Two components of individual differences in actively open-minded thinking standards: myside bias and uncertainty aversion

Abstract

The theory of actively open-minded thinking (AOT) implies standards for good thinking. Two broad aspects of these standards characterise individual differences in their acceptance: myside bias, in which thinking favours possible conclusions that are already strong; and uncertainty aversion, a belief that good thinking results in high confidence. Acceptance of AOT standards is often measured with short questionnaire scales. The present paper reports one study focusing on each of these two biases in the evaluation of the trustworthiness of sources, on the basis of short statements differing in signs of myside bias or of unjustified overconfidence. A 10-item scale measuring largely myside bias predicts sensitivity to both sources of difference in the statements. A third study examines the generality of the two standards across topics. The study used four questions about each topic and an 11-item scale with more items about uncertainty aversion. Individual differences are preserved across topic, although some results imply that individual differences are more pronounced when a standard is seen as more relevant to the content. The two standards are best characterised as two unidimensional factors that are correlated with each other. Thus, individual differences in AOT standards are not unidimensional.

Keywords:

• Actively open-minded thinking
• myside bias
• uncertainty
• individual differences

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

1 Attributed to the 19th-century writers Mark Twain and Artemus Ward but not found in the written works of either.

2 Beginning with Baron (1985, pp. 159–160) I have noted that two kinds of confidence can have opposite effects on whether thinking continues or stops. Confidence in the thinking done so far reduces the tendency to continue, while confidence in the effectiveness of further thinking can increase it.

3 They also suggest, implicitly, that the word “belief” as used in questionnaires was ambiguous between the logical sense in “believe that” and the sense in “believe in,” which represents a personal commitment and not just a logical attitude towards a proposition, even though the syntax in questionnaires is consistent with the logical sense. To avoid this possible ambiguity, I have replaced “belief” with “conclusion” for some of the items in the questionnaires that I use. This change is also consistent with the general theory of AOT, which implies that it is relevant to problem solving as well as development of general beliefs about politics, etc.

4 Stanovich (2022) also claims that myside bias has little generality and is unusual in not correlating with general cognitive abilities. The relation to general abilities is complicated and is not relevant here. The evidence for lack of generality is based on very few studies, most of which were not designed to test generality and which used possibly weak (unreliable) measures of individual differences, i.e., difference scores. But one cited study using a different measure of myside bias (Toplak & Stanovidh, 2003) found clear evidence of generality, as does Baron (2009, assuming that “conflicting responses” indicate absence of myside bias). Stanovich also defines myside bias as limited to situations involving disagreement between two groups of people. This is inconsistent with the items and scale measures used here, which assume that myside bias (perhaps a poorly chosen term) is relevant to all thinking, as assumed by Baron (1993 and in other writing after that). With this broader definition we might also count such papers as Russo et al. (2000) as showing generality.

5 If subjects strongly agree with a statement, they might think that the unqualified form is perfectly reasonable, thus pushing the Trust and Fair ratings of the unqualified form higher while not affecting or even reducing the ratings for the qualified form. It is the difference that matters here.

6 Note that coefficient alpha, as a measure of reliability of a scale, assumes unidimensionality and is thus an underestimate of the true reliability of the present AOT scale (which should be based on the correlation of the test with a parallel version of the same test). AOT is a multidimensional concept. The g6 measure reported by the alpha() function of the psych package (Revelle, 2024), or omega total, as reported by the omega() function, are probably more accurate estimates of reliability, especially when a reliability measure is used to correct for attenuation (in which case alpha will lead to over-correction and possibly to corrected correlations greater than 1).

7 Agreement might also affect Fair and Trust responses through a halo effect. However, such an effect should occur for both qualified and unqualified statements, so on the whole, we should expect no halo effect on the qualified/unqualified difference, which is the measure of primary interest.

8 Actually 15 pairs was presented, but data from the first item were lost due to a programming error. The results are thus based on 14 pairs.

9 An analysis of a combined measure yielded substantively identical results.

10 A quick scan suggested that this replacement made sense in every case, but, in retrospect, it did not make sense for the PEANUT items (see Appendix). Two subjects said that the items were not about a belief. And “asking” could easily refer to asking the server rather than “thinking about.” This item was omitted from all analyses.

11 The items overlap somewhat in meaning, but this division was what I had in mind when constructing the new scale. And it seemed to be the closest to the major meanings.

12 Canonical correlation finds a weighted average of each of two sets of variables such that the weights maximise the correlation between the two weighted averages. Then, in effect, it repeats the process with the residuals of these predictions, asking whether the prediction can be improved further with another set of weights, the second canonical correlate.

13 Note that this analysis applies to the 30 items as if they were different items in a scale. Such a scale would be analogous to any scale that asked roughly the same question about many situations, e.g., a happiness scale with items like, “I feel happy when I see …” for many different situations. It is thus concerned with the generality of individual differences across the 30 situations. It is not an analysis of reliability of each two-item “scales”, one for LackConf and one for AltNeg. These two items were constricted so that one was reverse scored, and responses to them were used to select subjects, so such an analysis would not be informative. We can assume that such a two-item test is not very reliable.

14 Because All4mean is the total of all items, including the item of interest, I used the r.drop function of the psych packages (Revelle, 2024) instead of the raw correlation with All4sum. This is an estimate of what the item-total correlation would be, on the basis of the correlation of the item with the sum of the other items.

15 Several attempts to find effects of differences in item variance also could not explain the results.