The Value of Rating Systems in Credence Goods Markets

We replicate the four main treatments in Angerer et al. (2022) with a neutral frame.
We employ a 2x2 factorial design to experimentally investigate the effect of a public rating system in two different expert markets: A market in which consumers can rely on their personal experience, and one in which they cannot. These two different market environments are implemented in the experiment as follows:

In the conditions without personal experience, in each period consumers choose one expert from a list of four without being able to identify them. All players are informed beforehand that consumers have no means of identifying experts from previous periods. Thus, although consumers observe their payoffs in each period and can partially infer expert behavior, they cannot attribute it to a particular expert and therefore cannot build up personal experience with a particular expert. In experimental conditions with personal experience, consumers can identify experts by a fixed ID and decide whether to interact with a particular identified expert. Over the 16 periods of play, they can thus learn from their personal experience (payoffs) with a particular identified expert.

In the conditions with the public rating system, consumers can choose to rate interactions with experts on a five-star rating scale after receiving their payoff in a given period. This rating is shown to the respective expert at the end of the period. Subsequently, ratings for each expert over all treated consumers are aggregated, averaged, and displayed to consumers. Consumers see these public ratings for all experts when they decide whether to interact and which expert to choose starting in period 5. In the condition without personal experience, as highlighted before, consumers cannot identify a particular expert and only see the public ratings. The public ratings of all experts are displayed to experts when they decide on the type of treatment and which price to charge in a given period.
In total, we run four experimental conditions: BASELINE (no experience, no rating), EXPERIENCE (experience, no rating), RATING (no experience, rating), EXP-RATING (experience, rating).

Reference:
Angerer, S., Glätzle-Rützler, D., Mimra, W., Rittmannsberger, T., & Waibel, C. (2022). The value of rating systems in healthcare credence goods markets. Available at SSRN. https://doi.org/10.2139/ssrn. 3965318

2023-11-13

2024-01-26

Undertreatment, and overcharging-rates

Undertreatment is defined as the consumer needing the major treatment q_H but the expert providing the minor treatment q_L. An expert might have incentives to do so since the costs for the major treatment are higher (6 ECU versus 2 ECU) and the expert can always charge the price of the major treatment (8 ECU). In the results section, undertreatment will be reported in percent of expert-consumer interactions in which consumers need the major treatment. In terms of information, consumers can detect undertreatment in a period ex-post via their payoff, as the problem is not solved. In particular, if the expert charged p_H, the consumer payoff from undertreatment is -8 ECU

Overcharging is defined as the expert charging the price of the major treatment p_H while only providing the minor treatment to a consumer who has a minor problem. Overcharging is accordingly reported in percent of expert-consumer interactions in which consumers need minor treatment. In terms of information, consumers cannot infer ex post whether they have been overcharged, as they might have had a major problem requiring the major treatment charged at p_H. Thus, an expert can 'hide' behind a major treatment problem when overcharging.

Consumer outcomes: Interactin rates, giving feedback, and rating.
Market outcomes: Market efficiency and consumer surplus.

On the consumer side, we record whether they choose to interact, and in the rating conditions whether they choose to provide a rating (captured by the variable feedback) and what the rating is (captured by variable rating).

We use two measures of market outcomes, overall market efficiency and consumer surplus. Market efficiency is driven by interaction (allowing surplus generation) and whether there is undertreatment, as undertreatment does not generate consumer value. Given our parametrization, we expect high levels of interactions, such that market efficiency is primarily determined by undertreatment. We normalize market efficiency, with 0% for no interaction and 100% for an interaction with the correct treatment.

Consumer surplus incorporates the prices paid by consumers and is thereby influenced by overcharging, which is not the case for market efficiency. Consumer surplus is reported in absolute value.

We plan to use a student sample from the University of Innsbruck and run each experimental condition with 48 subjects (as suggested by our power analysis). Therefore, we plan to run two sessions with 24 subjects each in every experimental condition. All sessions are run computerized using z-Tree and students are recruited using hroot. Participants do not know which experiment they are going to participate in when they register. They only receive information about the expected duration of the experiment (1:45h).
Our experiment is structured as follows for all our conditions:

Stage 1: The experimenter explains the experiment and participants read the instructions.
Stage 2: Participants answer several control questions to ensure they understood the game.
Stage 3: The computer randomly assigns roles and markets to participants.
Stage 4: Participants play the game for 16 periods.
Stage 5: Participants participate in additional games: an individual risk preference task, a dictator game, a lying task, and a trust game.
Stage 6: Participants fill out a questionnaire.

Randomization is carried out in the experiment by a computer.

at the session level

Yes

6 clusters á 8 individuals per experimental condition.

48 (6 x 8) individuals per experimental condition

192 (4 x 48) individuals (students at the University of Innsbruck).

Based on previous findings, we performed a power calculation, indicating that we need six clusters á 8 subjects per experimental condition when aiming for a power of 80%.