Contribution
It re-asks Pettigrew and Tropp’s question (LIT-1413) of the only studies a policy-maker can act on: those that randomly assign face-to-face contact and measure prejudice at least a day after it begins. There are 27, of which 8 were in the 2006 meta-analysis. Pooled, they say contact reduces prejudice. Looked at closely, the largest and best-designed studies say much less, and the population the hypothesis was built for, adults across racial lines, is not studied at all.
Key insight
A pooled average can be large while the studies that would settle the question are small in number and small in effect. Here the estimate shrinks toward zero as precision grows, and the three pre-registered field trials average 0.016. So the meta-analysis “offers qualified support”: it does not show that contact fails, and it does not show that contact works for the prejudice and population that motivated the hypothesis.
Assumptions
- Inclusion. Pettigrew and Tropp’s definition of contact, “actual face-to-face interaction between members of clearly defined groups”; random assignment; an outcome measured at least one day after the intervention began (a weaker filter than one day after it ended, which would drop ten studies); published or circulated by July 2016.
- One estimate per treatment arm, chosen by eight rules: the highest dose, the dominant group, the latest post-test, a linear estimator, the authors’ primary outcome, and the specification with the smallest apparent standard error.
- Effect size. Glass’s Δ, the treatment–control difference over the control group’s standard deviation, with Hedges’ small-sample correction. Three studies’ arms (four arms) lack variance information and enter only the sign test, leaving 25 studies and 27 arms.
- Random effects, the variance estimated by method of moments, since Cochran’s Q rejects homogeneity.
Key results
- Coverage. Thirteen studies use college students, six adults over 25, seven children or adolescents; 11 address racial or ethnic prejudice, all with participants from elementary school through college. Nineteen took place in the USA. Sixteen date from 2000 or later. Twenty-six of 27 had authority approval; 17 are classed as equal status, 14 as cooperative, 12 as having a common goal.
- Direction. 24 of the 27 estimates are positive (26 of 31 counting the arms without standard errors).
- Pooled. Random-effects Δ = 0.39, 95% CI 0.231–0.554; Q(26) = 173.178, I² = 0.85. Only 20 of 27 intervals cover 0.39, and four of the ten most precise fall below it.
- Precision (Table 3). Regressing Δ on its standard error: slope 2.086 (SE 1.007, p = 0.049), intercept −0.014 (CI −0.458 to 0.430), the expected effect of an infinitely large study.
- Target (F-test p = 0.01). Disability differs most from the base (p = 0.001). Ethnic, racial, religious and immigrant targets: 0.25, CI 0.08–0.42, with four of the five largest studies below it. Without the disability studies: 0.20.
- Robustness. Adding seven borderline quasi-experiments: 0.373 (SE 0.075), target heterogeneity p = 0.0024, precision slope p = 0.058. The three pre-registered studies (Broockman and Kalla 2016; Scacco and Warren 2018; Finseraas and Kotsadam 2017): 0.016, against 0.451 for the rest.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Most randomized contact studies with delayed outcomes find prejudice reduced | strong | 24 of 27 positive |
| C2 | The average effect is about 0.39 SD | weak | heterogeneity, the precision gradient and the pre-registered subset all undercut it |
| C3 | Effects vary by target, largest for disability, smaller for ethnic and racial prejudice | moderate | meta-regression, p = 0.01; contradicts Pettigrew and Tropp’s near-equality |
| C4 | Larger studies find smaller effects, consistent with a file drawer | weak | p = 0.049, and 0.058 with the borderline studies; “statistically equivocal” in the authors’ words |
| C5 | No randomized study with delayed outcomes tests interracial or interethnic contact on adults over 25 | strong | the inventory, Table 1 |
| C6 | Allport’s conditions have never been varied experimentally in such a study, so whether they are needed cannot be decided | strong | the inventory |
Concepts
- Policy-relevant study: randomized assignment of contact and an outcome measured at least a day after the intervention began.
- Glass’s Δ: the effect over the control group’s standard deviation, used because pre-treatment deviations are not always reported.
- Allport’s conditions: equal status, common goals, intergroup cooperation and the support of authorities, law or custom. Pettigrew and Tropp found them “not essential”; Allport warned that without them “the more contact the more trouble”.
Connections
An update of Pettigrew and Tropp’s 2006 meta-analysis (LIT-1413), with which it agrees that contact “typically reduces prejudice” and disagrees on two things: the near-equality of effects across targets, and their claim that “the most rigorous studies tend to provide the largest effects”. It draws studies from Lemmer and Wagner (2015) and from Paluck and Green’s 2009 review of prejudice reduction. Its sample includes Broockman and Kalla’s 2016 canvass, whose sequel is Kalla and Broockman (LIT-1407). Enos (2014), excluded as proximity without interaction, found prejudice increased.
Bearing on the record
- QUESTION-262. The question says “contact reduces prejudice and often generalizes (LIT-1413)”. On the randomized evidence with delayed outcomes the first half is qualified: positive in direction, 0.25 for ethnic and related targets, near zero in the pre-registered field trials. The second half is untested here: the review’s outcomes mostly concern the group met, and its pooling does not separate generalization to the group or to other groups. Nor does any included study isolate taking part from receiving information, the question’s first contrast.
- CLAIM-430. Its defeater, information of equal favourable content changing attitudes as much as taking part, has no test in this literature. Contact is compared with no contact.
- Pettigrew 1998 (LIT-1408). The review does not cite the 1998 review and does not use the word “deprovincialization”, so it does not bear on LIT-1408’s flag.
No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Small numbers. Twenty-seven arms, eleven on racial or ethnic prejudice, three pre-registered; every subgroup estimate rests on a handful of studies, and the 0.016 has no interval given.
- The selection rules are the authors’. Taking the specification with the smallest apparent standard error and the highest dose favours larger, more significant estimates; the direction of that bias runs against the paper’s sceptical findings, not with them.
- Within the interventions, little is known. Most reports describe the contact too thinly to say which of Allport’s conditions held.
- Errors and slips. None of substance found. “Hedge’s G correction” is Hedges’ g. Table 1’s cell layout did not survive text extraction; its margins (3, 4, 13, 1, 6 participants; 11, 2, 3, 3, 1, 4, 2, 1 targets) sum to 27 and agree with the prose. The text’s “four of the five largest studies come in below this average” is not checkable without the figure.
Open questions
- Does randomized contact reduce racial or ethnic prejudice in adults, measured weeks later?
- Do Allport’s conditions moderate the effect when they are manipulated rather than coded after the fact?
- Is the precision gradient a file drawer, or the difference between small intensive interventions and large naturalistic ones?