Resilience to climate change: diversification strategies and correlation neglect

The intervention is situated in Assis, Brazil, a region heavily reliant on agricultural activities. Our study employs a well-structured lab in the field experiment framed with real economic incentives, complemented by survey data. Participants will engage with a questionnaire in loco, individually, electronically, and with an anticipated completion time ranging from 90 to 120 minutes.
Our primary objective is to examine the presence of correlation neglect in the participants decisions, gauging not only its presence, but also the extent to which it manifests. In particular, we aim at understanding how correlation neglect relates with both the degree of correlation across decision problems and the individual degree of risk aversion among participants.
Next, we intend to evaluate the impact of integrated information provision on mitigating correlation neglect and/or fostering increased crop diversification. To discern such effects, participants will be randomly assigned to either the integrated or segregated information treatments. While the informational content remains consistent across both groups, the representation of assets’ returns differs—presented either jointly in a single graph or distinctly in two separate graphs.
As a secondary objective, we investigate whether the participants’ beliefs about the distribution of returns is more accurate when the information is provided jointly. For this reason, we include belief elicitation questions following the allocation decision problems.

2024-03-04

2024-04-05

The primary outcomes are the measures of correlation neglect. We use two different measures; (i) the share allocated to the dominated asset, and (ii) the distance between the theoretical optimal and the actual allocation to the dominated asset.

The two measures of correlation neglect indicate the extent to which individuals deviate from a diversified portfolio selection. The first measure is the share allocated to the dominated asset. When faced with a decision, participants will be presented with two investment options distinguished solely by their expected returns. The one with the lower expected return is considered dominated and should only be chosen if the individual seeks to increase diversification. This is a common measure of correlation neglect in the empirical literature (Laudenbach et al. 2021, Kalil and Sonsino 2009). The second measure is the distance between the theoretical optimal and the actual allocation to the dominated asset. This is an innovation proposed by our study, where the optimal allocation is calculated based on the degree of correlation between the assets and the participants’ individual degree of risk aversion (Markowitz 1952, Sandmo 1969).

The secondary outcome is a measure of belief about the returns of the investment options. We use an instrument that comprises three different belief elicitation questions.

The belief measure will be constructed using the obviously related instrumental variable (ORIV) approach. This method consolidates different questions and formulations to provide an estimator that is more efficient, as outlined by Gillen, Snowberg, and Yariv (2019). For this reason, we ask three types of questions concerning the perceived returns of the assets: qualitative, quantitative, and probabilistic, in order to ensure that the instrument is less prone to measurement errors (Haaland et al. 2023).

We use a between-within-subjects experimental design, wherein participants answer a questionnaire delivered in loco via tablets. Each participant is tasked with making a series of investment decisions involving varying degrees of correlation among asset returns, i.e., negative correlation, no correlation and positive correlation. In addition, participants are randomly assigned to one of two treatments:
• Integrated Information: assets returns' distribution data is presented jointly.
• Segregated Information: assets returns' distribution data is presented separately.
Randomization will be made at the individual level upon start of the questionnaire and it is computer-based, i.e., calculated by the experiment programming tool oTree (Holzmeister, 2017). The experiment has six parts.

Part 1: Mathematical and Financial ability
In this stage, we collect information on the participant’s mathematical ability by asking them to solve various mathematical operations with two digits and answer questions to elicit financial knowledge. Participants receive a fixed payment.

Part 2: Risk aversion (ex-ante)
We apply one task to elicit the participant’s degree of risk aversion, following the Binswanger (1980) experimental design. In the activity, each participant is presented with six lotteries, each offering, with equal probability, a high or a low payment. The participant’s task is to select their preferred lottery, thereby providing insight into their individual attitude towards risk.

Part 3: Training
Participants are assigned to one of two treatment conditions, that differ only on how the comparison if the assets’ return is displayed. The main goals of this part are to inform the participants and elicit measures of understanding. We train individuals to understand how investments returns are generated, how to read graphs, how to allocate resources between different decisions, and how the allocation among options affect total returns. We designed the questionnaire to induce an experience-based learning, similar to what has been done in the literature (Laudenbach et al. 2021, Kalil and Sonsino 2009). Throughout the training, there are comprehension questions to assess the participants’ understanding. Lastly, we included an attention checker, to capture whether the respondents’ take the time to read the instructions carefully (Kane and Barabas 2019; Berinsky, Margolis, and Sances 2014)

Part 4: Investment decisions
This is the first main experimental activity. Here, the participants are presented with four investment decisions, each requiring them to determine the allocation of a fixed endowment between two assets. In each decision, the correlation between assets’ returns varies and the joint distribution refers to the frequency of the four possible scenarios from the combination of the two outcomes. This activity allows for an individual assessment of the extent of correlation neglect. In addition, after the final investment decision, we elicit information on the perceived degree of correlation between the assets.

Part 5: External validity experiment
This is the second main experimental activity. Here, the participants are presented with information on two agricultural products and are asked to allocate them on a fictitious and exemplary plot of land. The distribution of returns of each crop is the only information provided to the participants. As with the general investment decisions, we elicit information on the perceived degree of correlation between the agricultural products. This stage allows for gaining additional insights on rural households learning from the task.

Part 6: Questionnaire
The last part is a survey where we ask participants about their socioeconomic information, land use, and other investment decisions. This stage allows us to assess the consequences of correlation neglect.

Upon completion of the questionnaire, one of the incentivized activities will be randomly selected as the basis for compensating the participant. The expected average compensation is 60 BRL, or approximately 11.50 EUR.

Randomization will be made at the individual level at the beginning of the questionnaire. The randomization is computer-based and will be calculated using the experiment programming tool, oTree.

Individual.

No

None.

200 farmers from the region of Assis, Brazil.

100 farmers will be assigned to the integrated treatment, and 100 to the segregated treatment.

Based on previous empirical results, we assume as reasonable values for the standard deviation of the main outcome variable 0.25. Considering 80% power, a 5% significance level in a two-tailed test scenario, and at least 100 participants allocated to each treatment independently, the smallest detectable effect size is estimated to be 0.0995, around 0.4 standard deviation of the outcome variable. Such an estimate is conservative, as it does not account for potential reduction in variance resulted with the inclusion of control variables, nor the expected variance reduction in one of the treatment groups.