We appreciate Crede et al.’s (1) attention to our research (2). We focus on the authors’ summary point: “Suffice it to say that, generally, suppression effects are considered statistical artifacts unless there is a strong theoretical explanation for their occurrence.” Crede et al. (1) suggest that (i) our interpretation lacks theoretical justification, (ii) suppression effects make results inherently uninterpretable, and/or (iii) our results are a statistical artifact.

First, as reviewed at length in our original article (2), our hypotheses are predicated on decades of theoretical and empirical support across the social sciences demonstrating that exposure to inequality results in negative emotions and antisocial behavior among high- and low-status individuals (recently, ref. 3, and see ref. 4). In this sense, our results may be seen as an application of well-grounded theory to a new setting.

Second, far from being settled, the interpretability of suppression effects remains a topic of considerable debate. Comprehensive examinations of suppression (5⇓⇓⇓⇓–10) treat suppression effects as potentially meaningful and interpretable—and even as a “fact of life” in applied social sciences (9, p. 1)—rather than discarding them by default. Consider one of many reasonable instances of interpretable suppression effects, the iconic example of the relationship between workers’ intelligence and performance on widget assembly (10) (Fig. 1A): The positive association between intelligence and performance is suppressed by intelligence’s association with boredom (which relates to lower performance). Without taking the suppression effect into account, researchers might draw erroneous conclusions regarding the relationship between intelligence and performance.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 1.

Theoretical relationships between variables demonstrating interpretable suppression in an example (A) and the present study (B).

Finally, although we agree that researchers’ indiscriminate use of covariates can result in artifactual relationships, we included only covariates that we expected to (and that common sense suggests would) predict air rage and that we thought might also relate to our two key predictors, presence of first class and boarding location: characteristics of planes (e.g., number of seats, size) and of flights (e.g., length, delays). The inclusion of our covariates was highly discriminate: They are essential controls for alternative explanations. (Note that we included correlation matrices to show the zero-order relationships among our variables in our article to be transparent about the nature of our data.) We further note that analyses that omit all but minimal covariates (i.e., number of seats and flight distance) offer robust support for our account (11), again suggesting that our results are not reliant on indiscriminate inclusion of covariates. To the issue of suppression effects in our analyses specifically, larger planes are likely to have more passengers and therefore more incidents of air rage, and are more likely to board from the middle (versus the front) of the plane. As a result, the actual negative association between middle boarding and air rage is suppressed by middle boarding’s association with plane size, given that plane size is positively associated with air rage (Fig. 1B). The suppression effect, in this case, is easily interpretable, is consistent with theory, and, as with the intelligence and widget example, researchers might draw erroneous conclusions regarding the relationship between boarding location and air rage without taking it into account.