Concerns and recommendations for using Amazon MTurk for eating disorder research

Abstract

Objective: Our original aim was to validate and norm common eating disorder (ED) symptom measures in a large, representative community sample of transgender adults in the United States. We recruited via Amazon Mechanical Turk (MTurk), a popular online recruitment and data collection platform both within and outside of the ED field. We present an overview of our experience using MTurk. Method: Recruitment began in Spring 2020; our original target N was 2,250 transgender adults stratified evenly across the United States. Measures included a demographics questionnaire, the Eating Disorder Examination-Questionnaire, and the Eating Attitudes Test-26. Consistent with current literature recommendations, we implemented a comprehensive set of attention and validity measures to reduce and identify bot responding, data farming, and participant misrepresentation.Results: Recommended validity and attention checks failed to identify the majority of likely invalid responses. Our collection of two similar ED measures, thorough weight history assessment, and gender identity experiences allowed us to examine response concordance and identify impossible and improbable responses, which revealed glaring discrepancies and invalid data. Furthermore, qualitative data (e.g., emails received from MTurk workers) raised concerns about economic conditions facing MTurk workers that could compel misrepresentation.Discussion: Our results strongly suggest most of our data were invalid, and call into question results of recently published MTurk studies. We assert that caution and rigor must be applied when using MTurk as a recruitment tool for ED research, and offer several suggestions for ED researchers to mitigate and identify invalid data.

Practice makes perfect: Commentary on Burnette et al. (2021)

“…instructed-response items and instructional manipulation checks (Kung, Kwok, & Brown, 2018). Burnette et al (2021) used only two instructedresponse items in their entire survey and used no instructional manipulations; 16 (0.5%) individuals failed. Perhaps this motivated their posthoc search for alternative strategies of filtering responses from people who passed a priori checks.…”

“…As such, this search was premised on a lack of trust in their a priori, evidence-based approach. Burnette et al (2021) queried participants twice about their gender. One strategy devised to detect suspected invalid responders was to exclude participants with inconsistent answers.…”

“…A related recommendation by Burnette et al (2021) to exclude supposed invalid responders is to use two similar items to "cross-reference" responses (recommendation's nos. 3, 6, 7, and 8).…”

“…Norms for commonly used measures of eating disorder (ED) psychopathology applicable to underrepresented groups are sorely needed. Burnette et al (2021) aimed to establish such norms for transgender adults using Amazon's MTurk, a crowd-sourcing platform. They were thwarted from achieving this laudable goal by what they understood to be invalid responses.…”

“…The context in which Burnette et al (2021) derived their recommendations is critical to consider: a failed attempt to recruit their population of interest. Accordingly, their recommendations were post-hoc constructions (e.g., "our data screening took on a life of its own"; p. 6), rather than a priori strategies subjected to empirical validation.…”