Information avoidance: Self-image concerns, inattention, and ideology

We report the results of an experiment on willful information avoidance regarding measures to address Covid-19. In the experiment, participants choose between two options, each associated with a contribution to the Corona Fund of the Red Cross USA and a payment to the participant. Depending on the treatment, either the participants’ payoff, the donation, both or none of these pieces of information were hidden, but revealable. With this design, we can separate motivated reasons for ignorance from non-motivated reasons, both of which are present in our data. Furthermore, we find evidence of both self-serving and pro-social information avoidance. These behavioral patterns correlate with the subjects’ political attitudes: while voters of the Democratic Party are prone to exhibit pro-social information avoidance, Republican voters rather engage in self-serving information avoidance.


Introduction
Since the outbreak of the Covid-19 pandemic, a polarization of views has been observable with respect to both, the health risks involved and the policies to address the pandemic ( Allcott et al., 2020;Gadarian et al., 2021 ). It has been surmised that differences in policy support do not solely hinge on differences in voters' preferences, but also result from selective search for and exposure to information that is driven by these preferences ( Romer and Jamieson, 2020 ). Explaining the voters' polarized worldviews via a human tendency for selective information search and its avoidance is indeed wellestablished. 1 Yet, while it is compelling to assume motivated reasoning as the main driver for information avoidance ( Epley and Gilovich, 2016;Golman et al., 2017;Gino et al., 2016 ), this phenomenon might very well be driven by non-motivated reasons ( Kessler and Exley, 2021 ).
For the purposes of this paper, we are interested in two different types of motives for information avoidance: (i) narrow self-interest and (ii) a propensity for pro-social behavior. The first type refers to individuals ignoring available information on the effects of their choice on others. In the context of Covid-19, these individuals would, for example, vote for policies that maximize their own payoff, while ignoring information on the consequences of these policies for society. The second type refers to a behavior where individuals avoid information on how their choice affects their own payoff. In this case, the individual decides using only information on the effectiveness of the policies, while ignoring available information on their own income. Thus, in this case, decision-makers avoid information due to pro-social motives. Note that both types of behavior are inconsistent with rational choice, as costless information would always be revealed by a homo economicus .
Type (i) of motivated information avoidance, which is also referred to as 'self-serving', has first been investigated via the use of incentivized experiments in Dana et al. (2007) . In a binary dictator game, they find that subjects willingly avoid information on the payoff of the recipient and then choose the option maximizing their own payoff. By remaining ignorant about the consequences of their choice for the recipient, they exploit moral wiggle room: An individual choosing not to know if their selfish choice harms another individual can act egoistically without incurring potential psychological costs from a bad conscience. Explanations for this type of information avoidance have been formalized by use of the dual-self theory by Grossman and van der Weele (2017) , as well as the theory of cognitive dissonance by Momsen and Ohndorf (2022) . Selfserving motivated information avoidance has been observed in a plethora of different incentivized experiments, such that its existence can be regarded as well-established within behavioral economics. 2 While self-serving motivated information avoidance is a well-established phenomenon, evidence for the inverse type (ii), i.e. avoiding information on own payoff consequences due to pro-social motives, has hitherto not been observed in incentivized experiments. Kandul and Ritov (2017) study a dictator game where information on the dictator's payoff is initially unobservable, but do not find evidence for pro-social information avoidance. 3 Similarly, Moradi (2018) reports null results on pro-social information avoidance. Hence, while motivated information avoidance seems to represent a convenient explanation for the polarization of voters' worldviews, thus far only the self-serving type has been established in incentivized experiments.
Another reason to remain skeptical towards the existence of motivated information avoidance is that it is clearly in contradiction with standard microeconomic theory, which would explain the phenomenon via unobservable information costs. While most studies investigating information avoidance in incentivized experiments use monetary information costs of zero, subjects' choices might be influenced by costs of attention or additional cognitive load ( Caplin and Dean, 2015;Gabaix, 2019 ). 4 Based on these considerations, Kessler and Exley (2021) conducted an online experiment using a variation of the dictator game used in Dana et al. (2007) , designed to separate these drivers of information avoidance. They find that information avoidance in such a setting is attributable to both inattention and motivated reasoning, with the former being more prominent.
In this paper, we present the results of an incentivized online experiment that was designed to address these issues in the context of Covid-19. Thus, we study if information avoidance can be attributed to either type of motivated reasoning or if other reasons, such as inattention, are the driving factors for this phenomenon. Moreover, we also investigate the effect of uncertainty about the revealed information being truthful. As shown in Momsen and Ohndorf (2022) , the type and extent of such uncertainty can have a significant influence on both, information revelation and exploitation of moral wiggle room.
In the experiment, subjects play a binary dictator game, with the Corona Fund of the American Red Cross taking the role of the recipient. In seven between-subjects treatments, we vary which type of information is hidden and whether complete revelation leads to certainty about the initially unobservable information. More precisely, we alter if the payoff, the donation, both or none of these pieces of information is initially unobservable, but revealable. Information revelation is designed as a process in two stages: In a first stage, subjects can click a button which reveals the correct information with a certain probability, while false information is displayed with the inverse probability. With the click of a second button, subjects can then fact-check the revealed information. The effectiveness of this fact-check varies over the different treatments. To account for potential effects of pivotality, the decision is framed as a democratic vote, with the size of the constituency being altered within-subject for each of the overall eight decisions.
In contrast to earlier experiments on information avoidance, we include treatments where both types of information are unobservable. We can hence compare the results for these treatments with those treatments where only one type of information -payoff or donation -is initially hidden. Via this comparison we can gain insights on the relative strength of self-serving motives in contrast to other explanations for information avoidance, like inattention. In our data, we find evidence for both, motivated and non-motivated explanations.
When revelation is certain, we observe both types of motivated information avoidance: self-serving and pro-social ignorance. Furthermore, we find that the subjects' preferences are a good predictor of what kind of motivated ignorance arises: while subjects supporting the Democratic Party tend to avoid information about their own costs, subjects supporting the Republicans tend to exploit moral wiggle room to behave more selfishly. A similar differentiation arises when considering the views on Covid-19 as a severe disease: subjects reporting higher levels of fear about Covid-19 tend to pro-social infor-mation avoidance, while self-serving information avoidance prevails among those who are more skeptical towards the risks associated with the disease.
The remainder of this paper is organized as follows: in the following section, we briefly put this paper in context of related strands of literature. Thereafter, we describe the experimental design. Our behavioral predictions are derived in Section 4 . We present our results in Section 5 and the last section concludes. Screenshots of the decision screens, the instructions, the questions from the questionnaire, as well as additional analyses are relegated to the Online Appendix.

Related literature
Our paper relates to several strands of literature. First and foremost, we extend the growing literature on motivated reasoning (see, e.g., Epley and Gilovich, 2016;Bénabou and Tirole, 2016 , for reviews), more precisely the strand on motivated information avoidance ( Golman et al., 2017 ). Our experimental design is inspired by the seminal paper by Dana et al. (2007) who coined the phrase "moral wiggle room": decision-makers make use of ways to justify their selfish behavior that allow them to maintain a positive self-image while acting egoistically. Building on their paper, an extensive body of literature has replicated and extended their insights on the exploitation of moral wiggle room ( Larson and Capra, 2009;Feiler, 2014;Matthey and Regner, 2011;Grossman, 2014;Grossman and van der Weele, 2017;Exley, 2016;Momsen and Ohndorf, 2022;Kessler and Exley, 2021;Kajackaite, 2015;Stüber, 2020 ). However, there is also research that does not find evidence of self-serving motivated ignorance in the absence of revelation costs ( Lind et al., 2019;Momsen and Ohndorf, 2020;Felgendreher, 2018 ), indicating that there are situational and behavioral boundaries to the occurrence of the phenomenon. In particular, experimental endeavors to find evidence for the pro-social type of motivated information avoidance were not successful thus far ( Kandul and Ritov, 2017;Moradi, 2018 ). By presenting evidence of the existence of both types of motivated information avoidance, this paper is meant to contribute to a better understanding of the conditions under which these phenomena occur.
Second, we contribute to explaining the different motivations behind the observed ignorance. With the notable exception of Kessler and Exley (2021) , the above-mentioned papers do not aim to investigate whether the observed rates of ignorance result from motivated considerations or are due to non-motivated reasons, like (rational) inattention ( Sims, 2003;Gabaix, 2014;2019;Caplin and Dean, 2015 ). By use of a different experimental design than Kessler and Exley (2021) , we seek to separate the driving factors of the subjects' ignorance via a comparison of the treatments with one-sided unobervability to those that feature the two-sided variant.
Third, our experiment extends the literature on the influence of political attitudes on economic decisions. Cappelen et al. (2017) investigate the relation between political views and moral behavior, Fosgaard et al. (2019) study cooperation and political attitudes, Kerschbamer and Müller (2020) analyze social preferences and Müller and Renes (2021) study fairness preferences in relation to political preferences. While previous work has mostly considered how political preferences relate to altruistic behavior, experimental evidence regarding the relationship between political preferences and (motivated) information avoidance to commit to (or to refrain from) behaving pro-socially is still lacking. This study is to provide such evidence in the context of financial contributions to address the pandemic.
Finally, we also add to the growing literature on Covid-19 and information avoidance. Up to the present, this strand was not concerned with the interaction of motivated ignorance and pro-sociality. Instead, the existing literature has rather followed the footsteps of Ganguly and Tasoff (2017) ; Oster et al. (2013) and, broadly speaking, has focused on the relation between stress, health-relevant information and its avoidance. In particular, this type of information avoidance seems to arise if people experience stress ( Soroya et al., 2021 ), information overload ( Soroya et al., 2021;Qu et al., 2023 ), and anxiety ( Siebenhaar et al., 2020 ) towards . While this type of information avoidance could also be the result of motivated reasoning, it lacks the pro-social dimension which is the focus of our paper.

Experimental design
We ran an online experiment to investigate the interplay of self-image concerns, attention and ideology in decision situations where the consequences for the decision-maker and the recipient may or may not be directly observable.
In the experiment, subjects make eight independent binary decisions between two options, A and B, that differ with respect to (i) their own payoff as well as (ii) an associated donation to the American Red Cross. In four of these decision situations, the option with the higher donation yields a higher monetary payoff for the subject (aligned interests), while in the other four decision situations, the interests of the subject and the American Red Cross are conflicting.
We vary the group size in which decisions are made in a within-subject manner: subjects either make their allocation decision independently of the other participants or they vote on the allocation to be implemented in a group. There are decision situations with groups consisting of three, nine and 27 participants. The option receiving the majority of votes is implemented for all group members, and groups are formed anew in each round. There is no feedback between rounds, and group members remain entirely anonymous.
For Option A, both the subject's payoff ($2) and the contribution to the American Red Cross ($1.50) are immediately observable in all treatments and remain fixed throughout the rounds. For Option B, the payoff can either be $1.50 or $2.50 and the contribution to charity varies between either $0 or $3. For both payoff and donation, each of the respective values is equally likely. Thus, depending on the actual realization of these values, the decision involves situations of either aligned  Table 2 Participants per treatment.

No of participants
FullInfo 108 Donation 108 Donation75 162 Payoff 108 Payoff75 162 Both 108 Both75 162 Our main treatment variations affect (i) the availability of information regarding own payoff consequences and donations associated with Option B and (ii) the process of revealing initially hidden information. In total, these two treatment variations yield seven treatments (see Table 3 ).

Treatment variation: information on payments
In a between-subjects design we vary the availability of information on the payments associated with Option B. In the FullInfo treatment, all payments are immediately observable, i.e. subjects know immediately if their own and the charity's interest are aligned or conflicting. In all other treatment variations, one or both payments associated with Option B are initially unobservable, but can be revealed at no cost via the click of a button. We implement two treatments where the donation is initially hidden, but the subjects' payoff is observable ( HiddenDonation ), two treatments where the subjects' payoff is hidden, but the donation is directly observable ( HiddenPayoff) and two treatments where both payoff and donation are initially hidden ( Both ). See Section D in the Online Appendix for screenshots of the decision screens.

Treatment variation: revelation
In the hidden information treatments, subjects have the option to reveal the information, but they can also make their allocation decision directly without gathering information on the payments related to Option B. In these treatments, revealing information is costless and consists of a two-step process: clicking the first "Reveal" button, subjects learn a number for the respective payment that is correct with a probability of 80%. In all other cases, the revealed information is false, e.g. if a donation of $3 is revealed, the actual donation is $3 with a probability of 80% and $0 with a probability of 20%. Analogously, if a payoff of $2.50 is revealed, the actual payoff is $2.50 with a probability of 80% and $1.50 with a probability of 20%.
Having clicked the first button, subjects have the option to click a second type of button labeled "Check". Depending on the treatment variation, this button reveals the true payment with certainty ( Certain ) or with a probability of 75% ( Uncertain ). In the treatments where both donation and payoff are initially hidden, each type of payment has its own set of "Reveal" and "Check" buttons. The two-step revelation approach serves two purposes: first, it assures better comparability of the treatments with certain and uncertain revelation. Second, this approach represents a device that may induce a higher level of inattention (as two buttons are to be pressed to reveal all possible information on an item). Revealing and checking the payments is optional, i.e. subjects can also make their allocation decision without observing the respective payments associated with Option B, and they can stop at any point of the revelation process and make their allocation decision.

Experimental procedure
The experiment was programmed in oTree ( Chen et al., 2016 ) and run online with US citizens recruited through Prolific in December 2020. 6 In total, 918 subjects completed the study (see Table 2 ). At the beginning of the experiment, subjects received detailed instructions that can also be found in the Online Appendix of this paper (see Section C). Subjects could only participate in the experiment after correctly solving a short quiz testing their understanding of the decision situation. Then, they made the eight main decisions of the experiment. After subjects had completed the study, it was determined randomly which of the eight decisions was relevant for the subject's payoff. We matched subjects into groups and computed their payoff as well as the resulting donation according to the option receiving the majority of votes after running the experiment. In total, we donated $1539 to the American Red Cross. Subjects earned a fixed payoff of $3.30 in addition to the variable payoff component that could either be $1.50, $2, or $2.50 depending on the round chosen for payment and the selected option.
Following one of the eight decision situations at random, we implemented a standard attention check by use of a short question. 7 At the end of the experiment, we elicited standard demographic variables (age, gender), the subjects' political preferences, and their opinion towards the American Red Cross as well as towards Covid-19. For an overview of the questions, see the Online Appendix.
To elicit the subjects' political preferences, we asked them for the party they voted for in the last presidential elections and which party came closest to their political views. In both cases, they could choose between the following options: 'Democrats', 'Republicans', 'Other parties' and 'I would rather not say'. The question targeting their opinion on the Corona Fund was formulated as 'To what extent do you consider the Corona Fund of the Red Cross USA useful to society?'. Here, subjects could indicate their opinion on a scale from 0 to 10. By asking 'To what extent do you consider Covid-19 to be a risk to your own health?' we sought to elicit their views on Covid-19 as a severe disease. Again, subjects could give their answer on a scale from 0 to 10.

Behavioral predictions
Ignorance may result from both motivated and non-motivated information avoidance, like inattention. In the standard experiments on willful ignorance, which are based on the design presented in Dana et al. (2007) , these two motives cannot be separated as only the aggregate effect is observed. Hence, for these experiments, we know that a substantial share of decision-makers remains ignorant, but we do not know if this is due to their desire to act selfishly while still being able to consider themselves a good person (i.e. motivated ignorance), or simply because they proceeded with the (selfish) decision due to other reasons. For example, subjects might not want to devote a sufficient share of their attention to the task and therefore do not reveal information. Instead, they might make their choices without gathering information and thus speed through the study to finish it quickly.
Motivated information avoidance has a solid theoretical basis, as predictions for the self-serving and pro-social types can be derived via formalized models based on the dual-self theory, as in Grossman and van der Weele (2017) , or on the theory of cognitive dissonance, as developed in Momsen and Ohndorf (2022) . In these models, both types of information avoidance are explained by the individuals' need to be consistent with their self-image. Thus, self-serving information avoidance allows the individual to choose the option with a higher payoff without having to deal with a potential conflict with their selfperception as a moral individual (which takes third-party consequences of own actions into account). Pro-social information avoidance is the inverse. Here, an individual avoids information on own payoffs to avoid conflicts with the self-perception as an individual that is responsible for its own well-being. Their decisions are hence exclusively based on the known effects of own actions on others.
Non-motivated explanations can be derived from models of (rational) inattention, ( Sims, 2003;Gabaix, 2014;2019;Caplin and Dean, 2015 ). In these models, willful ignorance is generally explained by non-monetary information costs, due to cognitive load or the opportunity cost of allocating time to a different task. As these costs arise independent of the type of information (on payoff or donation), non-motivated information avoidance can also arise if all characteristics of a choice option are initially unknown (but revealable).
Thus, while the theoretical foundations of motivated and non-motivated information avoidance seem to be wellestablished, an empirical differentiation between both explanations requires an extension to the experimental design used in Dana et al. (2007) . Kessler and Exley (2021) , for example, extend this design by including a treatment where own payoffs are substituted by payments to a third party, which eliminates potential self-serving motivations, but does not take the pro-social type into account.
In order to separate both types of motivated ignorance from non-motivated ignorance, we introduce the treatments where both payoff and donation are hidden ( Both ). If, in these treatments, subject do not reveal any type of information, they have no knowledge on the preferability of either option. Such a case of 'complete' ignorance in the Both treatments can be only explained by non-motivated reasons, as the subject has neither information on the payoff, nor the associated contribution of Option B. In contrast, the decision to remain uninformed in the treatments with HiddenDonation or Hid-denPayoff may result from both motivated (self-serving or pro-social) information avoidance and non-motivated reasons. In these treatments, decision-makers qualify as 'completely' ignorant if they do not reveal the respective hidden characteristic of Option B, i.e. if they do not use the available revelation button.
In case of self-serving motivated information avoidance, decision-makers remain ignorant about the consequences of their choices for the recipient in order to choose the option maximizing their own payoff without facing severe negative consequences for their self-image. This type would arise in the HiddenDonation treatments.
In case of pro-social information avoidance, subjects seek to avoid the temptation to behave according to their narrow self-interest, which might be the case if they were aware of a conflict of interests. If they remain unaware of such a conflict, their ignorance represents a commitment device to behave pro-socially. This type would arise in the HiddenPayoff treatments.
Thus, if no information is revealed in the Both treatments ('complete' ignorance) this can only be attributed to nonmotivated information avoidance, as motivated reasoning can only occur if at least part of the characteristics of Option B is known. Hence, we expect the share of subjects not revealing any information to be larger in the treatments with one-sided initial observability than in the Both treatments. The difference in complete ignorance between the Both treatments and the treatments where only one type of information is hidden can thus be attributed to motivated information avoidance. With this setup, we can hence establish a lower bound for non-motivated ignorance in the treatments where only one type of information is initially unobservable.
We summarize these considerations in the first hypothesis: Hypothesis 1. The share of complete ignorance is lower in Both than in HiddenDonation and HiddenPayoff.
If we find evidence supporting our first hypothesis, we can conclude that there are subjects who engage in motivated information avoidance. As shown in Momsen and Ohndorf (2022) , different information structures with stochastic revelation can also have an influence on motivated information avoidance. We can hence expect two countervailing effects: with uncertain revelation, the ex ante informational value of clicking the revelation button is lower, which might decrease the amount of revelation in these treatments. At the same time, even if a conflict of interests is revealed, uncertainty still offers an excuse not to choose pro-socially under an apparent conflict of interests, as it may be the case that, with a certain probability, interests are in fact aligned. While the former effect should lead to more information avoidance, the latter should mitigate potential self-serving concerns not to reveal hidden information and induce subjects to make use of the revelation button more often. Expecting the latter effect to prevail, we formulate the following hypothesis: Hypothesis 2. Subjects are more likely to make use of the revelation button under uncertain than under certain revelation.
Moreover, we are interested in the impact of the decision-makers' motivated ignorance on their final choices in conflict situations. We conjecture deliberate information avoidance to result in different shares of selfish choices under hidden information on donation or payoff compared to the full information case. In all treatments but FullInfo , subjects do not know ex ante if they face a situation with aligned or conflicting interests. Only when making use of the revelation buttons, subjects can become aware of all relevant decision variables and can hence identify a potential conflict of interests. Recall that in the treatments with HiddenDonation , they can immediately identify the option maximizing their own payoff, while in the treatments with HiddenPayoff, they know immediately which option is associated with a higher donation to the American Red Cross. Based on this one-sided lack of (revealable) information, different options arise for the subjects to make use of their ignorance.
In the treatments with HiddenDonation , decision-makers may prefer not to reveal information about a potential conflict of interests to circumvent the obligation to behave pro-socially. Remaining ignorant hence provides a convenient excuse to choose according to their narrow self-interest and still maintain a positive self-image. Such self-serving information avoidance would lead to a higher share of selfish choices under HiddenDonation than under FullInfo , as formulated in our third hypothesis: Hypothesis 3. In conflict situations, the share of selfish choices is higher in the treatments with HiddenDonation than in the treatment with FullInfo (self-serving information avoidance) .
In the treatments with HiddenPayoff, the opposite can arise. Ignorance about a potential conflict of interests can be used to circumvent the temptation to behave egoistically in case the pro-social option is revealed to be more expensive to the subject. The pattern of using ignorance about own payoff consequences as a commitment device not to engage in selfish behavior should manifest in a lower share of selfish choices under HiddenPayoff than under FullInfo . We therefore formulate the next hypothesis: Hypothesis 4. In conflict situations, the share of selfish choices is lower in the treatments with HiddenPayoff than in the treatment with FullInfo (pro-social information avoidance) .
The third hypothesis represents the prediction of subjects exploiting moral wiggle room, as defined by Dana et al. (2007) . The fourth hypothesis is the inverse, predicting some sort of pro-social information avoidance. Note that, again, both of these predictions are consistent both with formalized models of information avoidance based on the dual-self theory as in Grossman and van der Weele (2017) , and with models based on the theory of cognitive dissonance, as developed in Momsen and Ohndorf (2022) .
Note also that the effect of uncertainty on revelations ( Hypothesis 2 ) should carry over to the tendency to engage in motivated information avoidance ( Hypotheses 3 and 4 ). We hence expect less self-serving and pro-social information avoidance when revelation is uncertain than when it is certain.
Lastly, we consider motivated ignorance relative to the preferences towards the Corona Fund of the Red Cross. We expect to discover three behavioral types in the context of exploitation of moral wiggle room. 8 First, we expect to encounter decision-makers who tend to choose the selfish option regardless of their informational state. These decision-makers with weak preferences for giving ("Red Cross Skeptics") exhibit no motivated information avoidance as, for them, there is no conflict between their monetary payoff and their self-image. They hence do not need to engage in motivated reasoning to manage their self-image while acting egoistically, nor do they need to make use of their ignorance as a commitment device to behave pro-socially. Thus, we expect them to make the same choices under HiddenDonation , HiddenPayoff and FullInfo .
Second, we expect to observe decision-makers who tend to choose pro-socially if they face a situation where the tradeoff between their own payoff and the payoff for the recipient is immediately observable, but selfishly when they can remain ignorant about the externalities of their choice. Note that we refer to tendencies in behavior, i.e. to the fact that decisionmakers may be more likely to choose a specific option. Overall, we hypothesize that decision-makers with intermediate preferences for giving ("Red Cross Supporters") have a tendency to exhibit self-serving information avoidance to exploit moral wiggle room. We expect them to make the same choices under FullInfo and under HiddenPayoff, but to use their ignorance to exploit moral wiggle room under HiddenDonation .
In the third group, we summarize subjects with strong and with fairly strong altruistic preferences. These are the participants who either are very likely to choose the pro-social option regardless of the information condition, or those who tend to choose the pro-social option when the own-payoff consequences are hidden, but may be tempted to choose selfishly when aware of a conflict of interests. The latter subjects would hence be prone to pro-social information avoidance. Note again that we are discussing nuances in behavior here. Thus, in tendency, subjects falling into the third group may still be more likely to choose the pro-social option under FullInfo than decision-makers of the second group. Yet, concealing the own-payoff consequences increases their likelihood to choose pro-socially, even though their altruism is already at a comparatively high level under FullInfo . Overall, we expect pro-social information avoidance rather in this third group with relatively strong preferences for giving ("Red Cross Enthusiasts"). Hence, we hypothesize them to make the same choices under FullInfo and under HiddenDonation , but to choose more pro-socially under HiddenPayoff. We thus formulate the following hypothesis: Hypothesis 5. The subjects' preferences correlate with the variation in behavior across treatments.
1. The selfishness of choices made by "Red Cross Skeptics" does not depend on the treatment variation. 2. "Red Cross Supporters" act more selfishly under HiddenDonation than under Full-Info and HiddenPayoff. 3. "Red Cross Enthusiasts" behave more pro-socially under HiddenPayoff than under FullInfo and HiddenDonation .

Motivated vs. non-motivated information avoidance
As it is a priori unclear to what extent information avoidance is due to motivated reasons or other motives such as inattention, we first need to establish that subjects willingly avoid information. We therefore compare the share of subjects not revealing any information (i.e. not using any revelation button) in the treatments with HiddenDonation or HiddenPayoff to the respective treatments where Both payoff and donation are hidden. Fig. 1 presents the share of subjects who do not reveal any information. The length of the bars indicates the share of subjects remaining ignorant in the treatments where either the Donation or the Payoff is hidden. For each bar, the dark grey area represents the share of subjects remaining completely ignorant in the Both treatments under the respective revelation condition -certain or uncertain. Hence, the light grey part of the bars represents the difference in ignorance between the Both treatments and the respective HiddenDonation or HiddenPayoff treatment.
Recall that in the Both treatments, remaining ignorant cannot be due to motivated reasons, as in this case, subjects cannot know if their selected option is superior or inferior to the other option. Instead, the reason for remaining uninformed would be either inattention or confusion on the part of the subjects. A third explanation might be that subjects want to finish the experiment as quickly as possible without optimizing any type of payoff. While the same reasons to remain ignorant also exist in the treatments with HiddenDonation and in the treatments with HiddenPayoff, information avoidance in these treatments can additionally arise due to motivated reasoning. As can be seen from Fig. 1 , the share of uninformed subjects in the treatments with HiddenDonation and HiddenPayoff always exceeds the share of ignorant subjects in the corresponding Both treatment.
To test whether these differences in ignorance are statistically significant, we regress the share of ignorant subjects on indicator variables for the treatment condition, the decision number and, depending on the specification, control variables (see Table 4 ). Thereby, we take the respective Both treatment as reference category. When including controls, we use an exhaustive set of variables controlling for whether Option B had a lower payoff or a higher donation than A, the subject's age, gender, education and prolific score as well as the group size in which the decision was taken. We also include dummy variables indicating if the subject always chose Option A or B and if they passed the attention check. In addition, we control for the subject's political, time and risk preferences, their stated altruism, views on Corona, and their opinions towards the Red Cross. Notes: Dependent variable is an indicator variable taking the value of 1 if the subject made their allocation decision without revealing any additional information. Donation and Payoff are dummy variables taking the value of 1 in the respective hidden information treatments. Output from random-effects panel regressions. We control for whether Option B had a lower payoff or a higher donation than A, the subject's age, gender, education, political, altruism, time and risk preferences, attention measures, views on Corona, opinions towards the Red Cross, their prolific score as well as the group size in which the decision was taken. Cluster-robust standard errors in parentheses (clustered on subject-level). * p < 0 . 1 , * * p < 0 . 05 , * * * p < 0 . 01 . We find that there are significantly more ignorant subjects in the treatment where the donation is hidden and revelation is certain than in the corresponding Both treatment. Hence, ignorance in HiddenDonation must be due to self-serving considerations, in addition to motives such as inattention. When revelation is uncertain, however, the difference in ignorance between Both and HiddenDonation disappears -an observation that we will discuss in more detail later on. Comparing the HiddenPayoff treatments to those where Both payoff and donation are hidden, we observe a significantly higher share of subjects remaining ignorant regardless of whether revelation could yield certainty -an observation that we interpret as evidence for motivated ignorance in the treatments with HiddenPayoff. 9 Hence, we can formulate our first result: Result 1. The share of complete ignorance is lower in Both than in HiddenDonation when revelation is certain. It is also lower in Both than in HiddenPayoff irrespective of the revelation condition.

Uncertainty and ignorance
Building on the results in Momsen and Ohndorf (2022) , we hypothesized that different information structures may influence information avoidance. When comparing the treatments with certain and uncertain information revelation, two countervailing effects may occur. While the ex ante informational value of clicking the revelation buttons is lower under uncertain revelation, the remaining uncertainty after revelation may be used as an excuse not to act according to the revealed information when a conflict of interests becomes apparent. Hence, the former effect should decrease the participants' tendency to reveal information, while the latter should decrease their reluctance to make use of the revelation button. We present the effects of uncertainty in the revelation process on ignorance in Table 5 .
In the treatments with HiddenDonation , subjects are indeed significantly less likely to remain ignorant if the revelation does not yield certainty. As expected when deriving Hypothesis 2 , ex ante the subjects seem to consider the remaining uncertainty after revelation as a sufficient excuse to act selfishly when aware of a conflict of interests and hence do not to engage in motivated information avoidance. Whether they do take the revealed information into account ex post will be explored in the following analysis. For the treatments with HiddenPayoff, the uncertainty in the revelation does not have a significant influence on whether or not subjects decide not to reveal any information. As can be seen in Table 5 , the coefficients are positive, but insignificant. This leads to the conclusion that in the treatments with HiddenPayoff, the two countervailing effects of uncertain revelation -lower value of information and excuse to act selfishly -seem to level out. We summarize our findings in the second result: Result 2. When the donation is initially hidden, subjects are less likely to remain ignorant under uncertain than under certain revelation. There is no statistically significant difference when the payoff is hidden.

Motivated information avoidance
Having established that motivated ignorance does indeed occur, let us now investigate how ignorance translates into choices. We therefore take a look at the share of selfish choices in conflict situations across treatments. Note that, in the treatments with uncertain revelation, subjects did not always have the opportunity to become aware of a conflict of interests. Hence, we only look into situations where -after subjects had clicked all available buttons -a conflict of interests could be revealed. This approach ensures comparability of the decision situations across treatments, as we only consider situations where the same state of information can be reached as in the baseline treatment with full information. In these situations, differences in the share of selfish choices to the treatment with FullInfo indicate that ignorance is in fact due to motivated reasoning, providing an excuse for subjects to choose more or less egoistically than they would have under FullInfo . As can be seen from Fig. 2 , the share of selfish choices is lowest in the treatment where the payoff is initially hidden and can be revealed with certainty. It is highest in the treatment where the donation is initially hidden and can be revealed with certainty. The fact that the share of selfish choices under FullInfo lies in between these two treatments suggests that subjects may engage in motivated -self-serving or pro-social -information avoidance when payoff or donation is hidden. Furthermore, it seems that this tendency is mitigated by uncertainty in the revelation process. Besides, choices appear to be only slightly more selfish when both donation and payoff are hidden than under FullInfo , suggesting that in these treatments, subjects do not engage in willful information avoidance.
To investigate if these observations hold up to statistical scrutiny, we perform several regression analyses including and excluding control variables as well as using all data (1 and 2) or only data of those subjects who passed our attention check and whose answers cannot be considered questionable for various reasons (3 and 4). 10 The results are depicted in Table 6 . In each regression, we regress the share of selfish choices in revealable conflict situations on indicator variables for the respective treatments, taking the case with FullInfo as baseline. We further include the decision number to allow for potential time trends and, depending on the specification, the usual control variables.
We find that our initial observations are confirmed: 11 There is indeed statistically significant evidence of self-serving information avoidance in HiddenDonation and of pro-social information avoidance in HiddenPayoff. Note that for the latter, the result is only significant at the 10% -level, such that we interpret our findings rather as suggestive evidence for pro-social information avoidance.
When revelation is uncertain subjects do not seem to engage in systematic information avoidance that would materialize in a share of selfish choices statistically different from the FullInfo treatment. 12 We thus formulate our third result: Notes: Dependent variable is an indicator variable taking the value of 1 if the subject chose the selfish option in a revealable conflict situation. Output from random-effects panel regressions. We control for whether Option B had a lower payoff or a higher donation than A, the subject's age, gender, education, political, altruism, time and risk preferences, attention measures, views on Corona, opinions towards the Red Cross, their prolific score as well as the group size in which the decision was taken. Cluster-robust standard errors in parentheses (clustered on subject-level). * p < 0 . 1 , * * p < 0 . 05 , * * * p < 0 . 01 .

Result 3.
In conflict situations, the share of selfish choices is higher in the treatment with HiddenDonation and certain revelation than in the treatment with FullInfo (self-serving information avoidance). There is no evidence of self-serving information avoidance when revelation is uncertain.

Result 4.
There is only suggestive evidence that the share of selfish choices in conflict situations is lower in the treatments with HiddenPayoff and certain revelation than in the treatment with FullInfo (pro-social information avoidance). There is no evidence of pro-social information avoidance when revelation is uncertain.

Motivated information avoidance and preferences
We proceed by investigating whether the subjects' stated preferences correlate systematically with their tendency to engage in motivated reasoning through information avoidance. For the following analyses, we therefore split the sample according to the elicited preferences that may interact with the subjects' willingness to donate to the Corona Fund of the American Red Cross. More precisely, we group subjects according to their opinion towards the Corona Fund of the American Red Cross, their political preferences as well as their views on Covid-19 as a health risk and investigate if the specific groups differ in terms of their tendencies to engage in motivated reasoning through information avoidance. Note that the elicitation of these preferences was not incentivized, yet subjects had to answer the corresponding questions in order to finish the study and receive their payment.
For the question on their attitude toward the Corona Fund, subjects could indicate their opinion on a scale from 0 to 10. Based on their answers, we split subjects into three groups: 13 those replying with a number lower than 7 form one group -the Skeptics -, those who answered with a 7 or 8 into another -the Supporters -and those who answered with a 9 or 10 form the last group -the Enthusiasts. We then run the same regressions as before for each subgroup with and without control variables and report the results in Table 7 . As hypothesized in Hypothesis 5 , we do indeed find that the three different groups show differences in their tendency to avoid information. For the group of Skeptics, we find no evidence of motivated information avoidance, i.e. their choices do not differ across treatment variations. Hence, they do not make use of their ignorance to engage in motivated reasoning. Supporters, in contrast, are significantly more selfish when the donation is initially hidden, but can be revealed with certainty. They seem to strategically use their ignorance to maintain a clear conscience while acting egoistically. For the group of Enthusiasts, we observe the mirror picture: They exhibit prosocial information avoidance in the treatment where the payoff is revealable with certainty. In other words, they use their ignorance to avoid the temptation to be selfish in situations that might feature a conflict of interests. These results hence Notes: Dependent variable is an indicator variable taking the value of 1 if the subject chose the selfish option in a revealable conflict situation. Output from random-effects panel regressions. We control for whether Option B had a lower payoff or a higher donation than A, the subject's age, gender, education, political, altruism, time and risk preferences, attention measures, views on Corona, opinions towards the Red Cross, their prolific score as well as the group size in which the decision was taken. Cluster-robust standard errors in parentheses (clustered on subject-level). * p < 0 . 1 , * * p < 0 . 05 , * * * p < 0 . 01 . When controlling for multiplicity, the coefficient of Donation is significant at the 5% level in regressions 3 and 4, while the coefficient of Payoff is no longer significant in regression 5 and is only significant at the 10% level in regression 6.
confirm that motivated information avoidance can, indeed, go both ways: There is not only the possibility of self-serving motivated ignorance, but the pro-social type does, in fact, exist. These findings can be summarized in our sixth and last result: Result 5. The subjects' preferences correlate with how the selfishness of choices varies across treatments.
(i) The selfishness of choices made by "Red Cross Skeptics" does not depend on the treatment variation.
(ii) "Red Cross Supporters" act more selfishly under HiddenDonation than under Full-Info when revelation is certain. There is no significant difference in the share of selfish choices when revelation is uncertain. (iii) "Red Cross Enthusiasts" act more pro-socially under HiddenPayoff than under Full-Info when revelation is certain.
There is no significant difference in the share of selfish choices when revelation is uncertain.
In a more exploratory manner, we split subjects by their stated opinions on Covid-19, more precisely on whether they consider Covid-19 a health risk. For this question, subjects could again give their answer on a scale from 0 to 10. As the median answer was 7, we use this number to split subjects for our analyses into those who fear Covid-19 as a risk to their health and have thus given an answer above the median and those who are more skeptical and have given answers below or at the median. Again, we find that their behavior differs (see Table 8 ): Subjects who are skeptical of Covid-19 being a serious health risk engage in self-serving information avoidance. In contrast, subjects are more likely to engage in pro-social information avoidance, if they believe that Covid-19 is indeed a serious risk to their health.
For the analysis of the relation between political preferences and information avoidance, we only consider those supporting either the Democratic or the Republican Party. We hence exclude those identifying with other parties or preferring not to state their political preferences, since these are only a few observations, and we expect these subjects to be relatively heterogeneous. 14 We again perform the same regressions as before for the two subsamples, Democrats and Republicans (see Table 9 ): We find that subjects supporting the Democratic Party tend to engage in pro-social information avoidance in the HiddenPayoff treatment, while supporters of the Republicans are prone to engage in self-serving information avoidance in the HiddenDonation treatment. Again, there is no robust evidence of self-serving information avoidance when revelation is uncertain.  Notes: Dependent variable is an indicator variable taking the value of 1 if the subject chose the selfish option in a revealable conflict situation. Output from random-effects panel regressions. We control for whether Option B had a lower payoff or a higher donation than A, the subject's age, gender, education, political, altruism, time and risk preferences, attention measures, views on Corona, opinions towards the Red Cross, their prolific score as well as the group size in which the decision was taken. Cluster-robust standard errors in parentheses (clustered on subject-level). * p < 0 . 1 , * * p < 0 . 05 , * * * p < 0 . 01 . When controlling for multiplicity, the results for the Democrats remain significant at the 5% level (regression 1) and at the 1% level (regression 2), while those for the Republicans marginally fail to reach significance (p = 0.107).

Conclusion
In this paper, we present the results of an online experiment investigating the driving factors of ignorance in moral decisions. In the experiment, subjects could choose between two options that differed in their contribution to the Corona Fund of the American Red Cross as well as in the payoff the subjects would earn when choosing the respective option. Depending on the treatment, either the own-payoff consequences, the donations, both or none of these pieces of information were hidden. In case they were not immediately observable, subjects had the option to reveal the initially hidden pieces of information. Here, it varied between treatments whether or not revelation led to certainty about the payments associated with the options.
Our experimental design allows us to separate motivated reasons for ignorance from non-motivated reasons, both of which are present in our data. Furthermore, we find evidence of motivated self-serving, as well as pro-social, information avoidance. While the self-serving type has been found in previous studies (see, e.g., Dana et al., 2007;Grossman and van der Weele, 2017 ), pro-social information avoidance had proven elusive thus far ( Moradi, 2018;Kandul and Ritov, 2017 ). The confirmation of its existence in our data hence supports the conjecture that the polarization of voters' worldviews might be partially driven by a tendency to avoid information in opposite directions.
Interestingly, we find that these behavioral patterns strongly correlate with the subjects' political attitudes: while supporters of the Democratic Party are prone to exhibit pro-social information avoidance, Republicans rather engage in selfserving information avoidance. Hence, the different types of motivated information avoidance seem to reinforce the favored worldview and might lead exactly to the polarization mentioned above. The ideological determinants of the direction and strength of motivated information avoidance are certainly interesting subjects for future research.
Our findings may also contribute to explaining why some studies find evidence of willful information avoidance, whereas others do not. Our results suggest that wiggle room is exploited by decision makers with intermediate preference intensities, i.e. by individuals who tend to choose pro-socially under full information and selfishly under hidden information. If decision makers already choose fairly egoistically under full information or relatively pro-socially under hidden information, there is no scope for the exploitation of wiggle room. Whether or not willful ignorance can be detected thus depends on the preference intensities, which depend on the subject pool and the decision context. Results from previous experiments suggest that there are fewer individuals with intermediate preference intensities in environmentally relevant decision contexts ( Momsen and Ohndorf, 2020;Lind et al., 2019 ) and in reciprocal decisions ( Friedrichsen et al., 2022;van der Weele et al., 2014 ). 15 Moreover, we find that uncertainty in the revelation process induces decision makers to reveal more information. We surmise that the remaining doubt about the correctness of the revealed information still leaves some wiggle room that can be used to choose the self-serving option, despite the apparent conflict of interest. We also find that decisions under full information become more selfish when the constituency in which decisions are made increases, while we find no evidence of the group size influencing choices in the treatments where a subset of the payments is hidden.
Furthermore, our results speak to the literature on nudging ( Thaler and Sunstein, 2008 ), as the provision or withholding of information may have large consequences on choices. Choice architects may strongly influence behavior by hiding the payoff consequences of the available options: those who would like to be generous but tend to fall prey to the temptation to maximize their own monetary payoff if aware of a conflict of interests will remain ignorant and thus manage to better align their intentions with their actions. Those strongly interested in their own payoff, however, will reveal the hidden pieces of information and remain unaffected by the nudge. A huge benefit of this nudge would be that it is fairly innocent ( Thaler and Sunstein, 2003 ): decision makers are aware of the information structure and can thus decide if they want to make use of the commitment device by remaining uninformed.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability
Data will be made available on request.