Contingency and congruency switch in the congruency sequence effect: a reply to Blais, Stefanidi, and Brewer (2014)

INTRODUCTION The congruency sequence effect (CSE) is the observation that the congruency effect is reduced following an incongruent trial (Gratton et al., 1992). Generally, the CSE is interpreted in terms of conflict adaptation, the idea that participants decrease attention to the distracter and/or increase attention to the target after experiencing conflict (e.g., Botvinick et al., 2001). An alternative learning and memory account proposes that the CSE is instead due to basic learning confounds (for a review, see Schmidt, 2013). For instance, systematic differences in the types of feature repetitions that are possible in each cell of the design might produce a CSE (Mayr et al., 2003; Hommel et al., 2004). Schmidt and De Houwer (2011) considered two additional learning and memory biases: sequential contingencies and congruency switch costs. However, Blais et al. (2014) present data which they suggest argue against a role of these two biases. This article illustrates some issues with this work and suggests that contingency and congruency switch biases may play a role after all. CONTINGENCY ANALYSIS UNDERPOWERED Schmidt and De Houwer (2011) introduced the idea that sequential contingency biases might influence the CSE. Often, each distracter is presented more frequently in the congruent color than in each incongruent color. Unfortunately, this introduces a contingency, whereby words are predictive of the congruent response. Contingency biases are larger following an accurately predictive trial than following an incorrectly predictive trial (Schmidt et al., 2007). Thus, contingency biases can contribute to the CSE. Indeed, Mordkoff (2012) showed that, with feature repetitions removed, the CSE is present in a contingency-biased Simon task, but is not observed in a contingency-unbiased version of the same task. Blais et al. (2014) report a reanalysis of verbal Stroop data in which each participant performed several blocks of trials with varying contingencies (as manipulated with proportions of congruent trials) from 5 to 95% in increments of 5% (though only 10 to 80% could be analyzed). Overall, CSEs were not reliable for most contingency levels. Critically, the CSE did not significantly increase as a function of contingency in response times. With these data, the authors argued that contingencies are unlikely to play a role in the CSE. However, statistical power of the sample of 15 participants is a concern. Indeed, the slope was notably positive, but with considerable error, B = 0.438 ± 0.382, t(14) = 1.15, p > 0.25. The B parameter is the amount of change in the CSE for a 1% contingency increment (i.e., 2.2 ms for each 5% increment, and 31 ms overall). Though not significant, this represents a medium effect size (β = 0.315). As Figure 1 illustrates, the study only had high (0.8) power to detect a large effect size (β = 0.661). As a further concern, contingencies were manipulated between blocks. Contingency biases are known to transfer across blocks (Schmidt et al., 2010), causing contamination. Thus, contingency biases were probably underestimated. Curiously, the slope for the errors was significant and in the predicted direction, B = 0.038 ± 0.017, t(14) = 2.26, p = 0.040. Though seemingly confirming a role of contingencies, the authors reasoned that this slope is difficult to interpret given that none of the CSE estimates for the various contingency levels were significant. This argument does not seem particularly convincing, only feeding concerns about statistical power.