By Aaron Moss, PhD & Leib Litman, PhD
When researchers collect data online, it’s natural to be concerned about data quality. Participants aren’t in the lab, so researchers can’t see who is taking their survey, what those participants are doing while answering questions, or whether participants are who they say they are. Not knowing is unsettling.
Recently, the research community has been consumed with concern that workers on Amazon’s Mechanical Turk (MTurk) are cheating requesters by faking their location or using “bots” to submit surveys. These concerns originated and have been driven by reports from researchers that there are more nonsensical and low-quality responses in recent studies conducted on MTurk. In these studies, researchers have noticed that several low-quality responses are pinned to the same geolocation. In this blog, we’d like to add some context to the conversation, share the findings from our internal inquiry, and inform researchers what CloudResearch is doing to address the issue.
The recent concern about bots appears to have begun on Tuesday, August 7th, 2018, when a researcher asked the PscyhMap Facebook group if anyone had experienced an increase in low quality data. In just the third response to that thread, another researcher suggested, “maybe a machine?” Soon, other researchers were reporting an increase in nonsense responses and low-quality data, although at least a few reported no increase in junk responses to their studies. The primary piece of evidence causing researchers to suspect bots was that most of the low-quality responses were tagged to the same geolocation and a few places in particular—Niagara Square in Buffalo, NY; a lake in Kansas; and a forest in Venezuela. What’s more, many respondents from these geolocations provided suspicious responses to open-ended questions, often answering with “GOOD STUDY,” or “NICE.”
Although this activity raises concerns, the conversation, so far, has overlooked some important points. Most critically, while it is clear some researchers have unfortunately collected several bad responses, the research community does not yet know how widespread this problem is. Diagnosing the issue requires knowing how many studies don’t fit the pattern, as well as how many do.
At CloudResearch, we track the geolocation of all surveys submitted in studies run on our platform. In the last 24 hours, we have worked to determine whether there is a growing problem of multiple submissions from the same geolocation. In reviewing over 100,000 studies that have been launched on CloudResearch, we see that the rate of submissions from duplicate geolocations typically bounced from less than 1% to 2.5% within a study—a number that could be explained by people submitting surveys from the same building, office, internet service provider, or even the same city. Geolocations are not precise, an issue we will discuss in more detail in a future blog post.
Based on this analysis, we set 2.5% as the threshold for detecting suspicious activity. Over 97% of studies have not reached this threshold, showing that the overwhelming majority of studies have not been affected by data coming from the same geolocation.
However, when we look at the rate of duplicate submissions based on geolocation over time, we see that in March of this year the percentage of duplicate submissions began edging up. Clearly, this is a problem, but a problem that has emerged only recently.
At CloudResearch, we are developing tools that will help researchers combat suspicious activity. We have identified that all suspicious activity is coming from a relatively small number of sources. We have additionally confirmed that blocking those sources completely eliminates the problem. In fact, once the suspicious locations were removed, we saw that the number of duplicate submissions had actually dropped over the summer to a rate of just 1.7% in July 2018.
In the coming days, we will launch a Free feature that allows researchers to block suspicious geolocations. This means researchers will be able to block workers from suspicious geolocations, excluding submissions from those locations in their data collection. We will also launch a Pro feature that allows researchers to block multiple submissions from the same geolocation within a study. This feature will cast a wider net and may block well-intentioned workers using the same internet service provider, or working in the same library. This tool will give researchers greater confidence that they are not receiving submissions from anyone using the same location to submit junk responses.
Our data, and the work of multiple researchers, show there has been a recent increase in the number of low quality responses submitted on Mechanical Turk. Data from the CloudResearch database show that the vast majority of all studies, and the vast majority of recent studies, have never been affected by the current concern of bots. What we still don’t know about the recent issue, is whether these responses are coming from bots or foreign workers using a VPN to disguise their location and submit surveys intended to sample US workers. Either way, in the coming days CloudResearch will release tools that allow researchers to block workers from suspicious locations and to decide how narrowly they would like to set the exclusion criteria. Concerns about bots and low quality data on MTurk are not new. But at CloudResearch we will continue to look for ways to ensure quality data and to make conducting online research easier for researchers.