Deciphering moral intuition: How agents, deeds, and consequences influence moral judgment

Veljko Dubljević; Sebastian Sattler; Eric Racine

doi:10.1371/journal.pone.0204631

Abstract

Moral evaluations occur quickly following heuristic-like intuitive processes without effortful deliberation. There are several competing explanations for this. The ADC-model predicts that moral judgment consists in concurrent evaluations of three different intuitive components: the character of a person (Agent-component, A); their actions (Deed-component, D); and the consequences brought about in the situation (Consequences-component, C). Thereby, it explains the intuitive appeal of precepts from three dominant moral theories (virtue ethics, deontology, and consequentialism), and flexible yet stable nature of moral judgment. Insistence on single-component explanations has led to many centuries of debate as to which moral precepts and theories best describe (or should guide) moral evaluation. This study consists of two large-scale experiments and provides a first empirical investigation of predictions yielded by the ADC model. We use vignettes describing different moral situations in which all components of the model are varied simultaneously. Experiment 1 (within-subject design) shows that positive descriptions of the A-, D-, and C-components of moral intuition lead to more positive moral judgments in a situation with low-stakes. Also, interaction effects between the components were discovered. Experiment 2 further investigates these results in a between-subject design. We found that the effects of the A-, D-, and C-components vary in strength in a high-stakes situation. Moreover, sex, age, education, and social status had no effects. However, preferences for precepts in certain moral theories (PPIMT) partially moderated the effects of the A- and C-component. Future research on moral intuitions should consider the simultaneous three-component constitution of moral judgment.

Citation: Dubljević V, Sattler S, Racine E (2018) Deciphering moral intuition: How agents, deeds, and consequences influence moral judgment. PLoS ONE 13(10): e0204631. https://doi.org/10.1371/journal.pone.0204631

Editor: Rodrigo Ferrer, Universidad de Tarapaca, CHILE

Received: August 2, 2018; Accepted: September 11, 2018; Published: October 1, 2018

Copyright: © 2018 Dubljević et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: All relevant data are within the paper and its Supporting Information files.

Funding: Support for this work comes from the Banting Postdoctoral Fellowships Programme (VD, PI; 201310BAF-327654-235414; URL: http://banting.fellowships-bourses.gc.ca/en/home-accueil.html) and a grant from the Canadian Institutes of Health Research (ER, co-PI; EPP114801, URL: http://www.cihr-irsc.gc.ca). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Competing interests: The authors have declared that no competing interests exist.

Introduction

People perform quick moral evaluations regularly, whether driving to work or just reading the news. For example, when faced with a person they consider evil (e.g., a Nazi), they might blame them. Individuals witnessing an action they think is wrong (e.g., rape), may be more likely intervene and seek subsequent punishment of the perpetrator. Facing a situation in which many individuals suddenly die, people might request that similar tragedies be prevented. And persons that fail to exhibit appropriate moral responses are viewed as having questionable integrity as moral agents. There is now mounting evidence suggesting that such moral evaluations occur quickly following heuristic-like processes without the engagement of effortful and deliberative cognitive systems [1,2]. But what are the mechanisms triggering these responses? The most plausible answer to this question is ‘moral intuitions’. Jonathan Haidt, a pioneer of research into intuitive moral psychology, defines moral intuition as “the sudden appearance in consciousness of a moral judgment, including an affective valence (good-bad, like-dislike), without any conscious awareness of having gone through steps of searching, weighing evidence, or inferring a conclusion” ([3], p. 818). But how does this moral judgment appear? And why are some moral judgments considered ‘intuitive’, while others, which are deduced from normatively and cognitively demanding moral theories, are ‘counterintuitive’ [4]? For example, when asked if lying is wrong, most people will confidently say that it is, yet when faced with a situation in which considerable harm might come from telling the truth (e.g., when lying to a serial killer who asks for the whereabouts of his intended victims), they might reluctantly conclude that lying is morally acceptable in that particular case. This flexibility of moral judgment has frustrated efforts at reaching a consensus on a single moral theory that would adequately explain and justify moral judgment, and at the same time has led to concerns that moral intuitions are unstable and sensitive to seemingly irrelevant influences (see [5]). Thus far, moral psychology and theory have provided several competing explanations about the nature of moral intuition and underlying cognitive mechanisms: Moral Foundations Theory [6], Universal Moral Grammar [7,8], and the Agent-Deed-Consequence (ADC) model of moral judgment [9–11]. Whereas tests of the two prior explanations have not yielded definitive support for the postulated hypotheses and while the evidence is far from being unequivocal or conclusive (see S1 and S2 Texts), we want to focus on the latter model, which has not been directly tested yet.

Agent-Deed-Consequence model of moral judgment

The ADC-model [9–11] gives a fairly testable explanation for moral intuitions: it posits that moral judgment relies on positive and negative evaluations of three different components of moral intuitions: the character of a person (the Agent-component, A); their actions (the Deed-component, D); and the consequences brought about in a given situation (the Consequences-component, C). The theoretical approach underlying this model is grounded in the initial recognition that moral judgment involves multiple considerations that are difficult to compute. Moreover, most untrained individuals do not have explicit knowledge of philosophical ethics, and yet their intuitive moral judgments correspond to certain moral precepts implied in key ethical theories (e.g., see [12]). The ADC-model, following integrative approaches to moral theory [13,14] and empirical evidence noted above, considers the insights of three dominant ethical theories: virtue ethics, which focuses on the intentions and character of a person involved in a morally salient situation; deontology, which focuses on the analysis of certain actions that are prohibited or need to be undertaken as a duty; and consequentialism, which focuses on the balance of harms and gains resulting from the morally salient situation.

According to this integrative model, the moral evaluation of a situation can happen quickly and efficiently through the heuristic processing of morally salient cues. The heuristic processes substitute the overall moral judgment with more accessible information in distinct computations [9], and the outcomes of these are combined to form intuitive moral judgment. The ADC model suggests that this mostly happens unconsciously, but that conscious processes might monitor and correct judgments. The model predicts moral judgments to be positive if all three of the A-, D-, and C-components are positive and to be negative if all characteristics of the situation are evaluated as being negative. An important question, that will be explored in the following experiments, is what happens when the moral intuitions from the three components do not align. The model provides formulas for representing these situations, such as: [A+], [D-] and [C+] = [MJ+], where ‘+’ denotes ‘positive’, ‘-‘ denotes ‘negative’ and ‘MJ’ stands for moral judgment [9–11]. For instance, if the character and intentions of a person are good, and the action is good, individuals may be more likely to excuse bad consequences ([A+], [D+] & [C-] = [MJ+]). If the consequences are good and the action is done according to what duty demands, individuals may be more likely to forgive less than savory character traits and intentions ([A-], [D+] & [C+] = [MJ+]. Finally, the ADC-model explains how the three mentioned dominant moral theories rest on intuitive evaluations of the A-, D-, and C-components respectively. This implies that explicit moral precepts (i.e., concrete rules based on general principles) from these theories can be sufficiently dissociated from each other. The integrative approach also avoids “‘ethical blind spots’ and outlandish conclusions stemming from dogmatic application” of single-component theories ([9]:16), and provides a richer, multi-component, understanding of moral judgment.

Current study.

In this study, which comprises two experiments, we provide a first empirical investigation of some of the predictions yielded by the ADC-model using a factorial survey describing different moral situations (vignettes) across experimentally varied components of the ADC-model. Previous empirical support for the model has been indirect in terms of post-hoc explanations of numerous empirical findings (e.g., [15]), which were not testing the predictions of the model. We undertook a significant amount of work to develop a novel experimental paradigm, to avoid falling prey to the pitfalls of “trolley-ology” [16]. Most dilemmas used in experimental and survey research on moral judgment consist of trolley problems (see [17]), which are a poorly suited tool for teasing details apart from conflicting ‘consequentialist and ‘non-consequentialist’ responses (see [4,18]). Furthermore, much of the literature overlooks the predominantly mundane character of moral judgment [19]–the research has gravitated quickly to striking and artificial cases. Instead, we generated more realistic situations which were tested and piloted extensively. In both experiments, we were particularly interested in simultaneously testing the following predictions of the ADC-model where the moral acceptability was expected to be higher if the description of the:

Agent (A) is positive instead of negative;
Deed (D) is positive instead of negative;
Consequence (C) is positive instead of negative.

Apart from immediate hypotheses stemming from the ADC-model, this study hoped to elucidate additional important research questions. To gain a deeper understanding of moral judgment and its underlying processes, we explored whether and how the effects of the three components may affect each other, i.e., their potential interaction effects. If different components result in congruent moral intuitions, e.g., if the good intentions of an agent are congruent with his/her good deed, a positive interaction effect might occur. However, moral theories also generate antagonistic predictions, and individuals might hold antagonistic moral precepts, which could reflect the fluid boundaries of different components of moral intuitions, e.g., if a bad deed results in positive consequences, conflicts of intuitions are possible (see [20,21]).

The ADC model assumes that the A-, D-, and C-evaluations are not only characteristic of moral judgment, but also stable discrete components of moral intuition. However, a body of ethics literature is concerned with the distinctiveness of components or even domains of moral intuition (i.e., are they genuinely different from each other?) and their stability (i.e., do they remain the same across individuals or over time?) [1,2]. Thus, we examined whether the effects of the three components found in experiment 1 are similar in experiment 2 that uses another design, as well as to what degree the results could be generalized to another situation.

Finally, if moral judgments are influenced by the characteristics of the situation, namely variations in A, D, and C, as well as the conscious or unconscious evaluations they might trigger, then both experiments should explore whether these are the relevant influences—or if other variables (e.g., characteristics of moral decision-makers such as their age, sex, education, socioeconomic status) have additional effects on moral judgment as predicted by influential theories [22]. In addition, moral decision-makers might have developed their own explicit or implicit moral preferences for how to behave, judge, or solve moral problems [21,23,24]. With a newly developed instrument, we want to explore whether and how these preferences shape the way individuals react to variations of A, D, and C in order to further our knowledge about moral judgment. This will be tested in experiment 2.

Ethics statement

The ethics committees of the Institut de recherches cliniques de Montréal (approval number: 2015–13), and North Carolina State University (exempt status study number: 9362) approved this study.

Pretests

We conducted factorial surveys with vignettes, which are an experimental means of investigating causal effects. These vignettes are short descriptions of morally significant situations. This approach combines ideas from classical experiments and survey methodology to counterbalance the weakness of each approach: it provides high internal validity due to their orthogonal design; allows the investigation of multiple factors; provides an active mode of measurement; avoids multi-collinearity; allows for causal explanations; and acts as a good substitute for manipulations in the real world [25,26].

In this project, we initially developed seven preliminary vignettes featuring moral dilemmas of differing ‘stakes’. Stakes here refers to either the mundane or the more severe nature of a given moral transgression. We chose to develop both high- and low-stakes moral dilemmas because the mundane character of moral judgment is easily overlooked, and the fact that both the ethics and empirical moral psychology literatures have tended to gravitate toward striking and artificial cases [19], thus leaving a large proportion of actual moral judgment more or less unexplored [27]. With the help of a series of pretests (i.e., an expert review, a cognitive pretest, and a quantitative pretest), we selected the most suitable vignettes and undertook iterative refinement to ensure and increase the validity and quality of the instruments and design.

Expert review

We solicited written and oral feedback from experts (N = 4) with backgrounds in moral philosophy, applied ethics, survey methodology, and experimental moral psychology. We asked them for their thoughts on the aims of the project and the research design including the face validity of all measures, the plausibility of the vignettes, and the measurement concept. The feedback was assessed with structured questions (e.g., about the quality of each section of the study, with qualifications of the responses as well as suggestions). As a result of this process and through team discussions, we excluded one vignette due to its low plausibility. The experts indicated that the six remaining moral dilemmas were sufficiently plausible, had face validity in terms of measurement concept, and matched the overall goals of the study well. With the help of this feedback, several minor adjustments were made to the formulation of the instruments.

Cognitive pretest

To further refine the instruments, we conducted cognitive pretest interviews (N = 18) with a convenience sample of nine females and nine males of different age groups (18–50) and educational levels. We used a think-aloud-technique and probing questions to evaluate the understanding and clarity of our vignettes, questions, and instructions. The interviews were conducted by a research assistant under the supervision of the lead author, and verbal informed consent was obtained before each session. The assistant collected answers to the survey and took notes during each interview. Following team discussions, two vignettes were excluded because respondents consistently asked for clarifications, suggesting that the vignettes were not sufficiently comprehensible. Four vignettes were deemed clear enough for further testing.

Quantitative pretest

We tested the overall procedure and the scale properties with the help of a quantitative web-based pretest. Moreover, we wanted to test whether the experimental variations of the ADC-components in the vignettes are valid manipulations of the underlying concepts by testing their effect on moral acceptability. Furthermore, we wanted to test the convergent validity of our newly developed Preferences for Precepts Implied in Moral Theories (PPIMT) instrument which provides an operationalization of the preferences for specific precepts implied in the three key ethical theories listed above (namely, virtue ethics, deontology, and consequentialism) by examining their correlation with the self-identification of professional philosophers who endorse these theories. For this pretest, we invited members of several philosophy-related professional networks and newsletters (Philos-L, PHILOSOP, philosophy network on LinkedIn, and North Carolina Philosophical Society) to our study. Participation was completely anonymous and voluntary. Two hundred twenty-two people (N_Philos-L = 143; N_PHILOSOP = 47, N_LinkedIn = 13, and N_{North Carolina-Philosophical Society} = 19) visited the first survey-page and 209 (94.1%) consented to participate. Four participants were excluded because they either participated twice or did not respond to the question regarding prior participation. One-hundred-fifty-two (74.1%) of the participants, who consented and were not excluded, completed the survey. To maintain statistical power, we also use non-completers and those with missing values if they have valid values for the respective analysis.

Respondents received four vignettes with different moral problems (due to space limitations, we only show and discuss the results for the moral problems that were later used in the main study, see Table 1). We varied the information about the A, D, and C described in the dilemma in a 2x2x2x3-between-subjects design. Each of the factors A, D, and C had two levels (i.e., negative and positive valence), while the final factor indicates whether respondents had to evaluate the moral status of A, D, or C by stating whether each of the three components can be described as “bad”, “immoral”, “unethical”, “wrong”, and “horrible”. Response options for these five items ranged from 1 “not at all” to 10 “absolutely” (cf. [28–30]). To analyze the structure of these items, we used principal-component factor analyses with varimax rotation in each variation of the final factor. We found one-dimensional structures in each experimental arm (the Kaiser-Meyer-Olkin Measures ranged from 0.736 to 0.874, which indicated a good suitability of the data for structure detection; Cronbach’s α’s ranged from 0.90 to 0.98 (detailed results of the factor analyses and the reliability analyses are available upon request). For the items in each arm, we used regression factor scores for each factor (score 0 indicates an average moral evaluation of the A, D, or C, and 1 is the standard deviation). This has been done because some items are typically more important than others for explaining a certain construct. Thus, factor scores account for the different impact of each item—which would be not the case when using unweighted sum scores [31].

Download:

Table 1. Low- and high-stakes vignettes.

https://doi.org/10.1371/journal.pone.0204631.t001

The manipulation check analysis indicated that all negative valences of the treatments resulted in significantly higher numerical values regarding the judgment of A, D, and C (i.e., these components are judged as being more morally problematic, as in more “bad”, “immoral”, etc.) compared to the positive valences for one low-stakes vignette (M_A- = 0.72; M_A+ = -0.41; t(47) = 4.37; p<0.001; M_D- = 0.73; M_D+ = -0.70; t(51) = 7.25; p<0.001; M_C- = 0.38; M_C+ = -0.27; t(57) = 2.18; p = 0.032, see S1 Table) and one high-stakes vignette (M_A- = 0.35; M_A+ = -0.39; t(46) = 2.67; p = 0.011; M_D- = 0.45; M_D+ = -0.71; t(52) = 4.91; p<0.001; M_C- = 0.55; M_C+ = -0.47; t(59) = 4.50; p<0.001). Due to the failure of manipulation checks (results available upon request), we excluded the two other vignettes from the study.

Preferences for Precepts Implied in Moral Theories (PPIMT): To assess respondents’ preferences for the precepts implied in the three dominant moral theories (i.e., virtue ethics, deontology, and consequentialism), we developed a new instrument with 15 items (five for each theory, see Table 2). These items were presented after asking the following question: “When thinking about what is moral or immoral in a situation, it is important to me whether the involved persons…” Response options ranged from 1 “disagree very much” to 7 “agree very much”. We used principal-component factor analysis with varimax rotation (see Sample A in Table 2).

Download:

Table 2. Factor loadings of the items of the Preferences for Precepts Implied in Moral Theories (PPIMT) based on principal component factor analysis with a varimax rotation (Eigenvalues >1) (N_A = 140 and N_B = 786).

https://doi.org/10.1371/journal.pone.0204631.t002

We found a three-factor structure with five items that have substantial loadings on the theoretically implied factors and only low secondary loadings on other factors. The Kaiser-Meyer-Olkin Measure of 0.78 indicated a good suitability of the data for structure detection. Cronbach’s α’s indicate that the constructs can be measured reliably. For the analysis, we again used regression factor scores for each factor (score 0 indicates an average preference for precepts implied in the respective theory, and 1 is the standard deviation). Moreover, we tested the convergent validity of the scale. Therefore, we asked the members of several philosophy-related professional networks and newsletters how much they self-identify as virtue ethicists, deontologists, and consequentialists (on a scale from 1 “not at all” to 10 “absolutely”). We found that a person’s stronger self-identification with a certain theory was significantly correlated with their respective PPIMT underlying this theory (r _{(134) Self-identification as virtue ethicist, PPIMT-Virtue ethics} = 0.397, p<0.001); (r _{(137) Self-identification as deontologist, PPIMT-Deontology} = 0.477, p<0.001); and (r _{(137) Self-identification as consequentialist, PPIMT-Consequentialism} = 0.480, p<0.001), while all other correlations were statistically insignificant (see S2 Table), which indicates a high convergent validity of the PPIMT.

Experiment 1

Aims

As described above, the first experiment examined how variations in A, D, and C influence moral judgments in a low-stakes dilemma (featuring a case of possible syphilis contamination). Additionally, we explored the potential interaction effects of these variations and whether any differences in moral judgments occur across respondents of different age, sex, education, and socioeconomic status.

Methods

Participants and study design.

We invited participants via Amazon Mechanical Turk (AMT) to take part in a web-based study. AMT is a website that facilitates payment for completing surveys, and such samples have been shown to provide more reliable and representative data compared to student samples [32]. Participation was completely anonymous and voluntary. We offered .80 USD as monetary compensation after completion to motivate people to participate [33,34]. Six-hundred-seven people visited the first survey-page. Of these, 599 (98.7%) consented to participate, and of which 568 (94.8%) completed the survey. To ensure data quality, we excluded 21 persons who failed an attention check in which respondents are ostensibly asked about their favorite color [35]. Due to item non-response, the analytic sample comprises 525 respondents of which 46.5% are female (see Table 3 for all descriptive statistics).

Download:

Table 3. Descriptive statistics (N_{Experiment 1} = 525; N_{Experiment 2} = 786).

https://doi.org/10.1371/journal.pone.0204631.t003

Instruments.

Experimental design: In a 2x2x2 within-subject design, we varied the A, D, and C of the low-stakes vignette (see Table 1). Unlike previous work (e.g., [21]) which mostly contrasted actions and outcomes, or actions and omissions, our vignettes consistently provided positive and negative descriptions for A, D and C components. Every participant received all eight possible variations of this dilemma which allowed a test of the experimental treatments within each participant. The allocation of the eight vignettes was randomized for every participant.

Moral judgment: After each vignette, we asked the following question to assess respondent’s moral evaluation of the situation: “Taking all circumstances into consideration, [item 1: for you personally/item 2: for society], how morally acceptable is what the man did in this situation?” (for similar but distinct measures of moral acceptability, cf. [36–38]). Response options ranged from 1 “not at all” to 10 “absolutely”. Given the very high correlation of both items for the syphilis vignette (r = .941, p<0.001), we computed a mean score.

Socio-demographic information: We assessed gender (0 “female”; 1 “male”), age, education in years (degrees have been assigned to the average duration of this educational level), and subjective social status to capture one’s sense of their own place on the social ladder (measured by the MacArthur Scale of Subjective Social Status; 1 “lowest status”, 10 “highest status”; [39]).

Statistical analysis.

Each respondent rated eight vignettes, which resulted in a hierarchical structure of the data. Ratings of respondents have correlated error terms that can result in underestimated standard errors. Multilevel models were used to allow us to test for the effects of the experimental treatments (vignette level) and socio-demographics (respondent level). We used Wald post-estimation tests to explore statistical differences between the three ADC-components. We used a procedure for probing the three-way interaction between A, D, and C, including the conditional effects of A, D, and C, as well as all two-way interactions [40]. Simple slopes were graphically reported along with tests for differences of each pair of simple slopes by applying a slope difference test [41]. We plotted the interactions in three different ways (by exchanging the variables on the x-Axis across the plots), which is useful for the interpretation of such higher order interactions ([40]:52). We use an α-level of p<0.05 for the test of statistical significance.

Results of experiment 1

Panel 1 in Fig 1 shows the mean values of each treatment (lower part) and each treatment combination (upper part). An empty model without co-variates (not shown) shows that the vignette level accounts for 98.1% of the entire variance. Model 1 (M₁) in Table 4 shows that if the man (i.e., the agent) is loyal to his wife (A+), respondents found the situation to be more morally acceptable compared to a situation where the agent has been described as an adulterer (A-)–as indicated by the significant positive B-coefficient (1.858, p<0.001, see also S3 Table which summarizes all findings of the ADC-components). When the man tells the truth about the doctor’s prognosis (D+) instead of lying (D-), the moral acceptability was higher (p<0.001). When it turned out that the man is healthy (C+) compared to when the man was ill and his wife has the first symptoms of syphilis (C-), the situation was judged as being more positive (p<0.001). Post-estimation Wald tests show that the effect of A was significantly smaller than the effects of D, but significantly larger than the effect of C, while the effect of D was significantly larger than the effect of C (all p<0.001). The measured respondent characteristics (i.e., sex (p = 0.541), age (p = 0.196), education (p = 0.170), and subjective social status (p = 0.877) had no significant effects and did not influence the effects of A, D, and C (model not shown, results available upon request).

Download:

Fig 1. Mean values for moral judgments.

https://doi.org/10.1371/journal.pone.0204631.g001

Download:

Table 4. Multilevel models of the moral judgment regarding the low-stakes vignette on the experimental treatments and respondent characteristics (number of respondents = 525; number of vignettes = 4,200).

https://doi.org/10.1371/journal.pone.0204631.t004

In M₂, we tested all possible two-way interactions between treatments. We found positive interaction effects between A and D (p<0.001) and A and C (p = 0.014) as well as a negative interaction effect between D and C (p = 0.021). No three-way interaction between A, D, and C was found in M₃ (p = 0.546). However, this model shows that the (conditional) interaction effect between D and C is no longer significant (p = 0.230) if A is negative.

The S3 Table provides an overview of the complex interaction effects between A,D,C discovered in experiment 1 and which were further investigated in experiment 2. Fig 2 illustrates the effects of A, D, C, and their interactions and is accompanied by a simple slope analysis and slope difference tests to qualify and better understand the interaction effects. The simple slope analysis shows that A has significant positive effects in all treatment combinations (see lines 1–4 have positive ascents in Panel 1). Slope difference tests show that there is no interaction effect between A and C if D is positive (lines 1 and 2 are relatively parallel), but if D is negative, we find a positive interaction effect (line 3 is steeper than line 4). Thus, if D is negative, a positive C leads to a more positive effect of A (and vice versa). The positive interaction effect between A and D can be seen from the comparison of the following four lines: In situations with a positive C, the effect of A is larger if D is also positive rather than negative (line 1 is steeper than line 3), whereas in a situation with a negative C, the effect of A is also larger if D is positive rather than negative (line 2 is steeper than line 4). Moreover, when both D and C are positive, A has a larger effect than if D and C are negative (line 1 is steeper than line 4) and A has a stronger effect in a situation with a negative D and a positive C as compared to a positive D and negative C (line 2 is steeper than line 3), and that might be due to the stronger interaction effect between A and D compared to A and C.

Download:

Fig 2. Predicted values for moral judgments regarding the low-stakes vignette (syphilis) depending on the experimental treatments and their interactions based on a linear regression model.

https://doi.org/10.1371/journal.pone.0204631.g002

Panel 2 shows that D has positive effects in all treatment combinations (see positive ascents of lines 1–4). It also shows the positive interaction effect between A and D from the perspective of D (line 1 is slightly steeper than line 2, and line 3 is steeper than line 4). Thus, a positive A leads to a more positive effect of D (and vice versa). When A is positive, D loses some strength if C is positive rather than negative (line 1 is flatter than line 3) indicating a conditional negative interaction effect between D and C, while when A is negative, D has similar effects irrespective of C (line 2 and 4 are relatively parallel). When both A and D are negative, D has a smaller effect than if A and C are both positive (line 4 is flatter than line 1). In a situation with a positive A and negative C, D has a stronger effect compared to one in which A is negative and C positive (line 3 is steeper than line 2).

Panel 3 shows that C mainly has positive effects (lines 1, 3, and 4 are ascending), C has no effect if D is positive and A is negative (line 2 is almost parallel to the x-axis). Furthermore, the negative interaction effect between D and C is significant if A is positive (in Panel 3 line 1 is slightly flatter than line 3), thus if A is positive, a positive D leads to a smaller positive effect of C (and vice versa). But there is no such interaction if A is negative (lines 2 and 4 are almost parallel). The effect of C also does not differ in situations with D and C being both positive compared to being both negative or compared to situations with a D positive and negative A (line 1 is relatively parallel to lines 4 and 2). If D is negative, C has a stronger effect in situations with a positive rather than negative A (line 3 is steeper than line 4), again indicating a conditional interaction effect between A and C. C also has a stronger effect if A is positive and D negative compared to if A is negative and D positive (line 2 is flatter than line 3).

Discussion of experiment 1

The results of experiment 1 supported the most important prediction of the ADC-model: With a positive valence of A as compared to a negative valence of the A (being loyal vs. being an adulterer) the D (telling the truth vs. lying), and C (both the man and his wife wound up healthy vs. have syphilis), the situation was judged as being significantly more acceptable compared to when all three components had a negative valence. This is the first direct empirical corroboration of the ADC-model, and thus serves as ‘proof of principle’, which should be explored with more research to test predictions and additional hypotheses about the mechanisms underlying these different components of morality.

Differences in component effect strength: The strongest effect was observed in the variation of the D-component, followed by the A-component, whereas the C-component produced the smallest effects. This relative weakness of the C-component could be the result of, for example, a genuine difference in the intuitive strength of different components of moral judgment, or of the relative lack of moral impact for other people (i.e., the fact that it is a low-stakes scenario—it does not concern drastic outcomes, such as the death of innocent people).

Interaction effects: Our exploratory interaction analysis reveals a two-way interaction effect between A and D: If the agent was described as positive (i.e., loyal) rather than negative (i.e., an adulterer), the positive description of the action (i.e., telling the truth) had a stronger positive effect on moral judgment and vice versa. One interpretation of this effect might be that the description of the deed is congruent with the description of the agent (and vice versa). This supports a common belief that good people do good deeds (and that bad people do bad deeds). Qualitative data from the pre-test seem to indicate that this is the case at least with some responders. Namely, in the situation where the ‘loyal’ man decides to lie to his wife, several pretest participants verbalized their concerns along the following lines: ‘He might have been loyal because he didn’t have the chance to cheat; judging by the fact that he is lying, he ultimately would have cheated if able’. Therefore, a positive description of the agent confirms that the positive deed is not just a single instance of good behavior, but the agent’s overall stable disposition, which might lead to the stronger positive effect on moral judgment. Indeed, moral theories have long noted that congruence between intention and action is necessary for morality. For instance, Kant [42,43] famously argued that good deeds are not really moral if they are not motivated by good will.

We also found a conditional positive interaction effect between A and C: If A was positive rather than negative, the description of a positive C (i.e., that both the man and his wife are out of danger) had a stronger positive effect. But this interaction effect is conditional on the valence of D; it only occurs if D is negative, while no such effect exists if D is positive. This is in line with an opposing, but also common belief that good people sometimes do bad things to achieve significant good results.

Finally, we found a conditional negative interaction effect between D and C, i.e., if D was positive, the effect of a positive C was smaller (and vice versa), and this effect only occurred when A is positive but not if A is negative. However, when A was negative and D positive, we see that C has no effect. Generally speaking, deontology and consequentialism are moral theories which usually provide opposing ethical guidance, so some sort of reduction of effect is not surprising. The description of the agent could also be a ‘tie-breaker’ between C and D. For example, Dubljević & Racine [9] explain the difference between the trolley problem and the footbridge problem by taking into account implicit intuitions regarding the agent, which are not controlled (nor asked for) in most trolley dilemmas. In addition, intuitively deontological and consequentialist judgments are considered to be opposed in a large body of literature (see [4,21]). However, it has to be noted that the negative interaction effect between D and C is present only if A is positive—if A was negative, there was no two-way interaction between D and C, a finding for which we have no clear interpretation.

Lack of socio-demographics effects: We found no significant effects of the investigated socio-demographics (i.e., sex, age, education, and subjective social status) on moral judgment. It is important to note that the ADC-model only makes predictions regarding the relevance of A, D, and C. It makes no predictions regarding the effects of the individual characteristics of the moral decision-maker. However, some earlier models (e.g., [22]) have made such predictions. Our analysis, however, does not support this.

Potential limitations: This experiment used a within-subject design, and so every participant rated every variant of the situation. This could have produced learning effects, and contrast effects [44]. In addition, given that the vignette is low-stakes, the C-component might have had only a relatively small effect, which could change in a high-stakes situation. Experiment 2 aims to further explore the results of experiment 1 with another design, and test their generalizability to a more dramatic ethical problem.

Experiment 2

Aims

Experiment 2 was designed to address several potential limitations of experiment 1 and to explore the generalizability and robustness of its findings.

Between-subject design: In the previous within-subject experiment, every participant rated all eight variants of the low-stakes vignette. Exposing each participant to multiple vignettes might, for example, have elicited fatigue, learning, or contrast effects, and we minimized these effects in experiment 2 [25,26,44]. If the findings are similar in a between-subject design in which each participant reads only one vignette, this would rule out such effects.

High-stakes situations: Since it is assumed that the components of the model can play a role in different moral dilemmas, and our first experiment tested the model in a situation with relatively low-stakes, we wanted to test the ADC-model in a situation with higher stakes. Namely, based on the results of experiment 1, one might argue that the D-component is highly relevant in low-stakes scenarios, but that it could become irrelevant if the consequences were serious and harmful enough.

Interaction effects: We are also interested in whether the interaction effects that we found in our exploratory analysis are similar or different with this between-subject design across scenarios of varying levels of seriousness.

Effects of PPIMT: While situations have certain characteristics that are assumed to trigger moral judgment, certain characteristics of moral decision-makers such as their “ethical framework” [23] might also play a role in the perception of the situation and thus during moral judgment. This framework might include preferences for how to behave, judge, or solve moral problems. Such preferences can be seen as personality traits that are relatively stable over time and which are learned via socialization (see e.g., [45]). These preferences may embody the precepts implied by the three mentioned dominant moral theories: virtue ethics (e.g., “strive to be an honest person”), deontology (e.g., “obey rules, such as ‘never lie’”), and consequentialism (e.g., “maximize happiness and save lives with any means necessary”). Individuals might have preferences for different precepts of these theories which could serve as guidance for many morally salient situations in line with previous findings [21,24].

These preferences might work consciously or unconsciously: people with certain PPIMT might actively look for certain cues or significant symbols that help define the situation. Alternatively, they might be susceptible to implicit cues or symbols and thus, their preferences might create a certain accessibility, focus, or awareness for certain aspects of the situation, while other aspects might be less likely to be considered or might even be ignored. In this process, these preferences can help with processing information by structuring or framing the perception of characteristics of the situation (cf. [21,24]). We thus want to examine the effect of PPIMT on moral judgment, especially, the moderating role on the effects of the characteristics of the moral problem (i.e., the three ADC-components) on moral judgment. It can be argued that subscribing to such precepts or explicitly using them as tools moderates the effect of characteristics of the situation (cf. e.g., [21,23,24,46]). For example, stronger PPIMT_{Virtue ethics} might moderate the effect of the A-component (i.e., a virtuous agent might be judged more positively, while a non-virtuous agent might be judged more negatively). Moreover, someone with such strong preferences might tend to ignore variations of the D- or the C-components, or might pay less attention to them, thus giving D and C less weight in their overall moral judgment. Previous research explored a similar idea about how ‘ethical frameworks,’ in terms of a person’s predisposition towards either deontology or consequentialism, affects people’s moral judgment—and that it interacts with the parameters of the situation (see e.g., [46]) Another line of research has reported that a ‘consequentialist focus’ drives framing effects, whereas a ‘deontological focus’ decreases framing effects [21]. However, we do not know of any research that investigates the preferences for precepts implied in all three dominant moral theories at once.