Manipulation-Proof Machine Learning

How to cite

Blumenstock, Joshua; Bjorkegren, Daniel; Knight, Samsun. 2019. "Manipulation-Proof Machine Learning." AEA RCT Registry. AEARCTR-0004649. https://doi.org/10.1257/rct.4649-1.4000000000000001

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

Joshua Blumenstock* , Daniel Bjorkegren , Samsun Knight

* Primary investigator

Registry ID
AEARCTR-0004649
DOI
https://doi.org/10.1257/rct.4649-1.4000000000000001
First registered
Last updated
Status
Completed
Trial dates
2019-07-01 – 2021-01-01
Countries
Kenya (Nairobi)
Registry record
http://www.socialscienceregistry.org/trials/4649

Abstract

An increasing number of decisions are guided by machine learning algorithms. An individual’s behavior is typically used as input to an estimator that determines future decisions. But when an estimator is used to allocate resources, individuals may strategically alter their behavior to achieve a desired outcome. This paper develops a new class of estimators that are stable under manipulation, even when the decision rule is fully transparent. We explicitly model the costs of manipulating different behaviors, and identify decision rules that are stable in equilibrium. Through a large field experiment in Kenya, we test decision rules estimated with our strategy-robust method.

Intervention

The subject population consists of Kenyans aged 18 years or older who own a smartphone and are able to travel to the Busara center in Nairobi. A baseline survey will be conducted with all consenting participants in the Busara offices in Nairobi. At the time of enrollment, participants will install a "Sensing" app on their smartphone. More details on the app and the baseline survey are provided below.

While enrolled in the study, each participant will be dynamically assigned to a treatment for each week. Users will be sent both a text message (SMS) and an app push-notification on a weekly basis that directs them to the app. After a user opens the app, it will ask them to opt in to a "challenge". If they accept, they will observe a challenge which provides them some monetary incentive to change their behavior, as observed through the Sensing app.

Experimental design

The subject population consists of Kenyans aged 18 years or older who own a smartphone and are able to travel to the Busara center in Nairobi. We anticipate recruiting roughly 1,200 subjects for the main study, primarily through in-person solicitation in public spaces (e.g., public markets) by Busara research staff.

Informed consent is requested from each subject, following the procedures registered with the Committee for the Protection of Human Subjects at U.C. Berkeley. Subjects must actively decide to participate in the research; no pressure or undue influence will be given to induce subjects to participate.

A baseline survey will be conducted with all consenting participants in the Busara offices in Nairobi. At the time of enrollment, participants will install a "Sensing" app on their smartphone. While enrolled in the study, each participant will be dynamically assigned to a treatment for each week. Users will be sent both a text message (SMS) and an app push-notification on a weekly basis that directs them to the app. After a user opens the app, it will ask them to opt in to a "challenge". These challenges may use either a naive or manipulation-proof decision rule, and we will experimentally vary the level of transparency of the algorithm to the subject.

Design details

Randomization method
Randomization done in office by a computer
Randomization unit
Individual-week
Clusters
1500 individuals
Observations
21600 individual-weeks
Arms
Approximately 50 simple challenges and 10 complex challenges
Primary outcomes
Note that our RCT is not standard in the sense that our primary objective is not to estimate the average treatment effect of an intervention on primary and secondary outcomes. Instead, the RCT is designed to allow us to estimate: - The cost of manipulating different behaviors, as observed through the Sensing app (e.g., making a phone call, using multiple WiFi hotspots) - The ease with which subjects can "game" algorithmic decision rules - Whether transparent decision rules are easier to game than opaque ones

Papers from this trial

  1. Björkegren, D., Blumenstock, J. E., & Knight, S. (2020). Manipulation-Proof Machine Learning. ArXiv:2004.03865 [Cs, Econ]. http://arxiv.org/abs/2004.03865 https://arxiv.org/abs/2004.03865

Cited by

No citation data available for this registration yet.

Keywords