The experimenters' dilemma: inferential preferences over populations
Abstract
We compare three populations commonly used in experiments by economists and other social scientists: undergraduate students at a physical location (lab), Amazon's Mechanical Turk (MTurk), and Prolific. The comparison is made along three dimensions: the noise in the data due to inattention, the cost per observation, and the elasticity of response. We draw samples from each population, examining decisions in four one-shot games with varying tensions between the individual and socially efficient choices. When there is no tension, where individual and pro-social incentives coincide, noisy behavior accounts for 60% of the observations on MTurk, 19% on Prolific, and 14% for the lab. Taking costs into account, if noisy data is the only concern Prolific dominates from an inferential power point of view, combining relatively low noise with a cost per observation one fifth of the lab's. However, because the lab population is more sensitive to treatment, across our main PD game comparison the lab still outperforms both Prolific and MTurk.
Study specs
- Institution
- University of Cambridge,University of Verona,University of Oxford,University of Pittsburgh
- Discipline
- Social Science Research Methods
- Year
- 2021
- Human Data Platform
- Prolific
- Source
- View Source Google Scholar
Peer Review & Critical Discussion
Potential Selection Bias in 2023 Cohort
The participant pool shows a concerning overrepresentation of users from high-income demographics. Looking at Table 3, we can see that 78% of respondents had annual incomes above $75k, which significantly limits the generalizability of these findings to broader populations.
Non-naive Participants Issue
I've noticed a methodological concern regarding participant naivety. Given that Prolific users often complete multiple studies, there's a real risk that participants had prior exposure to similar experimental paradigms, which could confound the results.
RLHF Applicability to This Study Design
The implications for RLHF training pipelines are understated. If we accept the authors' conclusions about preference stability, this has direct consequences for how we should structure reward model training. The temporal decay effect described in Section 4.2 is particularly relevant.
Verify your expertise to join discussion
Create an account and verify your credentials to participate in peer discussions.