abstract - ghent university web viewusing mcmc chain outputs to efficiently estimate bayes factors....

Running Head: THE MORALITY OF ACTION 1

The morality of action:

The asymmetry between judgments of praise and blame in the action-omission effect.

Dries H. Bostyn a, Arne Roets a

Ghent University; Department of Developmental, Personality, and Social Psychology; Henri

Dunantlaan 2, B-9000, Ghent, Belgium.

[email protected], [email protected]

WORD COUNT: 7085

Biography:

X

X

X

X

X

X

Authors’ Note

Correspondence concerning this article should be addressed to Dries H. Bostyn,

Department of Developmental, Personality, and Social Psychology, Henri Dunantlaan 2, B-

9000, Ghent, Belgium. E-mail: [email protected], Tel: ++32(0)9 264 91 41, Fax.: ++32

(0)9 264 64 99.

THE MORALITY OF ACTION 2

Abstract

Actions leading to negative outcomes (i.e., harm) are seen as more blameworthy than

omissions of actions leading to the same negative outcomes. However, whether a similar

action-omission effect applies to judgments of praiseworthiness of positive outcomes is still

an open question. Drawing on positive-negative asymmetries found in other domains, we

hypothesized that positive events would not elicit an action-omission effect for judgments of

praise, because such positive events do not by default trigger the causal appraisal processes

that are central to the action-omission effect. Furthermore, we posited that when people are

explicitly asked to consider causality before or during the judgment, an action-omission effect

on judgments of praise could be obtained too. These hypotheses were verified in three

independent studies and a meta-analytic analysis. As such, the present set of studies provides

novel insights in the action-omission effect’s asymmetry for negative and positive outcomes,

as well as an increased understanding of the role of causality appraisal in this effect:

judgments of praise are less reliant on causal reasoning than judgments of blame, and

therefore also less susceptible to the action-omission bias.

Keywords: omission-bias; action effect; causal appraisal; praise; moral cognition


Actions yielding negative outcomes are judged to be morally worse than omissions of

actions resulting in the same negative outcomes (Spranca, Minsk, & Baron, 1991; Ritov &

Baron, 1999; Baron & Ritov, 2004; Cushman, Young, & Hauser, 2006; DeScioli, Bruening, &

Kurzban, 2011). The present research addresses whether this ‘action-omission’ effect is

generalizable to positive outcomes.

Although there is a rich research literature on the action-omission effect, no studies

that we are aware of have investigated whether judgments of praise, similar to judgments of

blame, demonstrate an action-omission effect. Intuitively, it would make sense that actions

leading to positive outcomes are deemed more praiseworthy than omissions leading to those

same outcomes. If it is more blameworthy to ‘kill’ than to ‘let die’ (Spranca et al. 1991) then it

is probably also more praiseworthy to actively ‘save someone’ than to ‘let someone be saved’.

However, there are some reasons to assume the effect may be slightly more complex and does

not display this kind of symmetry.

First of all, several studies have noted that negative events tend to elicit stronger and

different psychological reactions compared to positive events. Negative events and stimuli are

more salient, appear to be more potent and tend to trigger more deliberative thought than

positively valenced events do (Rozin & Royzmann, 2001; Baumeister, Bratslavsky,

Finkenauer & Vohs, 2001). This ‘negativity bias’ effect has been found in a wide variety of

domains spanning from loss aversion (Kahneman & Tversky, 1984) to impression formation

(Peeters & Czapinsky, 1990). Given the psychological ubiquity of this negativity bias it would

not be unreasonable to suppose that it might affect moral judgment as well. Indeed, some

research has suggested different evaluation standards for the morality of negative versus

positive actions. For instance, both adults and children tend to engage more frequently in

judgments of blame than then they do in judgments of praise (Ross & den Bak-Lammers,

1998; Wiessner, 2005) and legal systems serve to condemn criminals but do not reward the


virtuous (DeScioli & Kurzban, 2009). According to DeScioli & Kurzban (2013), moral

cognition seems inherently more attuned to judge blame than to judge praise and Cushman

and Greene (2012) showed that the presence of negative consequences is central to our

judgment of peoples’ actions.

If moral reasoning is indeed strongly affected by the presence of negative outcomes,

one could expect to find some asymmetries when it comes to blaming and praising behavior,

including the effect of action versus omission in the moral appraisal of such behavior. In this

regard, research on the action-omission effect has suggested a key role of the attribution of

causal responsibility. In their seminal work, Spranca et al. (1991) already reported that

individuals exhibiting an action-omission effect often referred to differences in causality when

asked for an explicit justification. Building on these initial findings, Kordes-de Vaal (1996)

did indeed find that actions lead to higher ratings of causal responsibility than omissions do.

Finally, Cushman and Young (2011) demonstrated that the action–omission effect was

especially prominent for judgments that had previously been shown to rely on an analysis of

causal responsibility, that is, judgments of blame and punishment.

Importantly, research has shown that negativity bias also impacts causal judgment. In

particular, negative events elicit more causal attribution than positive events do (Bohner,

Bless, Schwarz & Strack, 1988) and they trigger more counter-factual thinking (Roese &

Olson, 1997). Given the assumed importance of causal attribution to the action-omission

effect we expect that this might also affect whether or not an action-omission effect can be

found on judgments of praise. If positive events do not trigger causal attribution processes to

the same extent as negative events do, then an action-omission effect on judgments of praise

is likely to be absent or at least smaller in size.

Building upon these insights from different domains, we advance the following

hypotheses: First of all, negative events should trigger sufficient causal appraisal in and off


themselves, hence the existence of an action-omission effect on judgments of blame should

not be dependent on whether subjects are explicitly asked to reflect on causality. Even without

an explicit cue to consider causality, an action-omission effect on judgments of blame should

be present (as has often been demonstrated in the literature). However, when subjects are

presented with a positive event, this positive event in itself may not by default trigger the

causal attribution processes that are assumed to be necessary for the action-omission effect.

Thus when subjects are not explicitly required to reflect on causality, judgments of praise

should not be as susceptible to the action-omission effect. In contrast, if subjects are explicitly

asked to reflect on causality then an action-omission effect on judgments of praise may appear

(although not necessarily as strong as is the case for judgments of blame). Testing these

specific hypotheses will provide insight in the action-omission effect’s potential asymmetry

for negative and positive outcomes, as well as advance our understanding of the assumed role

of causality appraisal in this effect. These hypotheses were tested in a series of three

independent studies as well as a meta-analytic analysis on the combined results making use of

the full power of the combined data set to demonstrate the global pattern. For the meta-

analysis we used traditional methods as well as Bayesian statistics to test the alternative and

null hypotheses.

Study 1: Method

Participants

Before running the first experiment, power-analyses were conducted to determine a

sufficient sample size with the R package ‘pwr’ (Champely, Ekstrom, Dalgaard, Gill & De

Rosario, 2015). Based on previous research suggesting the action-omission effect is small to

medium in size, as per Cohen (1992), we deemed that one hundred participants per condition

should result in sufficient power to find the hypothesized action-omission effects.1

1 If we assume an effect size of Cohen’s d = 0.40 one hundred participants per condition is sufficient for a power of 80.3%.


Anticipating some drop-out, a total of four-hundred-fourteen participants (49% female, Mean

age: 37.7) were recruited through the online labor platform Amazon’s Mechanical Turk

(AMT). AMT has been demonstrated to be as reliable as traditional methods of recruiting

subjects (Paolacci, Chandler, & Ipeirotis, 2010; Rand, 2012). Participants were paid US$1.

Participation was limited to US-citizens with an AMT approval rating higher than 95%.

Procedure and Materials

After the completion of the demographic variables, participants started a judgment

task in which they were presented with six scenarios and were asked to rate the behavior of

the target. The experiment was designed as a 2 (Action) x2 (Outcome) x2 (Causality

Appraisal) between-subjects design. For each of the six scenarios, four different versions were

developed, each representing one of the between-subjects Action x Outcome conditions:

action-positive, omission-positive, action-negative, and omission-negative. All participants

were asked to rate the behavior of the target on a six point scale going from ‘extremely

blameworthy’ (1) to ‘extremely praiseworthy’ (6).

Importantly, the current study uses a slightly more subtle manipulation of actions and

omissions compared to previous studies. In particular, in most other studies on the action-

omission effect, the target is a bystander in the omission scenarios but the main actor/cause of

the outcome in the action scenarios (as is the case when contrasting ‘killing’ to ‘letting die’).

To eliminate this potential confound all scenarios were designed so the target was a bystander

who could influence the outcome of the event (although he was not the instigator of the

situation). An example scenario for the action-positive outcome condition reads:

“Joe is walking through his local fair. Joe notices a group of 3 children playing just a

little bit ahead of him when all of a sudden he hears a cry. Joe turns around to see a

teenager on a go cart storming in his direction. The go cart is heading straight for the

group of children and there is no way it will be able to brake in time. If Joe doesn’t


jump in front of the go cart it will surely hit the group of children. Joe realizes that if

he gets hit by the go cart he will get away with a few nasty bruises but if the go cart

hits the group of children they will undoubtedly be much worse off. Joe decides to

jump in front of the go cart. The outcome of this decision is that the children do not get

hurt.”

Conversely, in the omission version of the same positive outcome scenario, Joe is

already standing in the path of the go-cart and he decides not to jump out of its way. In the

negative outcome conditions of this particular scenario, Joe’s decision leads to the children

getting hurt, either by jumping out of the way (action) or deciding not to jump in front of the

go-cart (omission). After rating each scenario, participants answered two easy but crucial

comprehension questions to check if they had read and understood the scenario. An example

question is: “Will Joe get hit by the go-cart?”

In addition to the manipulations of Action (i.e., action vs. omission) and Outcome (i.e.,

positive vs. negative), we also manipulated causality appraisal. In particular, participants were

either only asked to provide merely a blame-praise judgment of Joe (no causality appraisal),

or they were asked to additionally also rate to what extent they felt Joe caused the outcome

(causality appraisal) on a four point scale anchored by: “Not at all” (1), “Maybe a little” (2),

“To a considerable degree” (3), and “Completely” (4). Including Causality appraisal as a

separate between subject factor allows us to investigate to what extent probing people to

explicitly make a causal analysis influences their moral judgments, and whether this affects

blame and praise ratings (differently). All scenarios and comprehension checks are presented

in the supplementary materials. All measures and variables that were part of the study are

included.


Study 1: Results

Data preparation

Participants who had responded incorrectly to the comprehension questions of more

than one scenario were deemed unreliable and removed from the sample (n = 59, spread

evenly across all conditions) (see, Oppenheimer, Meyvis & Davidenko, 2009). For

participants who failed the comprehension checks of only a single scenario, we only omitted

that specific trial (5.5% of all trials). Because, depending on the condition, more extreme

judgments are indicated by either lower (negative outcome condition) or higher (positive

outcome condition) ratings on the 6-point scale, a direct analysis of these raw scale scores

would yield confusing results when testing the full design. An action-omission effect in the

Blame condition would yield a negative parameter estimate for the effect of action whilst an

action-omission effect in the Praise condition would yield a positive parameter estimate. To

alleviate this problem, we split our scale along the midpoint and rescaled all judgment ratings

to a scale ranging from 1 to 3, with higher ratings meaning either higher blame (for the

Negative outcome condition) or higher praise (for the Positive outcome condition). As such,

this recode allows for a straightforward test for a main action-omission effect. Occasionally, a

participant rated a scenario with a positive outcome as “blameworthy” or a scenario with a

negative outcome as “praiseworthy”. Such responses were relatively rare (6.7% of all

judgments) and random across trials, but they did have a disproportionately large impact on

the cell means. In order to optimize the judgment measure, we deemed these responses to be

noise and omitted them when calculating the individual’s mean judgment score across the

scenarios.2 Three participants apparently misunderstood the scale and provided reverse scores

for all trials, and hence received no mean judgment score.2 It is important to note that this procedure did not lead to any substantial quantitative change in the

results reported in this study. When including these participants in the analyses, the overall action-omission effect on judgments of Blame was still significant; t(170) = 2.27, p = 0.02, whereas the action-omission effect on judgments of Praise remained non-significant; t(177) = 1.49, p = 0.14. However, in the follow-up contrast analysis, contrasts did not reach traditional significance levels. We belief this to be a power issue caused by the increased noisiness of the uncorrected version of our dependent measure.


Data analysis

For a straightforward representation of the findings, we first looked at the action-

omission effect in the traditional negative outcome scenarios in which participants’ ratings

reflect the degree of blame (Cronbach’s α = 0.64). ANOVA of the mean judgment scores

revealed a significant action-omission effect; F(1,168) = 8.94, p = .003, which was not

moderated by the causality appraisal manipulation; F(1,168) = 0.02, p = .635. The causality

appraisal manipulation itself showed no main effect; F(1,168) = 3.20, p = .075. Planned

comparisons revealed that the effect was somewhat stronger when subjects were asked to

provide causality ratings together with their judgment ratings; t(168) = 2.64, p = 0.009,

Cohen’s d = 0.53, compared to when they were not asked to provide causality ratings; t(168)

= 1.67, p = 0.097, Cohen’s d = 0.39.

A similar analysis for the effects on the positive scenarios where subjects had to rate

praise (Cronbach’s α = 0.67), revealed quite a different pattern. ANOVA of the mean

judgment scores revealed no significant overall action-omission effect; F(1,176) = 0.38, p

= .539, and no main effect of causality F(1,176) = 0.62, p = .432. Interestingly however, a

significant interaction between Action and Causality emerged; F(1,176) = 4.12, p = .044.

Further planned contrast analyses showed that the action-omission effect was absent when no

causality appraisal was required; t(176) = 0.92, p = 0.357, Cohen’s d = -0.22, whereas the

effect was significant when subjects were explicitly asked to assess causality as well; t(176) =

2.06, p = 0.041, Cohen’s d = 0.40 . These results are visualized in Figure 1.3

3 One reviewer noted that it might be worthwhile to rerun these analyses for each outcome condition with the six scenarios included as a within-subject factor. Although for both outcome conditions this within-subjects factor itself yielded a significant main effect (F(5, 885) = 24.21, p < 0.001, and F(5, 850) = 45.38, p < 0.001, for positive and negative outcomes, respectively), it did not affect the crucial action-omission effects. That is, similar to the reported analyses, the analyses revealed a significant action-omission effect in the negative outcome condition; F(1, 170) = 5.16, p = 0.03, but no overall action omission effect in the positive outcome condition; F(1, 177) = 2.21, p = 0.14. Importantly, none of these key findings showed any further interaction with the within-subjects factor; F(5, 885) = 0.304, p = 0.67, and F(5, 850) = 1.945, p = 0.09, for positive and negative outcomes, respectively.These analyses also suggested that varying levels of “self-sacrifice” in our positive outcome scenarios are unlikely to affect the findings with regard to the action-omission effect. A more specific test at the request of another reviewer, including amount of self-harm as a three level within-subject factor (two scenarios identified as including no self-harm, two pertaining to low self-harm, and two pertaining to high self-harm) demonstrated


Study 1: Discussion

The first study confirmed the overall action-omission effect on the blame ratings for

negative events, and most importantly, the results also provided evidence for our asymmetry

assumption on judgments of praise. Asking subjects to rate causality concurrently with praise

seemed to trigger an action-omission effect, whereas the effect was absent without the explicit

causality salience manipulation.

However, while these initial results are supportive of our hypotheses, the current study

has some limitations. Firstly, our dependent measure of praise and blame may be less then

optimal. To keep the procedure uniform for all participants we decided on using a single

bipolar scale to measure both blame and praise. However, this scale format appears to have

confused some participants (leading to some irrational judgments and occasionally even a

complete misunderstanding of the scale). Moreover, the recoding to two unipolar 3-point

scales also meant that each outcome condition effectively used only one halve of the original

scale, diminishing the overall variety captured by the scale. This may have artificially

suppressed the size of the effects.

Secondly, although we had a sample of sufficient size to detect medium sized main

and interaction effects, statistical power may not have been large enough to detect smaller

effects, especially when testing contrasts between specific individual cells. Therefore we

conducted a second study with more straightforward measures of blame and praise and an

increased sample size, in order to allow us to focus on the specific contrasts driving the

overall effects.

that self-harm indeed did not interact with the action effect; F(2, 354) = 1.18, p = 0.31, or the action-causality interaction; F(2,354) = 0.65, p = 0.52.


Study 2: Method

Participants

To increase power to find significant effects in each contrast we increased our desired

sample size; six hundred participants (49% female, Mean age: 37.9) were recruited through

the online labor platform Amazon’s Mechanical Turk (AMT). Participation-criteria were the

same as in Study 1 and participants were compensated $0.50 for their participation.


Procedure and materials were largely identical to Study 1 with two important changes.

First off all, instead of using a single bipolar scale to measure blame and praise, participants

in the positive outcome condition were presented with a five point rating scale going from

‘Not at all praiseworthy’ (1) to ‘Extremely praiseworthy’ (5) whereas the participants in the

negative outcome condition were presented with a similar five point scale going from ‘Not at

all blameworthy’ (1) to ‘Extremely blameworthy’ (5). Secondly, all participants were asked to

rate causality (on the same 4-point scale as in Study 1), but participants were presented with

each scenario twice in a blocked fashion, asking them to rate blame- or praiseworthiness in

the first block and causality in the second block, or vice versa. As such, halve of the

participants rated blame- or praiseworthiness for all six scenarios before rating causality

whereas the other half rated causality first and were only asked about blame- or

praiseworthiness in the second block. A single comprehension question was asked after each

presentation of a dilemma.

Study 2: Results

Data preparation

Data was prepared similarly as in Study 1; participants who had responded incorrectly

to the comprehension questions of more than one scenario were deemed unreliable and

removed from the sample (n = 53, evenly spread across conditions).4 For those participants

4 These participants appeared to be evenly spread between all conditions.


who provided an incorrect answer on the comprehension check for only for one judgment

scenario, scores on that specific trial were deleted (3.2%).

Data analysis

Action-omission effects on the judgment ratings.

As in Study 1 we looked for action-omission effects in each outcome condition

separately. ANOVA of the mean judgment scores in the negative outcome condition

(Cronbach’s α = 0.73) revealed a significant Action-omission effect; F(1,270) = 9.72, p = .002

but no effect of Causality condition; F(1,270) = 0.35, p = 0.552, nor a significant interaction-

effect; F(1,270) < 0.01, p = 0.962. As expected, planned contrast analyses revealed a

significant Action-omission effect both when causality was rated before the blame judgments;

t(270) = 2.16, p = 0.031, Cohen’s d = 0.38, and when causality was rated after the blame

judgments; t(270) = 2.26, p = 0.025, Cohen’s d = 0.37.

For the mean judgment scores in the positive outcome condition (Cronbach’s α = 0.81)

a similar pattern emerged. ANOVA of the mean judgment scores did reveal a significant

Action-omission effect; F(1,269) = 9.28, p = 0.003 but no effect of Causality condition;

F(1,269) = 1.54, p = 0.215, nor a significant interaction-effect; F(1,269) = 0.15, p = 0.699.

Crucially, planned contrast analyses did reveal that the Action-omission effect was present

when causality was rated before the praise judgments; t(269) = 2.36, p = 0.019, Cohen’s d =

0.42, but did not reach traditional significance levels when causality was rated after the blame

judgments; t(269) = 1.94, p = 0.053, Cohen’s d = 0.32, although the effect seemed not

completely absent either, as was the case in Study 1.

Study 2: Discussion

The results of this second study corroborated the general pattern obtained in Study 1.

Action-omission effects for blame ratings in negative events emerged regardless of the

experimental manipulation of causality salience. For praise ratings, we obtained an action-


omission effect when causality appraisal was made salient, but, similar to Study 1, no

statistically significant action-omission effect was found when participants were confronted

with a positive outcome while causality was not made salient, despite increased power, and a

more straightforward measurement of the judgments in this study. However, while this effect

was not technically statistically significant, it was undeniably present with a p-value that was

only slightly higher than the traditional .05 cut-off for significance. A skeptical reader might

therefore argue that this result, for all intends and purposes, does indicate the presence of an

effect. As such, this result does not provide the same level of evidence for our hypotheses

regarding the absence of an action-omission effect on praise ratings as the results from Study

1 did, and warrants further exploration. Therefore, to be able to state with confidence whether

or not the action-omission effect on praise ratings is dependent on the presence of a causality

salience manipulation we decided to run a third study. To maximize power, and because the

action-omission effect on judgments of blame is already well established, this third study only

looked for action-omission effects on praise ratings.

Study 3: Method

Participants

Before conducting this study we ran a new power analysis to determine the appropriate

sample size. A sample of 548 participants, or 137 per condition, would yield 80% power for a

hypothesized effect of a size Cohen’s d = 0.34; an effect size that is slightly smaller than the

effects we have uncovered so far in these studies (not including the hypothesized null effects

in the no causality salience, positive outcome condition). Furthermore, our previous two

studies had a drop-out of about 11% on average due to missed comprehension checks. Taking

this dropout into account we aimed to recruit a total sample of 608 participants, distributed

across four conditions.


All participants were recruited through Amazon’s Mechanical Turk as per the same

participation criteria of the previous two studies. Participants were compensated US$0.35. At

the suggestion of the action-editor we pre-registered this third study and its methodology at

Open Science Framework. This pre-registration can be found at https://osf.io/adt2r/.


The materials in this third study were identical to those of the previous two studies,

though we decided on using the superior measurement scale of Study 2. Furthermore, to keep

the length of the experiment to a minimum we reverted our Causality salience manipulation

back to the one of the first study (i.e., a concurrent Causality judgment).

Study 3: Results

Data preparation

Data was prepared in the same way as the previous studies. Once again,

participants who had responded incorrectly to the comprehension questions of more than one

scenario were deemed unreliable and removed from the sample, however out of the 608

participants recruited for this study, this time, this procedure led to a substantial data-loss of

119 participants,5 reducing our sample to 489 participants; which is 59 participants short of

our desired sample size. To attain the pre-set effective sample of n = 548, we collected an

additional 96 participants of which another 26 failed to correctly respond to the

comprehension questions of more than one scenario, leading to a total effective sample-size of

559 participants. This does mean that we deviated slightly from the procedure outlined in our

pre-registration. However, given that the aim of this study was to support the existence of a

5 These participants appeared to be evenly spread between all conditions. Given that the comprehension checks were the same as in the previous two studies we have no real explanation as to why this third study suffered from this increased participant drop-out rate. One possibility is that participants were less motivated due to the decreased financial compensation in this specific study. It is worth pointing out that this selection procedure did not impact the results of this study. Analyses on the full data-set revealed that when participants were asked to make a causality judgment concurrently with their praise ratings a small but non-significant action-omission effect was present; t(700) = 1.23, p = 0.220, Cohen’s d = 0.13, whereas the effect was fully absent when participants were not asked to make the causality judgments; t(700) = -0.114, p = 0.910, Cohen’s d = -0.01.

https://osf.io/adt2r/?view_only=e2b4adbaeb9f41bfb1751fa8096b819a


null effect, an increase in sample size certainly cannot be considered self-serving. Finally, for

those participants who provided an incorrect answer on the comprehension checks for only for

one scenario, as in Study 1 and Study 2, scores on that specific trial were omitted when

calculating the average mean praise scores (7.9%).

Data analysis

ANOVA of the mean praise scores (Cronbach’s α = 0.80) revealed no Action-omission

effect; F(1,555) = 0.55, p = .459; a significant main effect of Causality salience; F(1,555) =

21.19, p < 0.001, and no significant interaction-effect; F(1,555) = 0.56, p = 0.454. Planned

contrast analyses revealed that when participants were asked to make a concurrent causality

judgment, there was a minor action-omission effect, though it failed to reach significance;

t(555) = 1.06, p = 0.289, Cohen’s d = 0.13. Crucially, when participants were not asked to

make the concurrent causality judgment the effect was completely absent; t(555) = -0.01, p =

0.995, Cohen’s d = 0.

Study 3: Discussion

The results of this third study are somewhat mixed. On the one hand, with a t-value of

zero, the results of this third study are strong confirmation that the action-omission effect on

judgments of praise is indeed absent whenever participants are not explicitly asked to consider

the causal structure of the scenarios. However, in this third study, unlike the two previous

studies, we also did not find a significant action-omission effect when causality was made

salient, weakening the evidence from the previous studies that experimentally increasing

causality salience may yield an action-omission effect in positive outcome conditions as well.

There are considerable deviations in (the significance of) the individual effects across studies.

Therefore, we ran a meta-analysis, combining the results of all three studies to more

accurately gauge the size of all the respective effects studied in the current paper. Such a

meta-analysis not only provides a comprehensive insight in the general pattern, it also makes


maximal use of the combined statistical power of the individual studies. Furthermore, because

one of our main hypotheses involves a null-effect, we also used Bayesian analysis techniques

to quantify the strength of the evidence in favor, and against our hypotheses.

Meta-analysis

Effect sizes

Cohen’s d effect sizes of the action-omission effect for each outcome and causality

salience condition for all three studies were calculated through their respective t-statistic with

the “compute.es” R package (Del Re, 2013). Meta-analytic effect sizes were then estimated in

a random-effects model with the R package “metaphor” (Viechtbauer, 2010). Random-effects

models lead to meta-analytic effect size estimates that are valid beyond the subject

populations of the studies they are based upon (Hedges, & Vevea, 1998). Because there was

considerable variability in the effects (but not the significance levels) among the three studies,

we conducted homogeneity tests to see to what extent this variability is probable under a

model of stochastic sampling or should instead be explained through some other mechanism

such as differences in study design. Cochrane’s Q tests conducted for the results of the studies

within each outcome and causality salience condition separately revealed no significant

heterogeneity among our studies (Blame, No Causality: Q(1) < 0.01, p = 0.972; Blame,

Causality: Q(1) = 0.32, p = 0.571; Praise, No Causality: Q(2) = 4.02, p = 0.134; Praise,

Causality: Q(2) = 2.80, p = 0.246). However, given the limited number of studies within this

meta-analysis, these homogeneity tests are most likely underpowered and might thus

underestimate the total heterogeneity that is present. An I2 measure describes the proportion of

total variation that is due to heterogeneity, and as such more accurately describes to what

extent the results of a meta-analysis are influenced by heterogeneity (Higgins, & Thompson,

2002). For both blame conditions I2was equal to 0.00 indicating there was no heterogeneity at

all. For the no-causality and causality praise conditions, I2 of 49.66 and 34.84, respectively


indicated the presence of a moderate amount of heterogeneity (see Higgins, & Thompson,

2002; Higgins, Thompson, Deeks, & Altman, 2003).

Table 1 provides an overview of the effect sizes in the individual studies and the meta-

analytic effect size for each condition. When asked to rate negative outcome scenarios a meta-

analytic action-omission effect robustly emerged both when causality was not made salient; d

= 0.37, p = 0.006, and when causality was made salient; d = 0.44, p < 0.001. Crucially, when

asked to rate positive outcome scenarios, the emergence of the effect was fully dependent on

our causality manipulation and was significant only when causality was made salient; d =

0.27, p = 0.014, but virtually absent when it causality was not made salient; d = 0.05, p =

0.697. These meta-analytic effect size estimates not only can inform future research on the

size of the standard action-omission effect on judgments of blame, but also clearly

demonstrate the role of causality salience on action-omission effect. In particular, whereas

causality salience only somewhat strengthens an already considerable standard action-

omission effect for judgments of blame, it plays a far more crucial role in judgments of praise

by triggering an action-omission effect that is absent in default conditions for such judgments.

Bayesian analyses

Finally, we also performed Bayesian analyses. It is well known that standard

frequentist techniques can reject null hypotheses but cannot provide straightforward evidence

to accept the null hypothesis. Because a key aim of the current paper is to argue in favor of a

specific null effect, a Bayesian analysis estimating Bayes factors can provide more

appropriate information. Bayes factors weight the evidence in favor of the two competing

hypotheses (the null hypothesis and the alternative hypothesis). Typically this involves

calculating how likely the obtained data is given the alternative hypothesis relative to the

likelihood of the data given the null hypothesis. As such, Bayes factors for the alternative

hypothesis smaller than 1 mean that the data is more likely under the null hypothesis while


Bayes factors larger than 1 designate that the data is more likely under the alternative

hypothesis. Because Bayes factors quantify the likelihood under both hypotheses they also

indicate just how likely each of the two competing hypotheses are. For instance, a Bayes

factor of 3 for the alternative hypothesis means that the data is three times more likely to have

occurred under the alternative hypothesis than it is under the null hypothesis. To help with

interpretation, Jeffreys (1961) and Wetzelfs and Wagenmakers (2012) have suggested that

Bayes Factors spanning from one to three can be interpreted as providing ‘anecdotal evidence’

in favor of the tested hypothesis (either H0 or Ha), whereas Bayes Factors higher than three

provide “substantial evidence”.

A crucial element in the calculation of Bayes Factors are the specific distributional

assumptions one makes for both H0 and Ha. Typically, one models observed effects as

independent and identically distributed random variables with:

yi ∼ Normal(µ, σ2), i = 1, …, N

To calculate Bayes factors, the parameters µ and σ must be specified for both hypotheses.

The null model assumes that there is no effect, and thus states that μ0=0. The

alternative model on the other hand does assume that there is an effect and as such requires us

to specify a distribution of likely µ under the alternative hypothesis. Rouder, Speckman, Sun,

Morey and Iverson (2009) recommend using a minimally informative distribution function

that has more probability mass in the lower range of possible effect sizes than in the higher

ranges.6 By reparametrizing the problem in terms of effect size δ (with δ = µ/σ) we can use a

Cauchy distribution as a minimally informative prior for the possible effect sizes under Ha

6 Possible alternative approaches would be to specify a point estimate or a different type of distribution for µa. A point estimate requires firm prior knowledge about the size of the effect. However, for the size of the action-omission effect on judgments of praise there is no prior knowledge (from previous studies) and using the estimates obtained in the current study would be a circular approach. Hence, a point estimate approach is not a valid option. With respect to using a different type of distribution, one could suggest using a arbitrarily diffuse function that reflects that all µa (i.e., those in the normally expected range as well as impossibly large effects) are equally likely, but such an approach unwarrantedly favors H0. Therefore, the solution we chose is to use a minimally informative distribution function that has most of its probability mass wherever one expects to find effects. This approach is a well-established compromise between the ‘objective’ diffuse approach and the ‘subjective’ point estimate approach (see Bayarri, & Garcia-Donato, 2007).


(Jeffreys, 1961; see also Johnson, Kotz, and Balakrishnan, 1994). The Cauchy prior can be

scaled by a single scale factor that determines where most of its probability mass is located

and as such in this specific context, to what extent one expects to find either small, medium or

large effects.

Finally, a prior distribution needs to be specified for σ. Fortunately, σ is the same under

both the alternative and the null model and because a Bayes Factor is a ratio of likelihoods,

the influence of sigma cancels out. A high σ lowers the likelihood of the observed data under

H0 just as much as it lowers the likelihood of the observed data under Ha. The standard choice

is to use a non-informative Jeffreys prior as the prior on variance (Jeffreys, 1961). This

specific combination of priors is known as the JZS prior (Bayarri, & Garcia-Donato, 2007).

Bayes Factors were calculated for each of the action-omission effects as per the

procedure outlined in Rouder et al. (2009) with the “BayesFactor” package in R (Morey, &

Rouder, 2015). A scale factor of √22

for the Cauchy prior of the standardized effect was

applied as per Morey, Rouder, Pratte, and Speckman (2011) because we expected to find

small to medium effects. The individual Bayes factors were then combined into a Bayesian

meta-analysis following the procedure of Rouder and Morey (2011). These Bayesian analyses

further substantiate the results of our frequentist meta-analysis. For the action-omission effect

on judgments of blame in the standard condition (i.e., without causality salience), these

analysis revealed a Bayes factor of BFa = 5.78, suggesting substantial evidence in favor the

alternative hypothesis. When blame was rated under causality salience, the Bayes Factor

increased to BFa = 27.22, suggesting even stronger evidence for the action-omission effect.

With respect to judgments of praise, these analysis also revealed substantial evidence for the

existence of an action-omission effect on judgments of praise whenever causality was made

salient at a Bayes factor of BFa = 5.24. Perhaps most crucially however, the Bayes Factor for

the action-omission effect (i.e., the alternative hypothesis) on judgments of praise without the


causality manipulation was a mere BFa = 0.13, or put differently, the Bayes Factor favoring

the null hypothesis was BF0 = 7.96, suggesting substantial evidence for the null hypothesis.

These Bayesian analyses are especially helpful because they clearly demonstrate that there is

substantial evidence for each of our hypotheses, with the evidence for the hypothesized null-

effect in standard conditions of praise judgment being even stronger than the evidence for the

well-established action-omission effect on judgments of blame in standard conditions. .

General Discussion

Across three studies we tested the action-omission effect for judgments of blame and

praise on negative and positive outcomes respectively. Though some variations in the results

occurred across the individual studies, a clear pattern materialized and was verified through a

meta-analysis of our results. As hypothesized, an action-omission effect consistently emerged

for judgments of blame in negative outcome situations, but not on judgments of praise in

positive outcome situations, at least when these judgments were made without explicitly

triggering causal reasoning. This indicates that in default situations, actively pursuing good is

not deemed more worthy of acclaim than merely allowing good to happen. While this is in

line with previous studies that have highlighted other asymmetries between judgments of

blame and praise (e.g., Pizarro, Uhlmann & Salovey, 2003), this novel result with respect to

the action-omission effect is striking.

Furthermore, it is worth mentioning that in previous studies (focusing on negative

outcomes only) the target in the action scenarios was usually the main cause of the harmful

event whereas in the omission scenarios the target was generally a bystander and not the main

cause of the event (as per killing versus letting die). In the present studies, we explicitly aimed

to avoid this potential confound so we designed our scenarios in such a way that the target to

be judged had equal “status” in the action and the omission condition. That is, in both

conditions, he could easily influence the outcome but was not the original cause of the event.


Importantly we still observed a consistent action-omission effect for negative outcomes in our

studies demonstrating that effect is not merely due to objective actor-bystander status

differences, something that could not be assured in previous research. By excluding this

alternative interpretation, the present research hence also extends our understanding of the

nature of the standard action-omission effect on judgments of blame.

Finally, the current study suggests a mechanism that may explain why the action-

omission effect on judgments of praise does not occur. It appears that the effect is largely

dependent on causal attribution processes, and whereas negative outcomes seem to trigger

causality appraisals by default, this is not (or less) the case for positive outcomes. However,

by making the causal structure of the scenarios salient to our participants, an action-omission

effect could be obtained for positive outcomes as well. Interestingly, no matter the outcome

condition, the action-omission effect was increased by explicitly reminding participants of the

causal structure of the scenarios, though this increase was rather small in the negative

outcome conditions compared to the positive outcome conditions. This further corroborates

the idea that the action-omission effect is a side-effect of our causal attribution mechanisms

and is something scholars should keep in mind in future studies, as mere assessment of

causality appraisal may inflate the action-omission effect when already present, or induce it

where otherwise absent.

Acknowledgements

The authors would like to thank the University College West Flanders (Howest) for

their help with the data collection of the first study. In this regard, special thanks go to Van

Eynde Lien.


References

Baron, J., & Ritov, I. (2004). Omission Bias, individual differences, and normality.

Organizational Behavior and Human Decision Processes, 94, 74–85. doi:

10.1016/j.obhdp.2004.03.003

Baumeister, R. F., Bratslavsky, E., Finkenauer, C., & Vohs, K. D. (2001). Bad is stronger than

good. Review of general psychology, 5(4), 323. doi: 10.1037/1089-2680.5.4.323

Bayarri, M. J., & Garcia-Donato, G. (2007). Extending conventional priors for testing general

hypotheses in linear models. Biometrika, 94 , 135-152.

Bohner, G., Bless, H., Schwarz, N., & Strack, F. (1988). What triggers causal attributions?

The impact of valence and subjective probability. European Journal of Social

Psychology, 18(4), 335-345. doi: 10.1002/ejsp.2420180404

Champely, S., Ekstrom, C., Dalgaard, P., Gill, J., & De Rosario, H. (2015), ‘pwr’. Basic

functions for power analysis. R package version, 1.1-2.

Cohen, J. (1992). A power primer. Psychological Bulletin, 112(1), 155. doi: 10.1037/0033-

2909.112.1.155

Cushman, F. A., & Young, L. (2011). Patterns of moral judgment derive from nonmoral

psychological representations. Cognitive Science, 35(6), 1052-1075. doi:

10.1111/j.1551-6709.2010.01167.x

Cushman, F. A., Young, L., & Hauser, M. D. (2006). The role of conscious reasoning and

intuition in moral judgment: testing three principles of harm. Psychological Science,

17(12), 1082–1089. doi: 10.1111/j.1467-9280.2006.01834.x

Cushman, F., & Greene, J. D. (2012). Finding faults: how moral dilemmas illuminate

cognitive structure. Social Neuroscience, 7(3), 269-279. doi:

10.1080/17470919.2011.614000

Del Re, A.C. (2013) Compute.es: compute effect sizes. R package version 0.2-2.


Descioli P., & Kurzban, R. (2009). Mysteries of morality. Cognition, 112(2), 281-299. doi:

10.1016/j.cognition.2009.05.008

DeScioli P., & Kurzban, R. (2013). A solution to the mysteries of morality. Psychological

Bulletin, 139, 477-496. doi: 10.1037/a0029065

DeScioli P., Bruening R., & Kurzban, R. (2011). The omission effect in moral cognition:

toward a functional explanation. Evolution and Human Behavior, 32, 204-215. doi:

10.1016/j.evolhumbehav.2011.01.003

Hedges, L. V., & Vevea, J. L. (1998). Fixed-and random-effects models in meta-analysis.

Psychological Methods, 3(4), 486.

Higgins, J. P., & Thompson, S. G. (2002). Quantifying heterogeneity in a meta-analysis.

Statistics in Medicine, 21(11), 1539-1558.

Higgins, J. P., Thompson, S. G., Deeks, J. J., & Altman, D. G. (2003). Measuring

inconsistency in meta-analyses. British Medical Journal, 327(7414), 557.

Jeffreys, H. (1961). Theory of probability (3rd ed.). Oxford: Oxford University Press,

Clarendon Press.

Johnson, N. L., Kotz, S., & Balakrishnan, N. (1994). Continuous univariate distributions.

New York, NY: Wiley.

Kahneman, D., & Tversky, A. (1984). Choices, values, and frames. American

Psychologist, 39(4), 341. doi: 10.1037/0003-066X.39.4.341

Kordes-de Vaal, J. H. (1996). Intention and the omission bias: omissions perceived as

nondecisions. Acta Psychologica, 93(1), 161-172. doi: 10.1016/0001-6918(96)00027-

3

Morey, R.D., & Rouder, J.N. (2015). BayesFactor: computation of Bayes factors for common

designs. R package version 0.9.12-2.


Morey, R. D., Rouder, J. N., Pratte, M. S., & Speckman, P. L. (2011). Using MCMC chain

outputs to efficiently estimate Bayes factors. Journal of Mathematical Psychology, 55,

368-378.

Oppenheimer, D. M., Meyvis T., & Davidenko N. (2009). Instructional manipulation checks:

detecting satisficing to increase statistical power. Journal of Experimental Social

Psychology, 45, 867-872. doi: 10.1016/j.jesp.2009.03.009

Paolacci, G., Chandler, J., & Ipeirotis, P.G. (2010). Running experiments on amazon

mechanical turk. Judgment and Decision Making, 5(5), 411-419. Retrieved from

http://hdl.handle.net/1765/31983

Peeters, G., & Czapinsky, J. (1990). Positive-negative asymmetry in evaluations: The

distincion between affective and informational negativity effects. European Review of

Social Psychology, 1(1), 33-60. doi: 10.1080/14792779108401856

Pizarro, D., Uhlmann, E., & Salovey, P. (2003). Asymmetry in judgments of moral blame and

praise: the role of perceived metadesires. Psychological Science, 14, 267-272. doi:

10.1111/1467-9280.03433

Rand, D. G. (2012). The promise of mechanical turk: how online labor markets can help

theorists run behavioral experiments. Journal of Theoretical Biology, 299, 172-179.

doi: 10.1073/pnas.1108243108

Ritov, I. I., & Baron, J. (1999). Protected values and omission bias. Organizational Behavior

and Human Decision Processes, 79(2), 79–94. doi: 10.1006/obhd.1999.2839

Roese, N. J., & Olson, J. M. (1997). Counterfactual thinking: the intersection of affect and

function. Advances in Experimental Social Psychology, 29, 1-59. doi: 10.1016/S0065-

2601(08)60015-5

Rouder, J. N., & Morey, R. D. (2011). A Bayes factor meta-analysis of Bem’s ESP claim.

Psychonomic Bulletin and Review, 18, 682-689. doi: 10.3758/s13423-011-0088-7


Rouder, J. N., Speckman, P. L., Sun, D., Morey, R. D., & Iverson, G. (2009). Bayesian t tests

for accepting and rejecting the null hypothesis. Psychonomic Bulletin & Review, 16,

225-237. doi: 10.3758/PBR.16.2.225

Ross, H. S., & Den Bak-Lammers, I. M. (1998). Consistency and change in children’s tattling

on their siblings: children’s perspectives on the moral rules and procedures of family

life. Social Development, 7(3), 275–300. doi: 10.1111/1467-9507.00068

Rozin, P., & Royzman, E. B. (2001). Negativity bias, negativity dominance, and

contagion. Personality and Social Psychology Review, 5(4), 296-320. doi:

10.1207/S15327957PSPR0504_2

Spranca, M., Minsk, E., & Baron, J. (1991). Omission and commission in judgment and

choice. Journal of Experimental Social Psychology, 27(1), 76–105. doi: 10.1016/0022-

1031(91)90011-T

Viechtbauer, W. (2010). Conducting meta-analyses in R with the metafor package. Journal of

Statistical Software, 36(3), 1-48.

Wiessner, P. (2005). Norm enforcement among the Ju/’hoansi Bushmen: a case of strong

reciprocity? Human Nature, 16(2), 115-145. doi: 10.1007/s12110-005-1000-9

abstract - ghent university web viewusing mcmc chain outputs to efficiently estimate bayes factors....

Documents