Do People Engage in Motivated Reasoning to Think the World Is a Good Place for Others?
DDo People Engage in Motivated Reasoning to Think theWorld Is a Good Place for Others? ∗Michael Thaler † October 2020
Abstract
Motivated reasoning is a bias in inference in which people distort their updatingprocess in the direction of more attractive beliefs. Prior work has shown how motivatedreasoning leads people to form overly “positive” beliefs that also serve to bolster one’sself-image in domains such as intelligence, prosociality, and politics. In this paper,I study whether positivity-motivated reasoning persists in domains where self-imagedoes not play a role. In particular, I analyze whether individuals motivatedly reason tothink that the world is a better place for others. Building off of the design from Thaler(2020), I conduct a large online experiment to test for positivity-motivated reasoningon issues such as cancer survival rates, others’ happiness, and infant mortality. Ifind no systematic evidence for positivity-motivated or negativity-motivated reasoning,and can rule out modest effects. Positivity is not a sufficient condition for motivatedreasoning. ∗ This paper was previously circulated with the title: “The Limits of Motivated Reasoning When Self-Image Is Not at Stake.”I would like to thank Alberto Alesina, Roland Benabou, Christine Exley, David Laibson, Matthew Rabin, and numerous seminarparticipants for helpful comments. I am grateful for funding support from the Harvard Business School Research Fellowshipand the Eric M. Mindich Research Fund for the Foundations of Human Behavior. † Princeton University. Email: [email protected]. a r X i v : . [ ec on . GN ] D ec Introduction
There is a common intuition in economics that people find it more attractive to believethat they are in a “good” state of the world than in a “bad” state of the world, andthat this can lead them to form beliefs that are directionally distorted in favor ofgood states. However, tests of such over-optimism focus on states where self-image isat stake, such as about one’s future prospects, one’s ability, one’s altruism, or one’spolitics (e.g. Weinstein 1980; Mobius et al. 2014; Eil and Rao 2011; Thaler 2020).This paper argues that motivated reasoning — the distortion of new informationin the direction of more attractive beliefs — is not solely about “good” and “bad”states. I run an experiment to explore how people make inferences about states of theworld that are good or bad for others, and in which self-image does not play a role: positivity-motivated reasoning . I find evidence that this form of positivity-motivatedreasoning does not play much of a role in inference.In a large online experiment, I test whether people engage in positivity-motivatedreasoning on five topics: the survival rate of children with leukemia, global povertyrates, annual deaths in armed conflict, others’ happiness levels, and infant mortalityrates. The topics are in Table 1 below.To identify motivated reasoning, I conduct a large online experiment that builds offof the design of Thaler (2020). That paper found evidence of motivated reasoning inpolitical and performance domains. Forms of motivated reasoning have also been foundin other self-image domains such as about one’s altruism (Exley 2015; Di Tella et al.2015), intelligence (Mobius et al. 2014; Eil and Rao 2011), and attractiveness (Eil andRao 2011). Similar theoretical work looks at ego-motivated beliefs, identity-motivated
Topic Positive Motives Negative Motives
Infant mortality Low / Decreasing High / IncreasingOthers’ reported happiness High / Increasing Low / DecreasingLeukemia survival rate for children High / Increasing Low / DecreasingGlobal poverty rate Low / Decreasing High /IncreasingDeaths in armed conflicts Low / Decreasing High / IncreasingLatitude of US Neutral NeutralTable 1: The list of topics and positivity motives in the experiment. On the computer, eachtopic is a hyperlink that links to the exact question wording in Appendix B.2 eliefs, and other beliefs about the outside world (e.g. Kunda 1990; Benabou andTirole 2002; Brunnermeier and Parker 2005; Benabou 2013). In related experimentalwork, Barron (2020) does not find evidence for motivated reasoning in domains thatinvolve experimental payoffs but are not ego-relevant.The main result in this paper is that – across several settings in which self-imageis not relevant – there is no evidence for positivity- or negativity-motivated reasoning.I show that, aggregating across these questions, even modest effects can be ruled out.In fact, by comparing the magnitude of positivity-motivated reasoning to results inThaler (2020), we can rule out an effect of positivity or negativity that is half as largeas politically-driven or performance-driven motivated reasoning. This evidence showsthat positivity, by itself, is insufficient for motivated reasoning.The second result is that there is no evidence that subjects’ current beliefs arereflective of past positivity-motivated reasoning. That is, subjects whose beliefs areoverly positive are no more likely to engage in positivity-motivated reasoning in thisexperiment. This suggests that there is limited heterogeneity in subjects’ positivity-motivated reasoning. Relatedly, there are not substantial differences in motivatedreasoning by demographic factors like gender, education, or income.The third result is that people do not expect to see evidence for positivity-motivatedreasoning, but do expect positivity to affect happiness. In a separate survey, I ask par-ticipants what they expect the direction of motivated reasoning to be about positivity,politics, and own performance. While the majority of participants expect others to en-gage in pro-party and pro-performance motivated reasoning, they are similarly likelyto expect to see positivity-motivated reasoning, negativity-motivated reasoning, or nonotable difference. Yet a clear majority of participants expect positive news to makepeople happier.Taken together, these results are consistent with the notion that motivated reason-ing is not only driven by belief-based utility. Subjects may attain higher utility bylearning that the world is good for other people, and yet not systematically distorttheir inference process in favor of these beliefs. That is, the beliefs that people findmore attractive do not necessarily make them happier. Rather, it may be limited tobelief-based utility that relates to one’s self-image.The rest of the paper proceeds as follows: Section 2 discusses the main theory andexperimental design that identifies motivated reasoning, adapted from Thaler (2020).Section 3 discusses the data. Section 4 presents the main experimental results. Sec-tion 5 discusses interpretations of the main experiment and presents survey evidenceabout what people expect about others’ behavior and utility. Section 6 concludes andproposes directions for future work. The appendices provide a table that is omittedfrom the main text, and list the exact questions and pages that subjects see. Theory and Experimental Design
The theory of motivated reasoning follows Thaler (2020). Further details are in thatpaper. When a motivated-reasoning agent infers about the probability that an event istrue ( T ) or false ( ¬ T ), with prior P ( T ), the agent forms his posterior by incorporatingprior, likelihood, and a motivated beliefs term: P ( T | x ) | {z } posterior ∝ P ( T ) | {z } prior · P ( x | T ) | {z } likelihood · M ( T ) ϕ ( x ) | {z } mot. reasoning , We take log odds ratios to attain the additive form:logit P ( T | x ) = logit P ( T ) + log (cid:18) P ( x | T ) P ( x |¬ T ) (cid:19) + ϕ ( x )( m ( T ) − m ( ¬ T )) . (1)The motivated reasoner acts as if he receives both the actual signal ( x ) and a signalwhose relative likelihood corresponds to how much he is motivated to believe the stateis T . m ( T ) : { T, ¬ T } → R is denoted the motive function. The weight put on thissignal is ϕ ( x ) ≥
0, called susceptibility . When ϕ ( x ) = 0, the agent is Bayesian; when ϕ ( x ) >
0, the agent motivatedly reasons.This paper will assume ϕ ( x ) > m ( T ) − m ( ¬ T ) from the inference process. We willbe interested in the psychology of the motive function. In this paper, either T willcorrespond to positivity (the world being a better place) and ¬ T to negativity (theworld being a worse place), or vice versa.The experiment provides people with not-obviously-uninformative signals about theveracity of news sources. To fix ideas, consider the following question, taken verbatimfrom the experiment: Acute Myeloid Leukemia (AML) is a devastating illness in which cancerouscells emerge in the bone marrow, invade the blood stream, and may spreadto the rest of the body. Tragically, hundreds to thousands of children underthe age of 15 are diagnosed with AML each year; it is one of the mostcommon cancers among children.Of children under the age of 15 who are diagnosed with AML, whatpercent survive for at least 5 years?
This is a question for which higher-valued states are more positive. The main test of otivated reasoning then involves three steps:1. Beliefs:
Subjects are asked to guess the answers to questions like the one above.Importantly, they are asked and incentivized to guess their median belief (i.e.such that they find it equally likely for the answer to be above or below theirguess).2.
News:
Subjects receive a binary message from one of two randomly-chosen newssources: True News and Fake News. The message from True News is alwayscorrect, and the message from Fake News is always incorrect. This is the main(within-subject) treatment variation.The message says either “The answer is greater than your previous guess of[previous guess].” or “The answer is less than your previous guess of [previousguess].” Note that the exact messages are different for each subject since subjectshave different guesses. These customized messages are designed so that they havethe same subjective likelihood of occurring.For the cancer question above, “greater than” corresponds to Positive newsand “less than” to Negative news.3.
Assessment:
After receiving the message, subjects assess the probability thatthe message came from True News using a scale from 0/10 to 10/10, and areincentivized to state their true belief. This news veracity assessment is the mainoutcome measure. The effect of variation in news direction on veracity assess-ments is the primary focus for much of this paper.More formally, consider an agent with prior F ( θ ) about a state in Θ. Denote by µ ≡ F − (1 /
2) the median of F ( θ ). For simplicity, we assume that F has no atom at µ and that P ( µ = θ ) = 0. That is, the agent believes that the answer has probabilityzero of being exactly equal to µ , and the true probability is indeed zero. The agent receives a source that is either from True News ( T ) or Fake News ( ¬ T ).Both report one of two binary messages G or L : “The answer θ is greater than yourmedian µ ” or “The answer θ is less than your median µ .” Prior beliefs P ( T ) are fixed,and log (cid:16) P ( G | T ) P ( G |¬ T ) (cid:17) = log (cid:16) P ( L | T ) P ( L |¬ T ) (cid:17) = 0 by definition of a median. θ > µ θ < µ True News sends
G L
Fake News sends
L G
The agent has a prior about the news source p ≡ P ( T ) that does not depend on θ ,and infers about P ( T ) given the message received. In this experiment, zero answers are correct, so the assumption appears reasonable. e can now look at how a motivated reasoner updates his beliefs about the newssource after receiving G :logit P ( T | G ) = logit P ( T ) + log (cid:18) P ( G | T ) P ( G |¬ T ) (cid:19) + ϕ ( m ( θ | θ > µ ) − m ( θ | θ < µ ))= logit p + ϕ ( m ( θ | θ > µ ) − m ( θ | θ < µ )) . Therefore, if m is strictly monotonically increasing in θ , then P ( T | G ) > P ( T | L ),and if m is strictly monotonically decreasing in θ , then P ( T | G ) < P ( T | L ). By contra-position, if P ( T | G ) = P ( T | L ), then m is neither strictly monotonically increasing nordecreasing in θ .Additionally, if m is monotonic in θ for all agents but there is heterogeneity in itsslope, then the average slope may be zero because some agents have upward-slopingmotives (“positivity motives”) and some agents have downward-sloping motives (“neg-ativity motives”). In this case, if agents have received information drawn from thesame distribution in the past, then their current beliefs will reflect their motives. Apositivity-motivated reasoner will be more likely to hold a belief that µ > θ , and anegativity-motivated reasoner with a decreasing motive function will be more likely tohold a belief that µ < θ . This implies that an agent who believes µ > θ is more likelyto believe that P ( T | G ) > P ( T | L ) in the experiment, and an agent who believes µ < θ is more likely to believe that P ( T | G ) < P ( T | L ) in the experiment.That is, if motive direction is heterogeneous, subjects will trust Fake News morethan True News. For further details, see Thaler (2020). By contraposition, if subjectstrust Fake News and True News equally, then there is no evidence for heterogeneity inpositivity- versus negativity-motivated reasoning. The experiment follows the structure and incentive scheme of Thaler (2020), whichcontains further details. Screenshots of a version of each page in the experiment,including instructions and scoring rules, can be found in Appendix C.Subjects first see an Introduction page for consent, then a Demographics page,and then the instructions and point system for Question pages. On each Questionpage, subjects are asked and incentivized to give a median guess, a lower bound (equalto their 25th-percentile belief), and an upper bound (equal to their 75th-percentilebelief). The median is incentivized using a linear scoring rule, and the bounds usingpiecewise-linear scoring rules. For details, see Appendix C.Next, subjects see the instructions and point system for News pages. On each Newspage, subjects see the message that says whether the answer is greater than or less than heir previous median guess, and are asked and incentivized to assess the probabilitythat the message comes from True News versus Fake News using a quadratic lossscoring rule. Subjects are told that the ex ante probability of True News is 1/2. Theyare also asked to give an updated median guess after seeing the message, and are againincentivized with a linear scoring rule. For details, see Appendix C.Subjects see News pages after their corresponding Question page, in the order:Question 1, News 1, Question 2, News 2, .... At the end of the experiment, they see aResults page with details on all the correct answers, points scored, and money earned.At the end of the experiment, subjects’ points earned on each part of the experimentare averaged. Subjects are paid a $3 show-up fee and have a probability of winninga $10 bonus equal to their average score divided by 1000. This probabilistic bonus isdesigned to eliminate potential hedging and risk-aversion confounds. The experiment was conducted on Amazon’s Mechanical Turk (MTurk) platform.MTurk is an online labor marketplace in which participants choose “Human Intelli-gence Tasks” to complete. MTurk has become a very popular way to run economicexperiments (e.g. Horton, Rand, and Zeckhauser 2011; Kuziemko et al. 2015), andLevay, Freese, and Druckman (2016) find that participants generally tend to havemore diverse demographics than students in university laboratories on dimensions likeage and politics. The experiment was coded using oTree, an open-source software basedon the Django web application framework developed by Chen, Schonger, and Wickens(2016).Wave 1 was conducted on July 8-9, 2019, and asked about the leukemia survivalrates question. Wave 1 additionally included political and performance questions thatwere part of a separate experiment. Wave 2 was conducted on October 1-2, 2019, andasked about the other four questions. Both waves were offered to MTurk workers cur-rently living in the United States who had not previously taken one of my motivatedreasoning experiments. 522 participants from Wave 1 and 508 participants from Wave2 passed simple attention and comprehension checks. Wave 1 also included politicizedquestions and tested a debiasing treatment that is unrelated to this paper; only partic-ipants in the control group are included here, and only observations on the positivityquestions are kept.Subjects in Wave 1 answer one question about positivity; subjects in Wave 2 answer In order to pass these checks, subjects needed to perfectly answer the comprehension check question inAppendix B (by giving a correct answer, correct bounds, and answering the news assessment with certainty).In addition, many questions had clear maximum and minimum possible answers (such as percentages, be-tween 0 and 100). Subjects were dropped if any of their answers did not lie within these bounds. our questions about positivity and one neutral question. There are a total of 3,062guesses to these questions. Zero guesses were exactly correct. There are therefore atotal of 3,062 news assessments. 2,554 assessments are Positive or Negative, and 508assessments are on the neutral topic.The balance table for the Positive / Negative treatment is in Appendix A.1. Sincethis randomization is within subject, treatments are expected to be balanced acrossdemographics. The overall shares of Positive and Negative are not statistically signifi-cantly different, indicating that there is not substantially different attrition.
This subsection shows that the raw data does not support positivity- or negativity-motivated reasoning, and the following subsection shows the relevant regressions.The mean assessment of Positive news is 57.7 percent (s.e. 0.7 percent) and themean assessment of Negative news is 58.5 percent (s.e. 0.8 percent). The differencebetween these is -0.7 percentage points; this point estimate is statistically insignificantlydifferent from zero ( p = 0 . p = 0 . Due to subjects in the two waves seeing different questions, the main specification isbetween subjects. In particular, subjects in Wave 1 only see one positivity-relatedquestion, so the within-subject test essentially ignores this sample.In particular, the main specification for positivity-motivated reasoning is in Table 2,column 1. The regression looks at assessments a for subject i , question topic q , and This suggests, reassuringly, that subjects did not look up the correct answers. All standard errors are clustered at the individual level.
Notes:
Only Positive / Negative news observations, as defined in Table 1. Messages are customized sothat Bayesians give the same assessment for Positive and Negative news. ound r with fixed effects for q and r when all news is Positive or Negative: a iqr = α + β · iqr + γz i + δF E q + ζF E r + (cid:15) iqr z i is a vector of controls. The controls used are age, an indicator for political party,an indicator for race, an indicator for gender, log(income), years of education, and anindicator for whether the subject is part of a religious group.Column 2 uses the within-subject design; the standard error is larger, but thecoefficient does not substantially change. In order to test whether there are differencescompared to Neutral news, column 3 includes indicators for both Positive (versusNeutral) news and Negative (versus Neutral) news.Columns 4 and 5 regress assessments on an indicator for True News (as opposedto Fake News), with and without controls for positivity. Recall from Section 2 thatthis measures whether directional errors in current beliefs are partly explained by pastmotivated reasoning on these topics, and therefore whether we should expect muchheterogeneity in motive direction. Table 2: The Effect of News Direction and Actual Veracityon Perceived Veracity (1) (2) (3) (4) (5)Positive News -0.005 -0.017 0.007 -0.005(0.010) (0.016) (0.016) (0.010)Negative News 0.012(0.016)True News -0.002 0.000(0.010) (0.011)Neutral News No No Yes No NoQuestion FE Yes Yes Yes Yes YesRound FE Yes Yes Yes Yes YesSubject controls Yes No Yes Yes YesSubject FE No Yes No No NoObservations 2554 2554 3062 2554 2554 R ∗ p < . ∗∗ p < . ∗∗∗ p < . Notes:
OLS, errors clustered at subject level. Neutral News in-dicates that Positive / Negative news assessments are comparedto assessments on Neutral topics. These classifications are de-fined in Table 1. Controls: age, political party, race, gender,log(income), years of education, and member of religious group. very single coefficient in Table 2 is insignificant and each point estimate is within2 pp of zero. Modest effect sizes can be ruled out at the 95-percent significance level.There is no evidence for aggregate-level positivity-motivated reasoning or negativity-motivated reasoning. There is no evidence that subjects infer differently on positivity-related topics compared to the neutral topic. There is no evidence that subjects haveformed erroneous beliefs in the direction of their motivated reasoning.An alternative measure of motivated reasoning, which looks at subjects updatetheir beliefs about the original question, generates the same prediction. After seeingthe message, subjects’ updated median belief is elicited. 38 percent of the time, subjectsupdate in the positive direction (s.e. 1 percent). 38 percent of the time, subjects updatein the negative direction (s.e. 1 percent). 24 percent of the time, subjects stay withtheir original guess (s.e. 1 percent). Clearly, there is no systematic updating in thepositive or negative direction. The results above show that there is no aggregate evidence for positivity- or negativity-motivated reasoning. This may be because nobody engages in motivated reasoning orbecause there is some heterogeneity with mean zero. This section discusses two formsof heterogeneity: heterogeneity across people, and heterogeneity across questions.As shown in columns 4 and 5 of Table 2, there is no evidence that subjects haveformed their current beliefs on these topics because of past motivated reasoning. Thissuggests that the degree of heterogeneity across people is not likely to be large. Insupport of this, Table 3 shows that treatment effects are not especially heterogeneousacross demographic groups in systematic ways. (1) (2) (3) (4) (5) (6) (7) (8)Pos News x Male 0.03 0.02(0.02) (0.02)Pos News x (Age >
32) -0.04 ∗ -0.03 ∗ (0.02) (0.02)Pos News x White -0.01 -0.01(0.02) (0.02)Pos News x College -0.02 -0.01(0.02) (0.02)Pos News x (Inc > ∗ (0.03) (0.03)Pos News x Religious -0.00 -0.01(0.02) (0.02)Pos News Yes Yes Yes Yes Yes Yes Yes YesQuestion FE Yes Yes Yes Yes Yes Yes Yes YesRound FE Yes Yes Yes Yes Yes Yes Yes YesSubject controls Yes Yes Yes Yes Yes Yes Yes YesObservations 2554 2554 2554 2554 2554 2554 2554 2554 R ∗ p < . ∗∗ p < . ∗∗∗ p < . Notes:
OLS regression coefficients, errors clustered at subject level. FE included for roundnumber and topic. Only Positive / Negative news observations, as defined in Table 1. Re-ligious: subject affiliates with any religion. Controls: age, political party, race, gender,log(income), years of education, and religious.
The evidence is not precise enough to clearly argue in favor of or against between-topic heterogeneity. On one question (regarding global poverty rates), there is statis-tically significant evidence for negativity -motivated reasoning: the difference betweenNegative and Positive news assessments is 11 pp (s.e. 3 pp). The treatment effect onnews assessments for the other questions is neither large nor significant at 95 percent There are several potential explanations for this. It may be noise, and it may also be a question onwhich self-image does play a role. In particular, people may engage in social comparison and be motivatedto believe that many others are less well-off. onfidence levels. One direction for future work is to extend the domain of topics an-alyzed to determine whether the global poverty question is an outlier or in a differentmotivated category such as social comparisons.
Results from the experiment indicate that there is no evidence for positivity- or negativity-motivated reasoning. To better understand the precision of these results, it is helpfulto compare the results to those in Thaler (2020). Since the experimental design is thesame, the comparison has the same units. Treatment effects for the three categoriesare plotted in Figure 3.
Figure 3: Comparing Motivated Reasoning About Positivity to Politics and Performance
Notes:
Treatment effects for the effect of Positive versus Negative, Pro-Party versus Anti-Party, and Pro-Performance versus Anti-Performance on news veracity assessments. Pro-Party and Pro-Performancecoefficients from Thaler (2020). Error bars correspond to 95 percent confidence intervals.
It is easily apparent that the effect of Positive news is significantly different thanthe effect of seeing Pro-Party news or Pro-Performance news. The overall effect is still close to zero if we remove the global poverty question. In the between-subjecttest, the coefficient on Positive news is 0.019 (s.e. 0.11; p = 0 . p = . here are two sets of explanations for the experimental results. First, motives maybe systematically different from belief-based utility, leading people to distort how theyprocess information differently in the self-image-relevant domains from the positivitydomains. Second, people may not actually receive utility from holding beliefs frompositivity.As a suggestive test to separate these different hypotheses, I run two follow-up sur-veys among a new group of subjects — drawn from different Mechanical Turk samples— on January 8-9 and 13, 2020. There are 303 participants in Survey 1 and 167 in Sur-vey 2. The evidence is consistent with the hypothesis that motives are systematicallydifferent from belief-based utility, and that survey participants are aware of this.In Survey 1, participants were given a definition of motivated reasoning and asked topredict the direction of motivated reasoning about positivity, politics, and performance,and given sample topics on all three. On positivity, they were asked whether theythought that most people motivatedly reasoned in the direction of believing that theworld was a better place for others, most people motivatedly reasoned in the directionof believing the world was a worse place for others, or about the same. The exampletopics were infant mortality, happiness, and cancer survival rates.The results from this survey are shown in Figure 4. 65 percent of subjects ex-pect motivated-reasoning distortions in the Pro-Party direction, versus only 16 percentwho expect distortions in the Anti-Party direction. Similarly, 56 percent expect Pro-Performance distortions, and only 18 percent expect Anti-Performance distortions.Subjects, however, were similarly likely to predict distortions in the Positive (36 per-cent) and Negative (30 percent) directions; this difference is not statistically significant( p = 0 . happier . The resultsare shown in Figure 5. Unlike with motivated reasoning, a clear majority of 69 per-cent believes that positivity makes people happier, and only 10 percent believes thatnegativity increases happiness. Politics and performance have similar and statisticallyindistinguishable point estimates. This number does not include 16 subjects in Survey 1 and 5 subjects in Survey 2 who failed a simpleattention check question. Results do not qualitatively change with the inclusion of these subjects.
Notes:
The y-axis is the share of respondents who stated that they expect most people to have moti-vatedly reason in one direction, the other direction, or a similar amount in both directions. 0.4 percentof questions are left unanswered, and are coded as “Similar.” Error bars correspond to 95 percent confi-dence intervals. The differences between each of the Positivity bars and their corresponding bars in theParty and Performance columns are statistically significant at the 95 percent level.
Taken as a whole, the evidence is in support of the explanation that people doreceive utility (in the form of happiness) from believing that the world is good forothers, but not in support of this form of positivity influencing motives. That is,motives do not include other-regarding belief-based utility. This explanation may alsohelp explain action-induced motivated beliefs, where people distort their beliefs as anexcuse to not be generous to others (Exley 2015; Di Tella et al. 2015).
This paper has shown that people do not necessarily motivatedly reason in the di-rection of “good” states when their self-image is not at stake. When asked aboutpositive or negative news about others, this experiment finds no evidence of systematicdirectional distortions of how people process the information. It also does not indi-cate any evidence that people’s current beliefs are distorted due to such positivity- ornegativity-motivated reasoning.
Notes:
The y-axis is the share of respondents who stated that they expect most people to be happierwhen receiving news in one direction, the other direction, or a similar amount in both directions. Errorbars correspond to 95 percent confidence intervals. The differences between each of the Positivity barsand their corresponding bars in the Party and Performance columns are not statistically significant atthe 95 percent level.
Survey results additionally show that people think that believing in positive statesof the world does not induce motivated reasoning. However, people do think thatbelieving in these positive states leads to increased happiness. The results from thispaper also suggest that utility-maximizing beliefs do not necessarily explain why peopleform persistently inaccurate beliefs.One direction for future work is to better understand the relationship betweenthe belief-based utility function and the motive function. For instance, do particularemotions affect utility and motives differently? These results suggest that happinessis insufficient for motives; however, in many self-image domains, pride and identityconfirmation play larger roles than happiness. An alternative hypothesis in psychology,proposed by von Hippel and Trivers (2011), is that self-deception is a mechanism bywhich people can deceive others. Convincing others that the world is a good place canbe less impactful than convincing others that one is smarter, more altruistic, or morecorrect.Once the tether between the motive function and the utility function is removed, uture work can treat the motive function as a separate object for study. The objectiveis for this experimental design to then be used to elicit motives and analyze their rolein other economic contexts. References
Barron, Kai (2020). “Belief updating: does the ‘good-news, bad-news’ asymmetryextend to purely financial domains?” In:
Experimental Economics .Benabou, Roland (2013). “Groupthink: Collective Delusions in Organizations andMarkets”. In:
The Review of Economic Studies .Benabou, Roland and Jean Tirole (2002). “Self-Confidence and Personal Moti-vation”. In:
Quarterly Journal of Economics .Benabou, Roland and Jean Tirole (2011). “Identity, Morals, and Taboos: Beliefsas Assets”. In:
Quarterly Journal of Economics .Brunnermeier, Markus and Jonathan Parker (2005). “Optimal expectations”. In:
The American Economic Review .Chen, Daniel, Martin Schonger, and Chris Wickens (2016). “oTree – An open-source platform for laboratory, online, and field experiments”. In:
Journal ofBehavioral and Experimental Finance .Dana, Jason, Roberto Weber, and Jason Kuang (2007). “Exploiting moral wig-gle room: experiments demonstrating an illusory preference for fairness”. In:
Economic Theory .Di Tella, Rafael et al. (2015). “Conveniently Upset: Avoiding Altruism by Dis-torting Beliefs about Others’ Altruism”. In:
American Economic Review .Eil, David and Justin Rao (2011). “The good news-bad news effect: asymmetricprocessing of objective information about yourself”. In:
American EconomicJournal: Microeconomics .Exley, Christine (2015). “Excusing Selfishness in Charitable Giving: The Role ofRisk”. In:
Review of Economic Studies .Horton, John, David Rand, and Richard Zeckhauser (2011). “The online lab-oratory: conducting experiments in a real labor market”. In:
ExperimentalEconomics . 17unda, Ziva (1990). “The case for motivated reasoning”. In:
Psychological Bul-letin .Kuziemko, Ilyana et al. (2015). “How Elastic Are Preferences for Redistribution?Evidence from Randomized Survey Experiments”. In:
American EconomicReview .Levay, Kevin, Jeremy Freese, and James Druckman (2016). “The Demographicand Political Composition of Mechanical Turk Samples”. In:
SAGE Open .Mayraz, Guy (2013). “Wishful Thinking in Predictions of Asset Prices”. In:
Work-ing Paper .Mobius, Markus et al. (2014). “Managing self-confidence: Theory and experimen-tal evidence”. In:
Working Paper .Thaler, Michael (2020). “The “Fake News” Effect: Experimentally IdentifyingMotivated Reasoning Using Trust in News”. In:
Working Paper .von Hippel, William and Robert Trivers (2011). “The Evolution and Psychologyof Self-Deception”. In:
Behavioral and Brain Sciences .Weinstein, Neil (1980). “Unrealistic Optimism About Future Life Events”. In:
Journal of Personality and Social Psychology .18
Supplementary Appendix: Additional Tables
A.1 Balance Table
Negative News Positive News Neg vs. Pos p-valueDemocrat 0.479 0.468 0.011 0.567(0.014) (0.014) (0.020)Republican 0.195 0.193 0.002 0.911(0.011) (0.011) (0.016)Male 0.550 0.532 0.018 0.353(0.014) (0.014) (0.020)Female 0.442 0.458 -0.016 0.406(0.014) (0.014) (0.020)Age 35.353 35.732 -0.379 0.357(0.291) (0.290) (0.411)Education 14.919 14.873 0.046 0.538(0.052) (0.054) (0.075)Log(Income) 10.833 10.853 -0.021 0.520(0.023) (0.022) (0.032)White 0.745 0.730 0.015 0.374(0.012) (0.012) (0.017)Black 0.090 0.088 0.002 0.881(0.008) (0.008) (0.011)Hispanic 0.046 0.057 -0.012 0.180(0.006) (0.006) (0.009)Asian 0.087 0.090 -0.003 0.789(0.008) (0.008) (0.011)Religious 0.470 0.479 -0.009 0.649(0.014) (0.014) (0.020) N Notes:
Standard errors in parentheses. Education is in years. Religious is 1 ifsubject in any religious group. Positive and Negative News is defined in Table 1. Study Materials: Exact Question Wordings
Infant Mortality
The CDC provides statistics for mortality rates for infants. In 1997, there were 28.0thousand infant deaths in the United States.How many thousands of infant deaths in the United States were there in 2017 (themost recent year available)?(If you answer X, it means you think that there were X thousand deaths.)
Correct answer: 22.3.Source linked on results page:
Others’ Happiness
Many surveys ask the following question about subjective happiness:“Please imagine a ladder with steps numbered from zero at the bottom to ten atthe top. Suppose we say that the top of the ladder represents the best possible lifefor you and the bottom of the ladder represents the worst possible life for you. If thetop step is 10 and the bottom step is 0, on which step of the ladder do you feel youpersonally stand at the present time?”In 2006, the average subjective happiness level in the United States was 7.18 out of10. What was the average subjective happiness level in the US in 2018?
Correct answer: 6.88.Source linked on results page: https: // ourworldindata. org/ happiness-and-life-satisfaction
Cancer in Children
Acute Myeloid Leukemia (AML) is a devastating illness in which cancerous cells emergein the bone marrow, invade the blood stream, and may spread to the rest of the body.Tragically, hundreds to thousands of children under the age of 15 are diagnosed withAML each year; it is one of the most common cancers among children.Of children under the age of 15 who are diagnosed with AML, what percent survivefor at least 5 years?(Please guess between 0 and 100.)
Correct answer: 68.8.Source linked on results page: lobal Poverty Around the world, many people do not have enough money for basic necessities. TheWorld Bank defines extreme poverty as having less than the equivalent of $1.90 perday.In 1990, the World Bank estimated that 1897 million people around the world wereliving in extreme poverty.As of 2015 (the most recent year available), how many milllions of people aroundthe world live in extreme poverty?(If you answer X, it means you think that X million people live in extreme poverty.)
Correct answer: 731.Source linked on results page: http: // povertydata. worldbank. org/ poverty/home/
Armed Conflict
The Department of Peace and Conflict Research estimates that 45.8 thousand peoplewere killed per year in battles in the fifteen years from 1989-2003.How many thousands of people were killed per year in battles in the fifteen yearsfrom 2004-2018?(If you answer X, it means you think that X thousand people were killed per year.)
Correct answer: 48.12Source linked on results page:
Latitude of Center of the United States
The U.S. National Geodetic Survey approximated the geographic center of the conti-nental United States. (This excludes Alaska and Hawaii, and U.S. territories.)How many degrees North is this geographic center?(Please guess between 0 and 90. The continental U.S. lies in the Northern Hemi-sphere, the Equator is 0 degrees North, and the North Pole is 90 degrees North.)
Correct answer: 39.833.Source linked on results page: http: // bit. ly/ center-of-the-us
Comprehension Check: Current Year