Since launching in 2005, Amazon’s Mechanical Turk (MTurk) has become an increasingly important component to research in many social sciences. MTurk is an online platform composed of between 100,000 and 200,000 active people worldwide (Robinson et al., 2019). People on MTurk (i.e., Turkers) complete “Human Intelligence Tasks” (HITs; e.g., surveys or other computerized tasks) in exchange for immediate monetary compensation; Robinson et al., 2019). Since launching, the use of MTurk samples has been increasingly popular in many social science fields (Kennedy et al., 2020).
A major reason that MTurk has become a staple in modern research is due to the speed of data collection and access to a broader population compared to traditional undergraduate samples. While Turkers are somewhat younger, more liberal, and tend to be better educated than the general population as a whole, they are considerably more representative than college samples (Paolacci & Chandler, 2014). The ability to sample niche groups also aids in research that is difficult or impossible with more traditional recruitment techniques.
Despite MTurk’s advantages, there are several limitations. The current article is focused on four main issues that have been heavily discussed in past research: inattention, trolling, automated bots, and foreign workers (Kennedy et al., 2020). Each of these issues has displayed worrying trends over the last two years, with researchers increasingly finding that data collected through MTurk displays the tell-tale signs of being compromised. Recently, however, CloudResearch has instituted a more rigorous pre-screening procedure.
The current research aims to test whether data quality concerns may be mitigated by this vetting procedures. To this end, we compare data from an open sample of Turkers to data collected from CloudResearch’s “Approved Participants” list. We then examine how data screening solutions that measure in-survey performance might be used to further remove any problematic participants at the study-level.
Data Quality Concerns
It has been suggested that Turkers often give inadequate attention to study materials (e.g., Kennedy et al., 2020). Because they are not directly monitored, Turkers may divide their attention between the HIT and other stimuli. This can lead to random or repetitive response patterns that compromise the validity and reliability of the data (Meade & Craig, 2012). While this concern has drawn considerable attention, research has indicated that Turkers may actually be more attentive than online college samples (Hauser & Schwarz, 2016). Thus, this issue is not new and applies to traditional research populations as well.
Another issue is the possibility of “trolling”– i.e., intentionally providing insincere responses to be provocative or humorous (Lopez & Hillygus, 2018). Some researchers suggest that trolling in MTurk samples has increased recently, as endorsement of low base rate responses is implausibly high and increasing (Ahler et al., 2021). However, others suggest that it is difficult to distinguish trolling from other problems, such as inattention (Moss et al., 2021). Regardless, like inattentive responding, trolling would compromise both the reliability and validity of data.
“Bots” and Automated Programs
In 2018, concerns were raised that automated programs (i.e., “bots”) were being used to complete large amounts of surveys (Kennedy et al., 2020). Researchers remain concerned about this possibility, given that bots could not only compromise their data’s validity and reliability but also result in financial losses. Fortunately, recent research suggests that this concern may have been misplaced, as problematic responses continue even when reCAPTCHA technology is used to filter out automated programs, and many problematic Turkers show evidence of uniquely human processing (Moss et al., 2021). Instead, evidence suggests that the original “bot-panic” may have been due to foreign workers.
A major concern for MTurk researchers is their (limited) ability to restrict participation to a target location, such as the United States. While MTurk provides screening tools to limit IP addresses to specific locations, these restrictions can be circumvented (Kennedy et al., 2020). This is particularly problematic for researchers who are restricting studies to American samples, where pay rates are substantially higher than in other regions (Kennedy et al., 2020). This can compromise the data’s reliability and validity in multiple fashions. For example, some foreign workers may only have a limited understanding of English (Moss et al., 2021). Further, many psychological phenomena of course vary across cultures, and unknowingly sampling individuals from outside the target culture would introduce systematic error variance.
Consequences for Data Quality
Each of these issues can potentially compromise the reliability and validity of data. For example, inattention or poor English comprehension can result in repetitive response patterns. When scales include reverse-scored items, these items can become positively correlated with negative-scored items, compromising reliability.
An even more serious concern is that these issues can compromise validity – such that scales no longer measure the target construct. A scale’s construct validity is typically assessed by examining the surrounding nomological network (Cronbach & Meehl, 1955). If the above issues are severe, the established nomological network surrounding a measure can disappear. Consistent with this, Chmielewski and Kucker (2020) observed a significant loss in the reliability and validity of unscreened data, compared to data that had been screened for inattention.
Data Screening Solutions
Several data-screening methods have been developed to partially address the above issues, such as attention-check items (Oppenheimer et al., 2009). For example, researchers can directly instruct participants to select a particular response and discard participants who fail to provide it.
Another frequent recommendation is to include at least one open-ended question (Kennedy et al., 2020). Bots and participants who are not English-proficient will often provide non-sense answers to such questions. Inattentive participants will often leave such items blank or provide one-word answers, and thus researchers are typically encouraged to use questions that could not reasonably be answered in one word.
A more recent innovation is to include a number of questions with low base-rate response options (e.g., “Have you ever been convicted of a federal crime”; Lopez & Hillygus, 2018). The selection of any one low base-rate response is not necessarily indicative of the discussed issues, as it is of course possible that an occasional participant is telling the truth. Nonetheless, the selection of multiple low-base rate responses provides more definitive evidence of such problems.
Finally, reCAPTCHA software can be used to root out bots (von Ahn et al., 2008). With a mouse click, reCAPTCHA differentiates the movement patterns produced by human respondents from those produced by automated programs. However, this solution only screens out a small fraction of bad data (Moss et al., 2021).
While each of these solutions helps reduce common issues, they are not fool-proof. First, they are time-consuming to implement. If researchers recruit their target sample size and then discard 20% of the data, the resulting study is underpowered. They can of course replace these participants through additional data-collection, but this takes additional time. Past research also demonstrates that any single screening item is unlikely to remain effective for long, as Turkers learn what kinds of measures to look out for and show improvements in performance over time (Hauser & Schwarz, 2016).
CloudResearch’s Pre-Screening Solution
In 2020, CloudResearch’s MTurk Toolkit was developed to help researchers sample from participants on MTurk with added flexibility and demographic targeting capabilities (Litman et al., 2020). Specifically, CloudResearch has been vetting worker accounts on MTurk since spring 2020, implementing a platform-level solution to data quality problems. The goal of this vetting process is to identify Turkers who are likely to provide quality responses to surveys and separate them from Turkers who show signs of the issues reviewed above. As of this writing, CloudResearch has vetted more than 95,000 worker accounts (~90% of active accounts, Litman et al., 2020) and placed more than 60,000 Turkers onto an Approved List. Each month about 5,000 new worker accounts are vetted.
In Summer 2020, two of the current authors collected Sample 1 on MTurk without using the screening procedures described above. As we report below, it was apparent that the resulting data was highly problematic. After correspondences with CloudResearch, we decided to re-run the study using CloudResearch’s Approved Participants, to compare the quality of the data collected across samples. After additional improvements to the Pre-Screening procedure, we re-ran the study again roughly one year later to examine possible further improvements. Importantly, all studies used previously-validated measures of political ideology, whose psychometric properties and nomological networks have been established. As such, a comparison of the samples helps to illustrate the efficacy of CloudResearch’s screening procedure. We also included several in-survey data-screening solutions in Samples 2 and 3 to examine if data from participants who pass these checks are even more reliable and valid than the data resulting from CloudResearch’s pre-screening procedure alone.
Open Science Practices
We report all measures relevant to current concerns, how we determined our sample size, and data exclusions. Verbatim method files, data, and analytic code can be accessed at the Supplementary Materials section. A few additional measures were included that have not been as thoroughly validated in prior research. We disclose all such measures in full in the verbatim method files, but do not focus on them here. In all studies, we aimed to collect a final analyzable sample of 250 participants (Schönbrodt & Perugini, 2013). In all studies, we over-sampled in order to meet this goal after data exclusions.
In each sample, participants were recruited through MTurk. To be eligible, participants needed at least a 95% HIT approval rate. In Samples 2 and 3, all participants had to additionally pass CloudResearch’s vetting procedure (described below).
This sample was collected in June of 2020 and included 274 (192-males, 82-females; Mage = 37.11, SD = 10.56) participants. Our sample identified primarily as being either White/non-Hispanic (46.35%) or Black/African-American (45.26%) and conservative-leaning (M = 5.18, SD = 1.75). Please note that this does not match the typical demographics of Turkers (Paolacci & Chandler, 2014), raising concerns about the accuracy of demographic reports. Due to not being within the target population of the original study (i.e., current resident of United States), 17 participants were removed from the sample and not included in analyses.
This sample was collected shortly after the implementation of the prescreening procedure in July of 2020, and it included 306 (162-females, 143-males, 1-unidentified; Mage = 39.57, SD = 13.76) participants. Our samples identified primarily as being White/non-Hispanic (70.59%) and liberal-leaning (M = 3.69, SD = 1.87). Note that these demographics more closely resemble the typical demographics of MTurk samples (Berinsky et al., 2012). Due to not being within the target American population, 7 participants were removed from analyses.
This sample was collected in August of 2021, after improvements to the vetting procedure. It included 300 (170-females, 129-males, 1-unidentified; Mage = 42.27, SD = 13.14) participants. Participants primarily identified as White/non-Hispanic (72.67%) and liberal-leaning (M = 3.44, SD = 1.81). No participants were removed due to residing outside of the United States.
CloudResearch Vetting Process
Part of CloudResearch’s MTurk vetting process is an attention instrument (Litman et al., 2020), which checks for participant attention and basic English comprehension. Participants are presented with a target word, four response-option words, and asked which response-option is most similar to the target. The stimuli were created using an associative semantic network model, which assigns weights to word-pairs based on a large corpora of English language texts. The resulting questions are ones where the correct response option is highly associated with the target word (e.g., apologize-mistake) and the distractor response options have low associations with the target word (e.g., apologize-particle). Because items are randomly selected from a large library, participants are less likely to learn the correct responses, share them online, or create scripts that can respond automatically.
CloudResearch’s vetting process also includes items that check for mischievous responding, items that check for bots (e.g., reCAPTCHA), and technical checks that look for evidence of other automated solutions (Litman et al., 2020). In tandem, CloudResearch also monitors the number of times workers are added by MTurk Toolkit users to a “universal exclude” list and uses this data to further identify questionable respondents. This vetting process appears to effectively identify some of the most problematic participants on MTurk (Moss et al., 2021). In this study, we examine how well the Approved List improves data quality.
Procedures and Measures
Participants completed the following scales in a random order, with item order within each instrument also randomized. A full list of items and other related material can be found in the Supplementary Materials section.
Social Dominance Orientation (SDO)
Participants in both samples completed the 16-item SDO-7 scale (Ho et al., 2015). This instrument contains an equal number of positively-scored items and negatively-scored items. Participants rated their level of favorability towards each item on a 7-point Likert-scale (1–Strongly oppose to 7–Strongly favor). It has been found to be internally consistent in past research (α = .89; Ho et al., 2015).
Right Wing Authoritarianism (RWA)
Participants also completed the 22-item RWA scale (Altemeyer, 2006). Items includes 12 negatively-scored items. Participants were instructed to rate their agreement with each item on a 9-point Likert-scale (-4–Very strongly disagree to 4–Very strongly agree). This scale has been found to be internally consistent in past research (α = .90; Saunders & Ngo, 2017)
Participants also completed the ASC (Dunwoody & Funke, 2016) measure of RWA. It contains an equal number of positive and negatively-scored items. Participants rated their agreement with each item on a 7-point Likert-scale. It has been found to be internally consistent in past research (α = .85; Dunwoody & Funke, 2016)
Left-Wing Authoritarianism (LWA)
Participants also completed the 22-item LWA scale (Conway et al., 2018). It includes 10 negatively-scored items. Participants rated their agreement with each item on a 7-point Likert-scale. It has been found to be internally consistent in past research (α = .89; Conway et al., 2018)
Participants also completed the 10-item Dangerous World Beliefs scale (DWB) and the 14-item Competitive Jungle World Beliefs scale (CJWB) (Duckitt et al., 2002). These scales each contain an equal number of positively and negatively-scored items. All scales ask participants to indicate their level of agreement using a 7-point Likert-scale. The DWB scale (α = .80) and the CJWB scale (α = .84) have been found to be internally consistent in past research (Duckitt et al., 2002).
Fundamental Social Motives Inventory (FSMI)
Participants also rated their agreement with 12 items that measured self-protection and disease avoidance motives (Neel et al., 2016). The Self-Protection and Disease-Avoidance subscales contain 1 and 3 negatively-scored items (respectively). Participants rated that agreement with each item on a 7-point Likert-scale. These scales have been found to be internally reliable in past research (αs > .80; Neel et al., 2016).
Measure of Moral Motives (MMM)
Participants completed the 30-item Measure of Moral Motives Scale (Janoff-Bulman & Carnes, 2016), which measures 6 motives. Here, we focus on Social-Order and Social-Justice motives, which have been linked to Conservativism and Liberalism (respectively). Participants rated their agreement with each item on a 7-point Likert-scale. The MMM contains no negatively-scored items. The MMM scales have all been found to be internally reliable in past research (αs > .70; Janoff-Bulman & Carnes, 2016).
Moral Foundation Sacredness Scale (MFS)
Participants completed the 24-item Moral Foundations Sacredness Scale (MFS; Graham & Haidt, 2012). Items ask participants to indicate how much they would need to be paid to engage in actions typically viewed as immoral. It has 5 subscales (Loyalty, Authority, Purity, Harm, & Fairness). The Binding Foundations (Loyalty, Authority, & Purity) are related to Conservativism; while the Individualizing Foundations (Harm & Fairness) are modestly associated with Liberalism. It has no negatively-scored items. Each scale (αs > .65) has been found to be at least modestly reliable (Graham & Haidt, 2012).
Generalized Prejudice Items
To measure generalized dimensions of prejudice, participants responded to warmth thermometer items for 44 different social groups. Twenty-four items were modeled after items from Duckitt and Sibley (2007) to measure derogated groups (e.g., Immigrants), dangerous groups (e.g., Violent Criminals), and dissident groups (e.g., Protestors). We selected items that seemed to represent the same dimensions in a 2020 American sample, and added a few unique items. We also added items to measure prejudice against conservative groups (Brandt et al., 2014). Participants were instructed to rate their warmth towards each group along a 0–100 continuum (0–No Warmth to 100–Extremely Warm). We reverse-scored all items to form a measure of prejudice.
In-Survey Screening Items
Participants in Samples 2 and 3 were asked to complete four Attention Check items, which directly instructed them to select a particular response (e.g., I will select six now and show I'm paying attention). These were distributed across four scales.
Participants in Samples 2 and 3 were also asked two direct validation questions (i.e., “Were you paying attention when answering questions?”, “Were you telling the truth?”). Participants were assured that they would be paid regardless of their answers. Past research indicates that participants who indicate inattention or deception provide lower-quality data (Lopez & Hillygus, 2018).
In Samples 2 and 3, we also asked several open-response items pertaining to their height, state of residency, and annual income to screen for nonsense responses. We also added four low base-rate items (e.g., Are you in a gang?). The endorsement of two or more low base-rate responses is indicative of trolling or other problems (Lopez & Hillygus, 2018).
Finally, we asked Samples 2 and 3 participants to complete two reCAPTCHA items, at the beginning and end of the survey. This helped to screen out responses provided by automated software.
Initial Screening of Participants
We first screened Sample 2 and 3’s responses to the In-Survey Screening items. We used this to create two subsamples, Sample 2-Additional Screening (S2-AS) and Sample 3-Additional Screening (S3-AS), who displayed in-survey evidence of providing high-quality data.
In Sample 2, 43 of the 306 total participants (14%) were excluded—7 because they were outside the United States; 29 for failing two or more attention checks; 5 for directly indicating they provided low-quality data; and 2 for endorsing two or more low base-rate items. After their removal, S2-AS included 263 participants (Male = 114, Female = 148, Non-Conforming = 1; Mage = 40.98, SD = 14.05). This sub-sample identified primarily as being White/non-Hispanic (73.76%) and liberal-leaning (M = 3.67, SD = 1.90).
In Sample 3, 22 of the total 300 participants were excluded (7.3%). This suggests a significant improvement in the Approved Worker Pre-Screening procedure over time, as fewer participants were excluded than Sample 2, χ2 (1, N = 606) = 7.14, p = .008. 19 were excluded for failing two or more attention-check items, and 3 for directly indicating their data was low-quality. After their removal, S3-AS included 278 participants (Male = 118, Female = 159, Non-Conforming = 1; Mage = 42.89, SD = 13.16). They primarily identified as White/non-Hispanic (74.10%) and liberal-leaning (M = 3.47, SD=1.81).
Correlation Between Positive and Reverse-Scored Items
Perhaps the clearest indication of problematic responses by Sample 1 lies in the correlation between responses to positive and negative-scored items. Seven previously validated scales included negative-scored items, which typically correlate negatively with positively-scored items on the same instrument. In contrast to this typical finding, the average response to reverse-scored items was positively correlated with the average response to positively-scored items in Sample 1 (see Table 1). In Samples 2 and 3, these correlations returned to their typical negative direction (Sample 2 mean r = -.43; Sample 3 mean r = -.56). In the sub-samples that passed additional screening, these inverse correlations were slightly stronger (S2-AS mean r = -.52; S3-AS mean r = -.67).
|7. FSMI: Harm Avoidance||0.30*||-0.43*||-0.50*||-0.61*||-0.71*|
|8. FSMI: Disease Avoidance||0.35*||-0.57*||-0.62*||-0.60*||-0.71*|
Note. S1 = Sample. S2-Full = All Sample 2 participants. S2-AS = Sample 2 participants who passed additional screening.
*p < .05.
We next examined inter-item reliability. As noted above, negatively-scored items are included on 7 scales. When we properly scored these scales in Sample 1, many of them proved to be highly unreliable with alphas as low as .18 (see Table 2). When we improperly scored these questionnaires by leaving negative-scored items in their “raw” form, their reliability coefficients actually increased. This lends further supports the idea that many Sample 1 participants provided repetitive responses. In Samples 2 and 3, this issue was eliminated. All scales were internally reliable when properly-scored, with αs > .82 (see Table 2). Improperly scoring them lowered each scale’s reliability (Sample 2 mean α = .55; Sample 3 mean α = .45). Additional screening helped to make this pattern slightly more apparent, but the effect was relatively minimal (unscreened mean α = .889; additional screening mean α = .894). It is important to note that this issue is not apparent in scales containing no negative-scored items. For these scales, a consistent tendency to favor a certain side of the scale can result in an internally reliable scale. This underscores the utility of including negatively-scored items in online data collection.
||S2 - Full
||S2 - AS
||S3 - Full
||S3 - AS
|Scales Containing Negative-Scored Items
|Scales Without Negative-Scored Items:
|13. Social Order||0.83||--||0.89||--||0.89||--||0.86||--||0.87||--|
|14. Social Justice||0.73||--||0.90||--||0.90||--||0.90||--||0.90||--|
|18. Ingroup Loyalty||0.86||--||0.76||--||0.76||--||0.80||--||0.78||--|
The most serious problem is when the validity of instruments themselves is compromised, such that they no longer measure the target construct. Classically, construct validity is assessed by examining a “nomological network” that contains positive correlations with measures of similar constructs, and weaker correlations with measures of more unrelated constructs (Cronbach & Meehl, 1955). Fortunately, many aspects of the nomological network for the current measures are well-established. Table 3 summarizes the most important aspects of this network. Broadly, the typical pattern of correlations is not apparent in Sample 1. Instead, the pattern of correlations can be better interpreted as a consistent repetitive response tendency extending across multiple measures. In Samples 2 and 3, however, the more typical network re-appears.
|Variable||Political Conservativism (vs. Liberalism)
|1. MMM: Social Justice||0.15*||-0.50*||-0.56*||-0.45*||-0.52*|
|3. MFS: Harm||0.15*||-0.12*||-0.15*||-0.06||-0.09|
|4. MFS: Fairness||0.19*||>-.01||-0.02||-0.02||-0.04|
|10. MMM: Social Order||0.46*||0.49*||0.52*||0.47*||0.48*|
|11. MFS: In-Group Loyalty||0.20*||0.22*||0.25*||0.18*||0.18*|
|12. MFS: Authority||0.20*||0.20*||0.20*||0.16*||0.15*|
|13. MFS: Purity||0.15*||0.16*||0.18*||0.10||0.10|
|16. Generalized Prejudice: Illegal Behaviors||-0.32*||0.24*||0.27*||0.32*||0.30*|
|17. Generalized Prejudice: Derogated||-0.19*||0.18*||0.18*||0.15*||0.16*|
|18. Generalized Prejudice: Liberal Dissidents||-0.17*||0.60*||0.65*||0.64*||0.67*|
|19. Generalized Prejudice: Conservative Dissidents||-0.34*||-0.41*||-0.44*||-0.35*||-0.38*|
|20. Generalized Prejudice: Dangerous||-0.34*||-0.02||0.02||0.17*||0.19*|
|21. Generalized Prejudice: Foreign||-0.18*||0.34*||0.37*||0.32*||0.34*|
Note. See Table 1 notes for abbreviations.
*p < .05.
In contrast to previous research (Janoff-Bulman & Carnes, 2016; Sibley & Duckitt, 2008), Conservatism was related to greater Social-Justice motives in Sample 1; while Liberalism was related to greater prejudice against many groups (including liberal groups). This pattern is consistent with repetitive response tendencies (i.e., favoring the high side of the scale would result in reports of Conservatism, social justice motives, and less-prejudice). While other findings are more consistent with past research, they could also reflect a repetitive response tendency.
The typical nomological network re-appeared in Samples 2–3. Social-Justice motives were related to Liberalism (Janoff-Bulman & Carnes, 2016); while most forms of prejudice were related to Conservatism (Duckitt & Sibley, 2007). These samples also replicated an important exception to the latter pattern: Liberals actually expressed greater prejudice toward conservative groups (Brandt et al., 2014).
Finally, Sample 1 was unable to replicate more nuanced distinctions between RWA and SDO, but Samples 2-3 were able to replicate such distinctions (see Table 3). Compared to RWA, SDO exhibited stronger correlations in Samples 2–3 with being male (Pratto et al., 1997), lower concern with Harm/Fairness (Sinn & Hayes, 2017), greater prejudice against derogated groups (Duckitt & Sibley, 2007), and more Competitive Worldviews (Duckitt et al., 2002). By contrast, RWA exhibited stronger correlations (than SDO) in Samples 2–3 with Religiosity (Heaven & Connors, 2001), the ‘Binding’ Foundations (Sinn & Hayes, 2017), and Dangerous Worldviews (though its typical correlation with prejudice against dangerous groups was not replicated; Duckitt et al., 2002). None of these findings were apparent in Study 1.
Additional screening of Samples 2 (S2-AS) and 3 (S3-AS) made all of these patterns slightly more apparent, but the effect was relatively minimal.
MTurk has provided researchers with a reliable, efficient, and affordable participant pool that is more representative than college student samples. Recently, however, the quality of data provided by MTurk samples has come under scrutiny. To overcome these issues, CloudResearch implemented a vetting procedure.
To investigate the impact of these vetting procedures, three samples were collected using largely identical measures at different times: before implementation, immediately after implementation, and one year after implementation. Screening items were also added to Samples 2–3. While 12% of Sample 2 and 7% of Sample 3 were removed (e.g., because of failed attention checks), this is significantly lower than rates in unvetted MTurk samples (Kennedy et al., 2020). The reduction in exclusion rates from Samples 2 to 3 suggests that the vetting procedure has become more effective over time.
To examine the efficacy of the vetting procedures, we compared data quality across the three samples. Importantly, the vetted Samples 2–3 exhibited a marked increase in reliability and validity compared to the unvetted Sample 1. Additional screening resulted in only a very modest increase in reliability/validity.
Internal reliability indices indicate how strongly items on a scale are related to one another. Sample 1 exhibited problematically low internal reliability on previously-validated scales. Surprisingly, we observed increased reliability when negatively-scored items were improperly scored. Sample 1 responses appear to be driven by repetitive response patterns, rather than by veridical reports of the target construct. In the vetted samples, though, this pattern disappeared. All scales were reliable (all αs > .80), and improper scoring decreased reliability. This suggests that the vetting procedure reduced problems such as inattentive or non-English-proficient responses.
Beyond reliability, another important concern is whether a scale measures the intended construct. To assess this, we examined the established nomological network surrounding our measures. Within Sample 1, the established nomological network was not apparent. Social Justice Motives were related to Conservatism. Liberalism was related to prejudice against many groups (including liberal groups!), and more nuanced distinctions between RWA and SDO could not be made. After vetting, however, evidence of scale validity returned in Samples 2–3. Social Justice Motives resumed their typical correlation with Liberalism (Janoff-Bulman & Carnes, 2016), and Conservatism resumed its normal correlation with prejudices against many groups (Sibley & Duckitt, 2008), except for conservative groups (Brandt et al., 2014). More nuanced distinctions between RWA and SDO also reappeared (Duckitt & Sibley, 2007; Duckitt et al., 2002). Thus, the vetting procedure helps to ensure scales successfully measure their intended construct.
Impacts of Additional Screening
The removal of participants through in-survey screening items (e.g., attention checks, open-ended questions) led to an increase in reliability and validity. However, this increase was modest, at best. For example, the removal of problematic participants slightly increased the magnitude of convergent correlations with conservativism in Sample 2 (average before removal: r = .35; average after removal: r = .38) and Sample 3 (average before removal: r = .32; average after removal: r = .34). Thus, removal was generally beneficial, but extremely modest compared to the vetting procedure.
In summary, our results provide evidence for the utility of CloudResearch’s vetting procedures. Nonetheless, we still recommend that researchers screen out participants who fail to meet pre-specified criteria. In many contexts (e.g., with small effect sizes or low-frequency outcomes), even small numbers of problematic participants can lead to inaccurate conclusions (Litman et al., 2020). Moreover, this prescreening procedure removes only the most problematic and clearly fraudulent actors. Few participants are perfect all the time, and only in-survey screening items can exclude participants who are merely having an off day.
Given the current findings, what practical recommendations can we make? For example, some readers may be tempted to discount all MTurk research conducted during the same time period as Sample 1. However, Sample 1 is better regarded as an example of what can go wrong without proper precautions; and we cannot assume that such issues immediately generalize to all MTurk studies during this time. While CloudResearch’s vetting procedure is one effective solution, it is not the only one. If research conducted during this period contains evidence of proper precautions (e.g., attention-checks, reCAPTCHA, evidence of valid/reliable measures), then its data is likely more trustworthy than Sample 1. Without such evidence, however, readers should be attentive to the possibility that such data is indeed compromised.
Another important question is how long this vetting procedure method will remain effective. The process of maintaining data quality on MTurk is an ongoing process (Hauser & Schwarz, 2016). While CloudResearch’s solution has demonstrated continued effectively over the first year, this initial success does not necessarily speak to its continued success in the years to come. As such, the effectiveness of this vetting procedure should be periodically re-examined in the coming years, and it may be necessary to develop new procedures if Turkers’ behavior changes substantially in response to it.
Finally, future research should also examine the impact of this vetting procedure on other types of errors. Our focus here was solely on its impact on Measurement Error (i.e., reliability & validity). However, the Total Survey Error framework (e.g., Groves & Lyberg, 2010) clearly indicates that this is only one type of error. The impact of this vetting procedure on other types of error (e.g., sampling error) is currently unknown, and future research is needed to address this.
Since its founding, MTurk has provided social scientists with a consistent and relatively inexpensive data source. However, issues such as inattention and foreign workers can compromise data quality. CloudResearch instituted a vetting procedure aimed at identifying participants who provide more reliable data. The current results suggest that this vetting procedure increases reliability and validity of data. While we continue to recommend in-survey screening items, this vetting procedure is one viable means of collecting reliable and valid data.