Sentence reading behavioral data

In addition to the neural activation data, we collected and analyzed behavioral data in the form of comprehension question accuracy and whole-sentence reading times. The former allowed us to ascertain whether participants were reading and processing different parts of the sentences and thereby verify task compliance.

Whole-sentence reading times were computed from trials where the participants indicated reading completion and answered any accompanying comprehension question correctly. (Trials with no completion response and/or incorrect answers were excluded.) Thus, we restricted the analyses to trials where participants appeared to be following the task instructions. For each participant and run, reading times from the eligible trials were length-adjusted as follows. Linear regression was used to generate a predicted reading time for each sentence based on its length (number of letters). The intercept value was set to 0 (because a sentence with 0 characters would require 0 reading time). Length-adjusted reading times were computed as: (Actual reading time) − (Predicted reading time). A positive number indicates that a sentence was read slower than what would be expected based on length.

We used mixed-effects linear regression (lmer function in R version 1.1.463; Bates, Maechler, Bolker, & Walker, 2015) to evaluate the ambiguity effect and compare the effect between conditions. The models contained all applicable fixed effects (e.g., Ambiguity, Structure), random intercepts, and random slopes. We only simplified the random effects structure when the full model did not converge (described wherever applicable in the Results section).

To evaluate whether the ambiguity effect changed over the course of Run2 and whether that change correlated with Stroop performance, we used the same procedures as for the fMRI analyses. However, our predictions for the behavioral data were weaker than for the neural analyses. The use of whole-sentence rather than word-by-word or segment-by-segment presentation allowed us to examine relatively naturalistic reading. However, this design necessarily meant that the reading times were not restricted to the disambiguating regions within a sentence. They likely indexed the many cognitive operations involved in reading a sentence (Rayner, 1998; Rayner, Kambe, & Duffy, 2000), including those that are unrelated to conflict. One consequence of this could be that the lack of conflict-specificity of the reading time measure makes a correlation with Stroop less likely than in the fMRI analyses. A second consequence could be that the inclusion of non-disambiguating-region processing times in the reading time (e.g., wrap-up effects at the end of a sentence) makes evaluation of the change in the ambiguity effect less precise and predictable.

The critical fMRI analyses, by contrast, focused on regions known to be involved in cognitive control and could therefore prove more sensitive to the component of interest. Accordingly, we opted a priori to focus on neural predictions and therefore discuss the corresponding results at greater length below.