While research indicates that people are skilled causal reasoners, systematic deviations from the normative causal Bayesian network model have been observed. These include Markov violations, failures to ‘explain away’, and conservative responding. Different processes have been posited to account for these violations: sampling, associative reasoning, and heuristics. These processes entail effects of response time. To test the relationships between these theories, normative violations, and reasoning time we conducted a causal reasoning study employing time pressure manipulations and response time measurements. Our results show that time pressure decreases overall accuracy. Crucially, we find that time pressure does not affect the magnitude Markov independence violations. This is not what many existing explanations would predict. We find evidence that participants’ responses result from two separate cognitive processes and that time pressure modulates their relative contribution to responses. Hence we provide an explanation of non-normative reasoning patterns based on a mixture of cognitive processes.

A recent sampling-based model of causal cognition, the Mutation Sampler, can explain the average responses of participants on a wide variety of tasks (Davis & Rehder, 2020). Careful analysis, however, shows that the predicted response distributions do not match empirical data. As an explanation for the observed response distributions, we propose that people engage in mutation sampling when confronted with a causal query and subsequently integrate this information with a prior. Several theoretical arguments can be made for the notion that people use priors in causal judgements; it is rational and matches the phenomenology of making such judgements. We have modelled this process with a new model called the Bayesian Mutation Sampler (BMS). We fit the BMS to experimental data and find that, in addition to average responses, the BMS explains multiple ‘distributional phenomena’, such as the moderate conservatism of the bulk of responses and the lack of extreme responses.

The Interface Transparency Thesis (ITT) proposes that people tend to use a canonical interpretation of linguistic expressions, even when this interpretation is sub-optimal for the task at hand. The current paper sought to investigate this claim further by adding a time-pressure manipulation to a quantified sentence verification task and analyzing the results through a computational model of decision-making. The results indicate that time pressure -while effectively changing behavioral responses- does not alter cognitive processes associated with quantifier verification, thus supporting the ITT.

Experimental studies show that the tendency to derive Scalar Implicatures (SIs) varies considerably between individuals: some individuals accept sentences that are literally true but carry a false SI, while others systematically reject them. The question of what factors drive these differences is crucial to understanding the mechanisms involved in SIs and currently at the center of numerous discussions. To date, there is no agreement on how to quantify individual differences in SI rates. In this article, we show how a hierarchical Bayesian modelling approach can be used to quantify subjects' preferences observed in the results of a truth value judgement task that investigated intra-individual and inter-individual variability in the rates of upper-bounding and lower-bounding SIs associated with the -scale. The results provide further evidence that the robustness of an SI is modulated within individuals by certain linguistic features, such as the presence of negation.

Psycholinguistic studies have repeatedly demonstrated thatdownward entailing (DE) quantifiers are more difficult to pro-cess than upward entailing (UE) ones. We contribute to thecurrent debate on cognitive processes causing the monotonic-ity effect by testing predictions about the underlying processesderived from two competing theoretical proposals: two-stepand pragmatic processing models. We model reaction timesand accuracy from two verification experiments (a sentence-picture and a purely linguistic verification task), using the dif-fusion decision model (DDM). In both experiments, verifica-tion of UE quantifier more than half was compared to verifica-tion of DE quantifier fewer than half. Our analyses revealedthe same pattern of results across tasks: Both non-decisiontimes and drift rates, two of the free model parameters of theDDM, were affected by the monotonicity manipulation. Thus,our modeling results support both two-step (prediction: non-decision time is affected) and pragmatic processing models(prediction: drift rate is affected).