Recommender Systems and Consumer Choice - Experimental Evidence
How to cite
Schleef, Felix. 2024. "Recommender Systems and Consumer Choice - Experimental Evidence." AEA RCT Registry. AEARCTR-0014017. https://doi.org/10.1257/rct.14017-1.1
Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.
* Primary investigator
Abstract
Regulators are increasingly getting concerned with the power of large online platforms to bias consumer recommendations. In light of these concerns, I study the option of giving consumers control over the recommender algorithm as a potential avenue for regulation. In an online experiment, I let subjects make choices from lists, akin to an e-commerce shopping experience. The list is generated either by an algorithm that is optimizing consumer utility, or by an algorithm that includes both consumer utility and platform profit as part of its objective. My aim is to understand how subjects make choices in this abstract setting, whether recommender algorithms influence their choices, and how subjects choose between recommender algorithms.
Intervention
Each subject completes a risk elicitation task. Following this task, each subject is asked to make choices from ordered lists of three outcome lotteries for 30 rounds. The lists of lotteries are created by two different algorithms. Subjects are aware of the fact that two different algorithms create their choice lists. After 20 rounds of the choice task, subjects are asked to disclose their willingness to pay to increase the probability that one or the other algorithm creates the lists of lotteries for the remaining 10 choice rounds.
Experimental design
The experiment features three kinds of tasks. The first task is a risk preference elicitation task in the style of Johnson et al. (2021), where subjects have the prospect of gaining either the outcome of a lottery that pays 10$ with probability 50% and 0$ with probability 50% or a sure but unknown monetary amount X, that has been randomly drawn with equal probability between 0$ and 10$. Subjects are asked to declare a threshold
such that they would prefer receiving X if it is above the threshold, and they would prefer receiving the lottery if X is below the threshold.
The second task is a choice task, where subjects are asked to choose from lists of three-outcome lotteries that have been ordered and selected by one of two algorithms. These algorithms take into account the subjects risk aversion that has been calculated using the subjects threshold choice in the first task and assuming constant relative risk aversion. The outcomes of all lotteries are 10$, 5$, and 0$, and the respective probabilities are drawn to ensure that the probability of 10$ and 0$ does not exceed 50%.
Finally, I elicit subjects willingness to pay for one algorithm over the other. For this step, I give subjects an endowment of 1$. Using a slider, subjects can change the probability that they will face either algorithm in the next set of 10 choice rounds. Choosing equal probability between the two algorithms (50%-50%) is costless. Changing the probabilities by 1%, i.e. to 49% - 51% costs 0.02$. The cost increases linearly, so that the price of setting the probability to have either algorithm for the last 10 choice rounds is equal to the full endowment - 1$.
Design details
Cited by
No citation data available for this registration yet.