Authors: Michał Sikorski, Noah van Dongen, Jan Sprenger
Categories: Article
Source: Review of Philosophy and Psychology
Authors: Michał Sikorski, Noah van Dongen, Jan Sprenger
Indicative conditionals and tendency causal claims are closely related (e.g., Frosch and Byrne, 2012), but despite these connections, they are usually studied separately. A unifying framework could consist in their dependence on probabilistic factors such as high conditional probability and statistical relevance (e.g., Adams, 1975; Eells, 1991; Douven, 2008, 2015). This paper presents a comparative empirical study on differences between judgments on tendency causal claims and indicative conditionals, how these judgments are driven by probabilistic factors, and how these factors differ in their predictive power for both causal and conditional claims.
Indicative conditionals—that is, conditionals that do not involve the auxiliary verb “would”—are important linguistic structures. Among other things, we use them to predict and explain events, to formulate instructions, and to describe causal relationships. For example, we can describe a causal tendency, i.e., the general tendency of a cause to bring about an effect, using explicit causal A lot of rain causes the ground to be waterlogged.but also by means of an indicative conditional (henceforth “conditional”) If it rains a lot, then the ground will be waterlogged.Similarly, a tendency causal claim1 like Pressing the red button causes the fire alarm to go off.can be rephrased using the conditional (2b)If you press the red button, the fire alarm goes off.(example taken from Declerck and Reed 2012). In general, indicative conditionals seem to correspond systematically to tendency causal claims (see also Experiment 2 and 3 in Over et al. 2007). Whoever accepts (1a) or (2a) may also be inclined to accept the corresponding indicative conditionals (1b) and (2b), or vice versa.
In the above examples such as (1a/b) and (2a/b), the link between cause and effect was quite given the cause, the effect would occur almost certainly. However, some tendency causal claims express a weaker the cause raises the probability of the effect, but the effect may not be likely even in the presence of the cause. An example of a such case Smoking causes lung cancer.Most people would probably classify this sentence as true, but the corresponding conditional (3b)If one smokes, one will get lung cancer.seems false, or at least much less plausible than (3a).
So it seems that causal claims and conditionals are not always evaluated in the same way. Specifically, the difference between (3a) and (3b) suggests that true (tendency) causal claims are not always evaluated as true conditionals, where we seem to require that lung cancer will occur with high probability, too.
In this paper, we focus on tendency causal claims such as (1a), (2a) and (3a) and investigate how their evaluation as true (or highly acceptable) differs from the evaluation of the corresponding indicative conditionals. Specifically, we study the role of probability in driving such can judgments on the conditional and unconditional probability of the effect predict whether we classify the causal claim and/or the indicative conditional as true or false? And can the correlation between classifications and probability judgments also explain the differences between causal claims and conditionals (e.g., in terms of concepts such as statistical relevance and high conditional probability)?
This project is interesting for multiple reasons. First, many experiments have been devoted to conditionals and causal claims separately (e.g., Frosch and Byrne 2012; Sloman and Lagnado 2015; Douven 2016). However, as far as we know, there are no experiments devoted directly to testing the relation between both kinds of expressions. Specifically, we study how the assessment of causal strength affects the use of conditionals for reasoning, explaining and decision-making (for studies on the probabilistic aspects of causal conditionals, see Oberauer and Wilhelm 2003; Over et al. 2007; Over 2017).
Second, both in theories of causality and of indicative conditionals, probability plays an important role. Indeed, on the theoretical level, probability-raising (i.e., statistical relevance) has been identified as a key feature of tendency causal claims (e.g., Suppes 1970; Eells 1991) and in quantifying the strength of a causal connection (e.g., Cheng 1997; Fitelson and Hitchcock 2011; Sprenger 2018). On the other hand, probabilistic accounts of conditionals have analyzed their meaning in terms of the conditional probability of the consequent, given the antecedent (e.g., Ramsey 1926; Adams 1975; Edgington 1986; 1995). We analyze whether or not these theoretical differences are empirically traceable in the evaluation of causal claims and conditionals.
Third, the project may be of interest to psychologists and linguists working on conditionals. A large part of the empirical literature on probability and conditionals has focused on conditionals, which express a reasoning process in the speaker’s mind, for example, “if this paper was rejected, then it must have been bad” (e.g., Johnson-Laird and Byrne 2002). The consequent of these conditionals can typically be preceded by the phrase “then it means that” or the modal auxiliary “must”. A second category are causal conditionals where the antecedent expresses an actual or potential cause of the consequent, expressing a straightforward causal connection between antecedent and consequent, such as “if the students don’t prepare well, the exam will go badly”.2 Dancygier (1998, 2003) calls the first group of conditionals “inferential conditionals” while the second group falls under the heading (but is not identical to) “content conditionals”.3 The results of our paper contribute to these research programs by investigating the truth and acceptability conditions of causal conditionals.
Fourth, the project sheds, as a byproduct, light on a debate between two theories of the plausibility and acceptability of indicative accounts where the acceptability of the conditional “if C, then E” follows the conditional probability p(E|C) (Adams 1975; Evans et al. 2007; Over et al. 2007; Egré and Cozic 2011; Over 2017), and accounts such as Evidential Support Theory (EST: Douven 2008; 2015; Krzyżanowska 2015; 2017) which demand that for a conditional to be acceptable, (i) p(E|C) be high, and (ii) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ p(E|C) > p(E)
## The Hypotheses The baseline idea of our paper is that judgments on the truth value or acceptability of a conditional can be predicted on the basis of judgments on the corresponding causal claim and probabilistic factors. While this claim is consistent with most of the theoretical and empirical literature, it is too vague to be tested experimentally. We will therefore split it up into several more precise hypotheses. Our first hypothesis concerns the logical relationship between conditionals and the corresponding tendency causal claims. Two different kinds of relations are (Necessity) Conditionals are classified as true *only if* the corresponding tendency causal claim is classified as true.(Sufficiency) Conditionals are classified as true *if* the corresponding tendency causal claim is classified as true.In light of the discussed examples and previous empirical research (e.g., Skovgaard-Olsen et al. 2016a; b; Skovgaard-Olsen et al. 2017; Douven et al. 2018; Skovgaard-Olsen et al. 2019), we expect that H1.a will be supported. We contrast it with the less plausible H1.b. In the light of examples such as (3a/b), we would expect that H1.b fails in empirical investigation. In many cases, the truth of both expressions will co-occur but some conditionals are expected to be classified as false although the corresponding tendency causal claim appears true (e.g., “smoking causes cancer”). High conditional probability of the consequent given the antecedent is plausibly one of the conditions of acceptability of conditionals (Over et al. 2007; Douven and Verbrugge 2012), but it is not required for classifying a tendency causal claim as true, as we have seen in the (3a/b) example. This speaks against H1.b. We operationalize these hypotheses by demanding that of all conditionals evaluated as true, only a small percentage of the corresponding tendency causal claims are evaluated as false (H1.a). Similarly, for an overwhelming percentage of all tendency causal claims evaluated as true, the same claim in conditional form needs to be evaluated as true (H1.b). For the respective thresholds we consider a *strict interpretation* (5 and 95%) and a *lenient interpretation* (10 and 90%).4 The next hypothesis regards the question of whether *p*(*E*|*C*) is a strong predictor of the plausibility that a conditional is true (Handley et al. 2006; Evans et al. 2007; Over et al. 2007; Over 2017; Over and Cruz 2023; contrary evidence is presented in Douven and Verbrugge 2010; Skovgaard-Olsen et al. 2016b). However, even critics concede a *weak version* of Adams’s Thesis, i.e., *p*(*E*|*C*) is highly correlated with the classification of a conditional as true (e.g., Douven and Verbrugge 2010, p. 306). We should therefore expect that this probability predicts the classification of a conditional at least to some degree. H2.a(Weak Adams’s Thesis) The conditional probability *p*(*E*|*C*) predicts the classification of conditionals of the form “if *C*, then *E*” as true or false, and its degree of acceptability.Suppose, however, that H2.a is *not* confirmed, or only weakly so. It could then be of interest to investigate whether it holds *when restricted to causal claims classified as true*. Indeed, the experiments by Skovgaard-Olsen et al. (2016b) show strong correlations between judgments on conditionals and *p*(*E*|*C*) *only when there is a clear relevance between cause and effect*. H2.b operationalizes a “relevantist” account of conditionals along these lines, i.e., they are classified as true if and only if (1) the corresponding causal claim is classified as true; and (2) *p*(*E*|*C*) is “high enough”. H2.b(Restricted Adams’s Thesis) In the class of tendency causal claims classified as true/highly acceptable, the conditional probability *p*(*E*|*C*) predicts the evaluation of the conditional “If *C*, then *E*” as true/highly acceptable.As a criterion for evaluating H2.a and H2.b, we adopt the statistical significance of including conditional probability as a predictor variable, plus a non-negligible effect size. Effect size is measured by how much variance in the data can be explained by the predictor variables and expressed numerically by the squared correlation coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 $$\end{document}R2. For an effect size to be meaningful, we demand that it exceed the value \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.09 $$\end{document}R2=0.09, which is conventionally identified with the lower bound of a medium effect (Cohen 1988). The remaining hypotheses concern the role of statistical relevance in the evaluation of tendency causal claims and conditionals, as predicted by probabilistic accounts of causal strength, Evidential Support Theory and the various covariation proposals. In the context of conditionals and causal claims, statistical relevance is typically measured by a function of two arguments, increasing in the first argument (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ x = p(E|C) $$\end{document}x=p(E|C)) and decreasing in the second argument (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ y = p(E|\lnot C) $$\end{document}y=p(E|¬C) or \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ y = p(E) $$\end{document}y=p(E)). The most common ways of combining these arguments are as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \begin{aligned} d(x, y)&= x-y&\quad r(x, y)&= \log (x/y) \\ z(x, y)&= \frac{x-y}{1-y}&\quad l(x, y)&= \log \frac{x}{1-x} - \log \frac{y}{1-y} \end{aligned} $$\end{document}d(x,y)=x-yr(x,y)=log(x/y)z(x,y)=x-y1-yl(x,y)=logx1-x-logy1-ywhich are known, respectively, as the difference measure *d*, the log-ratio measure *r*, the *z*-measure or normalized difference measure, and the log-likelihood measure *l*. Dependent on whether the second argument is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ y = p(E) $$\end{document}y=p(E) or \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ y = p(E|\lnot C) $$\end{document}y=p(E|¬C), the difference measure *d* reads either \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ d = p(E|C) - p(E) $$\end{document}d=p(E|C)-p(E), or \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ d = p(E|C) - p(E|\lnot C) $$\end{document}d=p(E|C)-p(E|¬C), and both measures have been defended as quantification of causal strength (Suppes 1970; Pearl 2001; Sprenger 2018; Sprenger and Hartmann 2019). Crupi and Iacona (2021) propose *z* (with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ y = p(E) $$\end{document}y=p(E)) as a measure of the acceptability of an indicative conditional, and so do van Rooij and Schulz (2019) (but with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ y = p(E|\lnot C) $$\end{document}y=p(E|¬C)). The peculiar feature of the *z*-measure is that the probability raise is set in relation to the maximal possible raise, and so the degree of statistical relevance is always *at least as high as the conditional probability* \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ x = p(E|C) $$\end{document}x=p(E|C). In other words, the *z*-measure captures aspects of high conditional probability and statistical relevance in a single number.5 On the basis of these measures, we can examine a series of hypotheses about how statistical relevance affects judgments on the truth or acceptability of causal and conditional H3.aStatistical relevance measures predict the classification of tendency causal claims as true or false (respectively their degree of acceptability).H3.bStatistical relevance measures predict the classification of a conditional as true or false (respectively their degree of acceptability).H3.cIn the class of tendency causal claims classified as true/highly acceptable, statistical relevance measures predict the evaluation of conditional claims as true/highly acceptable.H3.dStatistical relevance and conditional probability are, taken together, better predictors for the classification of a conditional than conditional probability alone.We expect the first hypothesis, H3.a, to come out confirmed since the increase in probability (upon intervention of the cause) has a strong theoretical basis as a predictor of causal strength, as explained above. H3.b tests probabilistic accounts of conditionals centered on evidential support and statistical relevance, such as van Rooij and Schulz (2019) and Crupi and Iacona (2021). H3.c tests the same hypothesis, restricted to the class of tendency causal claims classified as true (i.e., when we know that the cause is relevant for the effect). Finally, H3.d tests whether taking into account statistical relevance on top of conditional probability improves the prediction of the classification of the conditional. If one of H3.b/c/d came out confirmed, it would give a boost to Evidential Support Theory and similar accounts that stress the importance of statistical relevance. If not, it might not affect their normative significance but diminish their predictive value. The hypotheses H3.a-H3.c are evaluated on the same basis as adding statistical relevance as a predictor variable needs to be statistically significant, and the effect size as measured by the correlation coefficient must exceed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.09 $$\end{document}R2=0.09. For H3.d, we ask an increase in explained variance by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ 9\% $$\end{document}9% i.e., and increase in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 $$\end{document}R2 of 0.09 over the results of hypothesis H2.a. ## Experiment 1 ### Participants Participants were recruited via Amazon’s Mechanical Turk ([www.mturk.com](http://www.mturk.com)). Mechanical Turk directed the participants to the experiment that was run on the Qualtrics platform ([www.qualtrics.com](http://www.qualtrics.com)). In return for their participation, subjects received a small monetary compensation. Seventy-four native English speakers participated in the experiment. Eighteen participants were excluded because they failed to give the correct response to at least one of the control questions. All participants indicated to have participated seriously. At the end of the experiment, participants were asked an open question about what they thought the experiment was about. None of the participants displayed clear knowledge of the purpose of the experiment. In total, 56 [39 female, mean age = 40.59 years, s.d. = 11.69 years] participants were included in the analysis. ### Design We used a within-subjects design were each participant evaluated 19 vignettes. These vignettes were presented in random order on the participants’ computer screen. The participants were instructed to answer questions with the requirement that each question needed to be answered to be able to progress to the next item (i.e., forced-choice). The entire experiment was conducted in English. Control questions were used as a check on participants’ attention and participation. Randomly dispersed throughout the experiment, participants had to give a correct answer to several repeats of the *elimination questions* “For quality control, please select answer category five. If you do not select five, the survey will be terminated.” In addition, subjects had to rate on a five-point Likert scale how seriously they participated in the experiment (1= “completely unserious”, 5 = “completely serious”). We included an open question on the purpose of the experiment because knowledge of the experiment could influence the behaviour of participants. Our exclusion criteria failure to give the answer 5 on one of the elimination questions; rating their seriousness in participation as 3 or lower; or describing the experiment as being about conditionals and causation or something similar. ### Material and Procedure Participants had to evaluate tendency causal claims and the corresponding conditional claims in hypothetical vignettes, where a certain situation was described (e.g., the effect of prohibiting alcohol on the crime rate). We used content conditionals which can be interpreted causally. Such conditionals are usually called *causal conditionals*—not in the strong, ontological sense that there is actually a causal connection between antecedent and consequent, but rather in the weaker, epistemic sense that such conditionals “can be justified by evidence about a possible causal relation or mechanism” (e.g., Over et al. 2007, p. 65). Experimental studies using such conditionals are both accepted and common in the literature (see e.g., Oberauer and Wilhelm 2003; Over et al. 2007; Over 2017).6 Through several pre-studies, 19 vignettes were selected on comprehensibility from a list of 60. When directed from Mechanical Turk to Qualtrics, participants first received instructions on the experiment. After the instructions they were presented with the vignettes in a randomized sequence. The 19 vignettes consisted of four questions each, eliciting probability judgments as well as dichotomous judgments on causal and conditional **(Unconditional) Probability of Consequent** This question elicits the probability of a certain development without making specific assumptions (e.g., “how likely is it that the crime rate will decline in the next five years?”). **Conditional Probability of the Consequent** This question elicits the probability of the same development under a specific assumption stated in the antecedent (e.g., “how likely is it that the crime rate will decline in the next five years if alcohol consumption is made illegal?”).7**Causal Claim** This question asks the participants to evaluate the truth or falsity of the causal connection between antecedent and consequent (e.g., “Making alcohol consumption illegal will cause the crime rate to decline in the next five years”). **Conditional Claim** This question asks the participants to evaluate the truth or falsity of the corresponding indicative conditional (e.g. “If alcohol consumption is made illegal, then the crime rate will decline in the next five years.”) The first two questions had to be answered on a visual analog scale of probability percentages from 0% to 100%. The third and fourth question had to be answered with either ”true” or ”false”. We reproduce the experimental material in Appendix A and B. This formulation of the stimuli implies that we will calculate statistical relevance measures with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ x = p(E|C) $$\end{document}x=p(E|C) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ y = p(E) $$\end{document}y=p(E); see Experiment 2 for an extension to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ y = p(E|\lnot C) $$\end{document}y=p(E|¬C). ### Results #### Hypotheses H1.a and H1.b: Conditional vs. Causal Claims Combined, the data consisted of 1064 entries; 56 participants responded to 19 vignettes. We evaluated H1.a and H1.b by simply checking the frequency statistics for the relevant categories (see Table 1). Of the 531 data points where a conditional was classified as true, 25 data points classified the corresponding tendency causal claim as false. This corresponds to a percentage of 4,71% and therefore confirms our hypothesis H1.a that perceived presence of a causal relationship is *necessary* for classifying a conditional as true, both for the strict 5% and the lenient 10% threshold. By contrast, Hypothesis H1.b that classifying a causal claim as true is a *sufficient* condition for classifying the conditional as true was not borne out by the of 611 data points where the tendency causal claim was evaluated as true, only 506 evaluated the corresponding conditional as true. This percentage of 82,82% is clearly below the thresholds of 90% and 95% necessary to establish sufficiency.Table 1Classification of causal and conditional claims as true and falseContingency TableConditional ClaimTotalTrueFalseCausal ClaimTrue506 [82.82%]105 [17.18%]611 [100.00%][95.29%][19.70%]False25 [5.52%]428 [94.48%]453 [100.00%][4.71%][80.30%]Total5315331064[100%][100%]Table 2The Generalized Linear Mixed Model (GLMM) for the dependent variable *Causal Claim* as a function of *Conditional Probability*Type of EffectVariableCoefficientStd. ErrorzpFixed effectsintercept-4.610.46-10.13< 0.0001Conditional Probability0.0760.00613.06< 0.0001Random effectsParticipant1.841.36Vignette0.640.80Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.34 $$\end{document}R2=0.34. Residual degrees of freedom = 1060. #### Hypothesis H2.a and H2.b: Weak and Restricted Adams’s Thesis To test H2.a and H2.b, a Generalized Linear Mixed Model (GLMM) was used. We used a logit link function, as the outcome variable for each hypothesis was binary (0 = False, 1 = True). We added participants and vignette number as crossed random effects, because of difference in content between the vignettes, and possible differences in their interpretation between participants. We used the R package *lme4* (Bates et al. 2015) to estimate the GLMM’s regression coefficients, variance components, and the amount of variance in the outcome explained by the predictors (i.e., *marginal* \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2_{GLMM} $$\end{document}RGLMM2, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 $$\end{document}R2 onwards; Nakagawa and Schielzeth 2013). For hypothesis H2.a, the results show a strong and positive association between the conditional probability attributed to the consequent of the conditional, and the log-odds of the corresponding conditional being considered as true. Specifically, with every percentage-point increase in conditional probability these log-odds are estimated to increase by 0.07 (an increase of 7 over the full 100 percentage points). Most importantly, the model explains 48% (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2=0.48 $$\end{document}R2=0.48) of the variance in the participants tendency to indicate conditionals as either true or false (see Table 2). In short, H2.a is supported by the observed data. To test hypothesis H2.b, only those data points were used where the tendency causal claim was indicated as true (see Table 1). Similar to the results for H2.a, results show a positive association between the conditional probability attributed to the consequent of the conditional, and the log-odds of corresponding conditional being considered as true (see Table 3). Although the amount of variance explained is greatly reduced, from 48% to 12%, it is still considered meaningful (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2>0.09 $$\end{document}R2>0.09) and thus supporting H2.b.Table 3The Generalized Linear Mixed Model (GLMM) for the dependent variable *Conditional Claim* as a function of *Conditional Probability* when *Causal Claim = “true”*Type of EffectVariableCoefficientStd. ErrorzpFixed effectsintercept0.890.62-1.140.15Conditional Probability0.050.0085.79< 0.0001Random effectsParticipant3.391.84Vignette0.450.67Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.13 $$\end{document}R2=0.13. Residual degrees of freedom = 607. #### Hypotheses H3.a—H3.d: The Impact of Statistical Relevance As above, a Generalized Linear Mixed Model (GLMM) was used to test hypothesis H3.a, H3.b, H3.c, and H3.d. For H3.a, all statistical relevance measures except *z* predict the classification of tendency causal claims as true or false (see Table 4). For the statistical relevance measures *d*, *r*, and *l*, the analyses show a positive and meaningful association (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 > 0.09 $$\end{document}R2>0.09) with the proclivity of participants to assess the tendency causal claim as true. Specifically, the coefficients indicate the estimated increase in log-odds (per unit of the relevance measure) of tendency causal claims being indicated as true versus false.8 The random-effects for vignette and participant, though non-zero, appear to be minor. Specifically, their coefficients are only slightly larger than their standard errors. Based on the test statistics (z-values) and amount of explained variance (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 $$\end{document}R2), the association between *d* and classification of tendency causal claims was the strongest (i.e., largest coefficient with respect to its standard error). The weakest association was with the *z* measure.Table 4The Generalized Linear Mixed Model (GLMM) for the dependent variable *Causal Claim* as a function of *Statistical Relevance*: the *d*-measure, *r*-measure, *l*-measure, and *z*-measureType of EffectVariableCoefficientStd. ErrorzpFixed effectsintercept-0.150.31-0.480.63*d*5.850.5311.04< 0.0001Random effectsParticipant1.141.07Vignette1.261.12Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.34 $$\end{document}R2=0.34. Residual degrees of freedom = 1060.Fixed effectsintercept0.260.350.750.46*r*1.160.148.36< 0.0001Random effectsParticipant1.021.01Vignette1.761.33Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.22 $$\end{document}R2=0.22. Residual degrees of freedom = 1060.Fixed effectsintercept-0.040.32-0.120.90*l*0.840.0810.08< 0.0001Random effectsParticipant1.101.05Vignette1.331.15Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.34 $$\end{document}R2=0.34. Residual degrees of freedom = 1060.Fixed effectsintercept-0.240.29-0.840.40*z*0.010.020.550.58Random effectsParticipant0.750.87Vignette2.391.55Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 < 0.001 $$\end{document}R2<0.001. Residual degrees of freedom = 1060. For hypothesis H3.b, all statistical relevance measures predict the classification of conditionals as true or false (see Table 5). For all the statistical relevance measures, the analyses show a positive and meaningful association (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 > 0.09 $$\end{document}R2>0.09) with the tendency of participants to assess the conditional as true, thus supporting hypothesis H3.b.Table 5The Generalized Linear Mixed Model (GLMM) for the dependent variable *Conditional* as a function of *Statistical Relevance*: the *d*-measure, *r*-measure, *l*-measure, and *z*-Type of EffectVariableCoefficientStd. Error*z**p*Fixed effectsintercept-0.690.31-2.190.029*d*5.270.4810.95< 0.0001Random effectsParticipant1.511.23Vignette1.181.09Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.28 $$\end{document}R2=0.28. Residual degrees of freedom = 1060.Fixed effectsintercept-0.260.33-0.790.43*r*1.020.128.27< 0.0001Random effectsParticipant1.171.08Vignette1.511.23Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.18 $$\end{document}R2=0.18. Residual degrees of freedom = 1060.Fixed effectsintercept-0.610.001-423.6< 0.0001*l*0.800.002553.8< 0.0001Random effectsParticipant1.451.20Vignette1.161.08Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.31 $$\end{document}R2=0.31. Residual degrees of freedom = 1060.Fixed effectsintercept-0.0240.34-0.070.95*z*0.220.082.630.009Random effectsParticipant0.940.97Vignette1.791.34Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.20 $$\end{document}R2=0.20. Residual degrees of freedom = 1060. To test hypotheses H3.c, only those data points were used where the tendency causal claim was indicated as true (see Table 1). The results show that none of the statistical relevance measures made a meaningful difference in explaining the variance in the participants’ tendency to indicate the conditionals as true or false (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2<0.09 $$\end{document}R2<0.09 in all cases, see Table 6), thus supporting our conjecture of a null effect and contradicting hypothesis H3.c.Table 6The Generalized Linear Mixed Model (GLMM) for the dependent variable *Conditional Claim* as a function of *Statistical Relevance* (*d*-measure, *r*-measure, *l*-measure, and *z*-measure) when *Causal Claim = “true”*Type of EffectVariableCoefficientStd. ErrorzpFixed effectsintercept1.890.384.99< 0.0001*d*2.040.623.280.001Random effectsParticipant2.801.67Vignette0.640.80Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.04 $$\end{document}R2=0.04. Residual degrees of freedom = 607.Fixed effectsintercept2.150.375.82< 0.0001*r*0.360.162.170.030Random effectsParticipant2.721.65Vignette0.650.81Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.02 $$\end{document}R2=0.02. Residual degrees of freedom = 607.Fixed effectsintercept1.890.384.99< 0.0001*l*0.370.094.000.0001Random effectsParticipant2.941.71Vignette0.640.80Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.07 $$\end{document}R2=0.07. Residual degrees of freedom = 607.Fixed effectsintercept1.630.384.31< 0.0001*z*0.090.051.820.07Random effectsParticipant2.991.73Vignette0.560.75Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.04 $$\end{document}R2=0.04. Residual degrees of freedom = 607. To test hypothesis H3.d, the change in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 $$\end{document}R2 is assessed when *conditional probability* *p*(*E*|*C*) is added to the models of H3.a. For all statistical relevance measures, none made a meaningful contribution over and above the conditional probability on the prediction of participants’ proclivity to indicate the conditional as true (see Table 7). Specifically, the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 $$\end{document}R2 of these models does not show a large enough increase (at most 0.01) over the models that only include the conditional probability as a prediction (see hypothesis H2.a; Table 2)Table 7The Generalized Linear Mixed Model (GLMM) for the dependent variable *Conditional Claim* as a function of *Statistical Relevance* (*d*-measure, *r*-measure, *l*-measure, and *z*-measure) and *conditional probability*Type of EffectVariableCoefficientStd. ErrorzpFixed effectsintercept-4.300.47-9.18< 0.0001*d*1.320.572.310.02Conditional Probability6.810.6710.23< 0.0001Random effectsParticipant1.891.38Vignette0.600.77Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.48 $$\end{document}R2=0.48. Residual degrees of freedom = 1059.Fixed effectsintercept-4.480.46-9.70< 0.0001*r*0.27600.151.890.06Conditional Probability7.260.6111.87< 0.0001Random effectsParticipant1.871.37Vignette0.620.79Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.49 $$\end{document}R2=0.49. Residual degrees of freedom = 1059.Fixed effectsintercept-4.270.47-9.07< 0.0001*l*0.210.092.290.02Conditional Probability6.800.6710.18< 0.0001Random effectsParticipant1.871.37Vignette0.600.77Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.49 $$\end{document}R2=0.49. Residual degrees of freedom = 1059.Fixed effectsintercept-4.600.46-10.03< 0.0001*z*0.010.0280.410.69Conditional Probability7.610.5912.91< 0.0001Random effectsParticipant1.851.36Vignette0.640.80Explained \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 = 0.48 $$\end{document}R2=0.48. Residual degrees of freedom = 1059. ### Summary of Results of Experiment 1 The experiment has confirmed that the truth conditions of indicative conditionals are more demanding than the truth conditions of the corresponding tendency causal claims (support for H1.a, no support for H1.b). It has also confirmed the weak, qualitative version of Adams’s Thesis—the conditional probability of the consequent predicts the judgment on the truth value of the conditional (H2.a)—, as well as its restriction to tendency causal claims evaluated as true (H2.b). The hypotheses about statistical relevance enjoy mixed statistical relevance predicts the classifications of causal/conditional claims as true (H3.a and H3.b), but this may simply be due to the fact that high statistical general relevance co-varies with high conditional probability *p*(*E*|*C*). Indeed, once we control for this effect, statistical relevance adds no further predictive value (null results for H3.c and H3.d). ## Experiment 2 The second experiment consisted in a conceptual replication of the first experiment, using only continuous scales for the dependent variables, and replacing one of the predictor variables. This experiment consisted of four separate parts (2.A, 2.B, 2.C, and 2.D) with independent samples for each part. Slight alterations were made in the phrasing of the vignettes. That said, 2.A was essentially the same experiment as Experiment 1 whereas in 2.B, 2.C and 2.D, important elements of the response variables were changed. These changes were motivated by two questions that we (and some reviewers) had outlined as targets for future First, whether our results would carry over from judgments about the *truth value* of a conditional to its *acceptability* (reasons for suspecting invariance are given byDouven and Krzyżanowska 2018). In other words, we replaced a dichotomous choice (“true/false”) for the evaluation of the causal claim and the conditional by a continuous scale. Second, given that various important statistical relevance measures such as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \Delta {}p = p(E|C) - p(E|\lnot C) $$\end{document}Δp=p(E|C)-p(E|¬C) are calculated on the basis of the probability of the effect given the *negation* of the cause (i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ p(E|\lnot C) $$\end{document}p(E|¬C)), we wanted to see whether such measures would predict judgments on causal claims and conditionals any better (or worse) than measures which depend on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ x = p(E|C) $$\end{document}x=p(E|C) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ y = p(E) $$\end{document}y=p(E). ### Participants Similar to the first experiment, participants were recruited via Amazon’s Mechanical Turk ([www.mturk.com](http://www.mturk.com)). Mechanical Turk directed the participants to the experiment that was run on the Qualtrics platform ([www.qualtrics.com](http://www.qualtrics.com)). In return for their participation, subjects received a small monetary compensation. 77 people participated in Experiment 2.a, 75 participated in Experiment 2.b, 75 participated in Experiment 2.c, and 72 participated in Experiment 2.d. In total 38 participants were excluded (6 from 2.a, 12 from 2.b, 5 from 2.c, 15 from and 2.d), because they failed to give the correct response to at least one of the control questions. All participants indicated to have participated seriously. None of the participants displayed clear knowledge of the purpose of the experiment on the open question about what they thought the experiment was about. Thus, the number of participants included in the analysis was 71 for Experiment 2.a, 63 for Experiment 2.b, 70 for Experiment 2.c, an 57 for Experiment 2.d. ### Design We used the same within-subject design were each participant evaluated 18 vignettes. These vignettes were presented in random order on the participants’ computer screen and they were instructed to answer questions with the requirement that each question needed to be answered to be able to progress to the next item (i.e., forced-choice) ### Material and Procedure Similar to Experiment 1, this experiment consisted of hypothetical vignettes, which the participants had to evaluate. With respect to Experiment 1, we altered the phrasing of the four questions in each vignette. One vignette had to be dropped, because it could not be fitted to the alterations. For a complete list of the altered vignettes, see Appendix D. The difference between the experiments is represented schematically in Table 8.Table 8Differences between Experiments 2.A, 2.B, 2.C and 2.D in terms of the quantities they elicitExperimentClassification of causal and conditional claims*p*(*E*|*C*)?*p*(*E*)?\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ p(E|\lnot C) $$\end{document}p(E|¬C)?2.A (=1)dichotomous scale (true/false)yesyesno2.Bdichotomous scale (true/false)yesnoyes2.Ccontinuous scale (0-100 agreement)yesyesno2.Dcontinuous scale (0-100 agreement)yesnoyes Across all parts of this experiment (2.A—2.D) the B question (“Conditional Probability of the Consequent” in Experiment 1) was rephrased to eliminate potential ambiguities with respect to the probability of the conditional (compare the discussion in footnote 3 on page 9). Specifically, this question was rephrased to have “Suppose *x*. How likely is it that *y*?”. For example, “Suppose that alcohol consumption will be prohibited. How likely is it then that the crime rate will decline in the next 5 years?” In Experiment 2.B and 2.D, the A question of Experiment 1 was replaced in order to elicit \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ p(E|\lnot C) $$\end{document}p(E|¬C). Instead of a question about unconditional probability of the vignette, an opposite to the B question was presented to the participants. For instance, “Suppose that alcohol consumption will stay legal. How likely is it then that the crime rate will decline in the next 5 years?” In Experiment 2.C and 2.D, answer options were changed in order to measure the classification of causal claims and conditionals on a continuous scale. In these experiments, the C (causal claim) and D (conditional claim) questions had to be answered on a visual analog scale from 0 to 100. Specifically, participants were asked the extent of their agreement with the claims. For “Making alcohol consumption illegal causes the crime rate to decline over the next 5 years.”To what extent do you agree with this statement? (0 = completely disagree; 100 = completely agree) ### Results #### Hypotheses H1.a and H1.b: Conditional vs. Causal Claims For this hypothesis, only Experiment 2.A and 2.B could be evaluated. In Experiment 2.C and 2.D, causal and conditional claims were no longer answered as ‘true’ or ‘false’, which precluded them from testing this hypothesis. For Experiment 2.A, 11.48% (94 out 819) of true conditional claims have a corresponding false causal claim and 86.21% (725 out of 841) of true causal claims have a corresponding true conditional claim. These results do not support hypotheses H1.a and H1.b. For Experiment 2.B, 9.12% (77 out of 844) of true conditional claims have a corresponding false causal claim and 87.06% (767 out of 881) of true causal claims have a corresponding true conditional claim (see Table 9). These results support hypothesis H1.a at the 10% level, but do not support hypothesis H1.b. Our findings are thus in general agreement with Experiment 1.Table 9Classification of causal and conditional claims as true and falseExperiment 2.AExperiment 2.BContingency TableConditional ClaimTotalConditional ClaimTotalTrueFalseTrueFalseCausal ClaimTrue725116841767114881False9436145577194271Total81947712968443081152 #### Hypotheses H2.a and H2.b: Weak and Restricted Adams’s Thesis Similar to Experiment 1, a Generalized Linear Mixed Model9 was used to test hypotheses H2.a and H2.b. In the case of hypothesis H2.a, the conditional probability explained a significant percentage of variance in the participants responses to the conditional claim across all experiments (2.A:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2=0.19 $$\end{document}R2=0.19, 2.B:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2=0.20 $$\end{document}R2=0.20, 2.C:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2=0.56 $$\end{document}R2=0.56, 2.D:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2=0.62 $$\end{document}R2=0.62, and all p-values < 0.05). In short, H2.a is supported by the observed data. To test hypotheses H2.b, only those data points were used where the tendency causal claim was indicated as true (Experiment 2.A and 2.B) or got an agreement score above 80 our of 100 (Experiment 2.C and 2.D). This threshold has been chosen because according to the so-called Lockean Thesis (e.g., Foley 2009) one *accepts* a proposition when it is highly probable, and 80% seems to us a natural (though in no way special) implementation of that criterion. Results show a positive association between the conditional probability attributed to the consequent of the conditional, and the log-odds of corresponding causal conditional being considered as true. However, the amount of variance explained is greatly reduced, and only two of the four experiments show an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 > 0.09 $$\end{document}R2>0.09 (2.A:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2=0.05 $$\end{document}R2=0.05, 2.B:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2=0.02 $$\end{document}R2=0.02, 2.C:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2=0.16 $$\end{document}R2=0.16, 2.D:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2=0.14 $$\end{document}R2=0.14, and all \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ p < .05 $$\end{document}p<.05). Thus H2.b is only partially supported. #### Hypotheses H3.a—H3.d: The Impact of Statistical Relevance To test H3.a/b/c/d, a GLMM was used analogous to Experiment 1. For hypothesis H3.a, all statistical relevance measures predict the participants response to the tendency causal claims (true or false for Experiments 2.A and 2.B, and level of agreement in Experiments 2.C and 2.D) to a certain extent. However, the amount of variance explained by the statistical relevance measures varies across the experiments. Hypothesis H3.a was generally not supported, because only in two out of 16 instances did the four statistical relevance measures meet the inference criteria across the four experiments. These cases were *r* (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2=0.12 $$\end{document}R2=0.12, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ p<0.0001 $$\end{document}p<0.0001) and *l* (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2=0.13 $$\end{document}R2=0.13, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ p<0.0001 $$\end{document}p<0.0001) in experiment 2.D. In general, the statistical relevance measures met the statistical significance criterion (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ p < .05 $$\end{document}p<.05; as it can be expected in a large sample), but failed the effect size criterion \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2>0.09 $$\end{document}R2>0.09. When the effect size criterion was not met, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 $$\end{document}R2 ranged from 0.0002 to 0.08. The *z* measure in particular fared poorly, contrary to theoretical expectations (van Rooij and Schulz 2019; Crupi and Iacona 2021). It delivered statistically significant results only for experiment 2.B and 2.D and had the lowest \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 $$\end{document}R2 values of all measures (0.0002, 0.05, 0.002, and 0.004 respectively). We made a similar finding for hypothesis H3.b: it was supported in two out of sixteen instances (four statistical relevance measures across four experiments). To test hypothesis H3.C, only those data points were used where the tendency causal claim was indicated as true (Experiment 2.A and 2.B) or got an agreement score above 80 our of 100 (Experiment 2.C and 2.D). In this case, none of the statistical relevance measures adequately predicted the participants response to the conditionals (true or false for Experiments 2.A and 2.B, and level of agreement in Experiments 2.C and 2.D). None of the relevance measures met the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2>0.09 $$\end{document}R2>0.09 criterion and the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ p<0.05 $$\end{document}p<0.05 criterion was only met by *d*, *r*, and *l* in experiment 2.A and *l* in experiment 2.C. To test hypothesis H3.d, the change in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 $$\end{document}R2 is assessed when *conditional probability* *p*(*E*|*C*) is added to the models of H3.a. Unfortunately, none of the statistical relevance measures is an adequate proxy for the probability effects. For experiment 2.C and 2.D, including a statistical relevance measure even a *negative* effect on the participants’ level of agreement with the conditional. ### Summary of Results of Experiment 2 The results agree with the findings of Experiment 1 regarding H1.a and H1.b: conditional claims have more demanding truth conditions than causal claims. Also, like in the previous experiment, Weak Adams’s Thesis (H2.a) was supported, and its restriction to causal claims (H2.b) enjoys partial support. Regarding the predictive value of statistical relevance, the findings are even more negative than in Experiment none of the four hypotheses has been confirmed, for H3.b–H3.d no effect is visible, and there is just very partial confirmation of H3.a (statistical relevance predicts the classification of the causal claim). This confirms that statistical relevance by itself should be treated with great caution as a predictor of causal and conditional judgments. ## General Discussion There is a natural mapping between tendency causal claims (“Smoking weed causes dizziness”) and conditionals in the indicative mood (“if somebody smokes weed, she will feel dizzy”). In the presented study, we tested various hypotheses about the classification of such sentences as true or false, especially with respect to predicting these classifications as a function of conditional probability and statistical relevance. This is a highly relevant research question since the influence of probabilistic factors on causal claims and conditionals has been studied extensively, but in separate literatures. We therefore conducted a study where participants classified a given causal claim and the corresponding conditional as true or false, and estimated in addition two probabilistic the conditional probability of the consequent, given the antecedent, and the probability of the consequent simpliciter. Our specific interest was in finding whether these probabilistic features could reliably predict the classification of causal claims/conditionals as true or false. Our informal discussion at the beginning of this paper has suggested that the truth conditions of indicative conditionals are more demanding than the truth conditions of corresponding tendency causal claims. This claim, expressed in hypotheses H1.a and H1.b, has been supported by our experimental results. This does not exclude that non-causal conditionals such as “if there is smoke, there is fire” can be true even when corresponding causal claims “smoke causes fire” are false.10 In line with theoretical expectations and previous research, conditional probability emerges as a reliable predictor for the classification of the conditional. This finding supports a weak, qualitative version of Adams’s Thesis according to which the conditional probability is correlated with high acceptability/classification as true (H2.a). The effect size in the GLMM is very remarkable (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 \in [0.19;0.62] $$\end{document}R2∈[0.19;0.62]). Effect size decreases when restricted to the set of causal claims classified as true (H2.b)—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 \in [0.12;0.16] $$\end{document}R2∈[0.12;0.16] in Experiments 1, 2.C and 2.D, and under the medium effect size threshold for Experiment 2.A and 2.B. However, this is to be expected given that in the set of true causal claims, most conditionals are classified as it is thus harder to achieve a high effect size than in the more heterogeneous baseline set. The result should therefore not be taken as an argument against the predictive performance of conditional probability. The third set of hypotheses in our paper concerns the role of statistical relevance. Here, we built on probabilistic theories of causal strength and on statistical relevance accounts of indicative conditionals in order to formulate four more hypotheses. In Experiment 1, we find that the classification of a causal claim and/or the corresponding conditional as true or false is usually—but not always—predicted by measures of statistical relevance (H3.a and H3.b). This finding has, however, not been replicated in Experiment 2. Finally, statistical relevance is not any more a relevant predictor if the corresponding causal claim is classified as true (H3.d): the relationship between the various measures and the target variable is still statistically significant, but this is to be expected for such a large data set and the effect size is too small to be of theoretical interest (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R^2 < 0.09 $$\end{document}R2<0.09 for all statistical relevance measures). In line with the partial confirmation of H3.a, this suggests that statistical relevance has little effect on top of causal it is a decent predictor of causal relevance, but when the latter has been established, statistical relevance does not lead to better predictions of the classification of the conditional and is, in any case, inferior to conditional probability as such a predictor. The following overall picture of the classification of causal conditionals emerges. They are judged as true or highly acceptable only if (not: if and only if) the corresponding tendency causal claim is accepted. However, the best *probabilistic* predictor of their classification is the conditional probability *p*(*E*|*C*), and not a statistical relevance measure. These results support, all in all, the “orthodox” line of psychological research that emphasizes the importance of conditional probability for predicting how people evaluate and reason with conditionals (e.g., Evans and Over 2004; Over et al. 2007; Over and Cruz 2023). Our results agree with Evidential Support Theory in so far as EST emphasizes the *relevance* of the antecedent for the consequent (see also Skovgaard-Olsen et al. 2016b). However, we do not observe support for a probabilistic operationalization of EST: statistical relevance can act as a proxy for classifying causal claims (full and partial confirmation of H3.a in the two experiments), but not for the classification of the conditional. Statistical relevance predicts the classification of conditionals neither overall nor in the category of tendency causal claims evaluated as true (see the failure of H3.b/c/d). In other words, we are skeptical that the truth or acceptability of a (causal) conditional can be reliably predicted by purely statistical factors such as the probability *p*(*E*|*C*) and the degree of statistical relevance of *C* for *E*. Something more substantive, which goes beyond statistical association, seems to be required, too. Whether our results support an inferential approach to the semantics of conditionals in general depends on the specific version and the chosen auxiliary assumptions (compare, for example Krzyżanowska et al. 2014; Douven 2015; Douven et al. 2021). One of the limitations of our study is the exclusion of counterfactual conditionals, whose causal character has been studied extensively in the literature (Lewis 1973a, b; Pearl 2000; Schulz 2017). We conjecture that analogical relations might hold between counterfactual conditionals and *actual* causal claims. Consider, for example, the sentences “Ben’s attending the party caused him to fail the exam” and “If Ben had not gone to the party, he would have passed the exam”. The relationship between such pairs of sentences strikes us as a valuable object for further research (see also Schulz 2011). All in all, the interface of conditionals, causality and probability emerges as an important and fruitful area for future research which eventually may lead to a unified theory for both types of expressions.