Journal of Risk and Uncertainty

, Volume 52, Issue 2, pp 117–136

Strategic self-ignorance

  • Linda Thunström
  • Jonas Nordström
  • Jason F. Shogren
  • Mariah Ehmke
  • Klaas van’t Veld
Article

DOI: 10.1007/s11166-016-9236-9

Cite this article as:
Thunström, L., Nordström, J., Shogren, J.F. et al. J Risk Uncertain (2016) 52: 117. doi:10.1007/s11166-016-9236-9

Abstract

We examine strategic self-ignorance—the use of ignorance as an excuse to over-indulge in pleasurable activities that may be harmful to one’s future self. Our model shows that guilt aversion provides a behavioral rationale for present-biased agents to avoid information about negative future impacts of such activities. We then confront our model with data from an experiment using prepared, restaurant-style meals—a good that is transparent in immediate pleasure (taste) but non-transparent in future harm (calories). Our results support the notion that strategic self-ignorance matters: nearly three of five subjects (58%) chose to ignore free information on calorie content, leading at-risk subjects to consume significantly more calories. We also find evidence consistent with our model on the determinants of strategic self-ignorance.

Keywords

Strategic ignorance Calorie information avoidance Guilt aversion Self-control 

JEL Classification

D03 D81 D83 

1 Introduction

In classical expected-utility theory, the value of information is non-negative (Machina 1989). A person should never be worse off gathering free information about a choice.1 Dana et al. (2007) find, however, that if the choice affects the well-being of other people, and if the person feels conflicted about doing what he wants versus “doing the right thing” (based on social norms such as fairness), he may exercise strategic ignorance: he reduces his internal conflict by choosing to avoid free information on what he “should” do.2

In this paper, we present evidence that a person may similarly exercise strategic self-ignorance when the choice he faces affects only his own well-being. We show that if a person engages in a pleasurable activity that may be harmful to his future self, and experiences feelings of guilt when over-indulging in that activity, he may avoid free information on the future consequences of his actions to reduce his guilt.

We consider a guilt-averse person who experiences an inner conflict due to present-biased preferences: he believes he should behave rationally (i.e., in a manner consistent with discounting utility at a constant rate over all future time periods), but since his true preferences are present-biased, he puts too much emphasis on today’s well-being.3 He over-indulges in activities that impose negative externalities on his future selves.4 We assume that feelings of guilt arise when the present-biased person gives in to immediate gratification, at the expense of future well-being.

We also assume that self-ignorance can be bliss. By ignoring information on the potential harmful future consequences of present activities, a person mitigates the inner conflict between what he should and wants to do and reduces feelings of guilt. As a result, he may use ignorance strategically, i.e., as an excuse to pursue his preferences for immediate gratification.5

People may exercise strategic self-ignorance to over-engage in a wide array of risky activities—impulsive spending, gambling, alcohol or drug abuse, extreme sports, unprotected sex—or to under-engage in protective activities—saving for old age, buying insurance, getting health check-ups, exercising. All these activities involve a tradeoff between transparent immediate pleasure and less-transparent future harm, which provides scope for ignoring information about that harm. Our lab experiment, designed to test the prevalence of strategic self-ignorance, uses restaurant-style ready meals; consumption of such meals, too, is transparent in immediate pleasure (taste) but non-transparent in future harm (calories). We find strong evidence of strategic self-ignorant behavior, and support for our model’s predictions about the determinants of such behavior.

2 Analytical framework

Consider a person with present-biased intertemporal preferences. Following self-control models (e.g., O’Donoghue and Rabin 1999), we represent his utility from time t onwards as
$$ {U}_t\left({u}_t,{u}_{t+1},\dots, {u}_T\right)={E}_t\left({u}_t+\beta {\displaystyle {\sum}_{\tau =t+1}^T{\delta}^{\tau -t}{u}_{\tau }}\right), $$
(1)
where ut is utility in period t, δ is a standard discount factor, and 0 ≤ β < 1 represents a time-inconsistent preference for immediate gratification, i.e., the present-bias. Assume the person is naïve, i.e., unaware that his present-bias will persist over time—he believes his present-bias will vanish tomorrow, but when tomorrow comes, he still has the same present-biased preferences. Being present-biased, the person is prone to over-indulging today in activities with harmful future consequences. Being naïve, he each time perceives this behavior as limited in scope (“just this once”). When tomorrow comes, however, he makes the same choice to over-indulge, and so on. Tomorrow’s self discounts future utility more than today’s self would prefer (and expect) tomorrow’s self to do.
The novel element we add to this standard self-control model is a feeling of guilt about the time-inconsistent behavior. Suppose at time t the person faces a choice at what level x to engage in a potentially harmful activity. To fix ideas, let x be the quantity consumed of a food that may be either “healthy” (low calorie) or “unhealthy” (high calorie). Assume his utility from the activity has three additively separable components, et, fτ and gt, each of which depends on x. Component et (x) represents his immediate “enjoyment.” We assume this component is strictly concave, with an interior maximum. If the person knows that the food is healthy, then this is the only component of utility. The person then optimally chooses xih (superscript ih for “informed” that the food is “healthy”) given by first-order condition
$$ {e}_t^{\hbox{'}}(x)=0. $$
(2)

If, however, the food is unhealthy, two additional utility components kick in. The first additional component, fτ(x), enters utility negatively, and represents the harmful “future consequences” of his consumption at some time τ > t. We assume this component is strictly increasing and weakly convex. We also premultiply it by a weighting factor ϕ that parameterizes the “severity” of the harm, allowing us to investigate below how the person’s behavior changes with that severity. To avoid uninteresting corner solutions, we assume ϕ\( {f}_{\tau}^{\hbox{'}} \)(0) < \( {e}_t^{\hbox{'}} \)(0).

The second additional component, gt(x − x*), also enters utility negatively, and represents the guilt experienced by the person if he consumes more of the unhealthy food than some reference level x* that he feels he “should” consume. We assume this component is strictly increasing and weakly convex in x − x* for all x > x*. We also premultiply it by a weighting factor γ that parameterizes the person’s “sensitivity” to guilt, again to allow us to investigate below how the person’s behavior changes with that sensitivity. To avoid uninteresting special cases, we assume \( {g}_t^{\hbox{'}} \)(0) = 0. Note our assumption, indicated by the subscript t, that guilt is experienced at the time of consumption. This is consistent with Baumeister et al.’s (2007) argument, based on an extensive review of the empirical literature in social psychology, that guilt operates as a feedback system: it affects behavior mainly by reminding a person of negative outcomes resulting from past transgressions, at times when he may transgress again.

A natural way to determine the reference level x* that the person feels he “should” consume is to consider how much of the unhealthy food the person would anticipate consuming at time t when pondering the decision at some earlier time s < t. Being naïve about his present-bias, the person would anticipate rationally weighing the immediate enjoyment et(x) against the future consequences ϕfτ(x), whereby the future consequences are discounted only by the standard factor δτ−t. For simplicity we normalize δ to unity. At time s < t, the person therefore anticipates facing the problem
$$ \underset{x}{ \max }{U}_t(x)={e}_t(x)-\phi {f}_{\tau }(x), $$
(3)
with solution x* given by first-order condition
$$ {e}_t^{\hbox{'}}(x)-\phi {f}_{\tau}^{\hbox{'}}(x)=0. $$
(4)

When period t actually arrives, however, the preference β for immediate gratification kicks in, and thereby the possibility of over-consumption and associated guilt gt(x − x*).

The basic idea is the same as in the existing literature on guilt aversion, in which guilt arises when a person betrays another person’s expectations. Here, a person experiences guilt if he betrays his own expectations about today’s consumption.6

The problem that the person actually faces at time t if he knows that the food is unhealthy becomes
$$ \underset{x}{ \max }{U}_t(x)={e}_t(x)-\beta \phi {f}_{\tau }(x)-\gamma {g}_t\left(x-{x}^{*}\right), $$
(5)
with solution xiu (superscript iu for “informed” that the food is “unhealthy”) given by first-order condition
$$ {e}_t^{\hbox{'}}(x)-\beta \phi {f}_{\tau}^{\hbox{'}}(x)-\gamma {g}_t^{\hbox{'}}\left(x-{x}^{*}\right)=0. $$
(6)

Comparing (4) and (6) shows that xiu > x*, so the person over-consumes and experiences guilt.

Suppose the person initially does not know whether the food is healthy. He only has a prior belief that the food is unhealthy with probability θ. If offered free information on whether the food is unhealthy, even a present-biased person subject to feelings of guilt would always choose to obtain the free information so that he can “do the right thing,” i.e., make a fully informed consumption decision. But if ignorance about the potential harm from consuming the food reduces his feelings of guilt, the person may gain from staying uninformed.

A variety of modelling approaches could be used to generate such guilt reduction. Consider three general approaches and how they correspond to the existing literature. First, if guilt is a strictly convex function of the expected violation x − x*, then since by Jensen’s inequality g(E[x − x*]) < E[g(x − x*)] = θg(x − x*), the person will feel less guilt from choosing a given consumption level under ignorance than under full information. This is analogous to the “information preference” approach used by Barigozzi and Levaggi (2010), building on Caplin and Leahy (2001) and Kőszegi (2003), to explain why patients might optimally avoid information about their health.7 Second, even if guilt is linear in the violation, ignorance may allow the person to downplay internally the probability θ that the food is unhealthy. This is analogous to the “optimal expectations” approach used by Oster et al. (2013), extending Brunnermeier and Parker (2005), to explain why patients might optimally avoid diagnostic tests. Third, ignorance may reduce the salience of the violation, and thereby of the induced guilt. This is analogous to the “elation and disappointment” approach used by Bell (1985) to explain why people might prefer delayed resolution of uncertainty.8

Rather than choosing among these approaches, we adopt a “reduced form” approach that captures elements of each.9 We posit that under ignorance, the guilt experienced by the person (conditional on his prior θ that the food is unhealthy) is not γgt(x − x*) but γ(1 − ρ)gt (x − x*), where ρ ϵ (0,1]. That is, for any of the three reasons mentioned above, staying ignorant reduces the intensity of his guilty feelings by a factor ρ. As with the weighting parameters ϕ and γ, we assume that ρ may vary across people: for some, the degree to which ignorance is bliss, i.e., how effectively information avoidance reduces guilt, is greater than for others. We therefore investigate below how the person’s behavior varies with ρ.10

The person’s optimization problem under ignorance becomes
$$ \underset{x}{ \max }E{U}_t(x)={e}_t(x)-\theta \beta \phi {f}_{\tau }(x)-\theta \gamma \left(1-\rho \right){g}_t\left(x-{x}^{*}\right), $$
(7)
with solution xn (superscript n for “non-informed”) given by first-order condition
$$ {e}_t^{\hbox{'}}(x)-\theta \beta \phi {f}_{\tau}^{\hbox{'}}(x)-\theta \gamma \left(1-\rho \right){g}_t^{\hbox{'}}\left(x-{x}^{*}\right)=0. $$
(8)

Comparing (6) and (8) shows that xn > xiuif the food is unhealthy, the person consumes more under ignorance than under full information.11

For a person to gain from ignoring free information, his indirect utility from doing so, which we can write (hereafter dropping time subscripts as understood) as
$$ {V}^n=\left(1-\theta \right)e\left({x}^n\right)+\theta \left\{e\left({x}^n\right)-\beta \phi f\left({x}^n\right)-\gamma \left(1-\rho \right)g\left({x}^n-{x}^{*}\right)\right\},\kern4.5em (9) $$
must exceed that from obtaining the information,
$$ {V}^i=\left(1-\theta \right)e\left({x}^{ih}\right)+\theta \left\{e\left({x}^{iu}\right)-\beta \phi f\left({x}^{iu}\right)-\gamma g\left({x}^{iu}-{x}^{*}\right)\right\}.\kern4.75em (10) $$
By revealed preference and strict concavity of the utility function, we have that e(xih) > e(x) for any x ≠ xih, and similarly e(xiu) − βϕf(xiu) − γg(xiu − x*) > e(x) − βϕf(x) − γg(x − x*) for any x ≠ xiu. It follows that if ignorance did not reduce the person’s guilt, so ρ = 0, Vi would strictly exceed Vn, making ignorance suboptimal. If ignorance does reduce guilt, however, a useful reference point is the indirect utility level
$$ \widehat{V}=\left(1-\theta \right)e\left({x}^{ih}\right)+\theta \left\{e\left(\widehat{x}\right)-\beta \phi f\left(\widehat{x}\right)-\gamma \left(1-\rho \right)g\left(\widehat{x}-{x}^{*}\right)\right\} $$
(11)
that the person would obtain if he were both fully informed and experienced the same reduction in guilt as a self-ignorant person. Since \( e\left(\widehat{x}\right)-\beta \phi f\left(\widehat{x}\right)-\gamma \left(1-\rho \right)\left(\widehat{x}-{x}^{*}\right)>e(x)-\beta \phi f(x)-\gamma \left(1-\rho \right)\left(x-{x}^{*}\right) \) for any \( x\ne \widehat{x} \), including both xn and xiu, we have that both \( {V}^n<\widehat{V} \) and \( {V}^i<\widehat{V} \). The comparison between Vn and Vi therefore depends on which falls further short of reference utility \( \widehat{V} \), which in turn depends on the person’s characteristics β, ϕ, γ, ρ, and θ.

Suppose, for example, that ρ = 1, so the person experiences no guilt at all if he remains ignorant and the guilt terms drop out of the expressions for Vn and \( \widehat{V} \). In this case, if additionally β = 1, so the person has no present-bias, the guilt term drops out of the expression for Vi as well, because without present-bias, xiu = x*. As a result, we have \( {V}^i=\widehat{V}>{V}^n \), so the person strictly prefers to become informed. If, at the other extreme, β = 0, the future-consequences term βϕf(x) drops out of all three indirect-utility expressions, and \( {x}^n=\widehat{x} \). As a result, we have \( {V}^n=\widehat{V}>{V}^i \), so the person strictly prefers to stay ignorant. Moreover, since by the envelope theorem ∂Vn/∂β = −θϕf(xn) < − θϕf(xiu) = ∂Vi/∂β, there must be a critical value of β between 0 and 1 where the preference switches.

In sum, we find that a guilt-averse, present-biased person may optimally use ignorance as an excuse to over-consume harmful goods. We label this behavior strategic self-ignorance.

When we take the model to our experimental data, we need to add a random term ε reflecting unobserved attributes of either the subject or the choice situation. The subject chooses to ignore free information if Vn − Vi + ε ≥ 0, i.e., with probability
$$ \Pr (n) \equiv \Pr \left(\varepsilon \ge {V}^i-{V}^n\right) = 1-F\left({V}^i-{V}^n\right), $$
(12)
where F is the cumulative distribution of ε. As a result, for any parameter y,
$$ \frac{d \Pr (n)}{dy}\circeq \frac{d\left({V}^n-{V}^i\right)}{dy}=\frac{\partial {V}^n}{\partial y}-\frac{\partial {V}^i}{\partial y}, $$
(13)
where “≗” denotes equality of sign, and the second step follows by the envelope theorem.
To understand the factors underlying the comparative statics, it is useful to decompose the change in utility Vn− Vi from staying ignorant into four probability-weighted terms:

The first term, Lh, represents the enjoyment lost when the food is healthy, by consuming xn rather than the optimal, higher level xih. The second term, Gu, represents the enjoyment gained by consuming xn also when the food is unhealthy, rather than the guilt-induced, lower level xiu that the person perceives to be optimal when informed. The third term, Lu, represents the loss incurred because the increase in consumption of unhealthy food also increases perceived future harm. Lastly, the fourth term, ∆g, represents the guilt avoided through self-ignorance. Paradoxically, it can be shown (by numerical example), that unless guilt reduction is complete, i.e., ρ = 1, this term need not be positive. Even though ρ must be positive for self-ignorance to ever be optimal (choosing ignorance must reduce feelings of guilt on the margin), the self-ignorant person may increase his consumption so much that he ends up feeling more guilt overall. As we show below, this implies also that a person with greater guilt sensitivity γ may paradoxically have less incentive to avoid guilt through ignorance.

Consider first how the incentive changes with β and ϕ. Using (14), we can write
$$ \frac{d \Pr (n)}{d\beta}\circeq -\theta \frac{\partial {L}^u}{\partial \beta }=-\theta \phi \left[f\left({x}^n\right)-f\left({x}^{iu}\right)\right]<0. $$
(15)

An increase in β, which implies a reduction in the person’s present-bias, increases perceived future harm from consuming unhealthy food in both the uninformed and informed states. Because consumption is higher in the uninformed state, however, the perceived increase in harm is larger as well. The incentive to stay ignorant unambiguously falls.

Also,
$$ \frac{d \Pr (n)}{d\phi}\circeq \theta \left\{\frac{\partial {\varDelta}^g}{\partial \phi }-\frac{\partial {L}^u}{\partial \phi}\right\}=-\theta \beta \left[f\left({x}^n\right)-f\left({x}^{iu}\right)\right]-\theta \gamma \left[{g}^{\prime}\left({x}^{iu}-{x}^{*}\right)-\left(1-\rho \right){g}^{\prime}\left({x}^n-{x}^{*}\right)\right]\frac{d{x}^{*}}{d\phi}\gtreqless\ 0. $$
(16)

An increase in the severity of future consequences, ϕ, has two effects. First, it increases perceived future harm, just as an increase in β does. All else equal, this increases the person’s incentive to become informed, since he can then reduce the harm if the food turns out to be unhealthy. Second, an increase in ϕ also intensifies the person’s guilt in the informed state, by lowering the benchmark level x* that he feels he ought to consume.12 If ρ = 1, so guilt reduction from ignorance is complete, this second effect unambiguously increases his incentive to stay ignorant. If ρ < 1, however, the second effect is ambiguous, depending among other factors on the curvature of the guilt function. Overall, the effect of an increase in ϕ is ambiguous even if ρ = 1.

Next,
$$ \frac{d \Pr (n)}{d\gamma}\circeq \theta \frac{\partial {\varDelta}^g}{\partial \gamma }=\theta \left[g\left({x}^{iu}-{x}^{*}\right)-\left(1-\rho \right)g\left({x}^n-{x}^{*}\right)\right]\gtreqless 0. $$
(17)

If ρ = 1, an increase in the person’s sensitivity to guilt, γ, unambiguously increases his gain from avoiding that guilt, and thereby his incentive to stay ignorant. If ρ < 1, however, the effect is ambiguous. Although Pr(n) must be small at low values of γ (since Vi > Vn in the limit as γ → 0), it may be relatively large at intermediate values of γ, where the positive Gu term in (14) dominates, but then relatively small again at large values of γ, where the ∆g term may be negative and dominate.

Lastly,
$$ \frac{d \Pr (n)}{d\rho}\circeq \theta \frac{\partial {\varDelta}^g}{\partial \rho }=\theta \gamma g\left({x}^n-{x}^{*}\right)>0. $$
(18)

An increase in the effectiveness of ignorance at reducing guilt unambiguously increases the person’s incentive to stay ignorant.

For ease of reference, we summarize our model’s predictions in Table 1.
Table 1

Model parameters and predicted comparative statics

Parameter

Symbol

Effect on Pr(n)

Present-bias1

β

Severity of future consequences

ϕ

?

Sensitivity to guilt

γ

?

Guilt reduction from ignorance2

ρ

+

1β = 1 implies no present-bias; β = 0 implies complete present-bias

2ρ = 0 implies no reduction; ρ = 1 implies complete reduction

3 Experimental design

We used restaurant-style ready meals in our experiment, offering subjects a choice between two meals: chicken and bulgur, containing 900 calories, and roast beef and glass noodles, containing 490 calories.13 In line with our goal of examining potential information avoidance behavior, we verified up front that the calorie content of these meals was difficult to guess: subjects in a focus group were unable to determine which meal was high calorie and which low calorie.14

A recruitment firm was hired to recruit 150 people in the Stockholm area of different age, education, and income levels and of both genders. Vegetarians and people with food allergies were excluded for practical reasons. Subjects were told that they were going to participate in a survey during lunch hour and that lunch would be provided on site. They were also told they would, privately, be measured and weighed.15 The experiment lasted for an hour and subjects received a gift card worth SEK 400 (approximately USD 60) for participating. Of the 148 subjects who showed up to participate, 55 were assigned to the control group and 93 to the treatment group. Subjects participated in groups of size 15–20 and were reminded on arrival that they had been recruited to fill in a survey. Subjects were asked not to communicate with each other during the session. The survey elicited background characteristics including health, label knowledge, nutritional knowledge, and nutritional interest, and also included a set of questions designed to measure general time preferences.

We followed a six-step procedure in the experiment.
  1. Step 1.

    Subjects were told that they could choose between a meal containing chicken and bulgur or a meal containing roast beef and glass noodles (portions of the lunch meals were displayed), and that one of these meals (at that point still unknown to the subjects) contained 900 calories, whereas the other contained 490 calories. Subjects were also informed that their preferred meal would have to be consumed on site.16

     
  2. Step 2.

    On private sheets of paper, subjects were asked to rate the expected taste of the two meal choices (from 1 = “very bad” to 5 = “very good”), and then state their choice of meal.

     
  3. Step 3.

    Subjects in the control group were visually (on a sheet of paper) and verbally provided with information on which meal was high calorie and which was low calorie. Subjects in the treatment group were asked to choose one of two folded sheets of paper in front of them. They were told that the paper to the right contained information on the calorie content of both meals, while the paper to the left was blank. It was equally costly to choose ignorance as it was to become informed.17 The decision on whether or not to become informed was visible to other subjects in the group, which, if anything, should reduce the incentive to choose ignorance.18

     
  4. Step 4.

    Subjects were given the option of revising their meal choice, based on the information they got, or, in the case of the treatment group, chose to get.19

     
  5. Step 5.

    Subjects were asked to complete the survey and eat the meal they had chosen.

     
  6. Step 6.

    Subjects were individually weighed and measured in a separate room, and leftovers from subjects’ meals were weighed. The fraction of the meal consumed (categorized as 25, 50, 70, 85 or 100%) was multiplied by the meal’s total calorie content to determine each subject’s calorie consumption.20

     

4 Results I: Existence of strategic self-ignorance

Table 2 reports our results pertaining to the existence of strategic self-ignorance.
Table 2

Average calorie consumption, meal choices, and meal switching behavior

 

All

Treatment

Self-ignorant

Self-informed

Control

N

148

93

54 (58%)

39 (42%)

55

Ave. calorie consumption

564

582

626

522

532

 Number who chose chicken initially

92

56

29 (52%)

27 (48%)

36

 For those, ave. calorie consumption

630

676

798

546

558

  Number who stuck with chicken

62

43

29 (67%)

14 (33%)

19

  For those, ave. calorie consumption

724

741

798

624

684

  Number who switched to beef

30

13

0

13

17

  For those, ave. calorie consumption

436

462

n.a.

462

417

 Number who chose beef initially

56

37

25 (68%)

12 (32%)

19

 For those, ave. calorie consumption

450

440

426

470

473

  Number who stuck with beef

53

35

25 (71%)

10 (29%)

18

  For those, ave. calorie consumption

426

418

426

397

444

  Number who switched to chicken

3

2

0 (0%)

2 (100%)

1

  For those, ave. calorie consumption

855

833

n.a.

833

900

Calorie consumption is missing for three subjects in the control group, all of whom chose beef initially and stuck with that choice; calorie averages in italics exclude those three subjects

Of all subjects in the treatment group, 58% (54 out of 93) chose not to learn the calorie content of the meals. However, this finding by itself need not imply that the information choice was strategic—some subjects may simply not care about calories, and hence not worry about harmful future effects or feel any guilt. Investigating strategic behavior requires examining the subgroup of 56 subjects in the treatment group and 36 subjects in the control group who indicated an initial preference for the high-calorie meal. This is the subgroup for which we can observe potential over-consumption due to self-ignorance. Any subjects in this subgroup who did not care about calories should consume similar amounts regardless of what group they were assigned to and, within the treatment group, regardless of their information choice. Our model predicts, however, that any subjects who did care, and who found themselves assigned to the treatment group, might avoid information strategically, using self-ignorance as an excuse to increase their calorie consumption.

For this subgroup—hereafter referred to as high-calorie-meal “lovers,” for the sake of brevity21—we find that having the option to stay ignorant of the meal’s calorie content indeed significantly increased average calorie consumption. Those with this option, i.e., those in the treatment group, consumed on average 676 calories, whereas those without this option, i.e., those in the control group, consumed on average 558 calories. A t-test strongly rejects the hypothesis of equal intake (p-value = 0.005). Figure 1 adds detail by comparing kernel density estimates of both groups’ calorie consumption. The density for the treatment group is heavily skewed towards higher values, and a non-parametric Kolmogorov-Smirnov test strongly rejects the null that the underlying distributions of calorie-intake levels are the same (p-value = 0.022).
Fig. 1

Kernel density estimates for subjects who preferred the high-calorie meal

The higher average calorie intake of the 56 high-calorie-meal lovers in the treatment group is moreover clearly driven by the behavior of self-ignorant subjects. Compared to the control group’s average intake of 558 calories, the average intake of 798 calories by the 29 subjects who ignored information is significantly higher (p-value < 0.001), whereas the average intake of 546 calories by the 27 subjects who chose information is not significantly different (p-value = 0.767).22 The kernel density estimates shown in Fig. 2 similarly indicate that the skew of the treatment group’s density shown in Fig. 1 comes from self-ignorant subjects, whereas the density for informed subjects seems similar to that of the control group. Kolmogorov-Smirnov tests add further support: the control-group distribution differs significantly from that for self-ignorant subjects in the treatment group (p-value < 0.001), but not from that for informed subjects (p-value = 0.571).
Fig. 2

Kernel density estimates for treatment-group subjects who preferred the high-calorie meal

We conclude that our data provide strong empirical support for the existence of strategic self-ignorance.23

5 Results II: Determinants of self-ignorance

We next explore the determinants of the choice to ignore calorie information. The analytical framework of Section 2 suggests that the probability of choosing self-ignorance should be increasing in a person’s present-bias, i.e., decreasing in β, and increasing in his ability to avoid guilt through ignorance, as captured by parameter ρ. The net effects of the person’s sensitivity to guilt, γ, and concern about future consequences, ϕ, are both ambiguous, however.24 Table 3 provides descriptive statistics of the covariates used in our analysis.
Table 3

Descriptive statistics of Probit covariates

 

All

Treatment

Control

Variable

mean

sd

min

max

mean

sd

mean

sd

t-test

β

1.005

0.032

0.848

1.157

1.002

0.031

1.011

0.035

(1.669)*

Female

0.534

0.501

0

1

0.538

0.501

0.527

0.504

(−0.121)

Age

39.950

12.704

20

61

40.153

12.748

39.630

12.748

(−0.236)

BMI

25.182

4.185

18

48

25.097

3.909

25.327

4.647

(0.323)

Health knowledge

6.483

1.528

2

10

6.407

1.584

6.611

1.433

(0.778)

Health concern

0.223

0.418

0

1

0.237

0.427

0.200

0.404

(−0.513)

Light exercise

5.905

8.277

0

50

6.413

9.085

5.055

6.707

(−0.963)

Moderate exercise

4.062

7.355

0

70

4.739

8.840

2.907

3.416

(−1.458)

Strenuous exercise

2.083

2.815

0

14

2.082

2.942

2.086

2.614

(0.0101)

Smoker

0.174

0.380

0

1

0.135

0.343

0.236

0.429

(1.565)

Above-ave. income

0.559

0.498

0

1

0.582

0.496

0.519

0.504

(−0.745)

College education

0.606

0.490

0

1

0.596

0.494

0.623

0.489

(0.318)

Exercise measured in hrs/week

*p < 0.10

To estimate subjects’ present-bias β, we use two hypothetical questions. The first asked subjects if they preferred receiving SEK 5,000 today or SEK X in 2 months, for a range of X values from SEK 5,010 to SEK 5,905. The second asked subjects if they preferred receiving SEK 5,000 in 1 month or SEK X in 3 months. The present-bias estimate β was calculated as the ratio of the two X values at which a given subject switched preferences (the second divided by the first). As shown in Table 3, β ranged from 0.85 to 1.16 in the sample, with a mean of close to 1. This implies that subjects were rational on average, but some were prone to over-consume in the present relative to what time-consistent behavior would dictate, while others were prone to under-consume. This is in line with previous empirical findings on self-control problems (Ameriks et al. 2007).

We are unable to directly measure subjects’ sensitivity to guilt from high-calorie food consumption, or their experienced guilt reduction from ignorance. Rather, we use Female and Age as variables that are associated with guilt in food consumption: Dewberry and Ussher (2001); Wansink et al. (2003); and Steenhuis (2008) find that women and younger people generally feel guiltier about unhealthy food consumption than do men and older people.

We include BMI, the subject’s Body Mass Index calculated from on-site height and weight measurements,25 as a variable that may affect the severity of future consequences from unhealthy food consumption. Wansink et al. (2003) find that perceived consequences are also affected by age: compared to younger people, older people are less likely to perceive consumption of high-calorie snacks as unhealthy.

Nayga (1996); Cowburn and Stockley (2005); and Drichoutis et al. (2005) find that people with greater knowledge of and interest in health issues more frequently pay attention to nutritional information, possibly because they are more aware of health impacts. We include a number of variables that plausibly capture these factors. Health knowledge is the number of correct responses by the subject on 11 health-related questions.

Health concern is a dummy based on subjects’ level of agreement (from 1 = “totally disagree” to 7 = “fully agree”) with the statement “I am very concerned about the food I eat being healthy.” The dummy takes value 1 if the level of agreement is 4 or higher. Light, moderate, and strenuous exercise are the average hours per week that the subject reported engaging in each. Lastly, we include income and education as additional demographic controls. Above-ave. income is a dummy taking value 1 if the subject’s stated income exceeded SEK 20,000/ month. College education is a dummy taking value 1 if the subject had any college or other post-secondary education.

Table 4 reports Probit estimates of the average marginal effects of our covariates on subjects’ decision to ignore calorie information.26 Note that we multiply the β variable by 100, so the reported effect is that of an increase by 0.01. Also note that the effects for the dummy variables Female, Smoker, Above-ave. income, and College education are discrete, i.e., estimates of Pr(n|y = 1) − Pr(n|y = 0) rather than of ∂ Pr(n) / ∂y.
Table 4

Determinants of self-ignorance

 

(1)

(2)

 

∂Pr(n)/∂x

∂Pr(n)/∂x

β*100

0.025* (0.013)

−0.078*** (0.026)

Female

−0.356*** (0.091)

−0.388*** (0.081)

Age

0.021*** (0.004)

0.022*** (0.004)

BMI

−0.047*** (0.012)

−0.034*** (0.013)

Health knowledge

−0.069** (0.034)

−0.094*** (0.028)

Health concern

−0.208** (0.094)

−0.295*** (0.080)

Light exercise

0.000 (0.006)

−0.004 (0.005)

Moderate exercise

−0.015 (0.011)

−0.008 (0.007)

Strenuous exercise

0.060*** (0.023)

0.050** (0.023)

Smoker

0.350*** (0.074)

0.408*** (0.056)

Above-ave. income

−0.210*** (0.081)

−0.219*** (0.066)

College education

0.189** (0.085)

0.310*** (0.063)

N

79

74

pseudo R2

0.42

0.56

Entries show average marginal effects for continuous variables, and average discrete effects—i.e., Pr(n|y = 1) − Pr(n|y = 0)—for dummy variables. Standard errors are in parentheses. Model (2) drops five β outliers

*p < 0.10, **p < 0.05, ***p < 0.01

The hypothetical nature of the questions used to estimate subjects’ present-bias may have contributed to five clear outlier values of β (one lower than 0.9 and four larger than 1.05) that turn out to strongly influence our estimate of β’s effect. A non-parametric smooth, using locally weighted regression (see Fig. 3 in the Appendix), indicates that, except at these outliers, self-ignorance has the negative relationship to β predicted by our model. Removing the outliers, as we do in column (2) of Table 4, does not materially change our results for any of the other covariates.

Focusing on column (2), we find that all covariates other than the exercise variables are highly statistically significant, and economically significant as well. An increase in β by 0.01, i.e., just one percentage point, reduces the probability of choosing self-ignorance by 8 percentage points; compared to men, women are on average 39 percentage points less likely to be self-ignorant; aging by just 1 year increases the probability of self-ignorance by more than 2 percentage points; for every unit increase in BMI, the probability of self-ignorance drops by more than 3 percentage points, etc.27

Our finding that women and younger people are much less likely to ignore information compared to men and older people is consistent with the above-mentioned findings in the psychology literature that the former tend to feel guiltier about unhealthy food consumption. However, we showed in Section 2 that, as long as ignorance leaves some residual guilt, the predicted relationship between self-ignorance and guilt sensitivity is non-monotonic; in particular, a person with relatively low guilt sensitivity (e.g., an elderly man) may gain from ignorance, whereas a person with high sensitivity (e.g., a young woman) may be better off choosing to become informed, and then reducing guilt by consuming less.

We also showed in Section 2 that an increase in future harm has an ambiguous effect on self-ignorance. First, future harm drives down the “guilt-free” reference consumption level x* thereby increasing the person’s guilt and possibly his incentive to avoid that guilt through ignorance (this effect is itself ambiguous, though, when guilt avoidance is imperfect). The counter-point is that it increases the person’s incentive to reduce expected harm by becoming informed. It seems plausible that the second effect will typically dominate, consistent with our finding of a strong negative relationship between self-ignorance and BMI. The negative relationship for the health-knowledge and health-concern variables is consistent with the second effect dominating as well.28

Two offsetting effects may explain the insignificance of the exercise measures. Subjects that exercise regularly are likely to be more interested in health issues, and more aware of high-calorie foods’ health impacts. However, they may also be less concerned about those impacts, given that they burn more calories anyway. Our finding that strenuous exercise (but not moderate or light) is marginally positively related to self-ignorance is consistent with this explanation.

Our finding that smokers are much more likely to avoid health information seems unlikely to be due to differences with non-smokers in either guilt sensitivity or future harm from overeating. More plausibly, it indicates a person’s high ability to reduce guilt through ignorance—a facility likely also applied to guilt about future consequences of smoking. In terms of our model, smokers may have a high value of ρ.

As for the strong effects of income and education (both unanticipated), it seems similarly unlikely that either of these variables correlates with guilt sensitivity or harm from overeating. We can speculate that these variables, too, might correlate with the ability to downplay guilt under uncertainty, ρ, a point which we leave for future research.

6 Concluding remarks

Classical expected-utility theory presumes the value of information is always non-negative. Herein we define a counterpoint to this view—strategic self-ignorance. People may avoid free information and use their ignorance strategically, i.e., as an excuse to over-indulge in activities that provide immediate pleasure and potential future harm. We show that a person with present-biased preferences who is conflicted about doing what he wants versus “doing the right thing” (as defined by time-consistent preferences) may benefit from ignoring free information on future consequences of his actions, if doing so reduces his internal conflict—for him, ignorance can be bliss.

We empirically test for the existence of strategic self-ignorance in relation to consumption of high-calorie food. Based on an experiment using restaurant-style ready meals, we find strong evidence of strategic self-ignorant behavior. Nearly 60% of our subjects chose to ignore free information on the calorie content of their meal, and these subjects consumed significantly more calories. We also find that people with stronger present-bias, men, younger people, people with a lower BMI, people with little knowledge of or interest in health, smokers, low-income earners, and highly-educated people are more likely to be self-ignorant.

Strategic self-ignorance may apply to a wide range of behavior, contributing not just to over-engagement in risky activities such as impulsive spending, gambling, alcohol or drug abuse, extreme sports, unprotected sex, or eating high-calorie foods, but also to under-engagement in protective activities such as saving for retirement, buying insurance, getting health check-ups, or exercising. People may use ignorance of the risks associated with their behavior to allow themselves to “enjoy the moment,” leaving their future selves to deal with debts, hangovers, broken bones, unwanted pregnancies, and health issues.

The central implication of strategic self-ignorance is that, because risk information is not free—it comes with a psychic cost of guilt—policies that rely on information to discourage over-indulgence in risky behavior may be inefficient. For instance, to combat the so-called “obesity epidemic” in the US, the Patient Protection and Affordable Care Act of 2010 (“Obamacare”) mandated that, starting in 2013, chain restaurants post calorie counts on their menus. But studies of a similar mandate enacted by New York City in 2008 have found that menu labeling has little or no impact on food consumption (see Borgmeier and Westenhoefer 2009; Elbel et al. 2009, 2011; Vadiveloo et al. 2011). Strategic self-ignorance could help explain why labeling does little to encourage healthier choices.

Overcoming the psychic cost of information may require either imposing additional costs on ignorance or providing additional benefits to being informed. Liability waivers, for instance, raise the cost of ignorance—they shift the responsibility of future harm to the person that engages in a risky activity, and could be used in areas ranging from extreme sports to food choices. Benefits from being informed could, for instance, consist of small rewards at schools and work places (e.g., financial rewards, extra credits, time off) associated with learning about risky or protective behavior. Increasing the salience of costs and benefits may also help.

Alternatively, strategic self-ignorant behavior may be pre-empted altogether by making use of people’s tendency to go with default or salient options, thereby encouraging optimal decisions without relying on people fully informing themselves. Examples of “nudges” (Thaler and Sunstein 2003) of this type are automatic enrollment in retirement or insurance plans and careful placement of healthy food options in supermarkets or cafeterias. We encourage future research on policy measures aimed at counteracting strategic self-ignorance, and research on strategic self-ignorance in the context of activities other than consuming tasty meals that stick around the waistlines of future selves.

Footnotes
1

Wakker (1988) and Schlee (1990) show that the controversial “independence axiom” underlying expected utility theory is key to this implication; if the axiom is relaxed, an agent may strictly prefer to avoid costless information.

 
2

In a dictator game, Dana et al. (2007) find that 74% of dictators choose the fairer distribution of money between themselves and a recipient when they are informed of the impact of their choice on the recipient. When dictators may choose to remain ignorant of that impact, however, only 47% of dictators both choose to be informed and choose the fairer option (also see the comment by Larson and Capra (2009)). Van der Weele (2012) finds that even when ignorance is associated with a small cost, more than 30% of subjects are strategically ignorant.

 
3

There is a rich literature on dynamically inconsistent, or present-biased, preferences: see, e.g., Strotz (1955); Thaler (1980); Akerlof (1991); Ainslie (1992); Loewenstein and Prelec (1992); Laibson (1997); and O’Donoghue and Rabin (1999, 2003).

 
4

This holds for a person who is unaware of his present-bias (is naïve). A person who is aware of his present-bias (is sophisticated) may commit himself to a decision path equivalent to that of a time-consistent person. For more on the implications of naïveté vs. sophistication, see O’Donoghue and Rabin (1999).

 
5

Key to this finding, too, is that the person is naïve about his present-bias. Carillo and Mariotti (2000) show that if a person is sophisticated, he may use ignorance strategically as a commitment device, to mitigate his future selves’ preferences for immediate gratification.

 
6

Guilt aversion has previously been studied in inter-personal conflicts: people experience a utility loss if they betray other people’s expectations, thereby letting them down (see, e.g., Charness and Dufwenberg 2006; Vanberg 2008; Reuben et al. 2009; and Ellingsen et al. 2010). To our knowledge, our analysis is the first to incorporate guilt in an intra-person conflict.

 
7

Closely related, but less applicable to the setting of our model, is Hoy et al.’s demonstration (2014), building on Snow (2010) and Klibanoff et al. (2005), that information avoidance may be driven by ambiguity aversion rather than by anxiety over outcomes.

 
8

Later studies that extend Bell’s approach include Gul (1991); Jia et al. (2001); and Delquié and Cillo (2006).

 
9

Karlsson et al. (2009) develop a model that similarly combines all three approaches. In their model of what they label the “ostrich effect,” selectively ignoring bad news about the likely return on an asset reduces the impact on the agent’s utility from that news (a salience effect), slows down the rate at which he updates his loss aversion reference point (an optimal expectations effect), and affects the local curvature of her utility function (an information preference effect). They find that, consistent with their model’s predictions, Swedish and American investors monitor their portfolios less frequently when markets are flat or falling.

 
10

An important difference between our model and the three other general approaches to modeling information avoidance is that the latter are driven by the notion that ignorance affects forward-looking, or “anticipatory” utility, i.e., anxiety or excitement about future outcomes. Because of this, agents need not be present-biased for information avoidance to occur, and present-bias may reduce agents’ proneness to avoid information. In contrast, in our model ignorance affects guilt, which as we noted above (referring to Baumeister et al. 2007) is inherently backward-looking, reminding a person of negative outcomes resulting from past transgressions. Because of this, we find that present-bias is needed for information avoidance (without present-bias, there would be no violations, and therefore no guilt to reduce), and that proneness to information avoidance unambiguously increases in present-bias.

 
11

In contrast, comparing (4) and (8) shows that xn < xih: if the food is healthy, the person consumes less under ignorance than under full information.

 
12

From (4),

$$ \frac{d{x}^{*}}{d\phi }=\frac{f\hbox{'}\left({x}^{*}\right)}{e^{{\prime\prime}}\left({x}^{*}\right)-\phi f\hbox{'}\hbox{'}\left({x}^{*}\right)}<0. $$
 
13

For comparison, Dumanovsky et al. (2009) report that the average calorie content of fast-food lunches is 823 calories.

 
14

This is consistent with research by Burton et al. (2006) showing that people are generally unable to accurately determine the calorie content of prepared meals served away from home.

 
15

The recruitment firm reported that subjects’ willingness to participate in the study was unaffected by the fact that they would be measured and weighed.

 
16

This experimental set-up differs slightly from the set-up of our theoretical model: rather than deciding how much to consume of a single meal, subjects were asked to choose between two meals, knowing up front that one meal was high calorie and the other low calorie, but not knowing which meal was which. The only implication for our theoretical analysis is that the function e(x) mapping calories to enjoyment presumably differed across meals. That is, if we use A and B to denote the high- and the low-calorie meal, subjects who initially chose A faced the optimization problem modeled in Section 2 with enjoyment function eA(x), whereas subjects who initially chose B faced it with enjoyment function eB(x).

 
17

The information provided was short—it simply stated which meal contained what number of calories. Subjects already knew that one meal contained 490 calories and the other 900 calories. We left the “no information” sheet blank due to the risk of any message on that sheet distorting the results (e.g., if subjects chose the no-information sheet out of curiosity).

 
18

Evidence suggests that being observed, even by people other than the recipient, increases generosity in dictator games, while anonymity decreases it (see Hoffman et al. 1996; Bohnet and Frey 1999a, b; Andreoni and Petrie 2004; Soetevent 2005; Charness and Gneezy 2008; and Andreoni and Bernheim 2009). If being observed similarly pressures subjects to “do the right thing” even for behavior that does not directly impact others, then it may reduce their incentive to choose ignorance.

 
19

Since subjects had no reason to anticipate this switching option, it does not affect the predictions from our theoretical analysis.

 
20

Six subjects (all in the control group) lacked a recorded amount of consumption. We assigned these subjects 100% consumption of their consumed meal. In doing so, we assured that, if anything, the calorie consumption of our control group would be overestimated.

 
21

Importantly, not everyone in this subgroup ended up actually consuming the high-calorie meal, since subjects were given the option of switching meals after learning the meals’ calorie content. As shown in Table 2, of the 56 high-calorie-meal “lovers” in the treatment group, 27 chose to find out the calorie content, and of those, 13 subsequently switched to the low-calorie meal; of the 36 high-calorie-meal “lovers” in the control group, 17 switched.

 
22

As shown in rows 5–9 of Table 2, the more specific behavior of informed subjects is quite similar as well. Of the 27 treatment-group subjects who chose information and learned that their initial meal choice was high-calorie, 14 stuck with their meal choice but responded by eating on average less than the uninformed subjects did, while 13 responded by switching to the low-calorie meal. Similarly, of the 36 control-group subjects who received the same information, 19 stuck with their high-calorie meal choice but ate less, while 17 switched to the low-calorie meal.

 
23

Importantly, our data do not support any broader claims about the importance that subjects attached to calories relative to other factors that might affect the healthiness of meals. In particular, some subjects—perhaps even the majority—may have viewed the high-calorie chicken meal as healthier than the low-calorie beef meal, for example because they perceived beef to contain more fat than chicken, or bulgur to be more nutritious than glass noodles. Given that we randomized the assignment of subjects to the treatment and control groups, and given that the only difference between these groups was whether or not subjects had the option to ignore calorie information, any differences between subjects unrelated to that option—including differences in their perceptions of meal healthiness for non-calorie-related reasons—should in expectation “wash out,” i.e., result in no systematic differences in observed behavior across the two groups. Conversely, any systematic differences in behavior that we did observe must have been driven by subjects’ perceptions about calorie content alone.

 
24

Our analytical framework predicts how these parameters affect calorie consumption levels as well. Unfortunately, because the predictions vary with the subjects’ information, testing them requires separate regressions for xn, xiu, and xih. The resulting sample sizes turn out to be too small to estimate effects with reasonable precision. We therefore do not report the results.

 
25

BMI is calculated by dividing a person’s body mass (weight in kg) by the square of his or her height (in meters). A person with a BMI between 18.5 and 25 is considered normal, a BMI between 25 and 30 overweight, and a BMI above 30 obese.

 
26

Of the 93 subjects in the treatment group, we dropped 14 from the regression analysis due to missing values for one or more of the explanatory variables.

 
27

To check for robustness, we estimated a number of alternative specifications. Using agreement level 5 or 6 rather than 4 as the cutoff for coding the Health concern dummy reduces the magnitude and significance of its estimated effect, without changing the sign or materially affecting the remaining estimates. A dummy for whether subjects perceived themselves to be overweight has the same negative effect as actual BMI, but is statistically insignificant. A continuous estimate of income, using the midpoints of the income intervals that subjects were asked about, has the same negative and statistically highly significant effect as the dummy for above-average income. Similarly, a continuous estimate of education, using reasonable guesses at the years required to attain subjects’ reported degrees, has the same positive and statistically highly significant effect as the dummy for college education.

 
28

The negative effect of health knowledge appears to be driven by correct answers to two questions in particular: one pertaining to the recommended daily calorie intake of middle-aged women and one pertaining to dietary guidelines for trans fats.

 

Acknowledgments

We thank David Granlund for helpful comments and suggestions, and seminar participants at the FOI, University of Copenhagen, and participants at the AAEA Annual Meeting 2012. Financial support is gratefully acknowledged from the Swedish Council for Working Life and Social Research, the Swedish Retail and Wholesale Development Council, and University of Wyoming’s College of Agriculture and Natural Resources’ Global Perspectives Funds. The collection of data from human subjects for this study has been approved by the Ethics Committee at Lund University. We thank Carin Blom for excellent research assistance.

Supplementary material

11166_2016_9236_MOESM1_ESM.pdf (396 kb)
ESM 1(PDF 396 kb)

Copyright information

© Springer Science+Business Media New York 2016

Authors and Affiliations

  • Linda Thunström
    • 1
    • 2
  • Jonas Nordström
    • 3
    • 4
  • Jason F. Shogren
    • 2
  • Mariah Ehmke
    • 5
  • Klaas van’t Veld
    • 2
  1. 1.HUI Research ABStockholmSweden
  2. 2.Department of Economics and FinanceUniversity of WyomingLaramieUSA
  3. 3.Department of EconomicsLund UniversityLundSweden
  4. 4.Department of Food and Resource EconomicsUniversity of CopenhagenFrederiksberg CDenmark
  5. 5.Department of Agricultural and Applied EconomicsUniversity of WyomingLaramieUSA

Personalised recommendations