Contribution
The first meta-analysis of the whole contact literature of the 20th century under explicit inclusion rules: contact must be actual face-to-face interaction between members of discrete groups, observed, reported or unavoidable, and prejudice must be measured on individuals. It establishes that contact and prejudice are negatively related across targets, settings, ages and countries, that the relation is not larger where participants chose their contact, and that it is larger, not smaller, in the more rigorous studies. It turns Allport’s conditions from necessary into facilitating.
Key insight
Contact reduces prejudice under most conditions, not only Allport’s. The authors’ explanation is that familiarity breeds liking (mere exposure) and that contact lowers intergroup anxiety and uncertainty, so the open theoretical question is not what makes contact work but what stops it working.
Assumptions
- Random-effects model on Fisher’s z, back-transformed to r; negative r means more contact, less prejudice. Effects reported only as “nonsignificant” set to 0 (17 samples, 2.4%), then dropped from moderator analyses. Very large studies capped (5,000 per study, 3,000 per sample, 2,000 per test). The analysed file is 696 samples.
- Allport’s conditions cannot be rated for most studies, which do not describe the contact. The global test is whether the contact was a structured program designed to meet them; the four conditions are rated singly only within the 134 structured samples.
- The causal question is handled by rating participants’ choice to engage in the contact, by the study type, and by the reciprocal-path literature cited, not by any new causal design.
- Most of the evidence is cross-sectional: surveys and field studies are 71% of studies, quasi-experiments 24%, true experiments 5%. The United States gives 71% of studies.
Key results
- Overall (Table 1). Random effects: r from −.205 to −.214 across studies, samples and tests; −.215 for the 696 corrected samples. 94% of samples are negative. Heterogeneity is large, Q(695) = 4,990, and remains after the fifth most extreme samples are removed.
- Choice (Table 2). No choice −.280 (k = 116), some choice −.190 (279), full choice −.218 (301). Selection by the prejudiced into avoiding contact would predict the reverse.
- Publication bias. Fail-safe N above 1,200 samples; sample size uncorrelated with effect size (r = .04); unpublished samples −.237 and published −.211 (QB = 2.17, p = .14); trim-and-fill estimates 72 missing samples and an adjusted mean of z = −.245. Vevea and Hedges’s model finds the file missing many small studies with small effects and, adjusted, no significant overall relationship (Z = −.02); for the 118 between-groups samples with strong controls the adjusted effect is still significant (z = −.109, one-tailed p < .02).
- Generalization (Table 2, tests). Within the situation −.231 (152 tests); whole outgroup −.213 (1,164); across situations −.244 (17 tests from 9 samples); to outgroups not in the contact −.190 (18).
- Rigor (Tables 3, 4). Experiments −.336 (k = 36), quasi-experiments −.237 (168), surveys −.204 (492). Controls without prior contact −.244, with extensive prior contact −.138. Observed contact −.246, assumed −.132. Reliable contact measures or manipulations −.296 against −.189. The 77 samples best on all three method ratings give −.323 against −.202.
- Allport’s conditions (Tables 5–7). Structured programs −.287 (k = 134) against −.204 (562); in a weighted regression with the method ratings the program term stays significant (β = −.099, p = .027; model R² = .10). Within programs, common goals, cooperation and equal status do not separately predict effects, are intercorrelated (.51 to .63), and 72% of programs have three of the four. Programs with only authority sanction (k = 31) give −.286, those with two or more conditions −.290.
- Racial and ethnic targets (Tables 8–10). The two subsets have the same mean (−.218 and −.220). But the rigor gradient by study type is carried by the other targets: experiments give −.221 (k = 9) for racial and ethnic samples and −.377 (k = 27) for the rest. Structured programs raise the effect significantly for the other targets (QB = 19.67) and not for racial and ethnic ones (QB = 2.62, p = .11); in regression the program term is marginal in both subsets (p = .069, .094).
- Moderators (Tables 11, 12). Largest for heterosexuals’ contact with gay men and lesbians (−.271); smaller for the mentally ill and the elderly (−.183); tourism smallest (−.113, 13 samples); recreational and laboratory settings largest (−.287). Samples after 1979 −.236 against −.184 before 1980, mostly but not wholly explained by rigor. No difference between U.S. and other samples.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Intergroup contact is reliably associated with lower prejudice across targets, settings and countries | strong | 696 samples, 94% negative, Table 1 |
| C2 | The association is not an artifact of participants’ choosing their contact | moderate | Table 2 choice ratings; experiments −.336 |
| C3 | It is not an artifact of publication bias | moderate | five of six checks; Vevea–Hedges dissents |
| C4 | The effect generalizes to the whole outgroup and beyond | moderate | Table 2, but 17 and 18 tests for the two widest kinds |
| C5 | Allport’s conditions enhance the effect but are not necessary | moderate | Table 5; β = −.099 with method controls |
| C6 | The conditions work as a bundle, not as separate factors | weak | non-significant single conditions within 134 samples; ratings intercorrelated |
| C7 | Mere exposure and anxiety reduction explain why contact works without the conditions | weak | argued from other literatures, not tested here |
Concepts
- Intergroup contact: “actual face-to-face interaction between members of clearly defined groups”; proximity and group proportions do not count, nor does information about an outgroup.
- Structured program: a contact situation the researchers designed to approximate Allport’s optimal conditions; the paper’s global test of them.
- Facilitating, not essential: Pettigrew’s 1998 distinction (LIT-1408), here given meta-analytic form.
Connections
It tests the reformulation in Pettigrew’s 1998 review (LIT-1408): the causal sequence, the facilitating-not-essential distinction and the three kinds of generalization are that review’s problems, and friendship as a contact measure (−.246 over 154 tests) its “friendship potential”. Its file of 515 studies is the base the authors’ 2008 meta-analysis of mediators extends (LIT-1405). Paluck, Green and Green (LIT-1420) re-run the question on the 27 randomized studies with delayed outcomes; their finding that ethnic and racial targets show weaker effects matches this paper’s own Table 8, where the experimental advantage disappears for racial and ethnic samples.
Bearing on the record
- CLAIM-275 (an intervention can come before the interpretation it makes available). The meta-analysis is the strongest support the record holds for contact changing attitudes, but most of its samples are cross-sectional surveys of reported contact, which cannot order the two. The ordered evidence is the 36 experimental samples and the no-choice samples; both give larger effects. The authors’ own mechanism, mere exposure and anxiety reduction, is an encounter acting before any revised view, which fits the claim.
- CLAIM-297 and CLAIM-330: nothing here bears on performance- weighted authority or on the linearized model; the coupling the exchange drew is an analogy the paper neither makes nor tests.
- LIT-1413’s own “why this is here” said the exchange’s three figures agree with the abstract. They agree with the paper. What the exchange omitted, and the abstract softens, is the Vevea–Hedges result and the absence of a rigor gradient for racial and ethnic targets.
No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Correlational at base. Only 5% of studies are experiments; the causal reading rests on the choice ratings, the rigor gradient and other work’s reciprocal-path estimates.
- Publication bias is not ruled out. The authors call the Vevea–Hedges result a reason for “caution”, and the abstract’s “appears not to result from … publication biases” is stronger than their own text.
- Allport’s conditions are tested crudely: one global indicator, and single ratings only for the 134 samples where they are nearly always together.
- Generalization beyond the outgroup rests on 17 and 18 tests.
- Errors and slips. The trim-and-fill adjustment is reported as raising the effect to z = −.245 from a mean near −.21; filling in missing studies on the small-effect side should lower it, and the paper does not explain the direction. “Z” is used both for the test statistic and for Fisher’s z effect size in the same paragraphs. The authority-sanction comparison gives the remaining samples as k = 564, where Table 5 has 562 (134 + 562 = 696).
Open questions
- What do the randomized, delayed-outcome studies give for adults’ racial and ethnic prejudice? Paluck, Green and Green (LIT-1420) found none had been run.
- Which negative factors stop contact from reducing prejudice? The authors name the question and do not answer it.