What You See Is All There Is ?
Result 11. In the Selected treatment, subjects’ naïveté regarding the selection problem ex- ex-hibits a strongly bimodal type distribution: people either fully take selection into account or
3.3 Learning Through Interaction
3.3.2 Exogenously Induced Disagreement
In a second step, I investigate how the different types respond in their updating behavior once they are forced to listen to those with different beliefs. To analyze this issue, I implemented treatment Exogenous. Here, subjects again solved the seven tasks from the Selected treatment. The treatment consisted of two steps. First, subjects solved three tasks by themselves, which again allowed me to compute an out-of-sample measure of individual’s naïveté to predict their subsequent behavior. In the remaining four tasks, subjects first again stated a belief.20 Then, they were shown the beliefs of two other randomly drawn subjects (“senders”) from the same experimental session. This random
19Appendix3.Cfurther analyzes the relationship between subjects’ inherent naïveté, their choice of advisor, and the subsequent belief patterns.
20In these four tasks, subjects did not decide on their group membership. Rather, the computer decided for them that whenever their private signal was higher (lower) than100, they entered the red (blue) group.
This was done to ensure that subjects indeed had symmetric information.
matching was not constant across tasks. Rather, in each task, subjects saw the beliefs of two new (and potentially different) randomly drawn subjects. Importantly, all subjects not only solved the same tasks, they also had access to the same information, and the presence of symmetric information was made clear to participants. Subjects were then asked to state a second belief. To ensure that laziness does not affect the findings, subjects had to explicitly type in this second belief, rather than, e.g., confirm their first guess. 96 subjects took part in this condition and earned 11.60 euros on average.
I again normalize the data across tasks by computing the naïvetéχ that is implied in each belief. The analysis begins by investigating the raw correlation between the naïveté implied in subjects’ first and second beliefs in each of the four tasks, i.e., the beliefs before and after they saw the beliefs of the two senders. I focus on cases in which the first belief of the receiver differs from the beliefs of at least one sender in a meaningful way; after all, studying how people revise their beliefs necessitates the presence of at least partial disagreement. I define disagreement as a binary variable which equals one iff the receiver’s belief differs from the belief of at least one sender in the sense that the implied naïveté of the receiver isχ ≤ 0.5 and that of at least one sender χ > 0.5, or vice versa. Despite this disagreement, Figure 3.5 shows that pre- and post-communication beliefs exhibit a strong raw correlation (ρ = 0.86), providing a first piece of evidence that subjects’ final belief was largely based on their own assessment of the available information, rather than the senders’ beliefs. If subjects had predominantly revised their beliefs, the sophisticated types should have adjusted upwards in Figure3.5, while the naïve types should have adjusted downwards. While the figure shows that the majority of adjustments indeed go in the expected direction, the large majority of people rarely revises their beliefs.
To provide a different perspective on this issue, I proceed by investigating how sub-jects revised their beliefs as a function (i) of the number of senders who state opposing views, and (ii) of the receiver’s type; after all, rational and naïve types may differ in how they respond to others’ solutions. Figure3.6presents histograms of subjects’ belief revisions as a consequence of the senders’ reports. To construct a measure of belief revi-sion, I compute by how much closer the receiver’s post-communication beliefs are to the average beliefs of the two senders, expressed as percentage of the pre-communication disagreement (measured as simple difference between the receiver’s pre-communication belief and the two senders’ average pre-communication belief). Thus, the belief revision measure describes by how much receivers altered their belief in response to the senders’
beliefs, relative to how much they could have changed their beliefs given the senders’
reports.
The figure provides an overview of belief revisions conditional on the receivers’ up-dating type as well as on the number of senders whose beliefs significantly depart from the receiver’s belief. To this end, I again use a coarser version of the naïveté parameter χ by calling receivers rational if both their (out-of-sample) median naïveté parameter from the first three tasks and the naïveté implied in the first belief of the respective tasks satisfyχ ≤ 0.5.21I define naïfs analogously withχ > 0.5. For instance, the top left panel
21I use both the out-of-sample measure and the first belief to classify subjects to ensure that I do not falsely classify them as, e.g., rational merely because they (perhaps due to random errors) stated a rational belief in the respective task. Appendix3.D.3reports robustness checks.
-.50.511.5Naïveté implied in belief after communication
-.5 0 .5 1 1.5
Naïveté implied in belief before communication Correlation b/w pre- and post-communication beliefs
Figure 3.5. Raw correlation between the naïvetéχimplied in first and second beliefs (ρ = 0.86).
To construct this figure, subjects’ pre- and post-communication naïveté is rounded to multiples of0.05. The ball size then represents the number of observations in the respective bin. The scatter only includes observations for which there was at least partial disagreement, see the main text for details. Appendix3.D.2illustrates the raw correlation including the cases in which
there was agreement. To ease readability, the scatter excludes 21 (out of 271) observations for which the implied naïveté of at least one belief is outside [-.5,1.5].
shows the belief adjustment of rational subjects who were confronted with one rational and one naïve sender.
The figure reveals that, consistent with the pattern reported above, subjects over-whelmingly abstain from adjusting their beliefs in response to the senders’ assessments.
When the senders report mixed beliefs (one rational and one naïve), the vast majority of both rationals and naïfs sticks with their own assessment, as indicated by the large spikes at belief revisions of 0%.22 Thus, for instance, seeing one deviating response does not induce naïfs to reconsider their solution strategy. On the other hand, when subjects see two consistent beliefs that contradict their own estimate, the updating behavior differs markedly across types. While a large majority of rationals does not adjust their beliefs at all (see the top right panel), most naïfs start moving towards the rational senders (bot-tom right panel). This suggests that the rationals know that they are right, while at least some naïfs exhibit doubts once the evidence becomes sufficiently strong.
To analyze the preceeding patterns more rigorously, in column (1) of Table 3.4, I regress the naïveté χ implied in subjects’ second belief (i.e., the belief subjects stated after they saw the beliefs of the senders) on the naïveté implied in subjects’ first belief, for each subject and task. Column (2) regresses subjects’ second belief on the average naïveté of the two senders.23 Results show that, on average, subjects react only very
22These results may be related to studies of overconfidence (e.g., C. Camerer and Lovallo,1999; Burks et al.,2013).
23I employ the average naïveté of the two senders for expositional convenience only. All results are robust to using the two measures separately.
weakly to the beliefs of their peers. While their own assessment of the available evidence explains 77.0% of the variation in the second beliefs, the beliefs of the peers only
ex-0.25.5.75
0 25 50 75 100 0 25 50 75 100
1 naïve sender 2 naïve senders
Fraction
Belief revision (in % of previous disagreement) Rationals
0.25.5.75
0 25 50 75 100 0 25 50 75 100
1 rational sender 2 rational senders
Fraction
Belief revision (in % of previous disagreement) Naïfs
Figure 3.6. Magnitude of belief revisions. Each histogram denotes the belief revision between the first and second belief (expressed as percent of the difference between the first belief and the
average belief of the two senders) conditional on the type of the subject (top / bottom panel) and on the composition of the senders. The top left panel shows the adjustment of rational subjects who face the beliefs of one naïve and one rational sender, while the top right panel illustrates the rational types’ belief revision if they faced two naïfs. The bottom left panel depicts
the adjustment behavior of naïfs when they faced one rational and one naïve belief, while the right panel illustrates adjustment in case of two rational senders. For a given subject and task, a subject (“receiver”) is classified as rational if both the out-of-sample median naïveté parameter from the first three tasks and the first belief statement in the respective task are “rational” (i.e., χ ≤ 0.5), and analogously for naïfs (χ > 0.5). Very similar results obtain when I define rationals and naïfs exclusively based on the out-of-sample naïveté measure or exclusively based on the
first belief in the respective task, see Appendix3.D.3. Adjustments> 100%and< 0%are excluded to ease readability (7 out of 185 obs.).
plain a miniscule 4.8%. Column (4) investigates whether the weight subjects put on other people’s beliefs depends on the degree of agreement among the senders. To this end, I regress subjects’ naïveté on the previously discussed variables as well as on (i) the degree of disagreement among the senders, and (ii) an interaction of the degree of disagreement with the average naïveté of the senders. Disagreement is defined as the ab-solute difference between the naïveté implied in the beliefs of the two senders. Results show that, consistent with intuition, subjects indeed place higher weight on the beliefs of their peers if disagreement is smaller: the negative and statistically significant interaction coefficient says that higher disagreement leads to a lower weight on the senders.
Columns (5)-(8) and (9)-(12) break these patterns down between rationals and naïfs.
Notably, as suggested by Figure 3.6, the rational subjects’ post-communication beliefs are not significantly correlated with the average naïveté of the senders, see columns (6)-(8). In addition, consistent with the visual evidence presented above, rationals never respond to the beliefs of their peers, regardless of whether they exhibit agreement or not. In contrast, naïfs partly respond to others’ beliefs, albeit to a rather small extent:
the variance in subjects’ beliefs that can be explained by the beliefs of their peers is only 17.6% (column (10)), compared to 42.3% for their own pre-communication beliefs (column (9)).
Appendix 3.D.5investigates learning over time. In particular, it is conceivable that those naïve subjects who revised their beliefs according to the beliefs of the senders state more rational beliefs in subsequent tasks. However, this is not the case, perhaps suggest-ing that while some subjects intuit that their strategy is incorrect, they are incapable of developing a better strategy themselves.
Result 13. People have a strong propensity to trust their own assessment of the available evidence, rather than that of their peers. In consequence, hearing other people’s beliefs does not induce meaningful convergence to a consensus.