How to communicate uncertainty? The effects of asymmetrical interval predictions on ambiguity attitudes.

There are 2x3 groups in this study. All of them read the same set of instructions and are exposed to the same set of historical weather data. As a consequence, these are treated equally before treatment. The treatments are introduced as weather forecasts that vary between subjects in two dimensions: The forecast contains either a point estimator, an asymmetric interval estimator, or both. In addition, we vary the forecasts moment of surprise as one set of forecasts is in line with the historical weather data, while the other set is not.

2021-04-26

2021-09-15

Baillon et al. ‘s ambiguity aversion index as well as their ambiguity-generated insensitivity index. Both pre and post treatment.

Baillon et al. (2018, p. 1840) describe the ambiguity aversion index as follows:
The ambiguity aversion index measures the well-known aversion to ambiguity and is often taken to be normative. The ambiguity-generated insensitivity index captures the degree of ambiguity, that is, the perceived level of ambiguity. The higher this level is, the less the decision maker discriminates between different degrees of likelihood, and the more these degrees are treated alike, as one blur. The second index thus also captures insensitivity toward likelihood changes

Baillon, A., Huang, Z., Selim, A., & Wakker, P. P. (2018). Measuring ambiguity attitudes for all (natural) events. Econometrica, 86(5), 1839-1858.

Control variables from a short questionnaire at the end of the survey:
• Demographic variables (age, gender, income, zip code, education level, number of children, marital status),
• Risk preferences (general, weather domain),
• Assessment of the forecast (in terms of accuracy and trustworthiness),
• Weather variables (use of weather forecasts, current outdoor temperature),
• Task comprehensibility.

This experiment will be conducted in the form of a representative survey in Germany. Before conducting a representative survey, we will also conduct a pilot project with about 120 participants drawn from a local adult convenience sample to test the participants' understanding and time requirements. The experiment (including the pilot) qualifies as an artefactual field experiment that will be conducted remotely and asynchronously.
After reading instructions and answering comprehension questions, respondents are exposed to some historical temperatures. With these information at hand, they decide at which probability of winning they prefer a risky lottery to a bet on temperature outcomes nine days in the future. Following Baillon et al. (2018), respondents make this decision for six different bets. Importantly, these six decisions are made before respondents are exposed to any treatment meaning that the provided information does not differ between respondents at this stage.
Before subjects make the same set of decisions once more, they are provided with a weather forecast, that we exogenously manipulate in a 2x3 between-subject design. More precisely, we provide the respondents with forecasts that differ with respect to its communication strategy as well as its momentum of surprise: The forecast contains either a point estimator, an asymmetric interval estimator, or both. In addition, we vary the forecasts moment of surprise as one set of forecasts is in line with the historical weather data, while the other set is not. The six treatments are randomly assigned at the individual level.
Respondents are then required to complete a short questionnaire to collect data on possible control variables. Questions include demographic variables (age, gender, income, zip code, education level, number of children, marital status), risk preferences (general, weather domain), assessment of the forecast (in terms of accuracy and trustworthiness), weather variables (use of weather forecasts, current outdoor temperature), and task comprehensibility.
Finally, the location, time, and realized temperatures corresponding to the historical and forecasted temperatures as well as the sources of these information are revealed, the payoffs are calculated, and the survey is closed.

To calculate the respondents’ payoffs immediately, we let them bet on historical temperatures and provide historical forecasts. Because we do not want them to game the task, we neither reveal the location nor the dates the provided information corresponds to until the end of the experiment.
We inform respondents that the provided information is real and ensure them that we’ll reveal all sources eventually.

Randomization is done applying python 3’s random.shuffle() function to two treatment arrays at the beginning of a session.

Individual.

No

120 individuals (pilot) + about 1500 individuals (the pilot was used to conduct a power analysis and therefore we increased the number of individuals from 1200 to 1500).

120 individuals (pilot) + about 1500 individuals (the pilot was used to conduct a power analysis and therefore we increased the number of individuals from 1200 to 1500).

Individuals will have a 50 percent chance of being randomly assigned to either the surprise or non-surprise treatment. In each of these arms they will have a 33.33 percent chance of being randomly assigned to one of the three communication strategies. As a consequence, there is an expected probability of 16.66 percent of being randomly assigned to one of the six treatments.

A power calculation will be run after the completion of the pilot. We plan to run the experiment with round about 1500 participants.

Forecast_Preregistration_Analysis_Plan.pdf