1 Introduction
More and more psychological and neuroscientific research on moral judgments appears each year. As fascinating results accumulate, the question arises of whether and how individual studies fit together to form a larger picture. In order to connect various studies and guide future work in this field, researchers need to determine which studies conflict, which support each other, and which are simply talking past each other.
Unfortunately, precise comparisons are hampered by the use of different moral terms across studies. Different researchers ask whether acts are wrong (e.g., Cushman, Reference Cushman2008; Haidt et al., Reference Haidt, Koller and Dias1993; Schaich Borg et al., Reference Schaich, Hynes, van, Grafton and Sinnott-Armstrong2006; Wheatley & Haidt, Reference Wheatley and Haidt2005), forbidden versus permitted (e.g., Cushman, Reference Cushman2008; Cushman et al., Reference Cushman, Young and Hauser2006), (in)appropriate (e.g., Greene et al., Reference Greene, Nystrom, Engell, Darley and Cohen2004; Moore et al., Reference Moore, Clark and Kane2008; Valdesolo & DeSteno, Reference Valdesolo and DeSteno2006) or deserve blame (e.g., Cushman, Reference Cushman2008; Pizzaro et al., Reference Pizzaro, Uhlmann and Salovey2003). Some researchers include the adverb “morally” before these terms (e.g., Moore et al., Reference Moore, Clark and Kane2008; Schaich Borg et al., Reference Paolacci, Chandler and Ipeirotis2006; Wheatley & Haidt, Reference Wheatley and Haidt2005), whereas others do not (e.g., Cushman, Reference Cushman2008; Cushman et al., Reference Cushman, Young and Hauser2006; Greene et al., Reference Greene, Nystrom, Engell, Darley and Cohen2004; Pizzaro et al., Reference Pizzaro, Uhlmann and Salovey2003; Valdesolo & DeSteno, Reference Valdesolo and DeSteno2006). It is unclear whether judgments of what is morally wrong vary in response to the same factors as do judgments of what is forbidden, inappropriate, or blameworthy. Some evidence comes from a meta-analysis on the asymmetry between “forbid” versus “allow” in attitudes research. This study demonstrated that people are reluctant to forbid but will readily not allow, even though these judgments are conceptually equivalent (Holleman, Reference Holleman1999). This asymmetry suggests that moral judgments, as well, may be influenced by subtle variations in wording. Further evidence from Cushman (Reference Cushman2008, Experiment 1) showed that harmless acts were judged as more wrong than blameworthy onlywhen the act was intended and believed to cause harm. This finding also suggests that people, in some circumstances, will draw fine distinctions between moral terms.
Until the effects of wording variations are understood, we cannot tell whether studies on similar moral issues couched in different terms really agree or disagree. In addition, some wording effects on moral judgments would undermine the search for a moral faculty. Some researchers suggest that moral judgments result from innate psychological mechanisms, or even a moral module that conforms to a universal moral grammar (Dwyer, Reference Dwyer1999; Harman, Reference Harman and Brinkmann1999; Hauser et al., Reference Hauser, Young, Cushman and Sinnott-Armstrong2008; Mikhail, Reference Hauser, Cushman, Young, Jin and Mikhail2007). Others propose dual-process models that build emotions or beliefs, desires, and consequences into the processes that form moral judgments (Cushman et al., Reference Cushman, Young, Greene and Doris2010; Greene et al., Reference Greene, Nystrom, Engell, Darley and Cohen2004). These theories and many more would be challenged if people judge acts in very different ways based on the moral terms used, because psychologically real mechanisms would be unlikely to vary markedly with such fine differences in wording within a particular natural language. Conversely, if certain patterns of moral judgments are robust enough to persist through non-substantial variations in wording, it would help defend the assumption that these studies are investigating distinctive psychological mechanisms.
In order to bring this research together into a coherent field and determine whether there are distinctive psychological mechanisms to be studied in moral psychology, we need to know whether, when, and how much phrasing questions in different terms may lead to different moral judgments. As an exploratory first step toward answering these questions, we tested the effects of four different moral adjectives across six different types of moral judgments.
2 Methods
2.1 Participants
Adult participants who had an internet protocol address within the United States were recruited through Amazon Mechanical Turk. This online participant recruitment system has been shown to produce quality data (Hsueh et al., Reference Hseuh, Melville and Sindhwani2009; Paolacci et al., Reference Paolacci, Chandler and Ipeirotis2010; see also Kittur et al., Reference Kittur, Chi and Suh2008, for a description of this system). A total of 845 participants received $3 for completing the study. Ninety-seven participants were removed for insufficient responding (withdrawal before completing the second block), seven for suspicious responding (predominantly entering the first available response across measures), and one for being younger than 18 years, leaving a final sample of 740 participants (716 participants provided complete data; see Table 1 for demographics). The sample was majority female (60.3%), White (70.1%), and ranged in age from 18 to 85 years (M= 33.5 years, SD = 11.38 years). To account for cohort effects, we controlled for age in all analyses.
2.2 Moral vignettes
Participants read 15 vignettes, each displayed on a new screen, which presented a hypothetical person’s morally ambiguous behavior. After each vignette, participants responded to a statement expressing disapproval of the behavior (e.g., “turning the train was wrong”) using a 9-point Likert-type scale (1 = strongly disagree; 5 = neither agree nor disagree; 9 = strongly agree). Low values, therefore, indicated acceptability and high values indicated unacceptability. The 15 vignettes were divided into six blocks of moral judgmentsFootnote 1:
Trolley — three vignettes in which the actor kills one person in order to save five others by either flipping a switch to divert a train (sidetrack, loop) or pushing a man in front of the train (footbridge) (Hauser et al., 2007).Footnote 2
Victimless — three vignettes describing taboo behaviors: brother-sister incest (Haidt, 2001)Footnote 3, cannibalism, and interspecies sex.
Harm versus offense — two vignettes comparing a private transgression, stealing money from a lost wallet (Greene et al., 2001)Footnote 4, with a public taboo, sexual intercourse.
Deceit — two vignettes comparing deception through lying versus omission.
Moral luck — three vignettes in which a drunk driver ignores a stoplight and either kills a pedestrian, misses a pedestrian, or there is no pedestrian present.
Disgust — two vignettes that compare sloppily eating unconventional foods privately versus publically (Feinberg, 1985).
2.3 Design and procedure
The experiment was a between-subjects 2 (Order) x 2 (Adverb) x 4 (Adjective) randomized full factorial. Participants were randomly assigned to respond to the 15 moral vignettes with 1 of 4 adjectives: “wrong”, “inappropriate”, “forbidden”, or “blameworthy”. Whether the adjective was preceded by the adverb “morally” was also randomly assigned. The judgment made by a given participant (e.g., “morally wrong”) remained constant across vignettes. Blocks were assumed to be independent and, thus, were presented in the same order across participants. Presentation order within each block, however, was randomly assigned to 1 of 2 conditions (Table 2).
A short description of the survey, including compensation, was posted online. Participants voluntarily clicked a hyperlink that directed them to our website, which served a multi-page Ruby on Rails application. After participants indicated they were at least 18 years old and provided informed consent, they made 15 moral judgments and provided demographic information.Footnote 5
2.4 Analysis and power
To test for general effects of wording variations on moral judgments, a repeated measures analysis of covariance (ANCOVA) controlling for age was performed, with between-subjects factors of Order (2 levels), Adverb (2 levels), and Adjective (4 levels); Vignette (15 levels) was the within-subjects factor. This analysis, however, ignored the distinction between different types of moral judgments. To determine, therefore, whether effects were limited to specific types of morality, repeated measures ANCOVAs were performed separately for each block. The within-subjects factor in these six ANCOVAs contained either 2 or 3 levels, depending on the number of vignettes in that block.
Because sufficient power is required to claim meaningful null effects (i.e., wording makes no difference), we conducted a sensitivity analysis using G*Power 3.1 (Faul et al., 2007) to determine how small of an effect we could detect in each ANCOVA. For a repeated measures ANCOVA with a within-between factor interaction, 16 between-subjects groups, 15 repeated measures (Cronbach’s α = .81), a Type II error probability of α = .05, and power equal to .80, we could find an effect size > .032 (i.e., a small effect; Cohen, 1992). We also performed this analysis for each block separately and found we could detect an effect > .06 for blocks with three judgments, and > .08 for blocks with two judgments, both small effects. After collecting the data, we confirmed that we achieved sufficient power to find these effects using a post-hocpower analysis for each block. Using a conservative estimated effect size of .08 and a Type II error probability of α = .05, power for 5 of 6 tests was > .94, the exception being the harm versus offense block, which had power of .81. These tests provided evidence that our analyses were sufficient for detecting a small effect in the data.
3 Results
3.1 Overall analysis
Main effects. The repeated measures ANCOVAFootnote 6 for all 15 moral judgments revealed significant between-subjects main effects for Order, F(1,622) = 5.18, p= .023, generalized eta-squared (η G2; Bakeman, 2005; Olejnik & Algina, 2003) = .040, Adverb, F(1,622) = 4.75, p= .030, η G2= .008, and Adjective, F(3,622) = 4.04, p= .007, η G2= .019, and a significant within-subjects effect of Vignette, F(14,8708) = 30.53, p< .001, η G2= .047 (see Table 3 for all means and standard deviations). Because Order is meaningless across blocks, it is explored below in further detail. For Adverb, participants were more accepting when “morally” was present (M= 5.57, SE= .06) versus absent (M= 5.76, SE= .06). For Adjective, participants judged acts as more wrong (M= 5.81, SE= .08) or inappropriate (M= 5.80, SE= .08) than either forbidden (M= 5.59, SE= .08) or blameworthy (M= 5.45, SE= .09). Figure 1 displays means and standard errors for the eight Adverb x Adjective conditions.
Note. SDs in parentheses. Scale: 1 = acceptable; 9 = unacceptable. W = “wrong”, I = “inappropriate”, F = “forbidden”, B = “blameworthy”.
Interactions There were no significant between-subjects interactions. Because the assumption of sphericity was violated, the Greenhouse-Geisser correction (1959) was applied to all within-subjects interactions. The Vignette x Order interaction, F(8,8708) = 13.83, p< .001, η G2= .012, and the Vignette x Adverb interaction, F(8,8708) = 3.06, p= .002, η G2= .003, were significant. To understand these interactions, they are described below for each block in which they achieved significance.
Mean moral judgments across 15 vignettes by Adverb and Adjective.
3.2 Analysis by block
Main effects. Each block showed a significant within-subjects Vignette effect. The Order main effect was found for the Trolley block, F(1,646) = 55.23, p< .001, η G2= .079: these behaviors were judged more acceptable when side track was presented first (M= 3.73, SE= .11) versus last (M= 4.90, SE= .11); and the Disgust block, F(1,643) = 10.31, p= .001, η G2= .016: unconventional eating was judged more acceptable when first described publically (M= 3.46, SE= .12) versus privately (M= 4.01, SE= .12). The Adverb effect only emerged for the Disgust block, F(1,643) = 17.68, p< .001, η G2= .027: these acts were more accepted when “morally” was included (M= 3.37, SE= .12) versus excluded (M= 4.09, SE = .12). The Adjective effect was found for the Victimless block, F(3,639) = 3.46, p= .016, η G2= .016, and the Disgust block, F(3,643) = 4.68, p= .003, η G2= .021. Participants judged victimless offenses as less blameworthy (M= 5.60, SE= .17) than either wrong (M= 6.17, SE= .17), inappropriate (M= 6.31, SE= .16), or forbidden (M= 6.17, SE= .16). Disgust acts were judged as more wrong (M= 4.00, SE= .17) or inappropriate (M= 4.10, SE= .17) than either forbidden (M= 3.41, SE= .16) or blameworthy (M= 3.42, SE= .18).
Interactions. Again, a Greenhouse-Geisser correction was used on all tests. The Vignette x Order interaction emerged only in the Trolley block, but was qualified by a significant Vignette x Order x Adverb x Adjective interaction that did not appear in the overall test, F(5,1292) = 2.78, p= .017, η G2= .011. When footbridge was presented first, footbridge was rated as more unacceptable than either sidetrack or loop across moral terms. When sidetrack was presented first, however, “blameworthy” (but not “morally blameworthy”) showed no significant differences between vignettes, F(2,70) = 2.36, p= .102. Finally, the Vignette x Adverb interaction emerged for the Disgust block, F(1,643) = 12.31, p< .001, η G2= .019: there was a smaller difference between judgments of private disgust when “morally” was included (M= 2.31, SE= .15) versus excluded (M= 2.66, SE= .14) than there was for public disgust (Ms = 4.44, 5.52, SEs = .14, .13, respectively).
4 Discussion
4.1 Implications
This study suggests that wording effects do not undermine psychological studies of moral judgments. For harm versus offense, deceit, and moral luck, we found no evidence of wording effects, indicating that these types of morality are robust against linguistic variations. We did find wording effects, though, for victimless offenses, disgust, and the trolley scenario. For the Victimless and Disgust blocks, we discovered scaling effects: participants judged victimless offenses as less blameworthy than wrong, inappropriate, or forbidden, and disgust as less blameworthy and less forbidden than wrong or inappropriate. In addition, participants were more likely to condone disgust when “morally” was included in their judgment. It is noteworthy, however, that the wording effects we found in both blocks did not cross the scale midpoint. Inasmuch as one can assume that responses above versus below the midpoint indicate global judgments of moral unacceptability versus acceptability, wording did not appear to change global judgments, only the strength or certainty of those judgments. In other words, people did not judge incest, for example, as “wrong” but not “blameworthy”; the tendency was only to judge it as morewrong than blameworthy.
The picture was more complicated for the trolley vignettes. When footbridge was presented first, we replicated the finding that pushing the man in footbridge is judged as less acceptable than flipping the switch in sidetrack or loop (Cushman et al., 2006), regardless of moral term used. When sidetrack was presented first, however, blameworthiness was rated similarly across all three vignettes, suggesting that wording effects do occur in some orders of trolley scenarios. Similar to the results of Cushman (2008), this finding suggests that judgments of blame are determined uniquely from global judgments of wrongness. The current finding that “morally blameworthy” demonstrated the expected order effect further supports this idea, suggesting that the inclusion of the adverb altered this judgment to be more abstract. We must acknowledge, however, that these differences showed very small effect sizes (η G2 < .05), indicating that the influence of wording variations on moral judgments was negligible.
4.2 Limitations
Online data collection limited our control over the participant population and the testing environment. We eliminated approximately one-eighth of the original sample for incomplete or inappropriate responding, but this rate was comparable to previous studies on the utility of Amazon Mechanical Turk (Kittur et al., Reference Kittur, Chi and Suh2008). Additionally, reaction times suggested that as many as 100 more participants may have rushed through the experiment, but analyses without these participants did not differ from the results presented. This rate of “gaming” the system was substantially lower, however, than in previously published reports (Kittur et al., Reference Kittur, Chi and Suh2008). In addition, we tested moral wording across an array of moral judgments in a within-subjects design: Having participants judge such varied situations without counterbalancing may have biased our results toward null effects. Future studies may be better served by focusing on a single type of judgment (e.g., the trolley problems) and examining a more comprehensive set of moral terms.
4.3 Conclusion and future directions
Our results indicate that participants in moral psychology studies are interpreting different moral terms in a similar manner, suggesting that researchers are studying a real psychological phenomenon, not a linguistic artifact. Our findings are also compatible with the possibility of a universal moral faculty or grammar. Although we did not examine every moral term used in previous studies or in our natural language, we believe the use of eight common terms makes our results sufficiently generalizable. It seems unlikely that people would process “wrong”, “inappropriate”, “forbidden”, and “blameworthy” in similar ways yet provide radically different responses to another related term.
Researchers should, however, still be cautious regarding the terms used in their studies. We found evidence that people are less apt to forbid (Holleman, Reference Holleman1999) or to lay blame (Cushman, Reference Cushman2008), so to the extent that the magnitude of judgments is relevant to one’s research question, similar terms should be used across studies. In addition, we also found evidence that judgments of purity may be more susceptible to these wording effects than harmful acts. These effects are likely to be found in any situation, like the disgust scenarios, in which the behaviors are governed by cultural norms but not formal rules or laws. It may be prudent, therefore, that studies of moral purity only be compared when the adverb “morally” is included in participants’ judgments. Finally, the trolley vignettes appeared to be interpreted differently when participants were asked to judge blameworthiness, but not moral blameworthiness: future research should take care not to treat wrongness and blameworthiness as interchangeable concepts, but acknowledge that they are likely derived from different processes (Cushman, Reference Cushman2008).
As the field expands, a meta-analysis on moral judgment research that examines wording as an independent variable will be necessary. From our findings, we do not expect large wording effects to emerge, but such research is yet to be completed. In addition, neither the types of morality studied nor the moral terms used were exhaustive, limitations that require follow-up studies. Related to this point, future research should also test whether wording may shift people between utilitarian and deontological perspectives. Some of the terms used in the current study were more relevant to a deontological framework (e.g., “forbidden”) than a utilitarian one, and follow-up work could compare these moral terms to judgments of whether behaviors should be done or would be best. These steps will facilitate comparison of studies in the field of moral psychology and help build a coherent picture of how people understand morality.
Open access
