A multilevel meta-analysis of effects of negative stereotypes of blacks in media on consumers’ attitudes

ABSTRACT

This meta-analysis of 49 studies, yielding 88 effect sizes (n = 10,215), examined the effect of negative stereotypes of Blacks in media on consumers’ attitudes. The results from the multilevel model (3-level) indicate that media stereotypes have a significant overall effect on consumers’ attitudes (r = .22, p < .001). This meta-analysis used multilevel multivariate models and meta-regression models to systematically investigate moderation of effect sizes by diversity in stimuli, dependent measures, research designs, sample demographics, and publication status. The results showed a consistent significant association between media stereotypes and attitudes across all moderator variables. Measures for attitudes in the subcategory of judgment showed larger effect sizes for the association between Black media stereotypes and consumers’ attitudes. Theoretical development, publication bias, and limitations are discussed. The entire database and codes for statistical procedures in R syntax is available from the Open Source Foundation (OSF) repository: https://osf.io/y8ndx/?view_only=dc6ecf5e18bf4fd883f0a396613022e0

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

1. The choice to include regression coefficient in meta-analysis is a controversial one. Aloe (2014, p. 62) warns very specifically that (a) “meta-analysts do not treat partial effect sizes as bivariate effects” and (b) “meta-analysts do not combine, in the same data set, bivariate correlations with partial effects.” While Aloe’s (2014) both concerns are valid, regarding the first one, Rosenthal and Rubin (2003) defend treating partial correlations as bivariate effects by arguing that r-equivalent “estimates of effect size are better than having no estimate at all,” (p. 493). Regarding the second one, since this meta-analysis did not have enough studies with just regression coefficients to do a separate meta-analysis of partial r that Aloe (2014) suggests, I chose to use Peterson and Brown’s (2005) formula and computed r equivalent as a conservative estimate for standardized regression coefficients. In cases where a study reported standardized regression coefficients but not zero-order correlation coefficients, r was computed using the formula r = .98β + .05λ, where λ = 1 when β > 0, λ = 0 when β < 0 (Peterson & Brown, 2005). When studies reported only unstandardized beta coefficients and standard errors, the effect size was estimated by calculating the t-value and substituting in the following equation r_equivalent = √ t²/√(t² + df) (Rosenthal & Rubin, 2003). These studies were compared with those that had zero-order correlations using sensitivity analysis. The average effect size for r_{_}zero-order = .23, k = 44, p < .001 was not significantly different (F (1, 100) = 0.11, p = .74) from r_equivalent = .20, k = 7, p < .01.