Temporal variability in moral value judgement

Alexandra Surdina (a.surdina@warwick.ac.uk)Department of Psychology, University of Warwick

Coventry CV4 7AL, United Kingdom

Adam Sanborn (a.n.sanborn@warwick.ac.uk)Department of Psychology, University of Warwick

Coventry CV4 7AL, United Kingdom

Abstract

Moral judgments are known to change in response tochanges in external conditions. But how variable aremoral judgments over time in the absence of environ-mental variation? The moral domain has been describedin terms of five moral foundations, categories that ap-pear to capture moral judgment across cultures. We ex-amined the temporal consistency of repeated responsesto the moral foundations questionnaire over short timeperiods, fitted a set of mixed effects models to the dataand compared them. We found correlations betweenchanges in participant responses for different founda-tions over time, suggesting a structure with at leasttwo underlying stochastic processes: one for moral judg-ments involving harm and fairness, and another formoral judgments related to loyalty, authority, and pu-rity.

Keywords: morality, moral foundations theory, consis-tency, variability

Introduction

Morality is a vital part of who we are. A person’s moralbeliefs are tied into their identity (Aquino & Reed II,2002; Aquino et al., 2009) – humans believe that if theirmoral values changed, they would change (Heiphetz etal., 2016). Are people’s intuitions about this correct?Are our moral values consistent over time?

Since moral beliefs tend to be associated with a per-son’s sense of identity, we should expect people’s under-lying moral values to largely endure over short time pe-riods. Yet, there have been many recent explorations ofmoral inconsistency. These have included manipulationsof two types – manipulations of response timing, or ma-nipulations by exposure to new information or decisions.In terms of timing, we now know that time-limited de-cisions appear to be more altruistic (Rand et al., 2012)and that choices can be influenced by forcing decisionsat a specific point in time (Pärnamets et al., 2015), indi-cating that the actual decision outcome is time-sensitive.Regarding information or decisions, dishonest behaviourincreases future dishonesty (Garrett et al., 2016; Engel-mann & Fehr, 2016). A morally good action makes asubsequent morally bad action more appealing and viceversa, effects known as moral cleansing and moral licens-ing (Merritt et al., 2010; Sachdeva et al., 2009). Expo-sure to a moral dilemma leads to belief revision in moraldecisions that persists for multiple hours (Horne et al.,2015).

The fact that changes in external circumstances caninfluence the outcomes of moral decisions is hardly sur-prising assuming morality evolved as an adaptive strat-egy (Machery & Mallon, 2010). Likewise, viewing moraljudgment as a decision process, we would expect theeffects of changed response timing on general decision-making (McClelland, 1979; Usher & McClelland, 2001)to transfer into the moral domain. But in the absenseof such manipulations, are our moral judgments funda-mentally noisy? Outside of the moral domain, there isevidence in decision making research that people’s de-cisions vary stochastically even in cases where externalconditions remain constant (Mosteller & Nogee, 1951).We are interested in exploring whether there is a cor-responding moral variability beyond the actual decisionprocess: are our moral values different from moment tomoment, even in the absence of new information or ma-nipulations of response timing?

Moral Foundations Theory (MFT) provides a way tolook at this. It is based on a dominant model of morality,the social intuitionist model, according to which moralchoices are made primarily intuitively and then justifiedpost hoc (Haidt, 2001). MFT maps out the moral do-main in terms of six fundamental hidden parameters thatappear to capture an individual’s moral judgment (Gra-ham et al., 2009), enabling us to distinguish betweenconservative and liberal political profiles on the basis ofan agent’s foundation weights. This idea that there arefoundational categories that guide intuitive moral judge-ment has the potential to explain people’s tendency todisagree on moral issues, and predict future moral judge-ment based on the individual scores. If we can find asystematic structure in the stochastic changes of differ-ent foundation scores beyond merely a layer of noise, thiswould point towards moral variability, rather than justmotor variability or variability in how the response scaleis used.

In line with the aforementioned results indicating tem-poral consistency, moral foundation scores appear sta-ble over longer time periods; Graham et al. (2011)tested participants again after approximately a monthand found that their moral foundation scores exhibitedtest-retest reliability. Yet, effects such as moral licensingand moral cleansing – where the outcome of an indi-vidual’s moral decision influences subsequent moral de-

3285

cisions, even decisions made by others in their ingroup(Kouchaki, 2011), over the course of single experimen-tal sessions and thus shorter timescales – suggest thepossibility of an interaction between moral foundations.Moreover, the list of known moral foundations is likelyincomplete – a view shared by moral foundations theo-rists (Haidt & Joseph, 2011).

Viewing moral decisions as a sampling process from adistribution that represents an agent’s moral values, wecan use the framework provided by MFT to investigatehidden parameters which predict an individual’s moralvariability. Conversely, observing within-subject vari-ability over time can help us understand to which extentindividual moral variability reflects between-individualvariability that has been used to support the existenceof MFT (Graham et al., 2011). Are we all sometimesa little bit more conservative and sometimes a little bitmore liberal in our moral judgments and values?

In this paper, we aim to discuss the extent to whichrandomness plays a role in moral judgment over time bycollecting responses to the moral foundations question-naire delivered repeatedly. We subsequently fit a set ofmodels to the data and compare them. If the variabilitywe observe stems merely from randomness in the deci-sion process, we expect variation in individual responsesto be explained by a single noise-generating process. Wefind evidence for at least two separate stochastic pro-cesses associated with different sets of moral foundations,indicating the existence of inherent variability in moralvalues.

Method

Participants

The participant pool consisted of 80 psychology under-graduate students (mean age 19 years, 90% female).14 participants were excluded from the analysis due towrong responses on the two ‘catch’ trials, as done byGraham et al. (2011).

Materials

The original moral foundations questionnaire (MFQ30)asks participants to respond using a 1–6 scale; to enhanceprecision and avoid subjects simply recalling previousanswers, the participants in our task had to use a sliderbar to indicate their responses instead:

not at all a lot

In addition, our version of the questionnaire containedfour further questions (see Figure 1). Those were cho-sen so as not to correspond in any obvious way to thefive foundations measured in the MFQ30, nor to the re-cent addition of the liberty foundation (Graham et al.,2012; Haidt, 2012). We added these questions becausewe wanted the same number of presumably neutral tri-als as the number of foundation-related questions – the

MFQ30 includes six question for each foundation butonly two neutral ‘catch’ items.

When you decide whethersomething is right or wrong,to what extent is the follow-ing consideration relevant toyour thinking?

Please read the following sentenceand indicate your agreement or dis-agreement:

Harm:

1. Whether or not someonesuffered emotionally

2. Whether or not someonecared for someone weakor vulnerable

3. Whether or not someonewas cruel

4. Compassion for those who aresuffering is the most crucialvirtue.

5. One of the worst things a personcould do is hurt a defenseless an-imal.

6. It can never be right to kill a hu-man being.

Fairness:

1. Whether or not somepeople were treated dif-ferently than others

2. Whether or not someoneacted unfairly

3. Whether or not someonewas denied his or herrights

4. When the government makeslaws, the number one principleshould be ensuring that everyoneis treated fairly.

5. Justice is the most important re-quirement for a society.

6. I think it’s morally wrong thatrich children inherit a lot ofmoney while poor children in-herit nothing.

Loyalty:

1. Whether or not some-one’s action showed lovefor his or her country

2. Whether or not someonedid something to betrayhis or her group

3. Whether or not someoneshowed a lack of loyalty

4. I am proud of my country’s his-tory.

5. People should be loyal to theirfamily members, even when theyhave done something wrong.

6. It is more important to be ateam player than to express one-self.

Authority:

1. Whether or not someoneshowed a lack of respectfor authority

2. Whether or not someoneconformed to the tradi-tions of society

3. Whether or not an actioncaused chaos or disorder

4. Respect for authority is some-thing all children need to learn.

5. Men and women each have dif-ferent roles to play in society.

6. If I were a soldier and disagreedwith my commanding officer’sorders, I would obey anyway be-cause that is my duty.

Purity:

1. Whether or not someoneviolated standards of pu-rity and decency

2. Whether or not someonedid something disgusting

3. Whether or not someoneacted in a way that Godwould approve of

4. People should not do things thatare disgusting, even if no one isharmed.

5. I would call some acts wrong onthe grounds that they are unnat-ural.

6. Chastity is an important andvaluable virtue.

Neutral:

1. Whether or not someonewas good at math

2. Whether or not someonetold the truth (*)

3. Whether or not someonemade a smart decision (*)

4. It is better to do good than todo bad.

5. If one’s children live a happy life,it is better to have children thannot to have children. (*)

6. Destroying beautiful things thattook long to create is worse thandestroying things that took lesstime. (*)

Figure 1: Moral foundations questionnaire. Questionsadded by us are marked with (*).

Procedure

The questionnaire was presented six times in randomisedorder, with a word search task before the last two tri-als. In each trial, one of the two question types wasdisplayed (see Figure 1, left and right side, respectively),along with one of the statements for that question type.Randomisation was implemented so that each statementwas shown to the participant exactly once in each block:The set of questions within each block was shuffled, andpresented within the block in randomised order, so noregular pattern in the order of foundations would occur.

3286

After four blocks, a word search task1 was shown for6 minutes to provide a timed break2: Participants hadto find and mark words in a 18x18 letter square filledwith a selection of words and random letters, based onthe WordFind.js library (Scheidel, 2012). With the ex-ception of the timed word search task, participants pro-vided responses at their own pace. The experiment tookapproximately 20-25 minutes to complete.

Results

Since participant responses are indicated using sliderbars, foundation scores change between the blocks (par-ticipants will be unable to recall the exact position ofthe slider for previous trials). But beyond the expectedvariation resulting from differences in participant’s slideroperation accuracy, is there a relationship between thesechanges in different moral foundation scores?

Means

As found by Graham et al. (2011), we anticipated andfound our psychology undergraduate subject pool in theUK to remain largely at the liberal end of the U.S. polit-ical spectrum. Welch’s t-test shows that the differencesbetween the means for harm and fairness (p=.16) andfor loyalty and authority (p=.44) are not significant. Allother pairs of means indeed differ significantly (p

encounter, there would be less need to for caution aboutnew options which are more or less morally upsettingthan the previous maximum or minimum, respectively.

We computed residual slider values by subtracting thewithin-subject mean for each foundation from the slidervalues for each trial. The two hypotheses above can berephrased as: The slider residual variance for each par-ticipant and block decreases as a reflection of the increasein certainty; and the average absolute residual value in-creases over time as a result of the decision drifting to-wards the extremes.

Neutral Purity

Harm Loyalty

Authority Fairness

1 2 3 4 5 6 1 2 3 4 5 6

0255075

100

0255075

100

0255075

100

Question number

Slid

er v

alue

Within−foundation changes

Figure 4: Changes over time, by foundation. The coloursrepresent the different blocks.

In fact, we found no significant effect of block num-ber on foundation variance: Again, an ANOVA onlyyields significant results for the variance hypothesis(F(5,325)=18.71, p

pled from the same distribution for each foundation (ran-dom noise model). The models for the slider residual yi jklof Participant i in Block j for a question or statement lrelating to Foundation k are:

yi jkl = ui j + ui jk + εi jkl , (M1-M3)

with ui j ∼ N (0,σ), and

ui jk = 0 (M1)ui j1ui j2ui j3ui j4ui j5

∼ N0,

σ11 σ12 σ13 σ14 σ15σ21 σ22 σ23 σ24 σ25σ31 σ32 σ33 σ34 σ35σ41 σ42 σ43 σ44 σ45σ51 σ52 σ53 σ54 σ55

(M2)

ui j1ui j2ui j3ui j4ui j5

∼ N0,

σ11 0 0 0 00 σ22 0 0 00 0 σ33 0 00 0 0 σ44 00 0 0 0 σ55

(M3)

The model M2 (χ2(10) = 26.98, p = .003) differs signifi-cantly from the baseline model M1. Comparing the mod-els M1-M3 to each other suggests that M2 (BIC=71058)has a lower BIC value than M1 (BIC=70960) and M3(BIC=70993). M2, which has a full covariance matrix,shows an interesting pattern of dependencies between thedifferent foundation types:

Foundation Harm Fair Loya AuthFair 1Loya -0.86 -0.86Auth -0.95 -0.95 0.95Puri -0.7 -0.7 0.64 0.80

Responses for harm and fairness appear to be posi-tively correlated with each other and negatively corre-lated with responses for the other foundations, and viceversa. This would be less surprising if it was merely cap-turing a between-participant relationship between foun-dation scores. Note however that this model describesthe slider residuals which add up to zero for each foun-dation and participant – yet, this result suggests thatparticipants who drag the slider bar a bit further to theright for harm-related questions than in the last blockwill do a similar thing with the fairness-question slider,but the opposite with sliders on loyalty, authority andpurity trials.

Is there some overlap between which property ofmorality harm and fairness on the one hand and loyalty,authority and purity on the other hand are measuring?Since the mean foundation scores for harm and fairness,and the scores for loyalty, authority, and purity seemsimilar to each other (see Figure 2), we introduced alter-native models that only distinguish between these twogroups instead of the individual foundations.

To find out if we could confirm the five-dimensionalmoral foundations structure, we fitted a set of linear

mixed effects models to the data. As an alternative,we dummy-coded two foundation types (the ‘individual-ising’ foundations harm and fairness and the ‘binding’foundations loyalty, authority, and purity (Graham etal., 2009)). Again, we fitted a full covariance modeland a diagonal covariance model to the data, adding thetwo models below to our list of candidate models. Theyare describing the slider residual yi jml of Participant i inBlock j for a question l of Foundation type m:

yi jml = ui j + ui jm + εi jml , (M4-M5)

with(ui j1ui j2

)∼ N

(0,[

σ11 σ12σ21 σ22

])(M4)(

ui j1ui j2

)∼ N

(0,[

σ11 00 σ22

])(M5)

We find that out of these, the model M5 differs signifi-cantly from the baseline model (χ2(2) = 27.85, p < .001).Comparing M4 (BIC=70960) and M5 (BIC=70951) tothe models above suggests that M5 is preferable to M2and M4. Thus, it appears that from a model comparisonperspective, the main distinction in the moral founda-tion framework lies in the two different foundation typesrather than the individual foundations, and that at thislevel of description, between-foundation correlations donot play a prominent role.

DiscussionWe found that people showed moral variability even inthe absence of new information or time pressure. Thismoral variability is distinguishable from response vari-ability because we found two random processes that wereassociated with different sets of moral foundations. Theevidence for MFT is based on an analysis of between-individual responses to the MFQ (Graham et al., 2011),and much of this may actually be due to the within-individual variability that we have found. This within-individual variability may also be what allows timinginterventions to have an effect (Pärnamets et al., 2015),and might potentially even allow to influence the out-comes of value-related decisions (such as election re-sults).

While for our dataset a simpler two-type model waspreferable to the more complex model including fivemoral foundations, we hesitate to draw general conclu-sions about the number of moral foundations due to thesmall size and relative cultural homogeneity of our sub-ject pool. Yet, our brief glimpse at candidates for ad-ditional foundations suggests the possibility of a widerunderlying structure of which MFT has captured but apart.

A common criticism of MFT is that the known moralfoundations are unlikely to capture moral judgment in itsentirety (Suhler & Churchland, 2011). We had expected

3289

our added questions to be rated similarly irrelevant tomorality as the more conservative moral foundations inour liberal subject pool. Somewhat surprisingly, the re-sponses to our added, ‘neutral’ foundation appear to beless neutral overall. We chose the four additional state-ments in the neutral foundation because we suspectedthat they might turn out to be morally relevant. Fig-ure 3 suggests that questions 2 and 5 in particular (seeFigures 1 and 4) indeed resonate with our participants’values. While the act of lying may arguably be relatedto the purity scale, it is remarkably more morally rele-vant than any of the purity questions. This particularlyutilitarian view on having children also appears to lieoutside of the given scales.

Interesting open questions remain that reach beyondrefining and expanding MFT. While we observe a rangeof scores for different moral foundations, we do not yetunderstand the actual decision process: How are differ-ent moral values integrated in a decision between optionsthat are morally relevant for more than one moral foun-dation, or options that are uncertain? Which impactdoes moral variability have on the kinds of moral deci-sions we face every day?

AcknowledgmentsThis research was supported by a Leverhulme Trust Doc-toral Scholarship within the Bridges programme at theUniversity of Warwick.

ReferencesAquino, K., Freeman, D., Reed II, A., Lim, V. K., &

Felps, W. (2009). Testing a social-cognitive model ofmoral behavior: the interactive influence of situationsand moral identity centrality. Journal of personalityand social psychology , 97 (1), 123.

Aquino, K., & Reed II, A. (2002). The self-importanceof moral identity. Journal of personality and socialpsychology , 83 (6), 1423.

Engelmann, J. B., & Fehr, E. (2016). The slippery slopeof dishonesty. Nature Neuroscience, 19 (12), 1543–1544.

Garrett, N., Lazzaro, S. C., Ariely, D., & Sharot, T.(2016). The brain adapts to dishonesty. Nature Neu-roscience.

Graham, J., Haidt, J., Koleva, S., Motyl, M., Iyer, R.,Wojcik, S. P., & Ditto, P. H. (2012). Moral foun-dations theory: The pragmatic validity of moral plu-ralism. Advances in Experimental Social Psychology,Forthcoming .

Graham, J., Haidt, J., & Nosek, B. A. (2009). Liberalsand conservatives rely on different sets of moral foun-dations. Journal of personality and social psychology ,96 (5), 1029.

Graham, J., Nosek, B. A., Haidt, J., Iyer, R., Koleva, S.,& Ditto, P. H. (2011). Mapping the moral domain.

Journal of personality and social psychology , 101 (2),366.

Haidt, J. (2001). The emotional dog and its rationaltail: a social intuitionist approach to moral judgment.Psychological review , 108 (4), 814.

Haidt, J. (2012). The righteous mind: Why good peopleare divided by politics and religion. New York: Pan-theon.

Haidt, J., & Joseph, C. (2011). How moral foundationstheory succeeded in building on sand: A response toSuhler and Churchland. Journal of Cognitive Neuro-science, 23 (9), 2117–2122.

Heiphetz, L., Strohminger, N., & Young, L. L. (2016).The role of moral beliefs, memories, and preferencesin representations of identity. Cognitive science.

Horne, Z., Powell, D., & Hummel, J. (2015). A singlecounterexample leads to moral belief revision. Cogni-tive science, 39 (8), 1950–1964.

Kouchaki, M. (2011). Vicarious moral licensing: theinfluence of others’ past moral actions on moral be-havior. Journal of personality and social psychology ,101 (4), 702.

Machery, E., & Mallon, R. (2010). Evolution of morality.The moral psychology handbook , 3–46.

McClelland, J. L. (1979). On the time relations of mentalprocesses: An examination of systems of processes incascade. Psychological review , 86 (4), 287.

Merritt, A. C., Effron, D. A., & Monin, B. (2010). Moralself-licensing: When being good frees us to be bad.Social and personality psychology compass, 4 (5), 344–357.

Mosteller, F., & Nogee, P. (1951). An experimentalmeasurement of utility. Journal of Political Economy ,59 (5), 371–404.

Pärnamets, P., Johansson, P., Hall, L., Balkenius, C.,Spivey, M. J., & Richardson, D. C. (2015). Bias-ing moral decisions by exploiting the dynamics of eyegaze. Proceedings of the National Academy of Sci-ences, 112 (13), 4170–4175.

Rand, D. G., Greene, J. D., & Nowak, M. A. (2012).Spontaneous giving and calculated greed. Nature,489 (7416), 427–430.

Sachdeva, S., Iliev, R., & Medin, D. L. (2009). Sinningsaints and saintly sinners the paradox of moral self-regulation. Psychological science, 20 (4), 523–528.

Scheidel, B. (2012). Wordfind. https://github.com/bunkat/wordfind.

Suhler, C. L., & Churchland, P. (2011). Can innate,modular ”foundations” explain morality? challengesfor haidt’s moral foundations theory. Journal of cog-nitive neuroscience, 23 (9), 2103–2116.

Usher, M., & McClelland, J. L. (2001). The time courseof perceptual choice: the leaky, competing accumula-tor model. Psychological review , 108 (3), 550.

3290