STUDY 2

In Study 1, we uncovered some cognitive and demographic predictors of conspiracy discrimination. In Study 2, we replicate and extend these findings with a larger sample. Moreover, to better understand what underlies conspiracy discrimination, we include additional predictors linked to conspiracy belief, including conspiracy mentality, trust in experts, political orientation and socio‐economic status (see Brotherton et al., 2013; Enders et al., 2024; Frenken et al., 2024; Imhoff et al., 2022; Van Prooijen et al., 2015; Vranic et al., 2022). We also re‐assess the relationship between cognitive sophistication and conspiracy beliefs using additional analyses.

Similar to Study 1, we disproportionately recruited people who endorse fringe conspiracies, preregistering several analyses comparing ‘Endorsers’ to ‘Non‐endorsers’. However, it became clear post‐data collection that comparing three groups is more appropriate: those who indiscriminately believe conspiracies, those who indiscriminately reject conspiracies (including verified ones), and those who accurately discriminate between the two. We therefore begin by replicating the Study 1 analyses, before comparing these three groups on cognitive sophistication, and then reporting the preregistered analyses.

Methods

Design, materials and procedure

Study 2's methods were similar to Study 1. Participants were pre‐screened in the same way. In the study itself, participants first responded to 32 conspiracy items presented in random order, but this time there was no intervening reasoning task. Conspiracy items were taken from Study 1, but some were reworded to improve clarity or replaced to avoid overlap with the pre‐screen. Participants then completed a general knowledge test, which we expanded to 16 items (8 true, 8 false), followed by the CRT and SRS.

This was followed by the Conspiracy Mentality Questionnaire (CMQ), a five‐item measure designed to assess differences in the tendency to engage in conspiracist ideation (Bruder et al., 2013). Responses were made on an 11‐point scale ranging from 0 (Certainly not) to 100 (Certain). An example item is ‘I think that many very important things happen in the world, which the public is never informed about.’ Scores were calculated by averaging across all five items (range: 0–11).

Participants then completed a five‐item scale measuring trust in expert authority (Robson, Faasse, & Martire, 2024), which has good internal consistency (α = .86). Responses were made on a seven‐point scale from 1 (Strongly disagree) to 7 (Strongly agree). An example item is ‘I tend to follow what experts say even if I don't fully understand their reasoning.’ We computed a mean score across all five items (range: 1–7).

Finally, participants indicated their political orientation on social and economic issues (1 = Left/Liberal, 7 = Right/Conservative), which we averaged to form a combined measure of conservatism (r = .76). They then answered demographic questions, which were the same as in Study 1 but with the addition of the MacArthur Scale of Subjective Social Status (Adler et al., 2000), which assesses perceived socio‐economic status (SES; range: 1–10).

Participants

We preregistered that we would recruit at least 600 participants, but aimed for as many as possible, with the stopping rule that recruitment would cease by the end of May 2024. We ended up with a sample of 663 participants after six were excluded for failing at least two attention checks (as preregistered). Post‐hoc sensitivity analyses indicate that this sample size provides 80% power to detect effect sizes of r = .11 and f 2 = .026.

Participants were primarily from the United States (22.2%) and the United Kingdom (16.7%), but also from other countries (10.1% from South Africa, 9.2% from Canada, 7.4% from Portugal, 6.8% from Mexico, 5.9% from Australia, and 21.7% from elsewhere). The mean age of participants was 39.0 (SD = 14.1). Most (56.0%) identified as male (42.1% female; 1.8% other; 0.2% did not say). Most (65.9%) identified as White/Caucasian (5.0% African American, 8.9% Hispanic; 9.7% Asian; 9.4% Other; 1.2% did not say). A majority (76.0%) also had tertiary level education or more (23.4% secondary education or less; 0.6% did not say), and most (59.9%) reported English as their first language (40.1% said it was not).

Results and discussion

Relationships between variables

Table 3 displays the zero‐order correlations. Like Study 1, there were strong positive intercorrelations between belief in unverified conspiracies, belief in verified conspiracies and response bias. Consistent with prior literature (Frenken et al., 2024), measures of cognitive ability (general knowledge, scientific reasoning and cognitive reflection) were associated with lower belief in unverified conspiracies, but unrelated to belief in verified conspiracies. In line with Study 1, distinguishing between verified and unverified conspiracies (AUC) was associated with higher scores on these cognitive measures. Better discriminators were also less conspiracy‐minded, more trusting of experts, more politically liberal and skewed male.

 | M | SD | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14
1. Unverified consp. belief | 3.59 | 2.17 | – |  |  |  |  |  |  |  |  |  |  |  |  | 
2. Verified consp. belief | 4.79 | 1.87 | 0.70** | – |  |  |  |  |  |  |  |  |  |  |  | 
3. Response bias (c) | −0.05 | 0.87 | 0.90** | 0.86** | – |  |  |  |  |  |  |  |  |  |  | 
4. Discriminability (AUC) | 0.61 | 0.16 | −0.61** | 0.08* | −0.34** | – |  |  |  |  |  |  |  |  |  | 
5. General knowledge | 73.23 | 13.36 | −0.20** | 0.06 | −0.08* | 0.32** | – |  |  |  |  |  |  |  |  | 
6. Scientific reasoning | 6.41 | 2.65 | −0.23** | −0.03 | −0.14** | 0.29** | 0.30** | – |  |  |  |  |  |  |  | 
7. Cognitive reflection | 4.05 | 1.99 | −0.19** | 0.01 | −0.10* | 0.28** | 0.36** | 0.43** | – |  |  |  |  |  |  | 
8. Conspiracy mentality | 6.68 | 10.67 | .67** | 0.49** | 0.64** | −0.41** | −0.16** | −0.25** | −0.19** | – |  |  |  |  |  | 
9. Trust in experts | 4.40 | 1.47 | −0.49** | −0.30** | −0.41** | 0.37** | 0.10* | 0.09* | 0.18** | −0.36** | – |  |  |  |  | 
10. Conservatism | 4.01 | 1.76 | 0.29** | 0.10** | 0.22** | −0.31** | −0.18** | −0.14** | −0.12** | 0.28** | −0.48** | – |  |  |  | 
11. SES | 5.31 | 1.64 | −0.04 | −0.08* | −0.06 | 0.00 | 0.06 | 0.07 | 0.04 | −0.07 | 0.06 | 0.00 | – |  |  | 
12. Age | 39.02 | 14.08 | 0.05 | 0.12** | 0.10** | 0.02 | −0.09* | 0.06 | −0.05 | 0.00 | −0.28** | 0.31** | −0.03 | – |  | 
13. Education | 2.98 | 0.74 | −0.02 | −0.02 | −0.01 | 0.02 | 0.05 | 0.05 | −0.01 | −0.06 | 0.11** | −0.07 | 0.21** | 0.02 | – | 
14. Gender | – | – | −0.08* | 0.02 | −0.05 | 0.16** | 0.06 | 0.06 | 0.19** | −0.07 | 0.10** | −0.03 | −0.02 | −0.15** | −0.05 | –
15. Native English | – | – | 0.10** | 0.12** | 0.13** | −0.03 | −0.13** | 0.02 | −0.12** | 0.11** | −0.24** | 0.26** | −0.10* | 0.45** | −0.07 | −0.13**

Predicting discriminability

We used linear regression analysis to explore unique predictors of conspiracy discriminability (see Table 4). The predictors together explained 34.6% of the variance in discriminability. Consistent with Study 1, general knowledge and scientific reasoning predicted better discrimination. Better discriminators also tended to be less conspiracy‐minded, more trusting of experts, more politically liberal and skewed male. In contrast to Study 1, age was a significant positive predictor of accurate discrimination.

Predictor | β | b | SE | t‐value | p‐value
Intercept | – | 0.39 | 0.06 | 6.74 | <.001
General knowledge | .21 | 0.00 | 0.00 | 5.98 | <.001
Scientific reasoning | .08 | 0.00 | 0.00 | 2.25 | .025
Cognitive reflection | .06 | 0.01 | 0.00 | 1.69 | .092
Age | .13 | 0.00 | 0.00 | 3.32 | .001
Education | −.02 | 0.00 | 0.01 | −0.53 | .597
Gender | .11 | 0.04 | 0.01 | 3.33 | .001
Native English | .07 | 0.02 | 0.01 | 1.95 | .051
Conspiracy mentality | −.22 | −0.02 | 0.00 | −6.22 | <.001
Trust in experts | .24 | 0.03 | 0.00 | 6.07 | <.001
Political conservatism | −.13 | −0.01 | 0.00 | −3.38 | .001
SES | −.03 | 0.00 | 0.00 | −0.97 | .331

Predicting cognitive sophistication

Like Study 1, we flipped the analytic direction and conducted linear regression analyses to examine whether AUC is a stronger predictor of cognitive sophistication than response bias. We conducted regression analyses with only response bias as a predictor and compared them to regressions with both AUC and response bias as predictors. Response bias significantly negatively predicted general knowledge (b = −1.20, p = .044), scientific reasoning (b = −0.43, p < .001) and cognitive reflection (b = −0.23, p = .010). When AUC was added as a predictor in each instance, however, response bias was no longer significantly predictive (_p_s > .05), whereas AUC was consistently a significant, positive predictor (general knowledge: b = 28.13, p < .001; scientific reasoning: b = 4.62, p < .001; cognitive reflection: b = 3.46, p < .001).

Comparing correlations with conspiracy mentality

Because conspiracy mentality is a widely used predictor of conspiracy beliefs, we conducted a dependent samples Fisher's r‐to‐z test comparing conspiracy mentality's association with response bias and with discriminability. Conspiracy mentality was more strongly associated with response bias than with discriminability, z = 32.25, p < .001, suggesting that conspiracy mentality primarily captures a tendency to endorse conspiracy claims rather than reduced discriminability.

Comparison of three groups

Consistent with prior research, we found that those higher in cognitive sophistication are less likely to believe unverified conspiracies. However, this relationship in isolation risks oversimplifying the picture because cognitive sophistication was also associated with discriminating between verified and unverified conspiracies. Among those who disbelieve unverified conspiracies, there may be two kinds of people: those who accurately differentiate between real and false conspiracies, and those who reject conspiracies indiscriminately. It is therefore unclear whether the link is driven by heuristic rejection or effective discrimination.

To test this idea, we categorized participants into three groups based on their conspiracy discrimination (AUC) and response bias (c) scores. ‘Indiscriminate Believers’ were those with a liberal response bias (c > 0) and poor discrimination (AUC < .7). ‘Indiscriminate Sceptics’ were those with a conservative response bias (c ≤ 0) and poor discrimination (AUC < .7). ‘Good Discriminators’ were those with high discrimination scores (AUC ≥ .7). We used these threshold values because an AUC of 0.7 is considered ‘reasonable’ (Swets, 1988) and a response bias of zero is the unbiased midpoint.

We compared these three groups on general knowledge, scientific reasoning, and cognitive reflection (see Figure 1). In each instance, the between‐groups one‐way ANOVA was significant. Follow up pairwise Tukey tests then revealed that, for each measure, Good Discriminators significantly outperformed the other two groups whereas Indiscriminate Believers and Indiscriminate Sceptics did not significantly differ from one another. Descriptive statistics and results are presented in Table 5.

FIGURE 1: Comparison of three groups on cognitive ability measures in Study 2. In Panel a, the three groups—Indiscriminate Believers (purple), Indiscriminate Sceptics (green) and Good Discriminators (dark blue)—are plotted according to their mean belief in unverified and verified conspiracies. The regression line represents the correlation between the two variables. General knowledge (b), scientific reasoning (c), and cognitive reflection (d) scores for each group are also depicted in the remaining panels. Rain clouds represent the distribution of scores, individual points depict each participant's score, and the orange diamonds represent group means. Good Discriminators significantly outperformed the other two groups on all three measures, whereas Indiscriminate Believers and Indiscriminate Sceptics did not significantly differ from one another on any measure.

Variable | Grouping method
Thresholds | 3‐profile LPA | 4‐profile LPA
IB (n = 258) Mauc = .51 Mc = .77 | IS (n = 205) Mauc = .56 Mc = −.77 | GD (n = 200) Mauc = .80 Mc = −.37 | IB (n = 258) Mauc = .48 Mc = .71 | IS (n = 94) Mauc = .53 Mc = −1.25 | GD (n = 311) Mauc = .75 Mc = −.31 | IB (n = 91) Mauc = .44 Mc = 1.36 | IS (n = 94) Mauc = .56 Mc = −1.35 | GD (n = 241) Mauc = .77 Mc = −.40 | AV (n = 237) Mauc = .54 Mc = .28
General knowledge | Mean (SD) | 71.05 (12.75) | 71.37 (13.54) | 77.93 (12.81) | 69.96 (12.78) | 71.60 (13.07) | 76.43 (13.21) | 69.56 (13.25) | 71.98 (13.45) | 77.52 (12.98) | 70.76 (12.66)
Overall F (df) Pairwise comparisons | F(2, 660) = 18.76, p < .001 IB vs. IS: Mdiff = 0.32, p = .961 IB vs. GD: Mdiff = 6.88, p < .001 IS vs. GD: Mdiff = 6.56, p < .001 | F(2, 660) = 18.24, p < .001 IB vs. IS: Mdiff = 1.64, p = .550 IB vs. GD: Mdiff = 6.47, p < .001 IS vs. GD: Mdiff = 4.83, p = .005 | F(3, 659) = 14.38, p < .001 IB vs. IS: Mdiff = 2.42, p = .584 IB vs. GD: Mdiff = 7.96, p < .001 IS vs. GD: Mdiff = 5.54, p = .003 IB vs. AV: Mdiff = 1.20, p = .877 IS vs. AV: Mdiff = 1.22, p = .867 GD vs. AV: Mdiff = 6.76, p < .001
Cognitive reflection | Mean (SD) | 3.67 (1.94) | 3.78 (2.02) | 4.82 (1.80) | 3.53 (1.95) | 3.82 (2.08) | 4.56 (1.88) | 3.55 (2.06) | 3.76 (2.05) | 4.69 (1.84) | 3.71 (1.93)
Overall F (df) Pairwise comparisons | F(2, 660) = 22.95, p < .001 IB vs. IS: Mdiff = 0.11, p = .815 IB vs. GD: Mdiff = 1.15, p < .001 IS vs. GD: Mdiff = 1.04, p < .001 | F(2, 660) = 20.75, p < .001 IB vs. IS: Mdiff = 0.29, p = .422 IB vs. GD: Mdiff = 1.03, p < .001 IS vs. GD: Mdiff = 0.74, p = .004 | F(3, 659) = 13.90, p < .001 IB vs. IS: Mdiff = 0.21, p = .888 IB vs. GD: Mdiff = 1.14, p < .001 IS vs. GD: Mdiff = 0.93, p = <.001 IB vs. AV: Mdiff = 0.16, p = .903 IS vs. AV: Mdiff = 0.04, p = .998 GD vs. AV: Mdiff = 0.98, p < .001
Scientific reasoning | Mean (SD) | 5.63 (2.49) | 6.17 (2.51) | 7 .66 (2.54) | 5.59 (2.39) | 6.09 (2.50) | 7.18 (2.68) | 5.64 (2.53) | 6.19 (2.53) | 7.41 (2.61) | 5.77 (2.47)
Overall F (df) Pairwise comparisons | F(2, 660) = 38.16, p < .001 IB vs. IS: Mdiff = 0.54, p = .055 IB vs. GD: Mdiff = 2.03, p < .001 IS vs. GD: Mdiff = 1.49, p < .001 | F(2, 660) = 28.37, p < .001 IB vs. IS: Mdiff = 0.49, p = .245 IB vs. GD: Mdiff = 1.59, p < .001 IS vs. GD: Mdiff = 1.10, p = .001 | F(3, 659) = 20.48, p < .001 IB vs. IS: Mdiff = 0.55, p = .448 IB vs. GD: Mdiff = 1.77, p < .001 IS vs. GD: Mdiff = 1.22, p = .001 IB vs. AV: Mdiff = 0.13, p = .973 IS vs. AV: Mdiff = 0.42, p = .529 GD vs. AV: Mdiff = 1.64, p < .001

A similar pattern of results emerged when grouping participants via a Latent Profile Analysis (LPA) based on discriminability and response bias scores. When we forced three profiles, the groups were similar to those above. A completely data‐driven LPA, however, suggested four profiles were appropriate, with the fourth profile capturing participants with poor discriminability and middling response bias (i.e. ‘Ambivalents’). Once again, Good Discriminators outperformed the other three groups on each cognitive measure, whereas the other groups did not differ significantly from one another on any measure.

Preregistered analyses

Similar to Study 1, participants were pre‐screened several months prior using four implausible conspiracy items rated on a 0–100 scale (Martire et al., 2020; Martire et al., 2023). However, these items were re‐administered during the study itself, and participants were formally grouped as either Endorsers (rated any claim >60 out of 100) or Non‐endorsers (rated all four claims <40 out of 100) based on in‐study responses. In all, there were 332 Endorsers and 295 Non‐endorsers (N = 627), with 36 no longer falling into either group. A priori power analyses indicated that a sample of 600 provides 80% power to detect small‐to‐medium group differences (Cohen's d = 0.23).

Our first prediction was that, compared to Non‐endorsers, Endorsers would more strongly believe in both unverified and verified conspiracies (Cohen's _d_s > .5), reflecting that those who believe false conspiracies also tend to believe true ones. As predicted, Endorsers (M = 4.86, SD = 2.05) believed unverified conspiracies more than Non‐endorsers did (M = 2.22, SD = 1.35), t(625) = 18.82, p < .001, d = 1.51. They also believed verified conspiracies (M = 5.38, SD = 1.82) more than Non‐endorsers did (M = 4.19, SD = 1.75), t(625) = 8.37, p < .001, d = 0.67.

Second, we planned to compare these groups on discriminability (AUC) but we had no a priori predictions. Endorsers were poorer (M = 0.54, SD = 0.14) than Non‐endorsers (M = 0.70, SD = 0.14) at discriminating between verified and unverified conspiracies, t(625) = −14.94, p < .001, d = −1.20. However, we later realized that the grouping method introduces a selection bias because Non‐endorsers are both more likely to include those with a conservative response bias and those who discriminate effectively. This is a key reason why considering three groups is more appropriate.

Finally, we expected Endorsers to report higher conspiracy mentality and lower trust in experts than Non‐endorsers,¹ and that these variables would correlate with response bias on the conspiracy items. As predicted, Endorsers had higher conspiracy mentality (M = 7.54, SD = 1.24) than Non‐endorsers (M = 5.72, SD = 1.60), t(625) = 16.01, p < .001, d = 1.28, and lower trust in experts (M = 3.51, SD = 1.36) than Non‐endorsers (M = 5.35, SD = 0.90), t(625) = −19.63, p < .001, d = −1.57. Table 3 also shows that response bias was moderately positively correlated with conspiracy mentality, and moderately negatively correlated with trust in experts. A conspiracist worldview and distrust of experts might therefore serve as heuristics people use to determine whether conspiracy claims are true.

Footnotes

1. We mistakenly preregistered that Endorsers would have higher trust, but we always anticipated that they would score lower given our preregistered rationale that ‘extreme trust in authority…may mean dismissing or overlooking nefarious actions committed by authorities.’ ↩