2) observer race and gender were not consistently related to pain assessment bias across all 3 studies, and (3) inclusion of observer race and gender would have resulted in certain cells having very low sample sizes (eg, n’s < 5). LSD corrections for multiple comparisons were employed for all pairwise comparisons. Following recommendations from Cumming and Finch,³⁹ we determined that any difference between a set of pairwise comparisons was significant at P < .05 if the point estimate for 1 pairwise comparison was not captured in the confidence interval for the other pairwise comparison.

Finally, we conducted a series of one-sample t-tests where post-intervention pain assessment bias scores across the various target groups and intervention conditions were compared against 0. The larger and more negative the Cohen’s d, the more underestimation of pain for specific target groups, indicating greater pain assessment bias.

All data analyses were performed in SPSS version 25. Alpha was set at .05 for all statistical tests and effect sizes are reported using either partial eta-squared statistic derived from a multivariate analysis of variance using Pillai’s trace to describe the effect size of the difference between 2 estimated marginal means or Cohen’s d. We interpret partial eta-squared values of .14 to be large, .06 to be medium, and .01 to be small and Cohen’s d values of .80 to be large, .50 to be medium, and .20 to be small.⁴⁰

Results

Study 3

Bias in Pain Assessment (Pre-Intervention Data Only)

Once again, we observed a large and significant main effect of target race (F[1, 165] = 60.2, P < .001, η_p² = .27), where observers underestimated targets of color’s pain (M = −2.82, SD = 1.70, 95% CI: −3.08, −2.56) significantly more than White targets’ pain (M = −2.25, SD = 1.91, 95% CI: −2.54, −1.96). Additionally, we observed a large and significant main effect of target gender (F [1, 165] = 218.06, P < .001, η_p² = .57), where observers underestimated women targets’ pain (M = −3.36, SD = 1.89, 95% CI: −3.65, −3.08) significantly more than men targets’ pain (M = −1.71, SD = 1.89, 95% CI: −1.99, −1.42).

As before, we also observed a large and significant target race by target gender interaction, F (1, 165) = 542.33, P < .001, η_p² = .77. As shown in Fig 1 (study 3 preintervention), the pairwise comparisons of the effect of target gender within target race revealed that observers’ underestimation of women target’s pain compared to men target’s pain was significantly greater among White targets (M_difference = −3.20, 95% CI: −3.44, −2.96; P < .001, η_p² = .81) than targets of color (M_difference = −.11, 95% CI: −.39, .17; P = .428, η_p² = .01). The other pairwise comparisons of the effect of target race within target gender similarly revealed that observers’ underestimation of targets of color’s pain compared to White targets’ pain was significantly greater among men targets (M_difference = −2.12, 95% CI: −2.28, −1.96; P < .001, η_p² = .80) than women targets (M_difference = .97, 95% CI: .75, 1.19; P < .001, η_p² = .31) (Table 2).

Although not hypothesized, we also observed a small 3-way interaction between target race, observer race, and observer gender, F(1, 165) = 6.62, P = .011, η_p² = .04. This interaction appeared to be driven by women of color observers showing similar pain assessment bias for White targets and people of color (M_difference = −.23, 95% CI: −.65, .20; P = .229, η_p² = .01), whereas all other observers (ie, men of color, White men, and White women) underestimated targets of color’s pain significantly more than White targets’ pain (Men of color observers: M_difference = −.88, 95% CI: −1.23, −.53; P < .001, η_p² = .13; White men observers: M_difference = −.54, 95% CI: −.68, −.41; P < .001, η_p² = .28; White women observers: M_difference = −.65, 95% CI: −.80, −.51; P < .001, η_p² = .32). There were no other significant (P < .05) interactions in the model.

Mini Meta-analysis of Pain Assessment Bias (Pre-Intervention Only).

The forest plot in Fig 2 displays fixed effects model results given the shared methodology across all 3 studies, and Table 3 presents both fixed and random effects, though results did not meaningfully differ when using a random effects approach. These results provide further support that women of color suffered from the most underestimation of pain compared to all other groups with a large effect size. The only target group that did not show a large effect in terms of the underestimation of pain was White men. White men targets still had their pain underestimated but the effect size was medium in size.

Figure 2.

Forest plot of mini meta-analysis results of observers’ pain assessment bias by target race and gender. Note. ES is effect size Cohen’s d.

Table 3.

Study 1, 2, and 3 (Preintervention Only) Mini Meta-analysis of a Series of One-Sample t-Tests Comparing Pain Assessment Bias Scores Across Different Target Groups to Zero (No Bias)

TARGET GROUP	RANDOM EFFECTS MODEL			FIXED EFFECTS MODEL
	COHEN'S D (SE)	95% CI	COMBINED Z	COHEN'S D (SE)	95% CI	COMBINED Z
White men	−.57 (.16)	[−.89, −.26]	−3.58*	−.48 (.04)	[−.56, −.39]	−10.86*
White women	−1.55 (.69)	[−2.91, −.20]	−2.25**	−1.07 (.06)	[−1.18, −.96]	−19.06*
Men of color	−2.07 (.39)	[−2.83, −1.30]	−5.29*	−1.73 (.07)	[−1.87, −1.59]	−24.25*
Women of color	−2.54 (.13)	[−2.80, −2.29]	−19.77*	−2.51 (.09)	[−2.68, −2.34]	−29.32*

Abbreviations: CI, confidence interval; SE, standard error.

^*P < .001.

^**P < .05.

Pain Assessment Interventions.

Given the clear underestimation of pain, especially for women and people of color, the primary goal of study 3 was to investigate the effectiveness of different types of interventions for decreasing pain assessment biases on post-intervention scores only. From the 2 × 2 × 4 ANOVA described in the analytic plan for study 3, we observed a large and significant main effect of target gender, where women’s pain (M = −2.32, SD = 1.24, 95% CI: −2.49, −2.16) was significantly underestimated in comparison to men’s pain (M = −.90, SD = 1.26, 95% CI: −1.06, −.73; F[1, 225] = 470.33, P < .001, η_p² = .68). Additionally, we observed a large main effect of target race, where targets of color’s pain were significantly underestimated (M = −1.81, SD = 1.10, 95% CI: −1.96, −1.67) in comparison to White targets’ pain (M = −1.41, SD = 1.30; F[1, 225] = 60.75, P < .001, η_p² = .21).

Again, these large main effects were qualified by a large and significant target gender by target race interaction (F[1, 225] = 1,129.30, P < .001, η_p² = .83). Similar to our preintervention results, pairwise comparisons of the effect of target gender within target race revealed that observers’ underestimation of women target’s pain compared to men target’s pain was significantly greater among White targets (M_difference = −3.11, 95% CI: −3.26, −2.97; P < .001, η_p² = .89) than targets of color (M_difference = .26, 95% CI: .82, .44; P = .004, η_p² = .04). The other pairwise comparisons of the effect of target race within target gender similarly revealed the opposite pattern that observers’ underestimation of targets of color’s pain compared to White targets’ pain was significantly greater among men targets (M_difference = −2.09, 95% CI: −2.20, −1.97; P < .001, η_p² = .85) than women targets (M_difference = 1.28, 95% CI: 1.12, 1.45; P < .001, η_p² = .51) (Table 2).

Of note here is how close, across training conditions, pain assessment bias for White men’s pain came to 0, signifying no bias (Fig 3). No other group had their pain assessed without bias. However, by collapsing across training conditions, as past work has done, we conflate potentially important training components that may be differentially impacting bias scores. Therefore, the effects we were most interested in on the post-intervention pain assessment bias scores were the between-subjects effect of the condition.

Figure 3.

Post-intervention pain assessment bias scores across various intervention conditions by target race and gender. Note. Error bars reflect standard error of the mean.

We found evidence of a significant and large main effect of condition, F(3, 225) = 31.68, P < .001, η_p² = .30, suggesting that the intervention conditions had differential effects on post-intervention pain assessment bias scores. Pairwise comparisons indicated that individuals who received the behavioral skills practice and feedback condition (M = −.40, SD = 1.15, 95% CI: −.69, −.11) were significantly less biased in assessing pain than individuals assigned to any other condition (raising awareness of self-biases M = −1.68, SD = 1.16, 95% CI: −1.97, −1.38; raising awareness of societal biases M = −2.11, SD = 1.16, 95% CI: −2.42, −1.80; control M = −2.25, SD = 1.16, 95% CI: −2.54, −1.95; P’s < .001, 1.11 < d < 1.60; 95% CI’s for d ranged from .72 to 2.01). Additionally, individuals who received the raising awareness of self-bias intervention were significantly less biased in assessing pain compared to individuals in the control condition (P = .008, Cohen’s d = .49, 95% CI: .12, .86). However, there was no significant difference between those who received the raising awareness in societal bias intervention and the control condition (P = .815, Cohen’s d = .04, 95% CI: −.25, .49). No two or three-way interactions were significant. Together, these results suggest that, consistent with past literature and our hypotheses, behavioral skills building of practice and immediate feedback were the most effective for reducing pain assessment biases.

While the above results revealed that the behavioral skills practice and immediate feedback intervention condition significantly decreased pain assessment bias scores compared to the other intervention conditions, we finally tested whether observers’ pain assessment bias scores were closer to 0 (ie, no bias) after being randomly assigned to complete this intervention compared to those randomly assigned to the control condition and explored any differences for each target group (ie, White men, White women, men of color, women of color). Results indicated that the Cohen’s ds all meaningfully decreased from the control condition to post-behavioral skills practice and immediate feedback for White women targets (control = −2.39, 95% CI: −2.89, −1.88; practice and immediate feedback = −1.32, 95% CI: −1.67, −.97), men of color targets (control = −2.16, 95% CI: −2.63, −1.68; practice and immediate feedback = −.69, 95% CI: −.95, −.40), and women of color targets (control = −1.67, 95% CI: −2.06, −1.27; practice and immediate feedback = −.36, 95% CI: −.61, −.09). Specifically, Cohen’s ds decreased by 1.28 points on average across these 3 groups. Only 1 group (ie, White men targets) increased in their Cohen’s d from the control condition (d = −.34; 95% CI: −.60, −.07) to post-behavioral skills practice and immediate feedback (d = 1.15; 95% CI: .82, 1.48). This increase reflected that pain assessment bias for White men was relatively close to 0 without any kind of intervention, and the behavioral skills practice and immediate feedback condition caused an overestimation of pain for this target group.

Discussion

Study 3.

In study 3, we replicated effects from studies 1 and 2, where women and people of color had their pain underestimated more than men and White targets. For the first time, we also tested the impact of several interventions individually (rather than grouped together) and examined their impact on pain assessment bias using a novel set of genuine and dynamic expressions of pain from target stimuli who varied in race and gender. Consistent with the literature, we found that behavioral skills building (ie, practice and immediate feedback) was the most effective at reducing observers’ pain assessment bias. However, while this intervention condition was effective at reducing pain assessment biases for marginalized targets, observers in this intervention condition began to overestimate the pain of White men. Therefore, we believe behavioral skills building to be a very promising new direction for reducing pain assessment bias but should only be applied to the targets (or patients) whose pain typically goes underestimated and undertreated, women and people of color.

General Discussion

Women’s pain and people of color’s pain often go undertreated and underestimated in medical care.^{1–14,16–19,23–29} The current work contributes to documenting the intersectional biases in pain assessment for people who hold multiple marginalized identities. In studies 1 and 2, observers underestimated the pain of women targets significantly more than that of men targets and underestimated the pain of people of color significantly more than that of White people. Importantly, targets’ race and gender interacted such that observers disproportionately underestimated women of color’s pain compared to any other group (by 3–5 points on a 0–10 pain scale). This pervasive intersectional bias in pain assessment for women of color adds to a growing literature documenting the disproportionate disparities faced by patients with multiple marginalized identities and helps contextualize their unique experiences in the health care system.³¹

While documenting disparities in pain assessment is important, the goal of doing so is to better understand and intervene on pain treatment disparities. Past work suggests that clinicians and laypeople who are more accurate assessors of others’ emotions and thoughts and feelings have better interpersonal and professional outcomes, such as providing more appropriate patient care and leaving patients feeling interpersonally respected and understood.^53–61 Thus, we believe that intervening on the level of observers’ pain assessment may be one path to help alleviate the undertreatment of pain and provide better quality health care for people of color and women.

Study 3 tested the effectiveness of 3 different interventions aimed at reducing observers’ pain assessment bias. We found behavioral skills building (ie, practice and immediate feedback) was the most effective intervention condition compared to raising awareness of self-biases, raising awareness of societal biases, or a control condition that received no information or feedback. Specifically, observers who were randomly assigned to the behavioral skills training had their pain assessment bias reduced by almost 2 full-scale points (M = 1.84) on a 0 to 10 pain scale for targets who were women or people of color. In practice, altering a provider’s pain assessment 2 points closer to a patient’s experience of pain could be the difference between a patient receiving the care they need or receiving inadequate pain management. These results present a promising direction to continue exploring to improve treatment outcomes for women and people of color.

Limitations and Future Directions

While the magnitude of the effectiveness of the behavioral skills intervention at reducing pain assessment bias specifically for targets with marginalized identities was large, a few questions remain about the generalizability of this type of behavioral skills training. First, in study 3, observers received the training in one laboratory session and were tested on their biases in the same 60-minute session. Thus, we still do not yet know how long the effects of this type of training might last and what types of periodic prompts or reminders might be needed to continue to reduce biases. Second, future research must address what changing biases in pain assessment causes in terms of patient care and communication. It is assumed that patients paired with providers who are less biased in their pain assessment likely get the treatment they need and thus should be more satisfied with their care;⁶¹ however, this is an open question that deserves future attention.

There are also several characteristics of the current study in terms of the sample characteristics of the targets used in the videos and the observers for which we garnered information about pain assessment biases that may limit the generalizability of our conclusions. While we included videos rather than images or vignettes of real people who varied in race and gender undergoing real pain, targets in the present work were all college-aged and otherwise healthy adults. Pain assessment biases likely differ for older adults compared to younger adults.⁶² Additionally, people who experience pain regularly or those who have their pain underestimated may be more likely to self-monitor their expressions⁶³ or underreport pain⁶⁴ thereby impacting pain assessment in clinical settings. Thus, future work should include age as an important moderator and intersectional variables of race and gender as well as diverse patients experiencing pain in clinical settings to fully understand biases. Across both targets and observers, we were limited in terms of understanding gender in the binary. Future work should consider including gender-diverse samples of targets and observers to understand the role of gender identity and intersection with other marginalized identities in terms of pain assessment biases and care.

Our decision to group people of different races together due to target sample size limitations and compare them to White people represents another limitation. These findings cannot speak to the additional burden that specific marginalized groups face.⁶⁵ There are likely important stereotypes and biases specific to different racial identities not captured in this study^66,67 and important heterogeneity within racial identities.⁶⁸ Therefore, future work should include enough targets to fully understand pain assessment biases for specific racial identities. Additionally, comparing pain assessment bias for White versus racialized targets is helpful in determining differences in bias, but this work is not able to establish why and how this bias emerges.⁶⁹ The current work also centers Whiteness by comparing White versus racialized groups. Future work should center marginalized people’s experiences in pain care through focus groups, interviews, and surveys designed for and with community advisory members with lived pain experience.

Finally, while this work highlighted the importance of considering intersectionality when examining biases in pain, we must examine the multiple layers of injustice and systems of oppression, beyond the intrapersonal and interpersonal, if we are to enact change in health care and at a societal level, as these factors are deeply rooted in historical and structural aspects of racism and sexism. Through interdisciplinary research and multi-level interventions that target injustice in policies, structures, institutions, and cultures, we can start to decrease biases and disparities in pain care.¹¹

Conclusions

Using real expressions of pain, the current work documented the striking underestimation of pain for women and people of color across observers and highlighted how targets with multiple marginalized identities (ie, women of color) had their pain underestimated by 3 to 5 points on a 10-point pain scale. To help combat these pain assessment biases, we tested a novel and promising intervention to reduce biases in pain assessment with implications for reducing pain treatment disparities in medical care. Together, these findings contribute to our understanding of how people with intersectional, marginalized identities have their pain underestimated and communicate the importance of reducing these biases to properly assess and treat women and people of color’s pain.