How redundant information can help overcome bias

How to cite

Nicholas, Aaron; Silva, Francisco da; González, Nicolás Figueroa; Jara, Ruben. 2026. "How redundant information can help overcome bias." AEA RCT Registry. AEARCTR-0018310. https://doi.org/10.1257/rct.18310-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Aaron Nicholas* , Francisco da Silva , Nicolás Figueroa González , Ruben Jara

* Primary investigator

Registry ID
AEARCTR-0018310
DOI
https://doi.org/10.1257/rct.18310-1.0
First registered
Last updated
Status
On going
Trial dates
2026-05-13 – 2026-12-31
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/18310

Abstract

A common response to the potential manipulation of data by data collectors is the obscuring or transforming of outcome information in order to protect the integrity of the process; e.g. through double blinding, or encryption. This experimental research aims to:
1) check if individuals are able to recognize incentives to manipulate information even when such data have been encrypted by having their labels removed ('outcome blinding').
2) check whether a specific form of encryption (the introduction of redundant labels) successfully reduces manipulation that arose despite outcome blinding.

Intervention

Two treatments to be conducted through Qualtrics, with recruitment from Prolific's panel.

*Baseline treatment*
This is a within-subject design, with three decisions in a fixed sequence.
Participants are given a hypothetical scenario where they are electoral officers attempting to forward the results of an election with two candidates A and B. However, the information they have has been encrypted, and they are uncertain whether the candidate they prefer to win has actually won. The encrypted message is of the form where they see two generic labels, of which one of them is crossed, but the participant does not know which candidate the crossed label belongs to. Participants have to decide whether to forward the information they receive, or to manipulate it by switching the cross to the other label. Participants are paid if the cross is located next to their preferred candidate.
Participants are told the probability that their preferred candidate won.
The three decisions vary as follows:
(1) The probability is set such that they have no incentive to manipulate the information.
(2) The probability is set such that they have incentive to manipulate the information.
(3) The probability is set equal to scenario (2); in addition, the set of labels are expanded to four, with participants told that one of them corresponds to their preferred candidate, and the remaining three correspond to the other candidate; this effective removes the incentive to manipulate the information.

*Do-no-harm treatment*
Identical to the baseline, except the following:
(1) The probability is set such that they have incentive to manipulate the information.
(2) The probability is set such that they have no incentive to manipulate the information.
(3) The probability is set equal to scenario (2); in addition, the set of labels are expanded to four, with participants told that three of them correspond to their preferred candidate, and the remaining one corresponds to the other candidate; while this increases the chance of being paid when manipulating information, the optimal decision remains not manipulating information

Experimental design

Please refer to intervention details above.

Design details

Randomization method
Randomization is done through Prolific's recruitment.
Randomization unit
By survey treatment.
Clusters
No clustering.
Observations
500-600
Arms
250-300 observations per treatment, 500-600 total.
Primary outcomes
Let B1, B2, B3 be the proportion of participants who choose to manipulate information in the three baseline treatment decisions, and D1, D2, D3 the corresponding proportions in the Do-no-harm treatment. We have two primary tests: (1) B2>B1; D1>D2 and (B2+D1)>(B1+D2). These are the tests that tell us whether participants recognize the incentive (to/to not) manipulate information in the presence of outcome blinding. These will be tested using within-subject non-parametric tests, as well as regressions. (2) Let B3* and D3* be the proportion of participants who choose to manipulate information in the respective third decision conditional on 'optimal' behavior in the first and second decision. (1-B3*)>(D3*) tells us that the expansion of the encrypted labels successfully reduces the incidence of manipulation, even after accounting for any increased manipulation arising from the expansion cueing or nudging participants. These will be tested using between-subject non-parametric tests, as well as regressions.
Secondary outcomes
We will also test B2>B3 which is a test of whether the expansion reduces manipulation (independent of whether they got B1 and B2 correct); and whether D3>D2, which is a test of whether expansion induces participants to manipulate even when they shouldn't. (B2+D2)>(B3+D3) then becomes a test of whether overall manipulation is reduced as a result of expansion. This is not a primary outcome because it potentially includes a sizable group of subjects who are confused by the incentives in even the first two decisions; one of the primary goals of the experiment is to test whether expansion reduces manipulation among those who recognize the incentive to manipulate even when there is outcome blinding. Hence even in these tests we will condition the sample to those who have appropriate understanding of their first two decisions (but not on their actual behavior).

Cited by

No citation data available for this registration yet.

Keywords