Racial/ethnic concordance between patients and researchers as a predictor of study attrition

Introduction

The differential attrition of various racial/ethnic groups in longitudinal research is a major roadblock to minimizing health disparities. If the individuals who complete longitudinal studies, particularly randomized clinical trials, are not representative of those experiencing poorer health outcomes, the validity of the study’s conclusions is threatened along with its generalizability and translational capacity (Robinson et al., 2007). While attention has been paid in the literature to issues of recruitment of racial/ethnic minority research participants, especially for clinical trials, research into the retention of these groups has been less studied, yet it remains an important concern (Yancey et al., 2006).

Despite continuing efforts, rates of participation into research do not mirror U.S. demographics, with research samples lacking in their representativeness of racial and ethnic minority populations (Winter et al., 2018). Almost 30 years after the National Institute of Health’s (NIH) Revitalization Act of 1993 (US Congress, 1993), which required NIH-funded clinical trials to include racial/ethnic minorities as participants and assess outcomes by race/ethnicity, this is still not the norm (Chen et al., 2014). A systematic review of NIH-funded RCTs found that only 13.4% analyzed or reported outcomes by race or ethnicity. Further, there had been no significant changes in inclusion, analyses or reporting by race or ethnicity in the period of 2009 to 2015 (Geller et al., 2018). In the specific case of attrition, differential rates have been found for racial/ethnic minorities, whose attrition rates are 7 to 12% higher than those of non-Hispanic Whites in various fields of study, including rheumatoid arthritis (Krishnan et al., 2004), psychiatric disorders (Lamers et al., 2012), and genomic medicine (Moore et al., 2017).

Although studies suggest variability in attrition rates among racial/ethnic minorities, the mechanisms underlying these differences are not yet clear. Prior studies have suggested that variables such as age, education and mental health status may at least partially explain attrition rates, as attrition has been associated with lower educational attainment (Krishnan et al., 2004) and older age (Brilleman et al., 2010), and is more frequent among participants endorsing symptoms of depression and anxiety (Lamers et al., 2012). Although many of these participant characteristics have direct effects on attrition, there are also synergistic effects with other variables such as race/ethnicity identification. For example, in a longitudinal study of cardiovascular disease prevention, the attrition rate for Black participants was twice that of White participants, and almost four times greater for Black participants with probable depression (Bambs et al., 2013).

Research Staff Demographics and Study Engagement

There is some evidence that the lack of diverse research staff can be a barrier to research participation and retention of a diverse sample. African American women participating in focus groups on recruitment and retention barriers expressed feeling discouraged when approached by White research staff (Smith et al., 2007). African American men taking part in a qualitative study exploring Seventh Day Adventist church members’ attitudes towards participating in research reported being excited by the presence of a “mixed” study team and the researchers’ commitment to involve Black staff members in all phases of the study (Herring et al., 2004). Similarly, in discussing the influence of the racial/ethnic composition of the research team in their participation in clinical trials of smoking cessation, Native Americans voiced their difficulty in sharing or speaking up in the presence of non-Native individuals (Fu et al., 2014). Although some studies have used “ethnic matching” between research staff and the community (Ejiogu et al., 2011, McSweeney et al., 2009) or research staff and the individual participants (Areán et al., 2003) as a strategy to improve recruitment and prevent attrition, whether this approach actually led to lower rates of attrition has not been systematically explored.

Matching patients with study recruiters based on racial/ethnic identity has been proposed as a strategy to minimize the attrition of racial/ethnic minority participants. This strategy has been suggested in studies across a variety of clinical conditions, including Alzheimer’s disease (Gilmore-Bykovskyi et al., 2019), genetic and genomic research (Johnson et al., 2011), HIV (Falcon et al., 2011), diabetes (Hu et al., 2012) and mental health (Areán et al., 2003). However, although these studies have reported on their efforts and success using these methods of matching research staff to research participants, they have not systematically compared their effectiveness; for example, comparing recruitment or retention rates of race concordant vs. discordant research staff.

Because of the limited research on the effect of racial/ethnic concordance on study attrition, we draw on the literature on the effects of racial/ethnic concordance within healthcare provider-patient relationships. This has been widely studied and has been associated with better adherence to treatment (Penner et al., 2013), greater satisfaction with care, and greater shared decision-making (Cooper and Powe, 2004). For example, in a study of mental health intake visits (Alegria et al., 2013), Latino patients in ethnically concordant relationships were more verbally dominant and engaged in more information seeking and disclosure of psychosocial and lifestyle barriers. In an urban sample of individuals with HIV, Latinos in ethnically concordant relationships with their physicians reported lower levels of mistrust in the healthcare system than those in discordant relationships (Sohler et al., 2007). Similar findings have been described in studies of Black and African American patients, who report lower trust in the health care system (Sohler et al., 2007) and poorer medication adherence (Schoenthaler et al., 2014) when in racially discordant relationships with their physicians. In a study of cardiovascular disease medication adherence, African American patients seeing White physicians were more likely to report poorer adherence than White patients seeing White physicians (Traylor et al., 2010). A similar pattern was reported in another study of African American primary care patients (Schoenthaler et al., 2014). Finally, Non-Hispanic White patients reported better medication adherence in concordant relationships in primary care than African American patients in discordant relationships (Schoenthaler et al., 2014).

Especially relevant to the issue of study attrition, concordance in patient-physician relationships also has been associated with continuity of care and patient engagement. For example, African American patients in racially discordant relationships with their primary care physicians were less likely to schedule follow-up appointments and more likely to postpone or delay scheduled visits (Major et al., 2013). In a study of mental health intakes at community outpatient clinics, both Latinos and non-Hispanic Whites were more likely to return for their next scheduled visit if they were in racially concordant vs. discordant relationships (Alegria et al., 2013).

Mechanisms Linking Racial/Ethnic Concordance to Health

The psychological process of social categorization has been used to explain the way in which racial/ethnic concordance may affect health outcomes and patient engagement in the case of patient-physician relationships. Because of the complexity of the social environment and limited cognitive resources, people categorize others into social groups as a way of efficiently navigating social interactions. Race and ethnicity are salient characteristics that allow for quick categorization on the basis of superficial characteristics (Dovidio and Gaertner, 1999). Once categorization takes place, people are more likely to see members of a group as more similar to one another, a process known as the outgroup homogeneity effect. And, once a person has been categorized as a member of a certain group, people access available stereotypes and attitudes about that group and adjust their expectations for that particular person according to available cognitions they have for the group as a whole.

Social categorization has been used to explain disparities in racially discordant physician-patient interactions in a number of studies, focusing on the construct of implicit bias—the unconscious attribution of certain qualities according to group membership that translates into attitudes and behaviors. Implicit bias towards racial/ethnic minority groups has been found in many studies and populations, including healthcare professionals (Blair et al., 2013, Sabin et al., 2009). Unconscious attitudes, subtle discrimination, and verbal and non-verbal behaviors during a healthcare encounter (e.g., eye blinking, eye contact, and friendliness) may influence clinical decision-making (van Ryn and Saha, 2011). This has been termed aversive racism, a subtle, automatic and unintentional form of implicit bias found in U.S. non-minority populations who believe they are not prejudiced but (unknowingly) exhibit discomfort, uneasiness, disgust and even fear around members of racial/ethnic minority groups (Dovidio and Gaertner, 1998).

Does et al. (2018) urged researchers to consider the demographic characteristics of the research staff on research processes and results. There is a tendency to ignore this, even though such factors introduce systematic variation into the methodology and are meaningful parts of the experimental environment. Drawing on the social categorization processes described above, demographic characteristics of the research staff may create ingroup vs. outgroup effects and bring into play stereotypes and implicit bias among study participants and research staff. As participant-research staff interactions may be colored by these processes, it is important to examine concordance as a predictor of study retention and attrition.

The Current Study

The primary aim of this study was to examine the influence of racial/ethnic concordance between study participants and research staff on attrition rates. As the “face” of the study and the participant’s point of contact, clinical research coordinators (CRCs) are likely to make a difference in study retention. CRCs describe the study, solicit participation, obtain informed consent and, in longitudinal studies, maintain regular contact with participants to minimize attrition. Congruent with the literature on patient-physician dyads, we hypothesized that there would be greater attrition among participants in a racial/ethnic discordant dyad than in a concordant dyad, controlling for the effects of other variables related to attrition (gender, age, education, depression, and health literacy).

This hypothesis is tested with studies of respiratory illness, which disproportionately affect racial/ethnic minority populations in terms of prevalence, morbidity and mortality (Cabana et al., 2007). Racial/ethnic minority populations experience a greater number of asthma-related emergency department visits and hospitalizations (Akinbami et al., 2014), and a decreased likelihood of use of anti-inflammatory therapy (Deshpande et al., 2016) than White individuals. African Americans are more likely to develop chronic obstructive pulmonary disease (COPD) at an early age (Foreman et al., 2011), to have more hospitalizations and lower rates of preventive physician office visits compared to Whites (Dransfield and Bailey, 2006), and less likely to receive lung transplants, die while on a transplant list, or be removed from a transplant list than Whites with equal insurance coverage, poverty level, age, lung function, pulmonary hypertension and cardiovascular risk factors (Lederer et al., 2008).

Despite the statistics presented above, a review of studies in respiratory medicine from 1993 to 2013 showed a marked and continuing underrepresentation of racial and ethnic minorities in clinical research studies of respiratory illness (Rosser et al., 2014): Racial and ethnic minority participants accounted for only 1.9% of participants in pulmonary samples, and less than 5% of all NIH-funded published studies of respiratory illness (Burchard et al., 2015). While efforts to increase engagement of African American and Latino patients have been made since 2013, especially in the case of patient-centered studies to reduce disparities in asthma outcomes, a pooled analyses of these data highlights the urgent need to focus on engagement of participants from racial/ethnic minority communities to address the continued lack of sample diversity in the case of asthma research (Kramer et al., 2016). Thus, pulmonary disease is a good exemplar for testing the study’s hypotheses. Understanding how to retain racial and ethnic minority participants with respiratory disease is an important step toward minimizing health disparities.

Method

Data were pooled from six longitudinal studies of individuals with asthma or chronic obstructive pulmonary disease (COPD) conducted at the Icahn School of Medicine at Mount Sinai in New York City. Each study had one or more CRCs who were responsible for recruitment and retention of participants.

Study Data

The six individual studies, conducted between 2009 and 2018, are described fully in the Appendix (available as a supplement to the online version of this article). Individuals with asthma (Studies 1–4) or chronic obstructive pulmonary disease (Studies 5–6) were identified through electronic health records. As enrollment is ongoing in Studies 3 and 4, data were included only from participants whose follow ups were completed at the time of these analyses. Because Study 1 was an intervention trial, we included only the data from the control arm (usual care) for consistency with the other five studies, which were longitudinal observational studies. All study protocols were approved by Mount Sinai’s institutional review board. CRCs assigned to the study had on average 2.7 years of experience in their position (SD = 2.2) and worked on an average of 4.2 projects at any given time as part of their duties (SD = 2.9). Training was standardized for all CRCs across projects and included regular observations with patients for quality control.

Patient measures.

Patient data were extracted from study records. Unless otherwise noted, the same variables were used across the six studies.

Attrition.

Attrition was defined as participants actively dropping out of the study, being lost to follow up, missing the interview, or death. Following Ribisl et al., (1996) study attrition was measured two ways, both dichotomous variables. Attrition at first follow-up refers to participants who did not complete the second study interview (the first interview after the baseline assessment). Across the six studies, attrition at first follow-up was measured 3–6 months after the baseline interview, with some studies collecting first follow-up data in person (Studies 3, 4 and 6) and other studies collecting data by telephone (Studies 1, 2 and 5). Attrition at one year refers to participants who did not complete a study interview at the 12-month mark. With one exception, all studies conducted a follow up assessment at 12 months after baseline; Study 6 conducted a 15-month follow-up, and this data point was used.

Sociodemographic information.

Age, gender, race, ethnicity and education were assessed using items adapted from the National Health Interview Survey (NHIS) (Pleis et al., 2010). For race/ethnicity, study participants were asked two questions: “What race do you consider yourself to be?” and “Do you consider yourself Latino?” The variable “race/ethnicity” was coded in the original datasets for all six studies into four categories: 1) Black, non-Hispanic, 2) White, non-Hispanic, 3) Hispanic, and 4) Other, non-Hispanic. Because the datasets combined individuals identifying as Asian or Other into a single category, and this category accounted for less than 6% of participants across the combined studies, we excluded study participants who identified as Asian/Other for these analyses. The choices provided to patients did not allow for the selection of multiple categories (e.g. multiracial participants).

Language.

In all six studies, participants were given the option to complete the interview in English or Spanish. Language refers to whether the study interviews were completed in English or Spanish.

Depressive symptoms.

Measures included the Patient Health Questionnaire 9 (PHQ-9; Kroenke et al., 2001; Studies 2,4,5, and 6), the Patient Reported Outcomes Measurement Information System Depression Scale (PROMIS SF 8; Cella et al., 2010; Study 1) and the Geriatric Depression Scale 15 (GDS 15; Sheikh, 1986; Study 3). To make these measures comparable, a dichotomous measure of probable depression was created using the validated cutoff scores for each measure (PHQ-9 ≥ 10; GDS-15 ≥ 10; PROMIS cutoff ≥ 60), with probable depression based on reports of moderate to severe symptoms. As only total scores were available for each instrument in our dataset, we were unable to calculate reliability for the scales. However, psychometric data on these measures have been published in the original studies that make up this pooled analysis (see Appendix for full study details) and studies of other populations show good reliability.

Health literacy.

Health literacy was assessed with the Newest Vital Signs (Weiss et al., 2005; Study 1), the Short Test of Functional Health Literacy in Adults (S-TOFHLA; Baker et al., 1999; Studies 2 and 5) and the Short Literacy Survey (SLS; Chew et al., 2004; Studies 3, 4 and 6). A dichotomous measure of inadequate health literacy was created using validated cut off scores for each measure (NVS cutoff < 2; S-TOFHLA cutoff < 17; SLS>10 [Higher scores indicate poorer literacy in the SLS]). As only total scores were available for each instrument in our dataset, we were unable to calculate reliability for these. Yet, details of these measures have been published in the original studies that make up this pooled analysis (see Appendix for full study details).

Measures administered to CRCs.

The CRCs of all six studies were invited to complete a brief self-administered questionnaire at the time of data analysis. CRCs received the same socio-demographic questions as the patients. We decided to include CRCs in the sample who reported their race/ethnicity as Asian or Other, even though the patient sample included only individuals identifying as Black, Hispanic or White. This decision allowed analyses to include the CRCs identifying as Asian/Other in discordant dyad counts.

Concordance among dyads.

Dyads were classified as concordant if the patient and CRC were of the same racial/ethnic group; we also refer to this as a match. Dyads were classified as discordant if the patient and CRC were of different racial/ethnic groups.

Statistical Analyses

Descriptive statistics were calculated to examine means, standard deviations, and frequencies of all variables. Comparisons of variables across studies were performed using chi-square statistics, t-tests, and Kruskal-Wallis tests (for non-normally distributed variables). Logistic regression and chi-square statistics were used to evaluate predictors of attrition rates in univariate models.

Hierarchical generalized linear modelling (HGLM) was used to analyze the effect of concordance on attrition, while accounting for nesting within CRCs (level 2). CRCs were treated as a contextual variable, which may introduce dependency in the data. Therefore, nesting by CRC accounts for the possibility that the attrition of subjects recruited by the same CRCs would be more similar than the attrition of subjects recruited by different CRCs. Adjusted models accounted for any covariates that were significant in univariate analyses. To adjust for the possible effects of any particular study on attrition, five dummy codes (for the six individual studies) were entered as a set. Models estimated the log odds (logit) of the probability of attrition at first follow-up and one year. Separate hierarchical models examined the relationship of concordance with attrition at first follow-up and attrition at one year. Three sequential models were tested at each time point. Model 1 included only the intercept and the nesting of study participants by CRC. Model 2 included the variables that were found to be significant predictors of attrition in univariate analysis (described above), including which study the participant was in. Model 3 added the concordance variable. All analyses were performed using two-tailed tests of significance, with significance set at α < .05. The R package (3.5.3, R Core Team) was used for multilevel models, and SAS 9.4 (SAS Institute, Cary NC) was used for all other analyses. Multiple imputation methods for missing data were not used as this study’s primary aims were to assess attrition (a type of missing data), and variables predictive of this type of missing data.

Results

Sample Characteristics

Participant characteristics at the first data collection point for each study can be found in Table 1. The sample consisted of 509 dyads, composed of 509 participants and 14 CRCs. Mean age was 67.1 years (SD = 9.7), and most participants were female (70.6%). The sample was racially and ethnically diverse, with most participants identifying as Hispanic (47.6%), or as Black or African American (31.7%). Data on CRCs race/ethnicity was obtained for 14 clinical research coordinators, or slightly over half (53.8%) of the CRCs who were involved in the six studies. CRCs were 50% female, with 42.6% identifying as Latino, 21.4% as Black or African American, 21.4% as non-Hispanic White, and 14.3% as Asian or Other. Combining the participant data and the CRC data into dyads resulted in 509 dyads for analysis. Dyads were mostly race/ethnicity discordant (82.5%; see Table 6 in Appendix for full breakdown). At the first follow-up, most interviews were completed by the same CRCs who had administered the baseline interview to participants (62%). At one year, most interviews were conducted by a different CRC than the one that had administered the baseline (96%).

Table 1.

Patient demographics and characteristics.

Characteristics	Total sample (n = 509)	Study 1 (n =51)	Study 2 (n =146)	Study 3 (n =50)	Study 4 (n =25)	Study 5 (n =155)	Study 6 (n =82)
	% or Mean (SD)	% or Mean (SD)	% or Mean (SD)	% or Mean (SD)	% or Mean (SD)	% or Mean (SD)	% or Mean (SD)
Age
M (SD)	67.1 (9.7)	69.5 (9.3)	67.7 (6.9)	68.9 (7.4)	45.8 (11.7)	68.4 (9.2)	67.1 (8.5)
Range	25–100	60–100	60–98	60–87	25–70	55–91	44–86
Number >90 years	6	1	1			4
Female, %	70.6	88.2	84.9	87.8	66.7	55.5	53.7
Race & Ethnicity, %
Hispanic	47.6	56.9	66.9	68.1	12.5	31.0	37.5
White, non-Hispanic	20.7	5.9	9.7	8.5	29.2	37.4	22.5
Black, non-Hispanic	31.7	37.2	23.4	23.4	58.3	31.6	40.0
Education Level, %
High school or less	60.0	70.6	74.3	59.6	54.2	50.3	47.5
At least some college	40.0	29.4	25.7	40.4	45.8	49.7	52.5
Interviews in Spanish	24.5	41.2	37.0	40.0	12.0	13.5	6.2
Probable Depression	29.7	56.9	28.6	19.0	12.5	29.2	23.7
Inadequate HL	54.0	51.3	46.7	66.0	52.0	46.7	75.0

Note. Cut-off scores for probable depression were defined as PHQ-9 cutoff ≥ 10; GDS-15 cutoff ≥ 10; PROMIS cutoff ≥ 60.

²Cut-off scores for inadequate health literacy were defined as NVS cutoff < 2; S-TOFHLA cutoff < 17; SLS>10.

Attrition Rates

The average attrition across studies was 11.8% at the first follow-up and 20.7% at one year. There were significant differences in attrition rates among the six studies (see Appendix, Table 3); Attrition rates at the first follow-up ranged from 2.7% (Study 2) to 25.5% and at one year, from 11.6% (Study 2) to 42.1% (Study 3). Therefore, study was included as a covariate in all multi-level models.

Demographic and Mental Health Predictors of Attrition

Significant associations were found for language, education, and depressive symptoms with attrition (see Table 2). For attrition at first follow-up, there was greater attrition for participants with a high school education or less compared to those with some college or more (OR = 1.9; 95% CI = 1.0, 3.7) and for those completing research interviews in Spanish (OR = 2.7; 95% CI = 1.5, 4.7). At the one-year follow-up, there was greater attrition among participants with probable depression compared to those without depression (OR = 1.7; 95% CI = 1.0, 2.8). Race/ethnicity was not a significant predictor of attrition at the first follow-up or at the one-year follow-up, nor were age, health literacy or gender significant predictors of attrition at any time point.

Table 2.

Predictors of attrition at the first follow-up and one year (bivariate analyses).

	Attrition	OR (95% CI)
Attrition at first follow-up
Gender, n (%)
Female	37 (10)
Male	21 (14)	1.4 (0.8 – 2.5)
Race & Ethnicity, n (%)
Hispanic	32 (13)	1.6 (0.7 – 3.6)
Black	14 (9)	1.0 (0.4 – 2.4)
Whit	9 (9)
Education Level, n (%)		1.9 (1.0 – 3.7)*
High school or less	38 (13)
Some college	14 (7)
Depression, n (%)
Probable depression	16 (11)	1.1 (0.6 – 2.1)
No depression	34 (10)
Health literacy, n (%)		1.0 (0.6 – 1.8)
Inadequate health literacy	27 (10)
Adequate health literacy	23 (10)
Language of interview, n (%)		2.7 (1.5 – 4.7)**
Spanish	25 (20)
English	33 (9)
Age (X2, p)	67.6 (12.1)	.51 (.48)
Attrition at one year
Gender, n (%)
Female	56 (19)
Male	31 (25)	1.4 (0.9 – 2.4)
Race & Ethnicity, n (%)
Hispanic	44 (22)	1.0 (0.6 – 1.9)
Black	23 (18)	0.8 (0.4 – 1.5)
White	20 (22)
Education Level, n (%)
High school or less	58 (23)	1.6 (0.9 – 2.6)
Some college	26 (16)
Depression, n (%)
Probable depression	32 (26)	1.7 (1.0 – 2.8)*
No depression	48 (17)
Health literacy, n (%)
Inadequate health literacy	41 (21)	1.0 (0.6 – 1.6)
Adequate health literacy	41 (21)
Language of interview, n (%)
Spanish	29 (27)	1.6 (1.0 – 2.7)
English	58 (19)
Age (X2, p)	66.8 (9.9)	.891 (.34)

Note.

^*p < .05

^**p<.01

All analyses used χ² tests except for age (Kruskal-Wallis).

The Relationship of Racial/Ethnic Concordance with Attrition

Racial/ethnic concordance was significantly associated with attrition at the first follow-up (b = 1.59, SE = .47, p <.001) and at one year (b = 1.35, SE = .48, p = .005). Introducing concordance in the model (Model 3) significantly improved model fit, χ²(1) = 11.12, p < .001 at first follow-up and at one year, χ² (1) = 6.20, p < .05. However, contrary to our hypothesis, there was greater attrition among concordant dyads than discordant dyads (see Table 3). Specifically, participants in concordant dyads of any race/ethnicity had 4.9-times the odds of not completing a first follow-up compared to those in discordant dyads (OR = 4.9, 95% CI = 1.95, 12.32), and 3.9 times the odds at one year (OR = 3.86, 95% CI = 1.51, 9.88). To further assess that the effects of concordance were not due to characteristics of one or a few specific CRCs, a subsequent multilevel model (not shown here) was tested in which we added a random effect for concordance. This model indicated that the effect of concordance did not significantly vary by CRC at the first follow-up or at one year. The degree to which attrition clustered within CRCs was explored through a slope-only random effects model with variance. There is no definitive approach to measuring intraclass correlations (ICC) (also known as a variance partitioning coefficient) with multilevel logistic regression models (Austin and Merlo, 2017, Goldstein and Rasbash, 1996). We chose to employ a latent variable approach. The proportion of variance in attrition at the first follow-up associated with CRCs was .171 or roughly 17%; likewise, the proportion of variance in attrition at one year associated with CRCs was .078 or roughly 8%. The results would suggest that short-term attrition (at the first follow-up) was more influenced by CRC than longer term (one year).

Table 3.

Multilevel results for any concordance and attrition at time two and one year (Aim 1).

	Model 1	Model 2	Model 3
Variables	logit (SE)	logit (SE)	logit (SE)	OR (95% CI)	p
Estimation of fixed effects
Attrition at first follow-up
Level 1: Participant characteristics
Study IDa	--	--	--	--	--
Some College		−0.46 (.37)	−.59 (.38)	.55 (.26 – 1.17)	.11
Spanish Interview		1.04 (.39)	0.34 (.45)	1.40 (.58 – 3.39)	.46
Any concordance			1.59 (.47)	4.90 (1.95 – 12.32)	.001***
Model fit
Deviance/df	272.06/7	263.81/9	252.68/11
∆ Deviance/∆df	18.58/5**	8.25/2*	11.12/1***
AIC	286.06	281.81	272.68
Attrition at one year
Level 1: Participant characteristics
Study IDa	--	--	--	--	--
Some College		−0.39 (.30)	−0.45 (.30)	.64 (.35 – 1.15)	.13
Spanish Interview		0.22 (.41)	0.18 (.36)	1.25 (.59 – 2.42)	.61
Depressive Symptoms		0.40 (.29)	0.36 (.29)	1.43 (.81 – 2.53)	.21
Any concordance			1.35 (.48)	3.86 (1.51 – 9.88)	.005**
Model fit
Deviance/df	377.38/7	372.35/10	366.16/11
∆ Deviance/∆df	7.34/5	5.03/3	6.20/1*
AIC	391.38	392.35	388.16

Notes.

^*p < .05

^**p < .01

^***p < .001

Model 1 = Dummy codes for study

Model 2 = Dummy codes for study + significant variables from univariate analysis

Model 3 = Dummy codes for study + significant variables from univariate analysis + any concordance

^aDummy codes for Studies 2, 3, 4, 5, and 6 were entered. The reference category was study 1.

We conducted exploratory analyses to compare attrition among racial/ethnic minority participants in dyads with a majority CRC (i.e., non-Hispanic White), vs. those in dyads with minority CRCs (i.e., non-White). For these analyses, dyads were classified as majority concordant if both participant and CRC reported their race/ethnicity as non-Hispanic White and minority concordant if both members of the dyad identified as a member of any racial/ethnic minority group.

Minority concordance at baseline was a significant predictor of attrition at the first follow-up (b = 1.69 SE = .48, p <.001) and attrition at one year (b = 1.48, SE = .54, p = .006) (see Table 4). Black or Latino participants in a dyad with a CRC who also identified as a racial/ethnic minority had 5.4 the odds of not completing a first follow up (OR = 5.42, 95% CI = 2.11, 13.88) and 4.4 times the odds at one year (OR = 4.39, 95% CI = 1.52, 12.66), compared to Black or Latino participants who were paired with a non-Hispanic White CRC. Introducing majority/minority concordance in the model significantly improved the model fit, χ²(2) = 11.79, p < .01 at the first follow-up and at one year, χ²(2) = 6.51, p < .05. Majority concordance was not predictive of attrition at any time point in the final model (Model 3).

Table 4.

Multilevel results for minority concordance and attrition at time two and one year (Aim 2).

	Model 1	Model 2	Model 3
Variables	logit (SE)	logit (SE)	logit (SE)	OR	p
Estimation of fixed effects
Attrition at first follow-up
Level 1: Participant characteristics
Study IDa	--	--	--	--	--
Some College		−0.46 (.37)	−0.56 (.38)	.57 (.27 – 1.20)	.14
Spanish Interview		1.04 (.40)	0.31 (.45)	1.36 (.56 – 3.29)	.49
Minority Concordance			1.69 (.48)	5.42 (2.11 – 13.88)	<.0001***
Majority Concordance			0.65 (1.31)	1.91 (.15 – 24.97)	.62
Model fit
Deviance/df	272.06/7	263.81/9	252.02/11
∆ Deviance/∆df	18.58/5**	8.25/2*	11.79/2**
AIC	286.06	281.81	274.02
Attrition at one year
Level 1: Participant characteristics
Study IDa	--	--	--	--	--
Some College		−0.40 (.30)	−0.42 (.31)	.66 (.36 – 1.21)	.16
Spanish Interview		0.22 (.41)	0.16 (.36)	1.17 (.58 – 2.38)	.66
Depressive Symptoms		.38 (.29)	0.36 (.29)	1.43 (.81 – 2.53)	.21
Minority Concordancea			1.48 (.54)	4.39 (1.52 – 12.66)	.006**
Majority Concordancea			0.87 (.98)	2.39 (.35 – 16.29)	.37
Model fit
Deviance/df	377.38/7	372.35/10	365.84/12
∆ Deviance/∆df	7.34/5	5.03/3	6.51/2*
AIC	381.38	392.35	389.84

Notes.

^*p < .05

^**p < .01

^***p < .001

Model 1 = Dummy codes for study

Model 2 = Dummy codes for study + significant variables from univariate analysis

Model 3 = Dummy codes for study + significant variables from univariate analysis + minority concordance (h/h, b/b) and majority concordance (w/w)

^aDummy codes for Studies 2, 3, 4, 5, and 6 were entered. The reference category was study 1. The minority and majority concordance variables were dummy codes using discordance as the reference group.

Discussion

The analyses revealed many surprising findings. Contrary to hypotheses, there was greater attrition among racial/ethnic concordant dyads than discordant dyads, and this effect was large: Black and Hispanic participants who were paired with a CRC who identified as a racial/ethnic minority were almost six times more likely to be lost at the first follow-up and almost five times more likely to be lost at one year, compared to those who had a non-Hispanic White CRC. This pattern was not observed for non-Hispanic White participants in concordant dyads. Because the effect of racial/ethnic concordance was only examined after the variance attributed to participant and CRC factors was removed, this points to an effect of concordance and not of any individual level factors. Moreover, factors previously identified as increasing study attrition—age, depressive symptoms, and education—were not significant predictors of attrition in adjusted models.

Several factors could explain the finding that concordance was associated with greater attrition, including stereotypes about what a scientist looks like, and internalized racism. First, it is possible that stereotypes of physicians and scientists as White may be strongly held by members of racial and ethnic minority groups; this stereotype may have challenged the authority of racial/ethnic minority CRCs. In studies exploring the role of race/ethnicity and microaggressions in the workplace, minority medical residents (Osseo-Asare et al., 2018) and faculty members (Constantine et al., 2008) reported having their qualifications questioned by colleagues, staff members, students, and patients. These experiences included being referred to by their first name or as “Miss” or “Mister” (i.e., instead of “Doctor” or “Professor”) and being confused for ancillary staff or service workers. Similar processes underlie social representations of scientists as “a White male, who wears a lab coat with a pocket full of pens and pencils. He’s middle aged and is either bald or has wild hair framing his myopic eyes…” (Long et al., 2001, p. 255). These beliefs have been reported in studies of elementary, secondary, and college students, as well as in the portrayal of scientists in the media. Research using the Draw-A-Scientist Test (Chambers, 1983)—a measure of children’s perceptions of scientists through drawings—showed that across a range of populations and countries, the results still find that the majority of the drawings are of White, male scientists (Steinke et al., 2007, Thomas et al., 2006). It is possible that the stereotype of physicians and scientists as White males is strongly held even by minority members, and therefore acts against the authority of non-White CRCs as they interacted with minority patients. Additionally, internalized racism—the acceptance by members of stigmatized races of negative messages about their abilities and intrinsic worth (Jones, 2000)—could have affected racial/ethnic minority participants’ perceptions of the competence of concordant CRCs. In a mixed-methods study among Arab and Jewish populations in Israel, Arab participants who voiced a preference for Jewish physicians attributed this choice to a perceived “lack of professionalism” by Arab physicians, as well as a belief that the dominant ethnic group (i.e. Jewish) were better qualified (Keshet, 2019). As has been suggested by these and other authors regarding patient-physician relationships (Schnittker and Liang, 2006), preferences for concordance may have been suppressed by internalized racism, such that racial/ethnic minority participants viewed discordant CRCs as more authoritative. Further research should examine whether a mismatch between the expectations of what a researcher should look like and the perceived race/ethnicity of the recruiting CRC affects study engagement.

Overcorrection on the part of discordant, and especially majority (i.e. non-Hispanic White) CRCs also may have accounted for the pattern of results observed in this study. Overcorrection occurs when individuals control and monitor their verbal and nonverbal behaviors in order to ensure that they appear unprejudiced in interactions with members of a racial or ethnic group different from their own (Mendes and Koslov, 2013). These overcorrection efforts can arise from both a desire to be egalitarian, as well as from impression-management concerns to not appear prejudiced against members of the outgroup (Mendes and Koslov, 2013). CRCs in discordant dyads could have engaged in overcorrection of their behaviors above and beyond those in concordant dyads. As keeping research participants engaged over time is part of CRCs’ job responsibilities, there is a motivation to appear unprejudiced, especially when completing study interviews that cover information on health behavior and illness that are of a highly personal nature. In an experimental study on dyadic interactions between Black and White college students, White participants who engaged in overcorrection were rated more positively by their Black dyad partner, who reported higher positive ratings for their interactions even if they were told the White partner was prejudiced (Shelton, 2003). Further, it has been suggested than when confronted with a member of a racial outgroup, the concern that one could be perceived and treated negatively because of one’s race can lead racial minority individuals to use social skills to make the interaction successful (Shelton, 2003). In the current study, when CRCs matched the research participant in race/ethnicity, it is possible that their concerns to appear unbiased were lower and therefore had little reason to engage in overcorrection. Future research should examine implicit attitudes and overcorrection efforts among study participants and CRCs to elucidate mechanisms of attrition.

The findings about racial/ethnic concordance are surprising and contrary to those found in the literature on patient-physician relationships. Although a “match” between the race/ethnicity of patients and that of their healthcare providers has been associated with improved continuity of care, satisfaction with care and communication, medical understanding, and patient engagement among Black and Latino patients (Alegria et al., 2013, Diamond et al., 2019, Hsueh et al., 2019, Major et al., 2013), we found the opposite pattern for study attrition. As such, the findings suggest different social categorization processes may underlie the relationship between researchers and study participants. The research relationship—where study participants are voluntarily agreeing to participate at times for little to no remuneration—differs from healthcare provider relationships, in which the patient relies on the provider for care. An important next step to explore this question might be to identify differences between clinical trials offering participants tangible health benefits from their participation (e.g., an experimental biologic therapy for people with advanced stage cancer), compared to observational studies in which factors other than health benefits are promised.

Limitations

The study results should be considered in the context of several limitations. First, the sample included only Latino, Black, and White study participants, as the small numbers of participants who identified as “Asian” or “Other” had been categorized into a single category in the original datasets. Second, our analyses could not elucidate mechanisms underlying the associations between racial/ethnic concordance and attrition, such as stereotypes, internalized racism, and overcorrection. Third, there was a relatively small number of CRCs; however, by using HLM with nesting and random effects multilevel modeling, individual differences should have been accounted for. Future approaches to analyzing study participant engagement should consider other analytical strategies, such as dyadic and one-to-many approaches. We chose to pool studies because it is improbable that any one study would provide sufficient data to test these hypotheses, as a large number of both research participants and CRCs are required to explore the issue of concordance and attrition. While the individual studies shared many characteristics (respiratory illnesses, recruited from same hospital and catchment areas, similar procedures in terms of interviews and questionnaires, participant burden, etc.), there may have been differences between studies that we could not account for even after covarying studies in adjusted models. Having the same CRC at all study time points may also be important, and in our dataset most follow up interviews at the first follow-up were conducted by the same CRC that patients had met at baseline; however this was not the case at one year. In our analyses, attrition at the first follow-up and at one year is attributed to the original participant-CRC dyad at baseline as this is the first in-person contact for participants with the study team. However, future studies should evaluate the impact of concordance at each time point on subsequent attrition. Ten participants did not complete follow-ups because they were deceased. We decided against removing them from the analyses as the guidelines posed by Ribisl et al. (1996) suggest attrition rates should include all preventable and non-preventable causes. Future studies with larger numbers of participant deaths may choose to censor this data to limit its impact on its findings. Despite these limitations, this study had several strengths, including the large number of CRC-participant dyads that allowed for the exploration of these hypothesis. Further, the sample of both CRCs and participants was diverse. Finally, the study also controlled for several other factors related to study attrition in prior studies (age, education, depressive symptoms, and health literacy), strengthening the findings.

Conclusions

This study is the first to systematically examine racial/ethnic concordance between researchers and participants on sample attrition and has several implications. Although the study should be replicated before firm conclusions are drawn, our findings question the assumption that a racial/ethnic match between participants and researchers may be the key to retention in longitudinal studies. Participant-research staff interactions may be influenced by processes that introduce systematic variation to study participation and should be considered important parts of the study environment.

As proposed by Does et al. (2018), bias against certain populations can influence research staff’s interactions with research participants, which in turn can affect the interpretation of research findings. Given our findings, it is plausible that concordance with research participants had effects on participants’ behaviors beyond attrition, including the manner in which they responded to study questionnaires, and further research should explore these hypotheses. Additionally, these findings highlight the fact that attrition may not be a characteristic of the participant, a result of motivation or interest, but rather influenced by interactions with the research team. Some qualitative evidence on recruitment and retention of racial/ethnic minorities in cancer clinical trials found an emergent theme of research staff not being trained to focus on how to retain patients from historically underrepresented minorities, such as specific trainings on cultural awareness and implicit bias (Niranjan et al., 2019). If our findings are replicated in other samples, then training efforts must be developed to address this curriculum gap. Although we were unable to explore the effect of CRC’s training on attrition, prior work on racial/ethnic discordance between patients and physicians (Chu et al., 2019) suggests there may be benefits of increasing awareness of racial and ethnic disparities and implicit bias. Current training in human subjects’ research is mandated for research assistants and covers historical injustices carried out by researchers against racial/ethnic minority participants; however, this may be insufficient and additional training may be needed.

If, as our findings suggest, there is differential attrition among racial and ethnic minority populations in longitudinal studies, then the capacity of researchers to address health disparities is severely constrained. Differential attrition among racial and ethnic minority populations has implications for translating study findings into interventions that can effectively address the health disparities prevalent in respiratory disease outcomes, and the characteristics of research staff have been traditionally disregarded as a potential pathway to ensure retention of diverse samples. Further research is needed to elucidate the mechanisms underlying this finding and why it is contrary to the patterns observed in patient-physician relationships. If the “why” is better understood, researchers can then address differential attrition and maximize internal and external validity, improving the capacity to generalize research findings to those disproportionately burdened with poor health outcomes.

Highlights.

• Attrition was greater for racial/ethnic concordant dyads of research staff and study participants

• Black and Hispanic adults recruited by concordant researchers were less likely to complete studies

• Racial/ethnic concordance does not always lead to retention in research studies