Misinterpreting yourself or rational updating about own self-control problems: Experimental evidence

Last registered on January 05, 2026

View Trial History

Pre-Trial

Trial Information

General Information

Title

Misinterpreting yourself or rational updating about own self-control problems: Experimental evidence

RCT ID

AEARCTR-0017539

Initial registration date

December 22, 2025

Initial registration date is when the trial was registered.

It corresponds to when the registration was submitted to the Registry to be reviewed for publication.

First published

January 05, 2026, 7:10 AM EST

First published corresponds to when the trial was first made public on the Registry after being reviewed.

Locations

Country

United States of America

Region

Primary Investigator

Name

Hannah Schildberg-Hörisch

Affiliation

Heinrich Heine University Düsseldorf

Contact Primary Investigator

Other Primary Investigator(s)

PI Name

Jill Amend

PI Affiliation

U Duesseldorf

Contact Investigator

PI Name

Flora Drucker

PI Affiliation

U Duesseldorf

Contact Investigator

Additional Trial Information

Status

In development

Start date

2026-01-11

End date

2026-02-14

Keywords

Behavior, Lab

Additional Keywords

self-control, sophistication, mislearning

JEL code(s)

C91

Secondary IDs

Prior work

This trial does not extend or rely on any prior RCTs.

Abstract

The aim of our study is to test two competing theories on what people learn about their own self-control problems in settings where they repeatedly perform an action (“task”) with immediate, known costs and delayed benefits. Agents with self-control problems are expected to do fewer tasks and therefore perform lower than they initially planned. While sophisticated agents predict this behavior correctly, naive agents predict they will stick to their initial plan. We vary across treatments whether the task’s benefit is known right from the start (control treatment) or uncertain (main treatment). In the latter case, people initially only know an interval that contains the task benefit. They get repeated signals about the task benefit that, in principle, allow for learning its exact realization. In our setup, we induce a lower than true prior belief on the task’s benefit. In such a setup, classic approaches to self-control problems of (partially) naïve agents such as Augenblick and Rabin (2019) and Le Yaouanq and Schwardmann (2022) predict the following: individuals who are initially (partially) naïve about their own self-control problems will update their belief about the task benefit rationally over time, converging to the correct belief, and learn about the extent of their own self-control problems from observing their past choices. By contrast, Heidhues et al. (2024) predict that people with self-control problems will misinterpret their past choices (low task performance due to self-control problems) as reflective of their “true” preferences. As a consequence, they will adjust their belief about the uncertain task benefit downwards over time to “justify” their past choices, resulting in a belief that is lower than justified by the signals they receive (misinterpreting oneself). This requires that people forget the signals about the task benefit that they have received and only remember their past performance – which our experiment design ensures. In both scenarios, people end up correctly predicting their future actions, i.e., seeming (in Heidhues et al., 2024) or actually being sophisticated about their own self-control problems (O’Donoghue and Rabin, 1999).

Augenblick, N. & Rabin, M. (2019). An experiment on time preference and misprediction in unpleasant tasks. Review of Economic Studies, 86(3), 941-975.
Heidhues, P., Kőszegi, B. & Strack, P. (2024). Misinterpreting yourself. ECONtribute Discussion Paper No. 317.
Le Yaouanq, Y. & Schwardmann, P. (2022). Learning about one’s self. Journal of the European Economic Association, 20(5), 1791-1828.
O’Donoghue, T. & Rabin, M. (1999). Doing it now or later. American Economic Review, 89(1), 103-124.

External Link(s)

Registration Citation

Citation

Amend, Jill, Flora Drucker and Hannah Schildberg-Hörisch. 2026. "Misinterpreting yourself or rational updating about own self-control problems: Experimental evidence." AEA RCT Registry. January 05. https://doi.org/10.1257/rct.17539-1.0

Sponsors & Partners

Experimental Details

Interventions

Intervention(s)

Participants will be randomly assigned to one of two treatments: the main or the control treatment that we describe in detail below (experimental design).

Intervention Start Date

2026-01-11

Intervention End Date

2026-02-14

Primary Outcomes

Primary Outcomes (end points)

The most important dependent variable (elicited before the trial part and parts 2-5) is participants’ belief about the true task benefit in the main treatment and the corresponding belief about a fundamental that is unrelated to the task benefit in the control treatment.

Primary Outcomes (explanation)

Secondary Outcomes

Secondary Outcomes (end points)

Moreover, we will measure participants’ task performance, i.e., the number of correctly solved encryption tasks in 8 minutes in the trial part and all five following parts.
At the beginning of part 1, we will elicit each participant’s ideal and predicted number of correctly solved encryption tasks (for a task benefit of 50, 60, 70, … 150 points in the main treatment and in the control treatment). At the beginning of part 5, we ask again for the ideal and predicted tasks at what the participants believe is the task benefit level in the main treatment and at the actual task benefit level of 120 in the control treatment.

Secondary Outcomes (explanation)

Experimental Design

Main treatment: Participants in the main treatment complete six parts (Trial and Parts 1–5). Participants are informed in the instructions that the true task benefit lies between 50 and 150 points per solved encryption task and is the mean of a distribution with 150 realizations, but they do not know that this underlying distribution is left-skewed with a true mean of 120.
• Trial: Before starting the trial round participants are asked about their prior belief about the true task benefit (θ0) only knowing that the payoff per task is between 50 and 150 points. Afterwards participants complete a 8-minute trial round of the encryption task (a0) to become familiar with the task itself and their task-specific effort costs.
• Part 1: Participants (i) are reminded of their previous number of solved encryption tasks, report (ii) their ideal number of solved encryption tasks for various hypothetical levels of the unknown task benefit, and (iii) their predicted number of solved tasks for those benefit levels. There is a two percent probability that participants actually have to work on as many encryption tasks in part 5 as they stated as ideal for the true task benefit level in part 1. After stating their ideal and predicted number of solved tasks, participants work on the encryption task for a maximum of 8 minutes (a1) or may instead choose to stop working and spend the remainder of the 10 minutes on an outside option (a distraction task such as using their cellphone with low compensation).
• Parts 2–4: Each of these parts follows an identical structure. Participants (i) are reminded of their previous number of solved tasks, (ii) are asked to state their updated belief about the task benefit (θ1, θ2, θ3), (iii) are shown a signal about the true task benefit (s1, s2, s3),—specifically, 50 out of the 150 possible benefit realizations (“balls”) drawn without replacement that are displayed for approximately 10 seconds—and (iv) again choose between working on the encryption task (a2, a3, a4) or the outside option for 8 minutes. Across Parts 2–4, participants see all 150 realizations once, allowing them to progressively learn the true value of the task benefit.
• Part 5: Participants (i) are reminded of their past action, (ii) state their final belief about the true task benefit (θ4), (iii) report their ideal and (iv) predicted number of solved tasks for what they believe the task benefit is. Then, (v) they complete a final 8-minute work period (a5), including the possibility of spending part of the time on the outside option. With two percent probability each, participants do not do as many tasks as they want in Part 5, but they have to do as many as they stated as ideal for the true task benefit level in either the first part or for what they believe the task benefit is in the fifth part.
After each part, participants complete a short “buffer” questionnaire to introduce a break between the different parts to foster forgetting of the precise signals on task benefit they have seen. The different questionnaires cover sociodemographic characteristics, last math grade in school, the digit span subtest by Wechsler (1939), the Big Five personality traits from John et al. (1991), measures of economic preferences from the Global Preferences Survey by Falk et al. (2018), the Brief Self-Control Scale from Tangney et al. (2004), an average calculation task, and questions regarding the experiment asking how good they remember the “balls” from the signals, and if they have taken a screenshot of them. Lastly, we included a hard to answer mathematical task to be able to detect AI agents.

Control treatment: The control treatment is structured in exactly the same way as the main treatment, except for the following difference. Participants know the exact, fixed task benefit of 120 points per correctly solved encryption task right from the beginning (already before the trial part). There is still a fundamental about which they are receiving signals and they have to report their beliefs about. However, the fundamental is unrelated to the task benefit. The signal structure and the belief elicitation process are identical to the main treatment. The purpose of the control treatment is to show (i) that / to which extent people can update beliefs correctly over time if there is “no reason to trick themselves” (i.e., to update too little on the task benefit to justify the own low performance in the encryption task) and (ii) that people possibly work harder if they cannot trick themselves (caveat: according to DellaVigna and Pope (2022), we often see maximum effort provision in real effort task experiments like ours that is largely insensitive to effort costs and benefits which is the reason why we consider beliefs as our key dependent variable). The control treatment also shows whether and how the cost of effort might change across the parts, allowing to control for a potential decrease in effort over time simply due to participants getting tired.

DellaVigna, S. & Pope, D. (2022). Stability of experimental results: Forecasts and evidence. American Economic Journal: Microeconomics, 14(3), 889-925.
Falk, A., Becker, A., Dohmen, T., Enke, B., Huffman, D. & Sunde, U. (2018). Global Evidence on Economic Preferences, The Quarterly Journal of Economics 133(4), 1645–1692, https://doi.org/10.1093/qje/qjy013.
John, O. P., Donahue, E. M., & Kentle, R. L. (1991). Big Five Inventory (BFI) [Database record]. APA PsycTests.
Tangney, J. P., Baumeister, R. F., & Boone, A. L. (2004). High self-control predicts good adjustment, less pathology, better grades, and interpersonal success. Journal of Personality 72, 271–322.
Wechsler, D. (1939). The measurement of adult intelligence. Williams & Wilkins Co. https://doi.org/10.1037/10020-000.

Experimental Design Details

Not available

Randomization Method

randomization done by computer

Randomization Unit

individual

Was the treatment clustered?

Experiment Characteristics

Sample size: planned number of clusters

no clusters

Sample size: planned number of observations

Our main hypothesis is that in the main treatment, beliefs about the task benefit will converge to a lower level than the true benefit level. The secondary test of this hypothesis is that beliefs in the main treatment in the last period will be lower than in the control treatment (regarding the unrelated fundamental in the latter). We will determine the sample size so that we can detect a difference between the average last-period beliefs in the main treatment and the control treatment at the 5 percent significance level and 80 percent power. Based on Cobb-Clark et al. (2024), we expect that around 60 percent of the participants (naive and partially sophisticated ones) will be prone to misinterpreting themselves, so these participants will drive the difference between the main treatment and the control treatment beliefs. After collecting the first 150 observations (75 in the main and 75 in the control treatment), we will calculate sample size N using the following Stata code after dropping the time-consistent, fully sophisticated, and overly pessimistic participants: power twomeans mean_control (mean_control-3), power(0.8) sd1(x) sd2(y) where mean_control is the mean belief for the control treatment observations, x is the standard deviation of ˆθ4 for the control treatment observations and y is the standard deviation of ˆθ4 for the main treatment observations after dropping time-consistent, fully sophisticated, and overly pessimistic individuals. We consider a deviation of means across treatments of at least 3 as economically meaningful. We will need to divide the resulting N by the share of naïve, partially naïve and overly pessimistic individuals to obtain the number of individuals we will collect data on.

Sample size (or number of clusters) by treatment arms

50-50 for control and main treatment

Minimum detectable effect size for main outcomes (accounting for sample design and clustering)

see above

Supporting Documents and Materials

IRB