Boosts and Nudges under GARP

In this research, we aim at assessing how decision support tools might improve economic decision-making. More precisely, we want to compare the impact of nudges and boosts on individuals’ deviations from utility maximization. Nudges had been considered since Thaler and Sunstein 2008’s book as cost-effective mechanisms to reduce the consequence of decision biases for individuals. More recently, some debate had risen about the persistence of nudges’ effects over time (Brandon et al 2022, DellaVigna and Linos 2022). Some authors have suggested that boost interventions (Hertwig and Grune-Yanoof 2017 ; Caballero and Ploner 2022, Banerjee and John 2020), -- ie, processes where elements of reflection and understanding for the mechanism are acquired by participants for « boosting » competence – might be more efficient and more persistent.
Our intervention consists in a laboratory economic experiment where participants make repeated allocation choices along a budget line using the induced budget experiment from Choi et al (2007a, 2007b). This method is based on testing the consistency of successive consumption choices by an individual with GARP (Generalized Axioms of Revealed Preferences). These allocation choices are incentivized and participants get a monetary payoff at the end of the experimental session that depends upon their personal choices during this task.
These participants are randomized into 6 experimental treatments (described below) : 2 control treatments, 2 nudge treatments and 2 boosts treatments. Each experimental session is divided in 8 steps for an expected duration of 1h45mn.

2023-05-23

2023-06-28

1) percentage of GARP-consistent participants
2) money-pump index.

1) GARP-consistency means the participant never violates the Generalized Axiom of Revealed Preferences, which is that whenever he chooses a bundle of goods A when a bundle of goods B is available, then he cannot choose the bundle B if A is available (Varian, 1982, Econometrica). 2) Average of maximum and minimum money pump index (a rationality index introduced by Echenique, Lee and Shum, 2011, Journal of Political Economy), as computed following the algorithm described in Smeulders et al, 2013 (Journal of Political Economy).

In our laboratory economic experiment, the main stage participants make repeated allocation choices along a budget line using the induced budget experiment from Choi et al (2007a, 2007b). This method is based on the consistency of successive bundle consumption choices by an individual to GARP (Generalized Axioms of Revealed Preferences). These allocation choices are incentivized and participants get a monetary payoff at the end of the experimental session that depends upon their personal choices during this task. This method is based on testing the consistency of successive consumption choices by an individual with GARP (Generalized Axioms of Revealed Preferences). These allocation choices are incentivized and participants get a monetary payoff at the end of the experimental session that depends upon their personal choices during this task.
These participants are randomized into 6 experimental treatments (described below) : 2 control treatments, 2 nudge treatments and 2 boosts treatments. Each experimental session is divided in 8 steps for an expected duration of 1h45mn.
In the control treatments, partcipants did not receive any help during the GARP sequence (nudges) or before the GARP sequence (boosts)
In our nudge treatments, participants get some help by the computer during each round. The computer uses previous choices made by each participant to indicate where the consistent-choices set lies for the current round.
In our boost treatments participants have the opportunity to learn during a 30' sequence about GARP principles and its implementation.

In our experimental design, 6 treatments are to be implemented. The basic task consists for a participant in choosing a consumption bundle along a given budgetary line that changes at each round of choice. There will be 50 rounds of choices using the design popularized by Choi et al. (2007a, 2007b). Answers to these tasks enable us to assess various rationality scores that are widely used in the GARP literature. Before the 50-decisions task, all participants play a 5-decisions task to acquire some experience with the interface.
For the nudge treatments, we implement 2 treatments. In the first one -- Nudge 1 --, the standard interface is augmented with colors that identify the segments of the budget line that are compatible with GARP. Specifically, relying on the sequence of past choices, the computer shades the segments of the budget line that are GARP-incompatible. The participant should therefore select an allocation from the 'right' line segment but remains totally free to make a choice along the 'wrong' line segment. The second one -- Nudge 2 -- is a variant of Nudge 1; Instead of indicating a 'right' and a 'wrong' line segment to each participant, the computer randomly selects a choice bundle that is consistent with their previous bundle choices as a default option. That is, the random bundle lies within the 'good' line segment but without materializing the segment in the computer graphical interface.
For the boosts treatments, we also implement 2 treatments. Both boosts occur before the 50-decisions sequence. The first one --- Boost 1 -- consists in a 30' duration teaching course about GARP. This course is divided in 2 phases, a 10' video teaching course that explains the basic principles of GARP theory + numerical illustrations about bundle choices and a 20' practice task where each participant should learn to identify choice inconsistencies in numerical exercices. Answers are collected through the computer interface and each participant have the opportunity to assess her level of understanding by obtaining immediate feedback regarding her answers to the numerical applications. The second one -- Boost 2 -- uses the Money Pump Argument developped in Echenique et al 2011. The idea is that a consumer that violates GARP is subject to being exploited as a "money pump". This money pump will be stronger if the severity of violations of GARP are more important and a Money Pump Index could be measured by the amount of money that can be extracted from the consumer by an arbitrager. In our boost 2, each participant is directly confronted to a computer AI that tries to extract money from her depending on her previous allocation choices during 30'. The participant makes a choice and has the opportunity to consider which amount could be extracted by the computer from choosing this particular bundle given her previous sequence of choices. Contrary to the 50-decisions sequence, boosts are non monetary-incentivized.
In order to compare cleanly our experimental treatments, we implement two control treatments: The first one -- Control 1 -- is the basic 50 rounds sequence of allocation choices to be completed by each participant, and that was implement in past economic experiments that are reported in the literature. The second control -- Control 2 --- consists in selecting a choice bundle at random along the budget line for each participant at each round. This control enables us toto obtain a more comparable benchmark for the Nudge 2 treatment where a choice bundle is selected within the « right » line segment.
Moreover, an important characteristic of our experiment is that we implement a 'fake' boost before the 50-decisions sequence in all treatments but boost ones. The aim of this « fake » boost is to control for the cognitive burden and induced fatigue implied by the « real » boosts. In this "fake" boost that lasts 30' (like in the other boosts), participants are exposed to a 10' duration teaching course about macroeconomic theory of growth and then should complete tasks with numerical applications aiming at computing growth rates and the contributions of production factors to it. In order to check the comparability of this fake boost to the GARP boosts in terms of induced fatigue, we will also use a psychometric score for measuring the fatigue level before and after the boost sequence (Multidimensional Fatigue Inventory measure, see Smets et al 1995). We use also the Cognitive Reflection Test (Frederick, 2005) to measure cognitive abilities for each participant, the BIG-5 personality test (Rammstedt and John, 2007) and a standard socio-demographic questionnaire at the end of the experiment.

Randomization is made within the experimental platform: Participants register for particular time slots for which they are available and treatments are randomized at the session level. Participants can only participate once to the experiment, which is ensured by the experimental platform registration system based on ORSEE (See Greiner 2017).

Session-level randomization: Participants register to participate on time slots and they are randomized in our various treatments. Individuals are randomly assigned to the treatment and comparison groups.

Yes

385 participants for individual choices, with 70 participants for all treatments but control 2 (robustness check) for which we plan to recruit only 35 participants. Given a even split of participants within the experimental sessions, the number of experimental sessions to complete is 17.

385 participants will make 50 allocation choices. Moreover they will answer to 2 fatigue questionnaires (2 scores per participant, before and after the boost sequence). We also observe their answers during the boost sequence. Last, they should complete the Big5 personality test and the Cognitive Reflection Test before the usual post experimental questionnaire.

- in Control 1 (Basic 50-decisions sequence based on Choi et al 2007) we will have 70 subjects and therefore 70 independent observations
- in Control 2, 35 participants
- In Nudge 1: 70 participants
- In Nudge 2: 70 participants
- In Boost 1 : 70 participants
- In Boost 2 : 70 participants