Contribution
The first meta-analysis restricted to contact interventions, that is, programmes deliberately run in schools, universities and other field settings, as against the mostly correlational surveys pooled by Pettigrew and Tropp (85% of its 73 documents were not in theirs). It adds delayed posttests, a comparison of direct with indirect programmes, and a comparison of regions with and without protracted conflict.
Key insight
For attitudes measured soon after a programme, it did not matter much whether participants met the other group or only met it at one remove. Extended contact programmes, in which children or students read or watch stories of friendship between a member of their own group and a member of the other, did at least as well as programmes that brought the groups together.
Assumptions
- Inclusion. A field programme based on contact theory that ensured contact actually occurred (desegregation, housing, exchange and tourism studies, and unstructured mixed camps excluded); imagined and parasocial (mass-media) contact excluded by definition; at least one measure of ethnic prejudice, cognitive, affective or conative (behavioural intention); English or German; documents to mid-2012.
- Designs. A primary cluster of randomized posttest-only and pretest–posttest designs with control groups (Hedges’s g, Morris’s g), and a secondary cluster of single-group pre–post designs (Becker’s g), analysed apart because they mix the intervention with time effects.
- Types. Direct: contact meetings (structured dialogue about the intergroup relation) and cooperative learning (joint work on a task not about the relation). Indirect: extended contact (“books, picture stories, radio plays, or films that explicitly display friendships or positive relations” across the line) and virtual contact (exchange between members of the groups by computer).
- Model. Random effects with REML for the means; mixed-effects meta-regression, one moderator at a time, for the hypotheses; the unit is the comparison, with dependent effects averaged within it.
Key results
- Overall, immediate (Table 2). Primary 0.28 [0.21, 0.35], k = 79, N = 9,212, I² = 71%; secondary 0.39 [0.31, 0.47], k = 44, N = 2,159.
- Overall, delayed (1–12 months). Primary 0.23 [0.13, 0.34], k = 8, N = 1,186, I² = 10%; secondary 0.35 [0.21, 0.50], k = 17, N = 464. No study measured beyond a year.
- Direct against indirect (Table 3), immediate only. Primary: direct 0.29 [0.21, 0.37], k = 63; indirect 0.23 [0.07, 0.38], k = 16; Q = 0.47, p = .49. Among majorities only, still no difference (p = .98). Secondary: direct 0.41, k = 37; indirect, all virtual, 0.33, k = 5.
- Subtypes, primary cluster. Contact meetings 0.31 [0.21, 0.41], k = 39; cooperative learning 0.25 [0.13, 0.38], k = 24; extended contact 0.42 [0.19, 0.65], k = 8; virtual 0.08 [−0.13, 0.28], k = 8. The subtype moderator is not significant (Q = 5.82, df = 3, p = .12).
- Conflict regions. Primary 0.31 (k = 11) against 0.27 (k = 68), no difference.
- Status. Primary: majorities 0.38 (k = 39), minorities 0.20 (k = 20), difference significant only one-tailed (two-tailed p = .0546, footnote 10). Only one indirect comparison concerns a minority.
- Generalization. Primary: individuals met 0.26 (k = 12), the whole target outgroup 0.32, unspecified “other ethnic groups” 0.21; no difference. No comparison measured attitudes toward a specific uninvolved outgroup.
- Quality. Randomized comparisons gave larger effects (0.43, k = 19) than non-randomized (0.24, k = 60), and lower attrition larger effects. No sign of publication bias by three tests.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Field contact programmes improve self-reported ethnic attitudes, about 0.28 SD in controlled designs | moderate | 79 comparisons, heterogeneous (I² = 71%); mostly non-randomized |
| C2 | The improvement persists for up to a year | weak | 8 controlled comparisons |
| C3 | Indirect programmes are as effective as direct ones | weak to moderate | between-study moderator test, 16 indirect comparisons, immediate posttests only |
| C4 | Extended contact programmes are effective | moderate | 0.42, 8 comparisons, CI excludes zero |
| C5 | Virtual contact programmes are effective | weak | null in the primary cluster; only single-group designs support it |
| C6 | Effects generalize to the whole outgroup and to unspecified outgroups | moderate | comparable effects by level of the outcome measure, between studies |
| C7 | More rigorous studies show larger effects | weak | two moderators (randomization, attrition), not adjusted for other features |
Concepts
- Interethnic contact intervention: a programme run under natural conditions to improve intergroup relations by establishing direct or indirect contact, which ensures the contact occurs.
- Extended contact: knowing, or here being shown, that a member of one’s own group has a close relationship with a member of the outgroup (Wright et al. 1997).
- Virtual contact: the authors’ term for interaction between members of the groups through a computer system, without meeting.
- Comparison: an intervention–control contrast for one sample, the unit of analysis.
Connections
It takes Pettigrew and Tropp’s 2006 meta-analysis (LIT-1413) from mostly correlational contact to field programmes and borrows its categories of generalization. It cites the 2008 mediator meta-analysis (LIT-1405) for anxiety and empathy and says the intervention literature is too thin to test mediators. Paluck, Green and Green (LIT-1420, NOTE-1393) draw studies from it and are stricter: randomized designs only, outcomes at least a day later. Its extended-contact arm rests on Wright et al. 1997, the triage’s fallback if this paper had not separated the arms. Scacco and Warren (LIT-1831) is the kind of behavioural study it found none of: cooperative learning with no peace content, measured by games.
Bearing on the record
- CLAIM-430, second defeater. The triage expected this paper to
supply the nearest test and to correct the claim’s note that the
defeater “has no test in the randomized contact literature”. It
supplies something nearer than the note allowed, and weaker than a
test. Nearer: it does estimate programmes with no taking part, the
eight extended-contact comparisons, in which participants read or
watch cross-group friendships, beside programmes in which they take
part, and the first are no less effective (0.42 against 0.31 for
meetings and 0.25 for cooperative learning). If anything, that weighs
toward the defeater. Weaker, on each of the defeater’s terms:
- equal favourable content: nothing matches content across arms; the comparison is between different studies of different programmes, populations and measures, and the subtype moderator is not significant;
- as large: immediate posttests only, self-reported attitudes only;
- as durable: the delayed analyses are not broken down by type, so nothing is known about extended contact’s persistence;
- taking part: “indirect” is not “without taking part”. The virtual subtype is interaction by computer, which is taking part at a distance, and it is the subtype that did not work in the stronger designs. Pooling it with extended contact as “indirect” hides the contrast the claim cares about. The note’s “no test” should become: no within-study test; between studies, observed friendships do about as well as taking part, on immediate self-reports.
- QUESTION-262. On the measurement demand, it is no help: “Due to a lack of appropriate research”, no behavioural outcome could be included, and the authors’ sixth direction for research is behaviour. On “local against general”, its outcomes reach the whole target outgroup and unspecified ethnic outgroups, but no comparison measured a specific outgroup not involved, so the question’s generalization to other members of the category is supported and generalization beyond it is untested, as NOTE-1393 found for its sample.
- CLAIM-275. Nothing: no included study orders conduct and interpretation.
No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Between-study comparisons. Every moderator result compares different studies; type of contact is confounded with age, setting and design (the extended-contact studies the text names, Cameron et al. 2006 and 2007 and Derbaum 1981, are with children). The authors run one moderator at a time and say why (footnote 5).
- Self-report only, and mostly short-term.
- Most comparisons are not randomized (60 of 79 in the primary cluster), and the authors’ reading of the randomization moderator as “higher quality, larger impact” is one of several.
- Discrepancies in the manuscript. These may be corrected in the
typeset article, which was not read.
- Table 1 gives indirect programmes in the primary cluster as 16 = 22.3%; 16 of 79 is 20.3%.
- Table 1’s secondary cluster lists 6 indirect comparisons (12.0%) but 5 virtual comparisons (also 12.0%), the only indirect subtype there.
- Table 2 marks the delayed primary-cluster Q = 9.11 with ***, while the text reports it as p = .25, homogeneous; and gives the secondary immediate Q as 150.60 against the text’s 150.85.
- Table 3’s secondary-cluster conflict estimates (0.47 [0.35, 0.60]; 0.34 [0.22, 0.46]) differ from the text’s (0.48 [0.36, 0.59]; 0.32 [0.20, 0.43]), and its primary “target outgroup” k = 38 from the text’s 40.
- The reference list gives Cernat (2011) and Ertmer et al. (2011) the same journal, volume and pages, Internet and Higher Education 14, 251–261; one of them is wrong. Connolly (1992) and Cook (2000) are listed twice; “Miles & Crips, 2014” is Crisp; “outgoups” in the abstract.
Open questions
- Does extended contact’s advantage survive a within-study comparison with cooperative contact on matched content, on behaviour, after a delay? That is the test CLAIM-430’s defeater asks for.
- Is the virtual null about the medium or about the studies (k = 8)?
Note of 2026-10-29: which of the two references is wrong
Limitations says the reference list gives Cernat (2011) and Ertmer et al. (2011) the same journal, volume and pages, Internet and Higher Education 14, 251–261, and that one of them is wrong. It is Cernat’s. Checked against Crossref and PubMed (LIT-1829’s history), Cernat’s article is in The Journal of Social Psychology 151(6), 737–753, doi:10.1080/00224545.2010.522622, and Crossref gives Internet and Higher Education 14, 251–261 to Ertmer, Newby, Yu et al., doi:10.1016/j.iheduc.2011.05.005. So Cernat’s entry carries Ertmer et al.’s venue. The typeset article’s list was not checked, so whether the error survives there is not known.
Open questions names a within-study comparison of extended contact with cooperative contact as the test CLAIM-430’s defeater asks for. Cernat, the candidate this reading named, is registered as LIT-1829, Unfetched. It is unread, and from its abstract it is not that test: all three arms read a story and none takes part; it compares a Roma exemplar’s friendship with a member of the reader’s own group against one with a member of another outgroup. Whether it is among the eight extended-contact comparisons was not checked, since the Supporting Information, which lists the comparisons, was not read.