Paradigms used to study AEI

Inspired by prior research (e.g., Cohen-Kdoshay & Meiran, 2007; De Houwer, Beckers, Vandorpe, & Custers, 2005; Wenke, Gaschler, & Nattkemper, 2007), Liefooghe and colleagues (2012, 2013) introduced the inducer-diagnostic paradigm to study AEI. On each run, participants are presented with two tasks that share stimuli and responses: the inducer- and the diagnostic-task. First, the participant is presented with two new stimulus-response mappings relevant to the inducer task (e.g., X → left, Y → right). These instructions have to be maintained throughout the diagnostic task where the participant has to ignore the identity of the letter and respond only to its orientation (e.g., upright → left, italics → right). The inducer instructions are only relevant after the diagnostic task, when a stimulus from the inducer task is presented and participants have to execute one of the instructed S-R bindings. The colour of the stimulus indicates which task to perform (e.g., black → diagnostic (orientation) task, green → inducer (identity) task). Performance in the diagnostic task is better on response congruent trials (where the stimulus affords the same response in the diagnostic and inducer tasks) relative to incongruent trials (where the stimulus afforded different responses in each task). Liefooghe et al. (2012, 2013) labelled this difference the instruction-based (task-rule) congruency effect.

Although the inducer-diagnostic paradigm has its merits for investigating AEI, Meiran, Pereg, Kessler, Cole, and Braver (2015a) pointed out that the instruction-based congruency effect measured in the diagnostic task, may not be induced by instructions alone. Their rationale is that participants might accidentally apply the Stimulus-Response (S-R) mappings of the inducer task on trials of the diagnostic task. Following theories of automatization through overt practice (e.g., Logan, 1988, 1992), such task-misapplication may lead to the formation of S-R associations in long-term memory, which may in turn lead to a congruency effect. As a solution, Meiran et al. (2015a) introduced the ‘NEXT’ paradigm. Each mini-block again starts with the presentation of two S-R mappings (e.g., X → left, Y → right) that need to be applied in a 2-trial ‘GO’ phase. These instructions have to be maintained through a short ‘NEXT’ phase where the participant has to simply press one of the two response keys to progress through several ‘slides’ displaying one of the two instructed stimuli. The relevant phase is indicated by the colour of the stimulus (e.g., red = NEXT, green = GO). Even though no classification is necessary during the NEXT phase, response times are faster when the GO response associated with the stimulus is the same as the NEXT response (compatible) relative to when the required responses are different (incompatible). This difference is termed the NEXT compatibility effect.

In comparison to the inducer-diagnostic paradigm, the NEXT paradigm reduces the risk for task-misapplication because the NEXT-phase does not include a two-choice reaction time task as is the case for the diagnostic task. In addition, Meiran et al. (2015a) observed a reliable NEXT compatibility effect on the very first NEXT trial following the instructions (i.e., before the instructed rules could be practiced), thus completely ruling out the potential contribution of task practice. For these reasons, the present study used the NEXT paradigm.

Specificity of the instructed stimulus and response

Several studies have observed that AEI are diminished under high working memory (WM) load conditions (e.g., Cohen-Kdoshay & Meiran, 2007; Meiran & Cohen-Kdoshay, 2012). This led to the conclusion that AEI are restricted to relatively simple S-R tasks or action plans (Cole, Braver, & Meiran, 2017; Meiran et al., 2017). When the retrieval of information has to occur in multiple steps (as is often the case for more complex action plans, including situations in which a large set of stimuli is mapped onto a specific response via their category membership), automatic execution of instructed responses may not occur. A study from Braem, Liefooghe, De Houwer, Brass, and Abrahamse (2017) provided some indirect support for this idea. They found that instructed S-R mappings produced AEI, but not the instructed task context (which can be considered to be a more abstract construct than a task stimulus). This could indicate that AEI are restricted to simple instructions that specify the S-R relationships.

The ‘specificity’ idea also receives support from the ‘implementation intention’ literature (see Martiny-Huenger, Martiny, & Gallowitzer, 2015, for a recent review). A large body of work indicates that goal attainment can be enhanced when people form simple action plans that link situations or cues in the environment to goal-directed actions (e.g., ‘If situation X arises, then I will do Y’). It has been argued that implementation intentions are effective because they allow the automatic execution of goal-directed actions. Importantly, in their review, Martiny-Huenger et al. argued that these plans have to be specific to be effective. For example, forming the intention to buy an apple (a specific item) in the canteen of your office building (a specific context) would be more effective than the intention to buy healthy food (a food category) for lunch (a vague context). In sum, it could be argued that instructions may have to be simple and specific to allow automatic implementation.

However, some recent research on practice-based learning and automaticity has questioned the relative importance of simple S-R associations (e.g., Hazeltine & Schumacher, 2015; Longman, Milton, Wills, & Verbruggen, 2018), and how a ‘stimulus’ and ‘response’ are represented (e.g., Henson, Eckstein, Waszak, Frings, & Horner, 2014; Horner & Henson, 2009, 2011; Longman, Kiesel, & Verbruggen, 2018). For example, Horner and Henson (2009, 2011) have demonstrated that at least two levels of stimulus representation (specific stimulus vs. abstract/semantic representation) can independently become associated with at least three levels of response representation (action, decision, classification). There is also some preliminary AEI evidence that instructions which link abstract stimulus and response codes can be implemented automatically. For example, Cohen-Kdoshay and Meiran (2007, 2009) investigated AEI for instructions describing category-response (C-R) bindings (e.g., first half of alphabet → left, last half of alphabet → right), suggesting that any stimulus from an instructed category might produce AEI (see also, Kiesel, Wendt, & Peters, 2007). Furthermore, using a variant of the inducer-diagnostic paradigm, Tibboel, Liefooghe, and De Houwer (2016; see also Tibboel & Liefooghe, in press) used words to describe items in the instruction phase, but images during the diagnostic phase. They also found AEI, indicating that semantic overlap may be sufficient. However, the words used in the instruction phase referred to very specific objects (e.g., ‘balcony’, ‘bomb’, ‘wall’), so it could be argued that the ‘specificity’ criterion was still met.

Similar findings have also been observed at the response level. Liefooghe et al. (2012) found AEIs even when the response effectors (e.g., index fingers on the left/right hand) or response modality (the spoken words ‘left’ and ‘right’) used in the diagnostic task did not match those used in the inducer task (e.g., middle fingers on the left/right hands). The latter suggests that the instructed response must also be represented at a more abstract level. This is consistent with previous research in the task-switching paradigm, which also observed task-rule congruency effects with responses that overlap only conceptually (e.g., Gade & Koch, 2005; Hübner & Druey, 2006; Schuch & Koch, 2003). Thus, even instructions that lack stimulus- or response specificity may elicit automatic effects.

In sum, the specificity of the response and stimulus representations included in task-sets induced on the basis of instructions remains unclear. On the one hand, some researchers propose that the automatic implementation of instructions is limited to instructions that specify the particular stimulus and action such that all relevant information (including specific exemplars) can easily be stored in WM (e.g., Cole et al., 2017; Martiny-Huenger et al., 2015). On the other hand, there is some preliminary evidence that AEI might also be found with more complex instructions that specify only categories of stimuli/responses and therefore rely on drawing additional information from longer-term memory stores. The available evidence, however, is dispersed across different studies and no systematic investigation has been conducted. Accordingly, the current experiments were designed to contrast AEI for instructions that varied in their degree of specificity, both at the stimulus (Experiment 1) and response (Experiment 2) level.

Method

Participants. 80 students from the University of Exeter participated for £3 or partial course credits (Experiment 1a: N = 40 (35 female), age: M = 20.4 years, SD = 6.3; Experiment 1b: N = 40 (31 female), age: M = 23.5 years, SD = 8.9). The target sample size and exclusion criteria were decided in advance of data collection (when N = 40, we could detect small to medium-sized differences at 80% power in the omnibus ANOVA given an alpha of 0.05). For all experiments from the present study, we obtained approval by the local research ethics committee at the School of Psychology, University of Exeter, and participants provided written informed consent after the nature and possible consequences of the studies were explained to them.

Apparatus and Materials. Stimuli were presented on a 21.5-inch iMac using Psychtoolbox (Brainard, 1997). In both experiments, the stimuli themselves were black drawings of everyday items presented on a white background (the complete set of stimuli used in all experiments are available to download along with the Matlab scripts, raw data files and R scripts from the Open Science Framework data repository: https://osf.io/4hmpv/). The images were collected by performing free images searches on the internet. Each miniblock included two ‘categories’ or ‘stimulus sets’ (pseudorandomly selected from the entire set – e.g., swordfish, guitars) and each category/set included two exemplars. In Experiment 1b, the exemplar stimuli were grouped into more abstract categories (e.g., ‘fruit’) than those used in Experiment 1a, which could be defined according to a single item (e.g., ‘apple’). This manipulation was not systematic, and the categories/stimuli used in Experiment 1b were selected after testing for Experiment 1a was complete.

The NEXT Procedure. In both experiments, each miniblock was divided into three phases: the instruction phase, the NEXT phase, and the GO phase. Figure 1 shows some example instructions from each condition and also the time line of a single trial in the NEXT and GO phases.

The instructions phase differed depending on the condition and experiment. In the exemplar-based instructions condition of Experiment 1a, the four exemplar stimuli used in the miniblock were presented to the left and right of the centre of the screen (2 on the left and 2 on the right). In the label-based instructions condition of Experiment 1a and in Experiment 1b, only the names of the two items were presented to the left and right of the centre of the screen (black Arial font size 30). The location of each exemplar stimulus or label indicated the correct response for that exemplar/set (left vs. right button press – ‘a’ or ‘l’ keys respectively). The instructions were visible until the participant pressed the space bar (but remained visible for at least 3 seconds).

The NEXT phase lasted for 0, 1, 2, or 3 trials (the proportion of each run was 16.66%, 33.33%, 33.33%, and 16.66% respectively to partially control for participants anticipating the start of the GO phase based on the number of NEXT trials and to encourage preparation to perform the GO rules from the outset; see Verbruggen, McLaren, Pereg, & Meiran, 2018). During the NEXT phase, one exemplar stimulus (pseudorandomly selected to balance the number of compatible/incompatible stimuli presented on the 1^st, 2^nd and 3^rd NEXT trial and with no immediate stimulus repetitions) was presented centrally surrounded by a thin black square (side = 50 mm; line thickness = 3 pt) to indicate that no categorization was necessary. Each NEXT trial started with a blank screen presented for 250 ms, followed by a fixation presented for 500 ms, after which the NEXT stimulus appeared. During the NEXT phase, the stimulus remained visible until the correct NEXT response was made (the total number of responses made was recorded for analysis). The NEXT response key (left or right) was counterbalanced over participants.

The GO phase started immediately after the NEXT phase ended and lasted for two trials only (stimuli were pseudo-randomly selected from the complete set of four exemplars used in the current miniblock, with no immediate stimulus repetitions either within the GO phase or from the last NEXT trial). The GO phase was identical to the NEXT phase except that the square surrounding the stimulus was bold (line thickness = 15 pt) to indicate that the stimulus should be classified according to the instructed rules. The time course of the GO trials was similar to the time course for NEXT trials, except that GO trials immediately ended when any response was made. The next miniblock started immediately after the second GO trial (in other words, there was no immediate feedback after individual miniblocks).

The experimental session of Experiment 1a consisted of 8 experimental blocks, including 12 miniblocks each (a total of 48 miniblocks per condition with a duration of 2–5 trials each; see above). Each block used only exemplar- or label-based instructions, and the format of instructions alternated on consecutive blocks (order counterbalanced over participants). The experimental session of Experiment 1b consisted of 4 experimental blocks, including 12 miniblocks each (a total of 48 miniblocks with a duration of 2–5 trials each). Mean GO performance (RT and accuracy) was presented at the end of each experimental block until the participant pressed the space bar to initiate the next experimental block.

Prior to the experimental session there was a brief familiarization phase. In Experiment 1a, this consisted of two practice blocks (three miniblocks per practice block, and each miniblock used novel stimuli). There was one practice block using exemplar-based instructions, and one using label-based instructions (order counterbalanced over participants). In order to provide the participant with some experience of the distribution of NEXT phase durations, each miniblock in each practice block used a different number of NEXT trials (0, 1 or 3; presented in a pseudorandom order). The familiarization phase of Experiment 1b consisted of only one practice block (consisting of three miniblocks). The data from the familiarization phases were not analyzed.

Category ratings. After completion of the experimental phase of Experiment 1b, the participants were asked to rate each stimulus pair. For completeness and as an extra manipulation check, we also asked them to rate the pairs from Experiment 1a. On each trial, we displayed the two exemplar stimuli from one pair (pseudorandomly selected) and the name of the category/item on the screen. The participant then had to rate the abstractness of the category by pressing one of the number keys (1–9) arranged across the top of a standard keyboard (1 = highly concrete, 9 = highly abstract). Participants were informed that a highly abstract category would be very general and could potentially include many examples from a wide range of sub-categories (e.g., ‘food’), whereas a highly concrete category would be very specific and might include very few examples with no obvious sub-categories (e.g., ‘Cox apples’).

Dependent variables and Analyses. All data processing and analyses were performed using R (R Core Team, 2017). In both experiments, the critical comparison was between NEXT performance on compatible trials (stimuli where the correct response during the GO phase was the same as the required response during the NEXT phase) vs. incompatible trials (stimuli where the correct response during the GO phase was different to the required response during the NEXT phase). Consistent with Verbruggen et al. (2018), we focused only on the first NEXT trial because AEI are largest on the first trial (Meiran et al., 2015) and performance on later NEXT trials could already be modulated by practice effects. Note that this decision was made prior to data collection, but it was important to include the remaining NEXT trials in the experiment to guarantee that the duration of the NEXT phase was unpredictable. This manipulation increases the likelihood that the instructions were maintained in a state that would allow their rapid implementation as soon as they became relevant.

Miniblocks where an incorrect response was made during the GO phase (Experiment 1a: 8.67%, Experiment 1b: 14.48%) were omitted from all NEXT analyses because this could indicate the instructions were not processed properly. We also excluded NEXT trials with RT <100 ms or >3000 ms resulting in a further 0.99% data loss in Experiment 1a, and 0.52% in Experiment 1b. The data from thirteen participants (Experiment 1a: 7, Experiment 1b: 6) were replaced because the above data cleaning procedures resulted in <10 (i.e., <50% of the maximum number of) observations per smallest cell (Instructions × Compatibility; see above). We decided on all selection and exclusion criteria before data collection had started.

Consistent with Verbruggen et al. (2018), we focused on three dependent variables. The primary dependent variable was the latency of the NEXT response (NEXT RT; equivalent to the analyses performed by Meiran et al., 2015). Because participants must press the NEXT key (see above) to progress through the NEXT phase, it is possible that NEXT RTs are slower on incompatible trials because participants press the incorrect key (i.e., the key associated with the presented stimulus) before pressing the NEXT key. We therefore also analysed the latency of the NEXT responses limited to those trials on which the first response was correct (Correct NEXT RT). Finally, we also analysed the proportion of errors made on NEXT trials (NEXT PE).

For Experiment 1a, we performed a separate Instructions (exemplar-based, label-based) by Compatibility (compatible, incompatible) ANOVA for each dependent variable. We also report Bayes factors and effect sizes (generalized eta squared) for all relevant effects/interactions. Bayes factors were calculated with the BayesFactor package, using the default JZS prior (.707; Morey, Rouder, & Jamil, 2015). To reduce the number of model comparisons, the interaction was only allowed if all constituent sub-effects were also included (see Morey et al., 2015). When this approach is used, Bayes factors <1 indicate that removing the effect/interaction from the full model is deleterious (i.e., it is a contributor to the fit of the full model). For all Bayesian analyses we report only the BF1 (i.e., the Bayes Factor for evidence in favour of the alternative hypothesis), and we interpret them using the classification discussed in Schönbrodt and Wagenmakers (2017). Finally, we performed paired-samples t-tests to directly assess the NEXT compatibility effect in each instructions condition. We applied a Holms-Bonferroni correction to correct for multiple comparisons. Bayes factors and effect sizes (Hedges’ g) are also reported for these analyses.

In Experiment 1b, we divided the categories into two groups (median split) based on the mean abstractness rating for each stimulus pair (concrete vs. abstract) and then performed Abstractness (concrete, abstract) by Compatibility (compatible, incompatible) ANOVAs on each dependent variable.¹ Note that we used the average group ratings of each category to determine if a category was concrete or abstract (i.e., we did not use subject-specific ratings). Due to a technical issue, we had to run Experiment 1b twice. The results reported here are from the second version of the experiment (in which the issue was resolved). However, because the technical issue did not affect the collection of abstractness ratings, we used all of the available data when calculating the mean abstractness rating of each stimulus pair.

We also analysed the performance data from the GO trials in both experiments. As with the data from the NEXT trials, we first omitted any trials with RT <100 ms or >3000 ms from all analyses, resulting in 0.60% data loss in Experiment 1a and 1.12% in Experiment 1b. Error trials were also omitted from the GO RT analysis. We submitted the GO RTs and proportion of errors on GO trials (GO PE) to paired-samples t-tests (equivalent Bayes factors and effect sizes are also reported) to determine whether the instructions manipulation affected GO performance.

Results and Discussion

The mean NEXT RTs, Correct NEXT RTs and NEXT PEs are plotted in Figure 2 as a function of Instructions and Compatibility. The results from the ANOVAs are reported in Table 1 and the results from the t-tests are reported in Table 2.

Experiment 1a. The ANOVA on the NEXT RTs found that participants were faster to respond on compatible (mean RT = 749 ± 24 ms) relative to incompatible trials (mean RT = 865 ± 32 ms; p < .001, BF < 0.001) – AEI. Participants also responded faster with exemplar-based instructions (mean RT = 785 ± 29 ms) relative to label-based instructions (mean RT = 829 ± 31 ms; p = .007, BF = 0.135) indicating that participants found the task easier when they had seen the stimuli prior to the start of the miniblock (even though the identity of the exemplars/items was irrelevant in the NEXT phase). The AEI were numerically larger in the exemplar-based instruction condition (mean difference = 141 ± 20 ms) than in the label-based instruction condition (mean difference = 90 ± 23 ms), but the Compatibility by Instructions interaction did not reach significance and the Bayesian analysis provided anecdotal evidence that removal of the interaction from the model would not materially impair its fit (Table 1). Critically, the planned contrasts found conclusive evidence of AEI or NEXT compatibility effects in both instruction conditions (see Table 2) indicating that AEI are not limited to conditions where the instructions display the relevant exemplar stimuli. As can be seen in Tables 1 and 2, the analyses on the Correct NEXT RTs and NEXT PEs broadly mirrored those of the NEXT RTs.

Finally, we correlated the NEXT compatibility effects from the two (within-subject) instruction conditions.² We found positive correlations for the NEXT RTs (r(39) = .32, p = .042), Correct NEXT RTs (r(39) = .27, p = .097; though the correlation was not significant in this measure), and NEXT PEs (r(39) = .48, p = .002). These findings indicate that those participants with a large NEXT compatibility effect in one instructions format were also likely to have a large effect in the other – i.e., there was little evidence that the instructions manipulation resulted in large individual differences in performance across conditions.

Mean GO RTs and GO PEs are reported in Table 3. GO RTs were considerably faster in the exemplar-based instructions condition (mean GO RT = 689 ± 22 ms) than in the label-based instructions condition (mean GO RT = 735 ± 27 ms; t(39) = 4.06, p < 0.001, BF = 116.136, g_av = 0.295), indicating that the GO task was easier when the instructions specified the particular exemplar stimuli. GO PEs were broadly equivalent in both instructions conditions (GO PEs for exemplar-based instructions = 5.0 ± 0.6%; label-based instructions = 5.3 ± 0.6%; t(39) = 0.56, p = 0.576, BF = 0.198, g_av = 0.087).

Experiment 1b. NEXT RTs were faster on compatible NEXT trials (NEXT RT = 782 ± 35 ms) relative to incompatible NEXT trials (NEXT RT = 848 ± 38 ms, p = .015, BF = 0.092 ± 5.6%). However, neither the main effect of Abstractness nor the Compatibility by Abstractness interaction approached significance, and the Bayesian analyses provided moderate evidence that removal of the effect/interaction would not materially impair the fit (Table 1). Thus, AEI were not influenced by the abstractness of the instructed categories. More generally, this finding suggests that abstract C-R mappings can also be implemented automatically. Inspection of Table 1 shows that this pattern was observed for all dependent variables.

GO phase. Mean GO RTs and GO PEs are reported in Table 3. GO performance was broadly comparable whether the categories were more concrete (mean GO RT = 869 ± 40 ms, mean GO PE = 8.6 ± 0.9%) or more abstract (GO RT = 880 ± 37 ms, GO PE = 9.7 ± 1.0%). The small difference between categories did not approach significance for either dependent variable (GO RT: (t(39) = 0.88, p = 0.382, BF = 0.246, g_av = 0.043; GO PE: (t(39) = 1.27, p = 0.212, BF = 0.359, g_av = 0.180).

Method

48 different students from the University of Exeter (39 female) with a mean age of 20.8 years (SD = 3.2) participated for £7 or partial course credits. The number was higher than in the previous experiments due to the counterbalancing.

The apparatus, design and procedure were largely identical to Experiments 1a and 1b, but with the following differences. The stimuli were letters, numbers and symbols selected from readily available fonts (e.g., Wingdings, Webdings, Hebrew alphabet, ideograms, etc.), presented in black on a white background. Each miniblock included two stimuli (pseudorandomly selected from the entire set), one associated with a left response and the other associated with a right response (see Figure 1).

The set of relevant GO response effectors changed on every experimental block (but remained constant within each experimental block): left vs. right index fingers (‘f’ and ‘h’ keys respectively); left vs. right little fingers (‘a’ and ‘l’ keys respectively); left vs. right foot response (using a foot pedal response box). The NEXT response was always a manual response (left or right response with the index or little finger) and remained constant throughout the experiment (counterbalanced over participants). Whether the GO response used the same set of fingers (‘same fingers’ condition), different fingers (‘different fingers’ condition), or different extremities (‘feet’ condition) was included as a factor in our analyses. The experimental phase consisted of 48 experimental blocks of 6 miniblocks each (a total of 96 miniblocks per condition, each consisting of 2–5 trials). In each experimental block, the GO response was performed by one set of effectors (index fingers, little fingers, or feet). The relevant effector for each experimental block alternated throughout the experiment (e.g., index fingers… little fingers… feet… index fingers…) and the order was counterbalanced over participants. The maximum number of observations available for analysis in each effector condition was 40 compatible and 40 incompatible NEXT stimuli.

Prior to the experimental session there was a brief familiarization phase consisting of three practice blocks of three miniblocks each (each miniblock used novel stimuli), one practice block requiring index finger GO responses, one practice block requiring little finger GO responses, one practice block requiring foot GO responses (order counterbalanced over participants). Each miniblock in each experimental block of the familiarization phase used a different number of NEXT trials (0, 1 or 3; presented in a pseudorandom order). The data from the familiarization phase was not analysed. The entire session (including the familiarization phase) lasted approximately one hour.

The same exclusion criteria as in Experiments 1a and 1b resulted in the omission of 13.30% (incorrect GO responses) + 0.21% (data trimming) of the trials. In the GO phase, 0.32% of the correct RTs were removed after data trimming. The data from 9 participants were replaced because the number of observations per cell after data cleaning was too low.³

To analyze the NEXT data, we performed an Effector (same fingers, different fingers, feet) by Compatibility (compatible, incompatible) ANOVA for each dependent variable. The Effector by Compatibility interaction can reveal if AEI effects are influenced by specificity at the response level. We also performed paired-samples t-tests and their Bayesian equivalents to directly assess the NEXT compatibility effect in each Effector condition. We submitted the GO RTs and GO PEs to one-way ANOVAs with Effector as the within-subjects factor (equivalent Bayes factors and effect sizes are also reported) to determine whether GO performance was modulated by the effectors used to respond.

Results and Discussion

NEXT phase. The mean NEXT RTs, Correct NEXT RTs and NEXT PEs are plotted in Figure 2 as a function of Effector and Compatibility. The results from the omnibus ANOVAs are reported in Table 4 and the results from the planned comparisons are reported in Table 5.

Participants responded faster on compatible (mean RT = 593 ± 19 ms) relative to incompatible trials (mean RT = 631 ± 20 ms; p < .001, BF < 0.001). The main effect of Effector did not reach significance (p = .233, BF = 6.364), but the difference in performance between compatible and incompatible stimuli during the NEXT phase was modulated by the effectors used to respond during the GO phase. Thus, in contrast to the findings of Liefooghe et al. (2012) and the current Experiment 1, specificity did seem to matter in Experiment 2. The NEXT compatibility effect was larger when the GO response was performed with the same fingers (mean difference = 79 ± 16 ms) relative to when the GO response was performed with different fingers (mean difference = 17 ± 11 ms) or with the feet (mean difference = 16 ± 11 ms; Effector by Compatibility interaction: p < .001, BF = 0.023). The planned contrasts found that the NEXT compatibility effect was reliable when the GO response was performed with the same fingers as the NEXT response (t = 5.03, BF > 1000). It did not reach significance when the GO response was performed with different fingers (t = 1.68, BF = 0.580) or with the feet (t = 1.43, BF = 0.403), but the Bayesian analyses in the latter two contrasts found only anecdotal evidence in support of the null hypothesis. Thus, even though we can conclude that AEI were substantially larger when the same effector was used and that specificity has an impact at the response level (the main aim of this experiment), further research is required to clarify whether the numerical differences in the expected direction for the ‘different fingers’ and ‘feet’ conditions represent abstract coding of instructed responses at all (cf. Liefooghe et al., 2012). Again, the analyses of correct NEXT RTs and NEXT PEs revealed similar patterns.

Taken together these results indicate that AEIs in the NEXT paradigm are largely limited to conditions when the GO and NEXT responses are executed with the same effectors (at best, AEI are markedly reduced when the NEXT and GO phases use different response effectors). This would seem to be at odds with Liefooghe and colleagues’ (2012) finding that the AEI in their paradigm were robust to changes in response effectors and even response modalities. There are at least three possible explanations for this apparent contradiction. First, it is possible that the frequent switching of response effectors in the current experiment might have induced shielding to reduce the response crosstalk between the two tasks. Second, response interference might be more likely to occur in two-choice tasks (in which multiple response options can be relevant), such as the diagnostic task of Liefooghe et al., compared to single-response tasks such as our NEXT task. Finally, it is possible that the generalization observed by Liefooghe et al. was the result of a practice effect – a possibility that we ruled out by only analysing performance on the first NEXT trial. We will come back to this issue in the General Discussion.

GO Phase. Mean GO RTs and GO PEs are reported in Table 3. GO RTs were fastest when the GO response was performed with the same fingers as the NEXT response (mean GO RT = 568 ± 15 ms), or with different fingers to the NEXT response (mean GO RT = 579 ± 15 ms) and was slowest when it required a foot response (mean GO RT = 630 ± 17 ms). The main effect of Effector was significant (F(2,94) = 25.69, p < 0.001, MSE = 1999.43, BF > 1000, η² = 0.353). The ANOVA on GO PEs found that participants made more errors when the GO response used the same fingers as the NEXT response (mean GO PE = 8.09 ± 0.7%) or different fingers to the NEXT response (mean GO PE = 8.72 ± 0.8%) than when the GO response was performed with the feet (mean GO PE = 6.99 ± 0.6%). This effect was also reliable (F(2,94) = 7.48, p < 0.001, MSE = 4.93, BF = 28.087 ± 0.6%, η² = 0.137), possibly indicating a speed-accuracy trade-off in GO performance – i.e., responding with the feet was slower, but more accurate than responding with the fingers.