Minimal Group Procedures and Outcomes

The Minimal Group Paradigm is characterized by three main features: 1) novel and arbitrary group categorization, 2) anonymity and no interaction among group members, and 3) no direct benefits to the participant in outcome measures (Otten, 2016; Spears & Otten, 2012; Tajfel et al., 1971). Design variations within these confines (and sometimes outside of them) have revealed features that can weaken or amplify minimal group effects.

The majority of research using the MGP has explored how outcomes vary according to features of the groups themselves rather than the induction procedures used for their assignment. This research has included manipulations of group characteristics such as group similarity (Diehl, 1988), number (Hartstone & Augoustinos, 1995; Spielman, 2000), status (Harvey & Bourhis, 2012; Otten et al., 1996; Rothgerber & Worchel, 1997; Sachdev & Bourhis, 1991) and size (Leonardelli & Brewer, 2001; Sachdev & Bourhis, 1991). Manipulations that target group norms (e.g., Lee et al., 2012; Maio et al., 2009) or identification (e.g., Gagnon & Bourhis, 1996; Stroebe et al., 2005) may have the largest moderating effects on minimal ingroup biases (Pechar & Kranton, 2017). However, minimal group induction procedures may also impact the magnitude of ingroup biases.

Classic minimal group inductions began with participants completing a trivial task such as estimating the number of dots in an array or evaluating paintings before they are assigned to groups (Tajfel, 1970; Tajfel et al., 1971). Participants believed that their group membership, although actually randomly assigned, resulted from the task. Brewer and Silver (1978) extended these findings to demonstrate that even blatant random assignment produced minimal group effects. Later studies confirmed that random assignment is an effective means of producing ingroup biases (e.g., Perreault & Bourhis, 1998) but perhaps to a lesser degree than classic procedures (Gaertner & Insko, 2000; Herringer & Garza, 1987). These researchers suggested that these relatively meaningful methods of categorization increased identification with and preferences for the ingroup.

Often MGP procedures include statements designed to reinforce identification or otherwise impose meaning on participant group membership. For instance, a painting preference induction may provide information about the perceptual processing style of people who have the same preferences as the participant (Ashburn-Nardo et al., 2001; Pinter & Greenwald, 2011). In fact, manipulations designed to enhance identification with the ingroup appear to increase discrimination in the MGP (e.g., Leonardelli & Brewer, 2001; Stroebe et al., 2005), and participants who report strongly identifying with their minimal ingroup discriminate more than those who do not (e.g., Gagnon & Bourhis, 1996). Participant control may also increase minimal group effects. When participants can choose their ingroup, they discriminate more against the outgroup than when they are randomly assigned (Perreault & Bourhis, 1999).

Induction procedures that do not include group assignment have also produced effects similar to those found in the MGP. For instance, Pinter and Greenwald (2011) found that imagined group membership produced biases in evaluation and identification. Additionally, research on implicit partisanship suggested that memorizing the names of novel group members creates biases toward that group relative to another on implicit measures of evaluation and identification (Greenwald et al., 2002). This memorization procedure has produced ingroup biases similar to those resulting from widely used MGP inductions (Dunham, 2013; Pinter & Greenwald, 2011). In a comparison of the memorization induction to painting preference and random assignment inductions, Pinter and Greenwald found equivalent minimal group effects on a resource allocation task and explicit evaluations and identifications and stronger effects on implicit measures of evaluation and identification. However, these results should be interpreted cautiously, due to insufficient power (i.e., only 25 participants per induction) and methodological limitations. For instance, the ingroup names that were memorized in the induction were used as stimuli for the implicit measures. The subsequent familiarity and fluency of those names during measurement may have contributed to the stronger effects for that procedure relative to other MGP inductions.¹

Irrespective of induction, the MGP appears to produce ingroup biases on a variety of dependent measures across several domains including evaluation, cooperation and resource allocation, and learning (Dunham, 2018). Minimal group effects have even been found on neurocognitive measures including attention, recognition, imagery, and processing of faces (e.g., Bernstein et al., 2007; Ratner et al., 2014; Ratner & Amodio, 2013; Van Bavel & Cunningham, 2010, 2012) and in brain activity (e.g., Molenberghs & Morrison, 2014). The present research focused on measures of discrimination, evaluation, and identification as outcomes.

Discrimination. Early minimal groups researchers used the Tajfel Matrices (Tajfel, 1970) as an assessment of discrimination for their primary dependent measure (e.g., Billig & Tajfel, 1973; Turner et al., 1979). In this task, participants distribute resources to ingroup and outgroup members across a series of allocation choices. Discrimination is inferred from which allocation strategies influence participants (see Bourhis et al., 1994, for comprehensive measurement details).

The Tajfel Matrices have received considerable criticism, in part, because the allocation strategies are not measured independently (Bornstein et al., 1983; Spears & Otten, 2012). Bornstein et al. (1983) developed the Multiple Alternative Matrices (MAM) to address this limitation. Unlike the Tajfel Matrices, the MAM provides separate response options for relative ingroup favoritism and maximization of ingroup profits. Although the original study failed to produce minimal group effects, the MGP has successfully produced discriminatory outcomes on the MAM in subsequent research (e.g., Gaertner & Insko, 2000; Pinter & Greenwald, 2011). Researchers have also used economic games as measures of discrimination in response to the MGP (for meta-analysis, see Balliet et al., 2014).

The magnitude of the MGP’s discrimination effects may depend on framing of the allocation tasks. Outcomes relevant to group status or group activities appear to produce increased discrimination (Forgas & Fiedler, 1996; Hartstone & Augoustinos, 1995; Reichl, 1997). Increasing interdependence among ingroup members may also amplify minimal group effects (e.g., Stroebe et al., 2005); however, the evidence is mixed (e.g., Gaertner & Insko, 2000; Gagnon & Bourhis, 1996). Linking ingroup outcomes appears to increase discrimination (e.g., Moscatelli & Rubini, 2013; Stroebe et al., 2005), but such expectations of reciprocity do not always affect outcomes (e.g., Gagnon & Bourhis, 1996).

The form of resource being allocated may also influence the magnitude of the MGP’s discrimination effects. Discrimination decreases when allocations are described as wages rather than bonuses (Gaertner & Insko, 2000). Similarly, participants demonstrate less discrimination when allocating monetary payments as opposed to non-monetary resources (Jost & Azzi, 1996). Discrimination also occurs less in the allocation of negative outcomes compared to positive outcomes (Gardham & Brown, 2001; Otten et al., 1996). In fact, increasing identification with minimal ingroups may only amplify discrimination for positive rather than negative outcomes (Hodson et al., 2003). These findings appear consistent with a positive-negative asymmetry in intergroup prejudice and discrimination (e.g., Mummendey & Otten, 1998). In line with this perspective, a meta-analysis of minimal group studies (Buhl, 1999) confirmed that participants show larger minimal group effects in positive domains than in negative domains.

Evaluation and Identification. Researchers have measured evaluative and identity-relevant outcomes of the MGP using both direct and indirect measurement instruments. Whereas self-report and other direct measurement instruments produce explicit measures, implicit measures are the result of indirect measurement instruments such as the Implicit Association Test (IAT; Greenwald et al., 1998). Researchers suggest that responses on implicit measures are produced by mainly associative or automatic cognitive processes while explicit measures also reflect propositional or controlled cognitive processes (e.g., Gawronski & Bodenhausen, 2006). Implicit and explicit measures may demonstrate distinct minimal group effects due to the differential involvement of these processes.

Research using the MGP has generally found intergroup evaluative and identification biases on both implicit and explicit measures. Participants explicitly evaluate minimal ingroups more positively as measured directly with self-report items (e.g., Dunham, 2013; Pinter & Greenwald, 2011), and implicit measures of evaluation also reveal minimal group effects (e.g., Ashburn-Nardo et al., 2001; Dunham, 2013; Pinter & Greenwald, 2011). The MGP similarly produces identification with ingroups over outgroups (e.g., Falk et al., 2014) on both implicit and explicit measures (Dunham, 2013; Pinter & Greenwald, 2011). In fact, minimal group effects may be larger for implicit identifications than for implicit evaluations (Dunham, 2013). However, researchers have focused on measures of discrimination in most previous investigations; thus, more evidence is needed to determine how the MGP and its variations impact implicit and explicit measures of identification and evaluation.

The Minimal Group Paradigm (MGP) continues to be a valuable tool in research on intergroup relations. Tajfel and colleagues’ (1971) seminal research using the MGP has been cited over 4,700 times in the last decade alone (Google Scholar, 2023). As reviewed above, previous research has demonstrated that intergroup bias effects produced by the MGP can depend on induction procedures. However, research systematically comparing the effect of these procedures across multiple outcome measures is largely absent from the literature on intergroup relations. In fact, Pinter and Greenwald (2011) report the only within study comparisons of multiple inductions across implicit, explicit, and behavioral outcome measures. Given the limitations of those studies in both power and scope, more research is necessary to identify exactly how methodological features influence minimal group effects.

In the present research, we manipulated induction and measurement procedures to determine how variations in the MGP impacted the magnitude of ingroup biases. Participants completed two randomly assigned outcomes from among six possibilities. Because participant responding to one measure may reduce the influence of the induction on subsequent measures, dependent measure order was counterbalanced and included in analyses to account for order effects. Our analyses compared the impact of six minimal group inductions on several outcome measures, including monetary allocations on the Tajfel Matrices and MAM and implicit and explicit evaluations and identifications. Simultaneously, we examined the moderating influence of other procedural elements (i.e., identification meaning and member name learning) and tested whether those effects differed by induction. Finally, we determined how minimal group effects varied according to outcome measure and measurement characteristics such as reciprocity expectation and IAT version. Regardless of the results, the present research reveals valuable insights for future researchers studying intergroup relations with the MGP.

All sample size determinations, data exclusions, manipulations, and measures are reported. We preregistered our methods (https://osf.io/h7d2r/) and analysis plan (https://osf.io/8m69a/). Study materials are available at https://osf.io/4t3k7/. The research protocol was approved by the Institutional Review Board at Southern Illinois University Carbondale (FWA#: 00005334).

Participants were recruited using Amazon’s Mechanical Turk (MTurk). We limited participation to residents of primary English-speaking countries over age 18 with at least 95% Human Intelligence Task (HIT) approval and more than 100 approved HITs on MTurk. Of the 9187 study sessions where participants provided consent, 6552 represented unique participants who successfully completed the HIT. While we intended to cease data collection after 5000 people were compensated, data quality was lower (i.e., more exclusions were necessary) than anticipated. Thus, we chose to collect an additional 1500 participants. After exclusions for data quality (see criteria below), 3926 valid observations remained in the analytic sample.

The majority of the sample lived in the United States (95.01%) and the average participant age was 40.10 years (SD = 12.81). Participant gender identification was 51.16% female, 47.6% male, and 1.34% non-conforming, nonbinary, or transgender. Racial identities among participants were 9.18% Black or African Origin, 5.31% East Asian, 74.90% White or European Origin, and 10.60% multiracial or one of several other identities. Hispanic or Latinx identifying participants represented 10.6% of the sample.

Power analyses were originally conducted in G*Power (Faul et al., 2007) assuming 10% exclusions from the planned sample of n = 5000. With approximately 1500 participants completing each dependent measure, most analyses would have had sufficient power (i.e., 90% at α = .05) to detect small interaction effects. Power sensitivity analyses with the analytic sample (total N = 3926) suggested sufficient power (i.e., 90% at α = .05) to detect fairly small interaction effects between induction and another factor in predicting outcome measures for most tests. The smallest detectable effect was slightly larger for the primary implicit measures analyses (n = 954-1036 with 72 groups; F[5, 882] = 2.22, η² = .017) than the primary explicit measures analyses (n = 1364 - 1438 with 48 groups; F[5, 1316] = 2.22, η² = .012) and Tajfel pull score analyses (n = 1331 with 96 groups; F[5, 1235] = 2.22, η² = .012).² Assuming a moderate correlation among repeated measures and a non-sphericity correction of 1, a mixed ANOVA had sufficient power to detect a very small 7 (item – within) by 6 (induction – between) interaction effect on the MAM ratings (n = 1351; F[30, 8070] = 1.46, η² = .002.³ Analyses including multiple dependent measures had the lowest power with only 167 to 272 participants per dependent measure combination; however, fairly small effects were detectable with 90% power and α = .05. For instance, the mixed ANOVA assessing differences in induction effects by implicit measure attribute (evaluation vs. identification) had enough power to detect a small interaction effect between induction and attribute type on ingroup bias (F[5, 157] = 2.27, η² = .036, with a moderate correlation between repeated measures).⁴

The study employed 6 (induction) x 2 (ingroup) x 2 (meaning) x 2 (procedure) x 2 (measure order) factorial design with two of six dependent measures randomly assigned to each participant. Two additional manipulations (i.e., reciprocity expectation and IAT stimuli) were included for relevant dependent measures. See the supplemental materials (Figure S1) for a graphic representation of this design.

HITs containing a link to the study were posted on MTurk. After selecting the HIT and providing consent, participants were randomly assigned to complete one of six induction procedures designed to induce membership with one of two groups (i.e., the Blue group or the Green Group). The six possible induction procedures included painting preference, dot estimation, blatant random assignment, member name memorization, imagined group membership, and participant choice. For ease of interpretation, we will use “ingroup” in reference to the participant’s assigned, memorized, imagined, or chosen group and “outgroup” in reference to the contrast group. In five of the induction conditions (i.e., all except the choice condition), participants were actually randomly assigned to an ingroup. Half of participants received group meaning information as part of their induction. Then, two-thirds of participants completed the Name-Group Association Task (NGAT).

Next, participants completed two of the following six possible dependent measures: an identification IAT, an evaluation IAT, explicit identification measures, explicit evaluation measures, the MAM, and the Tajfel Matrices. Of the participants assigned to complete one or both of the IATs, those in the NGAT condition were also randomly assigned to receive either the group name or member name versions of the IATs. All participants who did not complete the NGAT received the group name versions of the IATs. Participants assigned to complete both implicit measures received the same measure type for evaluation and identification. Of the participants asked to complete one or both of the discrimination measures, half were given information in the instructions designed to undermine norms of reciprocity. Participants assigned to complete both allocation tasks received the same instructions for each. Once both dependent measures were completed, all participants were asked the manipulation check and demographic questions, directed to create a unique identifier code, thanked for their participation, and debriefed. The codes were used to compensate participants on the MTurk website.

The study was programmed in JavaScript using Minno.js (Zlotnick et al., 2015) and hosted on the Project Implicit web server.

Participants completed one of six minimal group inductions. Possibilities included two classic inductions (i.e., painting preference and dot estimation), blatant random assignment, participant choice, member name memorization, and imagined group membership.

Painting Preference. For this induction, participants were told that their ratings of paintings would determine their group membership.⁵ They rated 12 paintings, 6 by Paul Klee and 6 by Wassily Kandinsky. This procedure was modeled after Tajfel and colleagues’ (1971) study that used Klee and Kandinsky painting preferences as their ostensible group induction. Paintings were displayed and rated individually on a 7-point Likert-type scale ranging from “Strongly Like” to “Strongly Dislike”. After rating the paintings, participants saw the text, “Calculating…”, appear on their screen with a 3 second countdown clock to maintain the illusion that their preferences would determine group membership. Then, they were informed that they were members of the Blue group or Green group based on their ratings.

Dot Estimation. This induction was also inspired by a classic induction used by Tajfel et al. (1971). Participants were presented with a series of twelve dot arrays and asked to estimate the number of dots in each one. Each array was displayed for 2 seconds; then, participants reported the number of dots they believed were in the image. In comparison to the original version, we presented fewer arrays (12 vs. 40) for longer (2 seconds vs. half a second or less) and included no practice trials. These alterations were made in the interest of time and improved participant experience. After the task, participants saw “Calculating…” on their screen for three seconds as described above before learning their group membership results.

Random Assignment. In this induction, participants were told, “For the purposes of this study, you will be randomly assigned to become a member of the Blue Group or the Green Group”. Then, after viewing a screen reading, “Assigning…”, for 3 seconds, participants were informed of their randomly determined group membership.

Participant Choice. In this induction, participants read, “For the purposes of this study, please choose to become a member of the Blue Group or the Green Group”. Then, they selected one of the two groups. The following screen confirmed their selection by informing participants of their group membership.

Name Memorization. This induction was adapted from Pinter and Greenwald (2011). Participants were asked to memorize a list of group member names from one of the two groups. Participants read, “The tasks that follow will be easier if you memorize the names of the members in one group.” Then, on the following screen, five group member names appeared under the heading of “Blue Group” or “Green Group” for 45 seconds. Names of the group members were either Alyssa, Michael, Rachel, Kevin, and Jordan or Kayla, Daniel, Amanda, Tyler, and Riley.⁶ Unlike the original study, the group membership of the names was counterbalanced.

Imagined Membership. This induction was based on the imagination induction used by Pinter and Greenwald (Experiment 1; 2011). Pinter and Greenwald reasoned that if imagined contact with outgroups can reduce prejudice (e.g., Crisp & Turner, 2009), then imagined group membership can create it. For this induction, participants were told that a small number of MTurk workers had been divided into two groups based on their art preferences.⁷ Participants were then instructed to imagine that they had rated a series of paintings and that they were placed in one of the two groups because of their art preferences.

Group Meaning. This manipulation was designed to increase identification with the ingroup and is loosely based on statements sometimes included as part of MGP inductions (e.g., Ashburn-Nardo et al., 2001) and manipulations with similar goals (e.g., Leonardelli & Brewer, 2001; Stroebe et al., 2005). In order for the same statement to cohere with all six inductions, we chose to use vague personality information in the tradition of the “Barnum effect” to achieve this effect (Meehl, 1956). Participants were informed that “past research on personality” has shown that members of their ingroup “are generally friendly people but at times can be quite shy.”

Name-Group Association Task. The Name-Group Association Task (NGAT; Pinter & Greenwald, 2011) was designed to familiarize participants with the group memberships of 10 people (e.g., five Blue Group members and five Green Group members). This task is necessary when participants respond to indirect measurement instruments that use the group member names as stimuli. The NGAT included two 30-trial blocks in which participants categorized names based on their group membership. In each trial, one of the ten names appeared in the center of the screen, and participants pressed the “E” or “I” key to sort the name as belonging to either the Blue Group or the Green Group. In the first block, the names appeared in bright blue and green to help participants learn their associated group. The group names switched sides in the second block, and the member names appeared in the same colors at 40% saturation. Text before each block informed participants of the task’s instructions and purpose. See Figure S2 in the supplemental materials for images of the NGAT as presented to participants.

Implicit Evaluation and Identification. Implicit evaluations and identifications were assessed using the Implicit Association Test (IAT; Greenwald et al., 1998). In the IAT, participants categorize pictures or words representing two concepts (e.g., “Green Group” and “Blue Group”) and two attributes (e.g., “Self” and “Other”). They sort stimuli that appear in the middle of the screen as belonging to categories represented on the right or left side of the screen using the “E” and “I” keys. In critical blocks of the task, participants simultaneously categorize concept and attribute stimuli. The response latencies during blocks with one pairing (e.g., “Blue Group” with “Self” and “Green Group” with “Other”) are compared to those during blocks with the opposite pairing (e.g., “Green Group” with “Self” and “Blue Group” with “Other”) to produce outcome measures.

The attribute categories for the evaluation IATs were “Good words”, with “laughter”, “happy”, “glorious”, “joy”, “wonderful”, “peace”, “pleasure”, and “love” as stimuli, and “Bad words”, with “awful”, “failure”, “agony”, “hurt”, “horrible”, “terrible”, “nasty”, and “evil” as stimuli. The attribute categories for the identification IATs were “Self”, with “I”, “me”, “mine”, “myself”, and “self” as stimuli, and “Other”, with “other”, “their”, “theirs”, “them”, and “they” as stimuli. The attribute category labels and stimuli were presented in orange font. “Blue Group” and “Green Group” served as group category labels for all IATs. The group name versions of the IATs used words representing the groups (i.e., “Blue”, “Blue Group”, “Green”, and “Green Group”) written in different sizes and fonts as stimuli. The member name version of the IATs used the names of the group members that were presented in the NGAT as stimuli. The group category labels and stimuli were presented in purple font. The IATs consisted of seven blocks of 20 or 40 trials following the recommendations of Nosek et al. (2005). See Figure S3 in the supplemental materials for images of the IATs as presented to participants.

Response latencies were analyzed as recommended by Greenwald et al. (2003) using the D algorithm. Latencies less than 400 ms and greater than 10000 ms were removed from calculations. Scores of participants who had latencies less than 300 ms for more than 10% of trials, greater than 30% error rates overall, or greater than 40% error rates in one block were excluded. The data were recoded so that positive D-scores indicated implicit preference for or identification with the ingroup.

Explicit Identification and Evaluation. Items measuring explicit identifications and evaluations were adapted from Pinter and Greenwald (2011). For each measure, participants responded to four statements on a 7-point Likert-type scale ranging from “Strongly Agree” to “Strongly Disagree,” with “Neither Agree nor Disagree,” as the midpoint. The explicit evaluation statements were “I like the [Blue/Green] Group,” and “The [Blue/Green] Group is good.” The explicit identification statements were “I identify with the [Blue/Green] Group,” and “I feel attached to the [Blue/Green] Group.” Reponses were recoded according to ingroup condition, and demonstrated high reliability: α = .892 for ingroup evaluation, α = .880 for outgroup evaluation, α = .855 for ingroup identification, and α = .874 for outgroup identification. For both explicit measures, the average of the two outgroup items was subtracted from the average of the two ingroup items to create difference scores. Positive scores indicated ingroup biases.

Participants assigned to complete explicit measures of evaluation or identification also responded to an additional single item measure. The explicit evaluation item asked participants to choose among seven statements that best described their evaluation, ranging from “I strongly prefer the Blue Group to the Green Group,” to “I strongly prefer the Green Group to the Blue Group.” The single item explicit measure of identification ranged from “I strongly identify more with the Blue Group than with the Green Group,” to “I strongly identify more with the Green Group than with the Blue Group.” These two measures were treated as Likert-type scales with the equal evaluation or identification serving as a meaningful zero; responses were recoded so that positive values indicated ingroup biases. Within each measure category, the four primary items were presented in a random order with the single item appearing first or last.

Allocation Tasks. We chose two allocation tasks to assess intergroup discrimination. The MTurk system allows the assignment of bonus payments as incentives or rewards for good performance. Participants were told that bonus payments would be made to other MTurk workers in a different version of the study based on their choices. Then, they were instructed to allocate small bonus payments to other workers based on the workers’ group membership.

Reciprocity. As a manipulation designed to undermine reciprocity expectations, half of participants were additionally informed before completing the tasks that they would not be receiving bonuses themselves. Specifically, they were told, “Unfortunately, you are not eligible to receive a bonus for your participation.” The other half of participants received no information as to whether they would receive bonuses.⁸

Tajfel Matrices. The Tajfel Matrices (Tajfel, 1970) served as one of the two potential discrimination measures. During this task, participants made six allocation decisions in a random order. Each decision matrix had 13 paired options, each with bonus amounts for a member of the Blue Group and for a member of the Green Group. The matrices were constructed based on descriptions by Bourhis et al. (1994) and adapted for use on MTurk (i.e., bonus amounts were presented as cents rather than as points). See the supplemental materials (Table S1) for allocation options by matrix type. Response were scored as instructed by Bourhis et al. to create “pull scores” for participants according to the strategies represented. Pull scores indicate the difference in ranked allocations between two paired matrices (i.e., when the two strategies are together and when they are opposed). Generally, positive pull scores suggest ingroup bias.

Multiple Alternative Matrices. A version of the Multiple Alternative Matrices (MAM; Bornstein et al., 1983) served as the other measure of discrimination. The allocation task instructions were adapted from those implemented by Gaertner and Insko (2000, Experiment 1) for use on MTurk. Participants first chose from seven possible distributions of bonuses to members of each group. Then, they rated the seven bonus pairs one at a time in a random order. See the supplemental materials (Table S2) for the allocation amounts according to the strategies they represent.

Two items were used as attention and manipulation checks. First, participants were asked to identify the minimal group induction instructions they were given at the beginning of the study from a list of seven options including, “I can’t remember.” Second, they were asked to identify their minimal ingroup from among “the Purple Group”, “the Blue Group”, “the Green Group”, “the Yellow Group”, and “I can’t remember”; the phrasing of the question depended on their response to the prior question (e.g., “What group did you imagine you were in?” if they selected, “I should imagine being assigned to a group”). Participants were also asked to report their age, race, ethnicity, gender identity, political orientation, and country of residence in a random order.

All dependent measures were recoded or calculated so that positive values suggest a preference for or identification with the participant’s ingroup. We coded the order of the dependent measures according to whether the focal measure was presented first or second and included it as a factor in most analyses.

Of the 9187 study sessions started, the 722 representing repeat participants were deleted. Remaining participants who had no valid data for any of the dependent measures (n = 1783) or who failed either manipulation check (n = 2756) were excluded from all analyses.⁹ IAT D-scores were excluded based on the criteria described above. Among participants who passed the manipulation checks and completed the evaluation IAT, 15.85% had their D-scores excluded, with exclusion rates of 10.48% for the group name version and 27.70% for the member name version. Identification IAT exclusion rates were 16.54% overall (i.e., 10.88% for the group name version and 28.73% for the member name version) among participants who passed manipulation checks.

The choice induction condition served as a positive control. As participants in this condition demonstrated ingroup bias on every dependent measure, analyses proceeded as planned. Because participants were randomly assigned to complete two of the six possible dependent measures, each analysis included a subset of participants. All participants with valid data for the relevant measure(s) in an analysis were included even if they had missing or excluded data on other measures. Our analysis plan and code were preregistered (https://osf.io/8m69a/), and both raw and cleaned data are publicly available (https://osf.io/6cbwf/). Deviations from our preregistered analyses were made in cases of errors and omissions and are noted below. See https://osf.io/p74zb/ for the implemented analysis code and output.

To determine whether the MGP produced ingroup biases in evaluation or identification, we compared implicit and explicit measures to zero with a series of one sample t-tests. Overall, participants demonstrated ingroup bias on explicit measures of evaluation (d = 0.60) and identification (d = 0.93) and implicit measures of evaluation (d = 0.37) and identification (d = 0.44). Within each induction condition, participants likewise demonstrated ingroup biases on all four measures (ps < .002). See Table 1 for full results.

We tested whether induction and other methodological features affected explicit evaluations and identifications with two 6 (induction) x 2 (meaning) x 2 (procedure) x 2 (measure order) factorial Analyses of Variance (ANOVAs). Analyses included main effects and two-way interactions. We examined any resulting effects using post-hoc pairwise comparisons with Bonferroni corrections. We predicted that analyses would reveal main effects of induction, meaning, and measure order for both measures. See Figure 1 for a comparison of ingroup biases by induction and measure and the supplemental materials for a comparison of ingroup biases by the procedural moderators and measure (Figure S4).

In the model predicting explicit evaluations (R²_adj = .021), we found main effects of induction condition (F[5, 1411] = 3.63, p = .003, η_p² = .013) and NGAT completion (F[1, 1411] = 4.73, p = .030, η_p² = .003). Participants in the choice induction condition demonstrated greater ingroup bias than participants in either the memorization (p = .002) or imagined membership conditions (p = .001); no other induction condition differences were reliable (ps > .145). Participants who completed the NGAT had less positive explicit evaluations compared to participants who did not, but this effect was qualified by an interaction with dependent measure order, F(1, 1411) = 14.30, p < .001, η_p² = .010. According to post hoc tests, participants who responded to the explicit evaluation measures first demonstrated weaker ingroup biases if they completed the NGAT than if they did not (M_diff = 0.44, SE = 0.11; p < .001). However, among participants who responded to the explicit evaluation measures second, NGAT completion did not influence ingroup bias magnitude (p = .226).

The model predicting explicit identifications (R²_adj = .041) revealed main effects of induction (F[5, 1337] = 10.80, p < .001, η_p² = .039), NGAT completion (F[1, 1337] = 5.94, p = .015, η_p² = .004), and dependent measure order (F[1, 1337] = 5.67, p = .017, η_p² = .004). According to post hoc tests, participants in the choice induction condition more strongly identified with their ingroup than participants in the painting rating (p = .002), imagined membership (p < .001), and memorization (p < .001) conditions. Participant identifications were also weaker in the memorization induction condition than in the random assignment (p < .001) or dot estimation (p = .001) conditions. Participants who completed the NGAT had weaker ingroup identification than those who did not (M_diff = 0.28, SE = 0.12). Participants who completed the explicit identification measure first also reported stronger ingroup identification than participants who completed the measure second (M_diff = 0.32, SE = 0.12).

Exploratory analyses of the single item explicit measures revealed similar patterns of results and can be found in the supplemental materials.

Next, we examined how induction and methodology affected implicit evaluations and identifications using 6 (induction) x 2 (meaning) x 3 (procedure & IAT type) x 2 (measure order) factorial ANOVAs including all main effects and two-way interactions. Procedure and IAT type, while separate manipulations, accounted for three groups of participants: 1) those who completed the NGAT and member name versions of the IAT, 2) those who completed the NGAT and group name versions of the IAT, and 3) those who completed the group name versions of the IAT without the NGAT. Thus, a single three-level factor was included in the model. Models were also fit including separate main effects of procedure and IAT type; see the supplemental materials for full results. For our implicit measure ANOVAs, we hypothesized the same main effects as for our explicit analyses but also predicted that participants in the memorization induction condition would demonstrate stronger ingroup biases on the member name versions of the IATs than on the group name versions. See Figure 2 for a comparison of ingroup biases by induction and implicit measure and the supplemental materials for a comparison of ingroup biases by the procedural moderators and implicit measure (Figure S5).

Despite the presence of ingroup biases in implicit evaluation, the model designed to explain variation in those biases performed poorly (R²_adj < 0), and we found no reliable main effects or interactions (ps > .092). Our model predicting implicit identifications was somewhat more successful (R²_adj = .019). We observed effects of induction (F[5, 919] = 2.53, p = .028, η_p² = .014), condition combination (F[2, 919] = 4.25, p = .009, η_p² = .009), and the interaction between them (F[10, 919] = 2.00, p = .030, η_p² = .021). Post hoc comparisons suggested that implicit identifications were weaker among participants in the painting rating induction condition than those in the choice induction condition (p = .034), but no other pairwise differences based on induction were reliable. The procedure - IAT type combination affected responding in the memorization induction only; implicit identifications were stronger when participants completed the NGAT and member name IAT (M = 0.37, SE = 0.07) than when they completed the NGAT and the group name IAT (M = 0.13, SE = 0.05, p = .017) or when they completed the group name IAT without the NGAT (M = 0.06, SE = 0.05, p < .001). See Figure 3 for visualization. Thus, our prediction that the member name version of the IAT would increase implicit biases in the memorization induction was confirmed for implicit identification but not evaluation.

Implicit vs. Explicit Measures. Descriptively, ingroup bias effects were stronger on explicit measures than on implicit measures (see Table 1). We examined whether the methodological factors differentially affected attributes (i.e., evaluation or identification) according to their measurement. We conducted mixed 2 x 6 x 2 x 2 x 2 ANOVAs including a within-participants factor of measure type (implicit vs. explicit) and between-participants factors of induction, meaning, procedure, and dependent measure order as predictors. Outcome measures were standardized using z transformations, which allowed for investigation of moderation effects, but rendered the main effect of measure type (implicit vs. explicit) as uninterpretable. The models were nested within participants and included all main effects, all two-way interactions, and all three-way interactions. When effects were identified, estimated marginal means were compared in post hoc tests with Satterthwaite adjusted degrees of freedom. In our analysis of ingroup evaluations (R²_marginal = .134) measure type did not interact with other factors in predicting bias (ps > .185). Our ingroup identification analysis (R²_marginal = .218) revealed an interaction between measure type and NGAT completion in predicting ingroup bias (F[1, 159] = 6.425, p = .012, η_p² = .023). Specifically, participants who did not complete the NGAT demonstrated stronger ingroup identification on explicit measures than on implicit measures (p = .043), but we found no such difference among participants who did complete the NGAT (p = .128), though trends were in the opposite direction. No other moderation effects were reliable.

Identification vs. Evaluation. Overall, explicit measures of identification and evaluation were positively correlated (r[312] = .596, p < .001), as were implicit measures of identification and evaluation (r[165] = .296, p < .001). To determine whether participants demonstrated stronger ingroup biases on measures of identification than evaluation as predicted, we performed mixed 2 x 6 x 2 x 2 x 2 (x 2) ANOVAs predicting explicit or implicit measures. The first factor, attribute, was within-participants and consisted of two levels: evaluation and identification. As listed, the other four factors were between-participants and included induction, meaning, procedure, measure order, and IAT version for the implicit analyses. The model was nested within participants and included all main effects, two-way interactions, and three-way interactions.

In our analysis of explicit bias magnitude (R²_marginal = .179), we found a main effect of attribute (F[1, 305] = 114.02, p < .001, η_p² = .204), such that ingroup identifications were stronger than ingroup evaluations. Attribute did not interact with any methodological factor in predicting bias. In our implicit bias analysis (R²_marginal = .144), the magnitude of ingroup bias did not differ according to attribute (p = .443), and attribute did not interact with any methodological factors in predicting implicit bias (ps > .085). Thus, our prediction that ingroup biases would be stronger for identification than evaluation was confirmed for explicit but not implicit measures.

Somewhat surprisingly, correlations were not reliable between implicit and explicit evaluations (r[210] = .133, p = .052) or between implicit and explicit identifications (r[184] = .140, p = .057). Still, we examined moderation effects of the methodological factors on these relationships using multiple regression as planned. Specifically, we investigated whether implicit evaluation (or identification) interacted with induction, meaning, procedure, dv order, or IAT type in predicting explicit evaluation (or identification). All main effects and two-way interactions were included in the models, which did reveal relationships between implicit and explicit measures. Participants’ implicit evaluations of their ingroups predicted their explicit evaluations, B = 1.27, F(1, 167) = 2.47, p = .049, η_p² = .023, although no moderation effects including implicit evaluations were significant (ps > .205). Participants’ implicit identifications scores also predicted their explicit identifications, B = 1.07, F(1, 141) = 3.97, p = .048, η_p² = .019. We found no interaction effects including implicit identification (ps > .103). Thus, none of the procedural variables moderated implicit-explicit relations; however, the power for these analyses was fairly low (i.e., 80% power to detect η² = .071 [evaluation] or η² = 0.083 [identification]).

The two allocation tasks implemented do not produce directly comparable measures of discrimination. We first analyzed their results separately and then compared a standardized measure. Across these analyses, we predicted main effects of induction, meaning, reciprocity, and measure order.

We examined both pull scores and overall group allocations from the Tajfel Matrices as measures of intergroup discrimination.

Pull Scores. First, we compared the pull scores calculated for each strategy to zero in a series of one sample t-tests.¹⁰ As a correction for familywise error, we used α = .008 (i.e., α = .05/6). These analyses were conducted overall and for each induction separately and are reported in Table 2. Nearly every strategy had a “pull” on participant allocation choices. The pulls of maximum ingroup profit and maximum differentiation (i.e., ingroup favoritism) on both maximum joint profit (d = 0.46) and parity (d = 0.43) were positive, both overall and within each induction group. However, we observed the strongest effects for the pull of parity on ingroup favoritism (d = 1.43) and the pull of maximum ingroup profit and maximum joint profit on maximum differentiation (d = 0.85). The pattern of results suggested that participants were generally pulled by strategies that benefited the ingroup, especially when the outgroup also profited. Correlations among pull scores can be found the supplemental materials (see Table S3). Exploratory pairwise comparisons with Satterthwaite adjusted degrees of freedom revealed that the parity pull score was stronger than any other pull score (ps < .001). See Figure 4 for a visualization of pull scores by induction condition.

Next, we examined the impact of methodological factors on pull scores in a series of 6 x 2 x 2 x 2 x 2 ANOVAs.¹¹ Main effects of induction, meaning, procedure, reciprocity, and measure order and all two-way interactions were included in models. When effects were found, we examined pairwise differences in post hoc comparisons with Bonferroni corrections. The model predicting the pull of maximum ingroup profit and maximum differentiation (i.e., ingroup favoritism) on maximum joint profit (R²_adj = .040) revealed main effects of induction (F[5, 1261] = 6.97, p < .001, η_p² = .026) and reciprocity (F[1, 1261] = 20.26, p < .001, η_p² = .015) and an interaction between induction and reciprocity (F[5, 1261] = 2.32, p = .042, η_p² = .009). Participants in the memorization induction condition had weaker ingroup favoritism pull scores than those in every other induction condition (ps < .006) except imagined membership. Undermining reciprocity also reduced the pull of ingroup favoritism on maximum joint profit (M_diff = -1.28, SE = 0.29), but post hoc tests suggested that this difference was only reliable in the choice, dot estimation, and random assignment induction conditions (ps < .034). The companion model predicting the pull of maximum joint profit on ingroup favoritism (R²_adj < 0) found no reliable effects.

The model predicting the pull of maximum ingroup profit and maximum joint profit on maximum differentiation (R²_adj = .022) revealed main effects of induction condition (F[1, 1261] = 2.582, p = .025, η_p² = .010), reciprocity (F[1, 1261] = 23.257, p < .001, η_p² = .018), and measure order (F[1, 1261] = 10.08, p = .002, η_p² = .008). These pull scores were lower when reciprocity was undermined (M_diff = -1.35, SE = 0.32) and when the Tajfel Matrices were the second dependent measure (M_diff = -0.93, SE = 0.32). Participants in the choice induction condition also had stronger pulls than those in the memorization condition (p = .032), but no other no pairwise differences between induction conditions were reliable. In the model predicting the pull of maximum differentiation on maximum ingroup profit and maximum joint profit (R²_adj = .006), we found main effects of the NGAT (F[1, 1261] = 6.00, p = .014, η_p² = .005) and an interaction between reciprocity and measure order (F[1, 1261] = 6.68, p = .010, η_p² = .005). Post hoc tests suggested that pulls were stronger when participants did not complete the NGAT (M_diff = 0.38, SE = 0.18) and that measure order impacted pulls when reciprocity was not undermined (M_diff = -0.59, SE = 0.25; p = .019) but not when it was (p = .223).

In the model predicting the pull of parity on maximum ingroup profit and maximum differentiation (i.e., ingroup favoritism; R²_adj = .017), we observed main effects of induction (F[1, 1261] = 2.61, p = .023, η_p² = .010) and reciprocity (F[1, 1261] = 7.20, p = .007, η_p² = .006), and an interaction between induction and reciprocity (F[1, 1261] = 2.87, p = .014, η_p² = .011). No pairwise comparisons between induction conditions were reliable. Undermining reciprocity increased the pull of parity for participants overall (M_diff = 0.87, SE = 0.33), but post hoc tests suggested that this condition difference was only reliable within the choice (p < .001), dot estimation (p = .023), and random assignment (p = .043) induction conditions. The model predicting the pull of ingroup favoritism on parity (R²_adj = .019) revealed main effects of induction (F[1, 1261] = 3.86, p = .002, η_p² = .015) and reciprocity condition (F[1, 1261] = 20.14, p < .001, η_p² = .015). Undermining reciprocity reduced the pull of ingroup favoritism (M_diff = 1.31, SE = 0.30), and pulls were stronger in the choice condition than the memorization condition (p = .033); no other pairwise comparisons between induction conditions were reliable. See Figure 5 for a visualization of pull scores by reciprocity condition.

Group Allocation Totals. To determine how the total bonus money awarded to ingroup and outgroup members varied according to methodological factors, we first summed allocations to each group across the six matrix items. Then, we included the group (ingroup vs. outgroup) as a within-participants variable in a 2 x 6 x 2 x 2 x 2 x 2 mixed-model ANOVA predicting bonus allocations. The other five factors (i.e., induction, meaning, procedure, reciprocity, and dependent measure order, respectively) were between-participants. The model included all main effects and all two- and three-way interactions nested within participant. This model initially produced a singular fit, which we identified as caused by the near-zero estimate for the random effect of participant. We removed this random intercept from the model to eliminate the issue, but results should be interpreted cautiously due to the resulting model assumption violation (i.e., non-independence of observations). When effects were identified, estimated marginal means were compared in post hoc tests with Bonferroni corrections.

The model (R²_adj = .160) identified main effects of group (F[1, 2556] = 443.30, p < .001, η_p² = .148), reciprocity (F[1, 2556] = 14.79, p = .001, η_p² = .006), and dependent measure order (F[1, 2556] = 6.83, p = .009, η_p² = .003). As expected, participants allocated less bonus money (in cents) overall to outgroup members (M = 85.20, SE = 0.43) than to ingroup members (M = 97.30, SE = 0.43). Overall, allocations also decreased when participants knew they would not be receiving bonuses themselves (M_diff = -1.96, SE = 0.62) and when they completed the task second (M_diff = -1.43, SE = 0.62). We also found interaction effects between group and reciprocity (F[1, 2556] = 31.201, p < .001, η_p² = .013), group and NGAT completion (F[1, 2556] = 4.505, p = .034, η_p² = .002), group and induction (F[5, 2556] = 8.32, p < .001, η_p² = .016), and group, reciprocity, and induction (F[5, 2556] = 3.16, p = .007, η_p² = .006). The reciprocity manipulation appeared to influence ingroup allocation totals (M_diff = -5.24, SE = 0.86, p < .001) but not outgroup allocation totals (p = .126). The interaction between group and induction suggested that discrimination varied by induction. Post hoc tests indicated that participants allocated larger bonuses to ingroup members than outgroup members in all induction conditions. Outgroup allocations did not differ according to induction (ps > .850), but participants in the memorization induction condition allocated less money to ingroup members than participants in any other induction condition (ps < .002) except for imagined membership (p = .288). Finally, examining the three-way interaction suggested that reciprocity affected ingroup, but not outgroup, allocations for participants in the choice, dot estimation, and random induction conditions (ps < .001) but did not influence total allocations to either group among participants in the other induction conditions. See Figures S6 and S7 in the supplemental materials for visualizations of total allocations by group, induction, and reciprocity.

We also assessed discrimination as measured by the MAM, which included a choice item and seven rating items.

MAM Choice. We first analyzed the single item choice using a chi-square goodness of fit test to determine whether ingroup favoring allocations were chosen more often than would be expected by chance. This analysis was repeated for each induction separately and is reported in Table 3. Overall and within each induction condition, participants were more likely to choose parity, or equal allocations, than any other choice. However, ingroup favoring allocations were chosen more frequently than those favoring the outgroup. See Figure 6 for visualization.

Then, we used multinomial logistic regression to test the effects of methodological factors on allocation choice with equal allocation as the reference group. Induction, reciprocity, meaning, procedure, measure order, and all two-way interactions were included as predictor variables. The model (χ²[6] = 17.201, p = .009) indicated that the probability of choosing maximum joint profit favoring the ingroup decreased when participants were told they could not receive bonuses themselves (p = .031). See Figure 7 for illustration of this effect. Although we observed minor variations in the patterns of probabilities within particular condition combinations, we found no other noteworthy effects.

MAM Ratings. Next, we compared the MAM allocation ratings to the scale midpoint in a series of one-sample t-tests. We tested each item with α = .007 (i.e., α = .05/7) to correct for family-wise error. Overall, participants positively rated allocations that maximized ingroup absolute earnings (d = 0.25), maximized joint earnings with ingroup advantage (d =0.36), and minimized group differences (d = 1.48), while they negatively rated allocations with outgroup advantages whether they maximized relative earnings (d = -0.83), maximized absolute earnings (d = -0.50), or maximized joint earnings (d = -0.26). Positive ratings for maximizing the ingroup’s relative earnings were also given by participants in the choice (d = .18) and picture rating (d = .25) induction conditions, but not overall or in the other induction conditions. See Table 4 for the full results. Interestingly, correlations between ratings were positive across the non-equal ratings (see Table S3 in the supplemental materials). Seemingly, participants tended to respond negatively or positively to unequal items regardless of the group benefit, though correlations were strongest among allocations favoring the same group. See Figure 8 for a visualization of MAM ratings by induction condition.

We analyzed the effect of condition and allocation strategy on the MAM ratings with a 7 x 6 x 2 x 2 x 2 x 2 mixed ANOVA. The first factor was allocation type (within-participants) and has seven levels, one for each MAM item. The subsequent between-participants factors were induction, meaning, procedure, reciprocity, and dependent measure order. All main effects and two-way and three-way interactions were included in a model nested within participants. We chose this analytic approach to account for individual variability in responding across items in our estimates of procedural moderation effects (i.e., whether these variables influence the relationship between allocation type and ratings). However, any main effects of the moderators themselves are not very meaningful because of this approach. For effects identified, we compared estimated marginal means in post-hoc tests with Satterthwaite adjusted degrees of freedom.

In this model (R²_marginal = .296), we identified main effects of allocation type (F[6, 7878] = 729.45, p < .001, η_p² = .349), induction (F[5, 1279] = 2.24, p = .049, η_p² = .001), NGAT completion (F[1, 1279] = 7.87, p = .005, η_p² = .001), and reciprocity (F[1, 1010] = 6.65, p = .010, η_p² = .001). Pairwise comparisons indicated that ingroup favoring allocations were rated more positively than outgroup favoring allocations. Further, ratings overall were more positive when participants completed the NGAT than when they did not (M_diff = 0.20, SE = 0.07), and less positive when they learned they could not receive bonuses themselves than when no such information was given (M_diff = -0.18, SE = 0.07). These latter two effects are not interpretable in terms of ingroup bias since the allocations rated varied in their group benefits. More notably, we found interactions between allocation type and induction (F[30, 7878] = 6.70, p < .001, η_p² = .024), NGAT completion (F[6, 7878] = 2.34, p = .029, η_p² = .002), and reciprocity (F[6, 7878] = 17.88, p < .001, η_p² = .013). While the pattern of ratings was similar across induction condition, the magnitude of bias varied across induction. For instance, ratings of participants in the memorization condition were less biased toward the ingroup than those of participants in at least one other condition for every unequal allocation. When reciprocity was undermined, ratings of all strategies favoring the ingroup decreased in positivity (ps < .001) and ratings of maximizing relative profit in favor of the outgroup decreased in negativity (p = .022). See Figure 9 for a visualization of MAM ratings by reciprocity condition. NGAT completion decreased bias on some ratings, reducing negativity toward allocations favoring the outgroup (ps < .010), but increasing positivity toward maximizing absolute profit for the ingroup (p = .010). These effects were further qualified by three-way interactions between allocation type, NGAT completion, and reciprocity (F[6, 7878] = 2.20, p = .040, η_p² < .001) and allocation type, induction, and measure order (F[30, 7878] = 1.77, p = .006, η_p² = .004). Seemingly, participants who completed the NGAT rated all non-equal allocations more positively regardless of group benefit than those who did not, but only if they were in the no reciprocity condition (ps < .008). Otherwise, the NGAT had no impact (ps > .111). Pairwise comparisons revealed no meaningful patterns in how induction and dependent measure order impacted ratings depending on allocation type. We also observed interactions between NGAT completion and reciprocity (F[1, 1279] = 8.20, p = .004, η_p² = .001), as well as induction, meaning, and dv order (F[5, 1279] = 3.73, p = .002, η_p² = .002). However, these effects are not interpretable given the setup of the analysis.

In order to examine whether the two discrimination measures revealed equivalent minimal group effects, we first standardized task responses. For participants assigned to complete both allocation tasks, the ingroup and outgroup allocations on the single choice item from the MAM and the total ingroup and outgroup allocations across all the Tajfel Matrices were calculated and z-transformed to create four scores per participant.¹² Then, we used a 2 x 2 x 6 x 2 x 2 x 2 x 2 Mixed ANOVA to test whether the within factors of group and task and the between factors of induction, meaning, procedure, reciprocity, and outcome order affect standardized bonus allocations. This model included all main effects, two-way interactions, and three-way interactions nested within participant.

The model (R²_marginal = .114) yielded a main effect of group (F[1, 1822] = 43.10, p < .001, η_p² = .022) but not of task (p = .393). Participants made larger allocations to ingroup members than outgroup members. Group did not interact with task in predicting allocations (p = .681); seemingly, the magnitude of discrimination effects was similar for the two tasks. We also observed interactions between group and induction (F[5, 1822] = 2.44, p = .032, η_p² = .006), group, induction, and procedure (F[5, 1822] = 2.40, p = .035, η_p² = .006), group, induction, and reciprocity (F[5, 1822] = 2.47, p = .031, η_p² = .006), and induction, meaning, and procedure (F[5, 1822] = 3.37, p = .006, η_p² = .008). Using marginal mean comparisons with Satterthwaite adjusted degrees of freedom, we confirmed that these effects paralleled those found in the analyses of the individual measures. No interaction effects including task were reliable (ps > .054).

Decades of research using the Minimal Group Paradigm (MGP) has revealed unique insights into the sources of intergroup biases. However, implementation of the MGP and measurement of relevant outcomes varies across studies, and systematic investigation of how those features impact minimal group effects is lacking. In the present research, we examined how methodological variations in the MGP influence ingroup biases as assessed by two allocation tasks and implicit and explicit measures of evaluation and identification. We tested six induction procedures, including group choice as a positive control, and other procedural variations, such as manipulations designed to increase the meaning of the groups or to undermine assumed reciprocity in the allocation tasks.

Participants demonstrated ingroup biases on measures of evaluation, identification, and discrimination both overall and in most condition combinations. On most outcome measures, ingroup biases varied according to induction, and the choice induction consistently produced the largest effects. The meaning manipulation generally did not impact biases, while the procedural moderator of completing the Name-Group Association Task (NGAT) influenced biases in some condition combinations. Eliminating assumed reciprocity also reduced discrimination as measured by both the Tajfel Matrices and the Multiple Alternative Matrices (MAM). Notably, both allocation tasks also revealed strong preferences for equitable outcomes in addition to those benefiting the ingroup. Overall, our results suggest that many variations of the MGP are sufficient to produce the expected ingroup bias effect, but both theoretically relevant and irrelevant features of the procedure and outcome measures moderate that effect.

The present research included three types of dependent measures (i.e., evaluation, identification, and discrimination). We found evidence of ingroup bias across measure types, and many of the tested outcomes demonstrated expected relationships with one another.

Prior investigations have found consistent evidence of minimal group effects on implicit and explicit measures of identification and evaluation (e.g., Ashburn-Nardo et al., 2001; Dunham, 2013; Pinter & Greenwald, 2011). Our results largely align with this work. All four measures, including multiple versions of the implicit measures and alternative explicit measures, demonstrated ingroup bias effects. Consistent with Dunham (2013), we observed relationships between explicit measures of identification and evaluation and between implicit measures of identification and evaluation, but the correlations between explicit and implicit measures of the same attribute were not reliable. These very small or non-existent relationships between implicit and explicit measures may be best explained by features of the attributes being assessed (e.g., Nosek, 2005; Schmidt et al., 2022); for instance, the evaluations and identifications produced by the MGP are not well-elaborated, a feature that increases implicit-explicit correspondence.

A few discrepancies with prior work are worth noting. While Pinter and Greenwald (2011) found consistently large ingroup bias effects for explicit and implicit measures, our effect sizes were medium to large on explicit measures and small to medium on implicit measures. Their large effects on implicit measures may be explained in part by a measurement artifact; we discuss this possibility below. Additionally, Dunham (2013) found stronger effects for implicit identifications than for implicit evaluations. While minimal groups effects were larger for explicit identifications than explicit evaluations in the present research, we found no difference in bias magnitude between the two implicit measures. Dunham likewise found larger ingroup biases in explicit identifications (d = .74) than explicit evaluations (d = .52), but the magnitudes of those effects were not directly compared in analyses. Overall, however, our research coheres with and adds to the evidence that the MGP produces implicit and explicit ingroup biases in identification and evaluation.

We found consistent patterns of ingroup biases on our discrimination measures, which were derived from the Tajfel Matrices and the MAM. The pull scores of ingroup favoritism on parity and on maximum joint profit and the ratings of allocations maximizing the ingroup’s absolute earning and maximizing joint earnings with an ingroup advantage suggested small to medium ingroup bias effects. As strategies, maximum differentiation on the Tajfel Matrices and maximizing relative ingroup profit on the MAM were not favored, however. These strategies both maximize the differences between the outgroup and ingroup amounts while sacrificing ingroup profit.

Another notable finding was the general support for equitable outcomes on both allocation tasks. The pattern of results from both tasks suggest that equity was as important or more important than ingroup profit for participants. In responding to the Tajfel Matrices, participants appeared to have been pulled more strongly by parity than any other strategy. Likewise, minimum differentiation on the MAM was chosen most often and rated most highly. These findings are consistent with some previous research using the Tajfel Matrices (Perreault & Bourhis, 1998; Sachdev & Bourhis, 1991) and MAM (Gaertner & Insko, 2000). While the MGP did produce patterns of discrimination in participant allocations, ingroup bias was overshadowed by egalitarianism.

Our research compared six previously implemented minimal group induction procedures. Consistent with previous research demonstrating variation in biases based on induction (e.g., Gaertner & Insko, 2000; Perreault & Bourhis, 1998, 1999; Pinter & Greenwald, 2011), we found a main effect of induction condition in several outcome analyses. As can be seen in Figure 10, the choice condition generally produced the strongest ingroup biases while the name memorization and imagined membership conditions generally produced the weakest ingroup biases. Post hoc pairwise comparisons typically only found differences involving one or more of these three induction conditions. However, we detected evidence of minimal ingroup bias in almost every combination of induction and outcome measure. Even creating associations with a group through imagination or the memorization of member names, instead of actual group membership, was sufficient to produce minimal group effects. These robust findings suggest that while varying induction procedures can influence the magnitude of effects, minimal groups produce ingroup biases regardless of the ostensible basis of their formation.

Consistent with previous work (e.g., Perreault & Bourhis, 1998), random assignment proved to be an effective induction procedure. However, we found no evidence that participants in the choice, dot estimation, or painting preference conditions had stronger biases than those in the random assignment conditions. This finding contradicts previous research demonstrating stronger ingroup biases in response to meaningful inductions (e.g., choice or painting preference) than in response to a random assignment induction (Gaertner & Insko, 2000; Herringer & Garza, 1987; Perreault & Bourhis, 1999).

Our findings regarding the memorization induction somewhat challenge the results reported by Pinter and Greenwald (2011), who found stronger or comparable effects of this induction compared to random assignment, imagined membership, and painting preference. In our study, the memorization induction produced weaker ingroup biases than other inductions on many outcome measures. Most notably, unlike Pinter and Greenwald’s findings, memorization did not produce the strongest minimal group effects on implicit outcomes. The procedural moderators included in our study may partially explain these discrepancies. Specifically, we found that participants in the memorization condition who completed the NGAT and the member name IAT had stronger implicit identifications than participants who completed the group name IAT both with and without the NGAT. Using the member names as stimuli for the IAT, as was done by Pinter and Greenwald (2011), increased implicit identification among participants in the memorization induction condition. Ingroup bias effects were strong for memorization induction participants only on the outcome of implicit identification and only within that condition combination. Perhaps the increased familiarity and fluency of the stimuli themselves influenced responding as an artifact. Thus, the memorization induction in and of itself was a relatively ineffective means of producing ingroup bias.

We tested several procedural moderators in addition to group induction that were derived from or inspired by previous work using the MGP.

Our manipulation to increase group meaning by providing personality information was inspired by the perspective that minimal group effects depend on participant ingroup identification (e.g., Bourhis et al., 1997; Brown, 2000; Gagnon & Bourhis, 1996). Minimal group induction procedures sometimes include information designed to imbue the minimal groups with additional meaning or distinctive characteristics (e.g., Ashburn-Nardo et al., 2001). Manipulations designed to foster ingroup identification by suggesting shared qualities appear to increase discrimination (Leonardelli & Brewer, 2001), though such effects may not be reliable (Stroebe et al., 2005). Although, in the present research, more ostensibly meaningful inductions (e.g., choice) appeared to produce stronger ingroup biases than less meaningful inductions (e.g., memorization) on some outcomes, our manipulation designed to imbue group membership with additional meaning had little effect on ingroup bias. Interestingly, an exploratory analysis of our single item measure of identification did reveal stronger identifications among participants who received the meaning information than among those who did not. However, we found no evidence that this effect occurred on the other outcomes. Perhaps the manipulation was too brief or generic to increase participants’ connection to their minimal group. To our knowledge, our investigation represents the first study to examine whether increasing the meaning of minimal group categorizations affects ingroup biases across multiple inductions. The failure of this manipulation to influence outcomes suggests that such information is unnecessary to produce ingroup biases and may have little impact on them.

Some researchers have argued that reciprocity norms and expectations underlie discrimination resulting from the MGP (e.g., Rabbie et al., 1989; Yamagishi & Kiyonari, 2000). Consistent with previous work (e.g., Gaertner & Insko, 2000; Stroebe et al., 2005), undermining reciprocity did influence discrimination measures from both allocation tasks. When participants were told that they were not eligible to receive bonuses, they discriminated less than when they received no such information. This manipulation impacted most of the Tajfel pull scores, the overall allocations on the Tajfel Matrices, the MAM choice, and many of the MAM ratings. Our results suggest that undermining reciprocity reduced ingroup favoritism, specifically. The manipulation appeared to have more influence on MAM ratings of the ingroup favoring allocations than the outgroup favoring allocations and impacted the ingroup allocation totals but not the outgroup allocation totals across the Tajfel Matrices.

In an exploratory analysis not reported above, we further tested the strength of this manipulation by examining the measures of evaluation and identification when they were completed after a measure of discrimination. Undermining reciprocity on the allocation task did not impact ingroup biases on subsequent evaluations and identifications (ps > .75). Thus, while discrimination in favor of the minimal ingroup may arise in part from the expectation of reciprocity (e.g., Gaertner & Insko, 2000; Rabbie et al., 1989), other minimal ingroup biases likely have alternative mechanisms (e.g., see Dunham, 2018), and reciprocity expectations alone do not explain discrimination effects.

Pinter and Greenwald (2011) devised the NGAT as a procedural necessity so participants could complete the indirect measurement tasks used in their research. To respond to the names of the ingroup and outgroup members as stimuli, participants had to learn to which group each name belonged. Dunham (2013) similarly employed this task. We tested the effect of the NGAT procedure on all outcomes, including alternative group name versions of the IATs.

As previously discussed, the memorization induction was uniquely impacted by the procedure - member name IAT combination by increasing bias magnitude. We also found that completing the NGAT decreased ingroup biases in explicit evaluations and identifications in some conditions. Analyses of the discrimination outcomes likewise revealed evidence that the NGAT reduced bias in some cases. Perhaps learning member names of both groups through the NGAT dampens bias because it identifies outgroup members. Alternatively, the delay between the induction and outcome measure during completion of the task may have decreased the impact of the inductions.

The apparent difficulty of the member name IATs is also notable. Nearly 30% of IAT D-scores were excluded from participants who received this version of the IAT. These rates were nearly triple those of the group name IAT versions. Seemingly, the NGAT was not sufficient for participants to learn the group members’ names, which, in turn, did not serve as effective IAT stimuli. Taken together, our results suggest that the member name version of the IAT, which necessitates the use of the NGAT, should be avoided in research using the MGP.

Participants demonstrated stronger biases for outcome measures they completed first instead of second on some measures, such as explicit identification. In combination with our NGAT findings, these effects suggest that outcome proximity to the induction may influence ingroup bias. Alternatively, having expressed their bias on one measure might alter how participants respond to a second measure, perhaps due to fatigue or carry-over effects.

Our study is not without its limitations. First, we chose to focus on procedural moderators in the present research, only manipulating features of the MGP and outcome measures. Many theoretically relevant moderators have been examined in prior research that did not fit within those constraints (e.g., group size and status). If and how those moderators interact with those examined here, or whether they differentially affect outcome measures, was outside the scope of our investigation.

Second, our chosen procedures and outcome measures by necessity limit the generalizability of the research. Induction condition did sometimes impact the magnitude of ingroup bias; the inductions included and their specific implementation could have influenced results. For instance, recent research suggests that group category labels relevant to the induction (e.g., overestimator vs. underestimator groups in a dot estimation induction) may impact some minimal group effects (Hong & Ratner, 2021). This finding illustrates the potential influence of incidental methodological features not considered in the present research.

We implemented relative measurement tasks rather than absolute ones for our implicit and explicit measures of evaluation and identification. This choice eliminated our ability to determine whether the minimal group effects stem from ingroup positivity and attachment or outgroup negativity and attachment. However, bias is by definition relative, and this approach allowed for clear comparisons across measures of evaluation and identification. For our explicit measures, we employed previously implemented indices (Pinter & Greenwald, 2011), and exploratory analyses of their single item alternatives revealed the same general pattern of results. For our implicit measures, we used two versions of the IAT. Though previous research implemented the member name version of the task (Dunham, 2013; Pinter & Greenwald, 2011), we had to exclude three times as many scores from this version as from the group name version according to preregistered performance exclusions. This high exclusion rate may be concerning, but IAT version did not influence the magnitude of ingroup bias as a main effect on either implicit evaluations or identifications. The ability of the MGP to impact implicit measures of intergroup bias is fairly well established; research using alternative indirect measurement tasks, such as priming tasks (e.g., Otten & Wentura, 1999; Xiao & Van Bavel, 2019), have also found implicit ingroup biases. Though the IAT has arguably the best psychometric properties of mainstream indirect measurement tasks (Bar-Anan & Nosek, 2014), alternatives to the IAT may have been more or less sensitive to our procedural manipulations.

For our discrimination measures, we elected to use both the Tajfel Matrices and the MAM due to their prominence in the MGP literature. The MAM was developed in part to address the limitations of the Tajfel Matrices in isolating allocation strategies (Bornstein et al., 1983). To be sure, interpreting the pattern of Tajfel Matrix pull score results was difficult because of the strategy combinations. Notably, though, we found similar patterns of results on the measures derived from the Tajfel Matrices and the MAM, which suggests that they may capture similar biases in resource allocation. Economic games have commonly been implemented as discrimination measures in previous MGP research (e.g., Yamagishi & Kiyonari, 2000; Yamagishi & Mifune, 2008; for meta-analysis, see Balliet et al., 2014); such alternatives may have produced a different pattern of results. However, our choice of tasks allowed us to examine effects across a variety of derived metrics, including pull-scores, total allocation amounts, allocation choices, and allocation ratings. The reference to bonus allocation in the present research may also have affected results; researchers have found stronger discrimination in allocations of positive outcomes rather than negative outcomes (Gardham & Brown, 2001; Otten et al., 1996).

Finally, our high rates of participant exclusions are also worth noting. Though we determined exclusion criteria a priori based on failures to pass manipulation checks, we did not anticipate how extensively they would apply to our sample. Despite collecting some additional data to increase the sample size, our power was still lower than anticipated. Consequently, our analyses including multiple outcome measures should be interpreted cautiously due to their comparatively low power. Low power may also partially account for the null findings and poor model performance in some analyses. However, most of our models were designed to examine the impact of procedural factors on ingroup bias magnitude. Model performance would have been much better if ingroup was included as a predictor of outcomes that were not scored to index ingroup bias. Thus, null effects do not suggest that minimal group effects were not reliable but, rather, that procedural variations did not impact their magnitude. Further, our sample sizes were still much larger than most prior investigations using the MGP.

The present research evaluated the effects of six minimal group inductions and several procedural manipulations on measures of discrimination and explicit and implicit evaluation and identification. We found consistent evidence of ingroup bias regardless of procedure or outcome. Of the procedural moderators, induction condition and the allocation reciprocity manipulation had the most consistent effects. Future researchers can use our findings to inform their methodological decisions in implementing the MGP. A unified account of minimal group effects must consider and interpret how variations in the MGP impact multiple outcome measures in order to sufficiently explain their psychological bases.

Contributed to conception and design: KS

Contributed to acquisition of data: KS

Contributed to analysis and interpretation of data: KS and RGD

Drafted and revised the article: KS and RGD

Approved the submitted version for publication: KS and RGD

KS was affiliated with Southern Illinois University when the study was designed and the data were collected.

KS was supported by a grant from the John Templeton Foundation (#61825). The funding source was not involved in the study design, in the collection, analysis, and interpretation of the data, in the writing of the report, or in the decision to submit the article for publication.

KS has no potential competing interests to declare. RGD is a consultant for Project Implicit, Inc., a nonprofit organization with a mission to “educate the public about bias and provide a ‘virtual laboratory’ for collecting data on the internet.”

Supplementary results, tables, and figures can be found at https://osf.io/frb3p/.

Study preregistrations, materials, raw and clean data, and analysis code and output are available on the Open Science Framework at https://osf.io/k5pv4/.