Experimental Design Details
Also see attachment:
I. DATA COLLECTION:
In the BonnEconLab at the University of Bonn, we conduct laboratory experiments. We conduct individual belief formation experiments as well as market experiments.
Individual
The basic structure of the experimental setup is as follows. Subjects are asked to estimate the quality of 14 hypothetical companies. Each company is either good or bad. The computer determines, independently for each company and with 50-50 probability whether the company is good or bad. Participants do not learn about the actual realization of the computer draw, but receive multiple pieces of news. News are symmetric and correct with 65%, so Pr(news=good|company=good) = Pr(news=bad|company=bad) = 65%. News are independently drawn conditional on the state of the company (good or bad).
The experiment consists of the following stages:
Part 1:
• For each of the 14 companies, a subject first observes a sequence of news. The news appear sequentially on a subject’s computer screen. The number of signals per company is 0, 1, 2, 3 or 4.
• The subject is then asked to guess whether the company is good or bad (we will henceforth refer to this as “first belief”). This belief is financially incentivized through a binarized scoring rule.
• This procedure is repeated for all 14 companies.
• The subject completes real effort tasks for 15 minutes.
Part 2:
• A subject observes a final piece of news about the company.
• The subject then guesses whether the company is good or bad (we will henceforth refer to this as “second belief”). This belief is financially incentivized through a binarized scoring rule.
• This procedure is repeated for all 14 companies.
Part 3: Questionnaire.
Markets
The basic structure of the market experiments is identical to the individual experiments, except that now subjects participate in a version of a parimutuel betting market. Subjects are matched into groups of three. In both part 1 and part 2, after seeing the pieces of news and stating their individual estimates, the three matched subjects bet in a market. Subjects know that all participants in their market saw exactly the same pieces of information as they did. For their betting decision, subjects obtain a budget of 10 Euros. Betting occurs in two steps:
1. Subjects state which part of the total amount they are willing to bet they want to bet on the company being good. 0% means the entire amount they bet is bet on the company being bad, 100% means the entire amount they bet is bet on the company being good.
2. Subjects state the total amount they want to bet (maximum 10€, minimum 0€).
II. TREATMENTS:
There are three pairs of treatments, Main Individual and No Cue Individual, Recall Main and Recall No Cue, Main Market and No Cue Market, where each pair is implemented in a within-subject design.
1. Treatment Main Individual
In treatment Main Individual, the pieces of news are communicated on subjects’ computer screens along with a context. The context consists of a story and an image. Importantly, there is a one-to-one mapping between type of news for a given company and context. That is, every positive news for company A is communicated with the same context (image and story). Likewise, every negative news for company A is communicated with the same context (albeit a different one than the positive news of course). The same logic holds for all other companies.
2. Treatment No Cue Individual
In treatment No Cue Individual, the setup is exactly the same as in Main, except that each piece of news is communicated with a different context. That is, a given context (image and story) never appears twice, even if the company and type of news is identical.
3. Treatment Recall Main
In treatment Recall Main, the setup is exactly the same as in Main Individual, except that, before eliciting beliefs in period 2, we ask subjects to recall the pieces of news they saw.
4. Treatment Recall No Cue
In treatment Recall No Cue, the setup is exactly the same as in No Cue Individual, except that, before eliciting beliefs in period 2, we ask subjects to recall the pieces of news they saw.
5. Treatment Main Market
As in treatment Main Individual, there is a one-to-one mapping between type of news for a given company and context. That is, every positive news for company A is communicated with the same context (image and story). Likewise, every negative news for company A is communicated with the same context (albeit a different one than the positive news of course).
6. Treatment No Cue Market
As in treatment No Cue Individual, each piece of news is communicated with a different context. That is, a given context (image and story) never appears twice, even if the company and type of news is identical. As described above, after stating beliefs, subjects participate in a betting market.
III. NATURE OF ANALYSES
Individual
We analyze our experimental data by means of OLS regressions:
• The main dependent variable will be log belief odds of subject’s second period belief.
• The main independent variable is the realization of the last signal, as well as its interaction with a treatment dummy (Main versus No Cue).
• Because we have multiple observations per subject, we cluster the standard errors at the subject level.
Market
We analyze our experimental data by means of OLS regressions:
• The main dependent variable will be log belief odds of second period market prices.
• The main independent variable is the realization of the last signal, as well as its interaction with a treatment dummy (Main Market versus No Cue Market).
• Because we have multiple observations per market, we cluster the standard errors at the market level.
IV. HYPOTHESES:
Individual
1. Overreaction in beliefs
Restrict attention to treatment Main Individual. Take as dependent variable log belief odds of a subject’s second belief, divided by ln(0,65/(1-0.65)). Regress on the last piece of news, controlling for number of positive news from stage 1 minus the number of negative news from stage 1. We hypothesize that the OLS coefficient on the last piece of new is significantly larger than one.
2. The role of associative memory
Restrict attention to treatments Main Individual and No cue Individual. Take as dependent variable log belief odds of a subject’s second belief, divided by ln(0,65/(1-0.65)). Regress on (i) the last piece of news, (ii) a treatment dummy, and (iii) the interaction of the value of the last signal and a treatment dummy (where treatment Main is coded as 1), controlling for number of positive news from stage 1 minus the number of negative news from stage 1. We hypothesize that the OLS coefficient of the interaction term is significantly larger than zero.
3. Variation in the number of cued signals
Restrict attention to treatment Main Individual. Take as dependent variable log belief odds of a subject’s second belief, divided by ln(0,65/(1-0.65)). Regress each of these on (i) the last news, (ii) the number of signals in the first part that took on the same value as the signal in the second part, and (iii) a corresponding interaction term, controlling for number of positive news from stage 1 minus the number of negative news from stage 1. We hypothesize that the OLS coefficient of the interaction term is significantly larger than zero.
4. Weight on cued versus non-cued signals
Restrict attention to treatments Main Individual and No cue Individual. Compute the sum of cued signals as well as the sum non-cued signals from stage 1. Take as dependent variable log belief odds of a subject’s second belief, divided by ln(0,65/(1-0.65)). Regress on (i) the last piece of news, (ii) a treatment dummy, (iii) the sum of cued signals, (iv) the sum of non-cued signals (v) the interaction of the sum of cued signals and a treatment dummy (where treatment Main is coded as 1), and (vi) the interaction of the sum of non-cued signals and a treatment dummy (where treatment Main is coded as 1). We hypothesize that the OLS coefficient of the interaction term with cued signals is significantly larger than zero, whereas the OLS coefficient of the interaction term with non-cued signals is not statistically different from zero.
Markets
5. Overreaction in market prizes
Restrict attention to treatment Main Market. Take as dependent variable ln(market price/1-market price), divided by ln(0,65/(1-0.65)). Regress on the last piece of news, controlling for number of positive news from stage 1 minus the number of negative news from stage 1. We hypothesize that the OLS coefficient on the last piece of new is significantly larger than one.
6. The role of associative memory
Restrict attention to treatments Main Market and No Cue Market. Take as dependent variable ln(market price/1-market price), divided by ln(0,65/(1-0.65)). Regress on (i) the last piece of news, (ii) a treatment dummy, and (iii) the interaction of the value of the last signal and a treatment dummy (where treatment Main is coded as 1), controlling for number of positive news from stage 1 minus the number of negative news from stage 1. We hypothesize that the OLS coefficient of the interaction term is significantly larger than zero.
7. Variation in the number of cued signals
Restrict attention to treatment Main Market. Take as dependent variable ln(market price/1-market price), divided by ln(0,65/(1-0.65)). Regress each of these on (i) the last news, (ii) the number of signals in the first part that took on the same value as the signal in the second part, and (iii) a corresponding interaction term, controlling for number of positive news from stage 1 minus the number of negative news from stage 1. We hypothesize that the OLS coefficient of the interaction term is significantly larger than zero.