Evaluative priming in a semantic flanker task: ERP evidence for a mutual facilitation explanation
In semantic flanker tasks, target categorization response times are affected by the semantic compatibility of the flanker and target. With positive and negative category exemplars, we investigated the influence of evaluative congruency (whether flanker and target share evaluative valence) on the flanker effect, using behavioral and electrophysiological measures. We hypothesized a moderation of the flanker effect by evaluative congruency on the basis of the assumption that evaluatively congruent concepts mutually facilitate each other’s activation (see Schmitz & Wentura in Journal of Experimental Psychology: Learning, Memory, and Cognition 38:984–1000, 2012). Applying an onset delay of 50 ms for the flanker, we aimed to decrease the facilitative effect of an evaluatively congruent flanker on target encoding and, at the same time, increase the facilitative effect of an evaluatively congruent target on flanker encoding. As a consequence of increased flanker activation in the case of evaluative congruency, we expected a semantically incompatible flanker to interfere with the target categorization to a larger extent (as compared with an evaluatively incongruent pairing). Confirming our hypotheses, the flanker effect significantly depended on evaluative congruency, in both mean response times and N2 mean amplitudes. Thus, the present study provided behavioral and electrophysiological evidence for the mutual facilitation of evaluatively congruent concepts. Implications for the representation of evaluative connotations of semantic concepts are discussed.
KeywordsEvaluative priming Affective priming Flanker task ERP N2
Positive and negative attitudes toward objects or events are almost automatically activated—for example, when humans detect objects in the environment (e.g., Öhman, Flykt, & Esteves, 2001), when they remember studied material in a recognition memory task (e.g., Zajonc, 1980), or even when they are confronted with new verbal or pictorial stimuli (see Duckworth, Bargh, Garcia, & Chaiken, 2002). This phenomenon has led some researchers in the field of cognition and emotion to postulate that the evaluative dimension has a privileged status, in comparison with semantic, nonevaluative dimensions (e.g., Bargh, Chaiken, Raymond, & Hymes, 1996; Murphy & Zajonc, 1993).
One crucial question in this context concerns the memory representation of the evaluative features enabling preferential processing of the evaluative dimension. While a variety of paradigms have been used to investigate this issue (e.g., the evaluative priming task [see Fazio, Sanbonmatsu, Powell, & Kardes, 1986], the evaluative Simon task [see De Houwer & Eelen, 1998], and the emotional Stroop task [see Pratto & John, 1991]), the evaluative priming task is the most popular one, examining the evaluative relation (i.e., evaluative congruency or incongruency) between semantic concepts (see Klauer & Musch, 2003, for a review). In a typical evaluative priming task, a positive or negative prime precedes a positive or negative target, with only the target requiring a response. Shorter response times (RTs) and fewer errors in evaluatively congruent (i.e., both prime and target are either positive or negative), as compared with incongruent (i.e., the prime is positive and the target is negative, or vice versa) conditions are taken as evidence for evaluative priming (see Klauer & Musch, 2003, for a review).
Regarding the interpretation of the evaluative priming effect, the type of task is a highly important factor. Early evaluative priming studies applied an evaluation task, in which the target is evaluated as positive or negative (Fazio et al., 1986); hence, evaluative congruency and response compatibility of the prime and target are fully confounded in this task. Evaluative priming effects in the evaluation task have been explained by the assumption that the target evaluation leads participants to also evaluate the prime, which, in turn, either supports the target evaluation or interferes with it (e.g., Klauer, Roßnagel, & Musch, 1997; Wentura, 1999). We label this type of evaluative priming effect the stimulus–response (S–R)-based evaluative priming effect, since S–R associations—that is, the evaluative response compatibility of the prime and target—can easily account for the priming effect (see De Houwer, 2003).
Deeper insights into the mnemonic representation of the evaluative connotations, however, can be gained by using tasks that do not confound evaluative congruency and response compatibility. A prominent example is the naming task, in which the target words are simply uttered as quickly as possible, while primes, as well as targets, are positively or negatively valenced. In the remainder of this article, we label this type of evaluative priming effect the stimulus–stimulus (S–S)-based evaluative priming effect, because the similarity of stimulus features—that is, the inherent evaluative congruency of the prime and target—can account for the priming effect (see De Houwer, 2003). Since S–S-based evaluative priming effects are not explainable by response processes, any explanation needs to involve the mental representations of the prime and target’s evaluative dimension. Specifically, an evaluatively congruent prime is thought to facilitate target encoding, resulting in faster target responses (see, e.g., Bargh et al., 1996; Hermans, De Houwer, & Eelen, 1994; Spruyt, Hermans, De Houwer, & Eelen, 2002).
A puzzling evaluative priming effect and its explanation by mutual facilitation
Some studies have questioned the actuality of S–S-based evaluative priming effects (Klauer & Musch, 2001; Spruyt, Hermans, Pandelaere, De Houwer, & Eelen, 2004), but the consensus is now that this type of evaluative priming effect is a genuine phenomenon (e.g., De Houwer, Hermans, & Spruyt, 2001; De Houwer & Randell, 2002, 2004; Everaert, Spruyt, & De Houwer, 2011; Spruyt, De Houwer, Everaert, & Hermans, 2012; Spruyt, De Houwer, & Hermans, 2009; Spruyt, De Houwer, Hermans, & Eelen, 2007; Spruyt & Hermans, 2008; Spruyt et al., 2002; Wentura & Frings, 2008). Still, the results of some evaluative priming studies using the naming task remain puzzling. In particular, Glaser and Banaji (1999) reported negatively signed effects for extremely valenced prime words but positive effects for moderately valenced prime words.
Trying to explain the inconsistent findings in S–S-based evaluative priming tasks, Schmitz and Wentura (2012) proposed a three-process model of evaluative priming (see also Wentura & Rothermund, 2003). The first process characterizes the parallel activation of prime and target representations (i.e., a precondition for S–R-based evaluative priming effects). The second process is an additional facilitative activation of prime and target concepts in the case of evaluative congruency. We call this process mutual facilitation because it may operate in both directions—that is, from prime to target, as well as from target to prime. On the one hand, prime processing may facilitate subsequent target processing because an evaluatively congruent prime facilitates the encoding of the target. This target-encoding facilitation is the component traditionally used to explain S–S-based evaluative priming effects. On the other hand, target processing may serve to uphold prime activation via the evaluative congruency of both concepts. This prime activation maintenance elicited by an evaluatively congruent target is typically not addressed, whereas this component may be of great importance for the interpretation of evaluative priming effects. Presumably, experimental specifics (e.g., the stimulus onset asynchrony [SOA]) determine whether the prime-to-target or the target-to-prime component has a relatively stronger influence. We will soon return to this issue.
The third suggested process is response related and may result in response facilitation (given response compatibility between the prime and target) or response conflict (given incompatibility). As was discussed above, these response-related processes are traditionally used to explain S–R-based evaluative priming effects. Some have argued that this third response-related process is irrelevant in the naming task because primes are not associated with a task-relevant response. However, this assumption can be challenged (e.g., Wentura & Frings, 2008). One might assume that in the context of naming target words, prime words also trigger a naming response, which is incompatible with the target response.
The three-process model predicts that this naming response conflict is larger in the case of evaluative congruency because a congruent target will help maintain the activation of the prime (due to process 2, discussed above). In other words, the target-to-prime activation maintenance in the case of evaluative congruency can increase response conflict and, hence, target RTs, which can result in a negatively signed evaluative priming effect (longer target RTs in the case of evaluative congruency vs. incongruency). On the other hand, prime-to-target encoding facilitation speeds up target RTs and, hence, can result in positive evaluative priming effects. It follows that one might obtain positively or negatively signed evaluative priming effects (or even null effects when both components cancel each other out), depending on whether the prime-to-target or the target-to-prime component (at process 2 of the model) is dominant.
This three-process model elegantly explains the puzzling effects of Glaser and Banaji (1999), insofar as we can assume that target-to-prime processes and response conflicts played a major role in their experiments (for details, see Wentura & Frings, 2008; Schmitz & Wentura, 2012). Wentura and Frings (2008) as well as Schmitz and Wentura (2012, Experiment 1) provided a priori hypothesized evidence for this explanation of the naming task data. In the present article, we want to focus on an important implication of the three-process model.
Implementation of mutual facilitation in “classic” priming explanations
The central claim of the three-process model is that evaluatively congruent primes and targets are activated in parallel and mutually facilitate each other. To explain the importance of this claim, we first need to consider the following question: How do we explain S–S-based evaluative priming effects? The “classic” (and to anticipate, unsatisfactory) answer involves a semantic network featuring valence nodes (see Bower, 1991). In such a model, a prime word (e.g., love) will activate its internal representation, symbolized as a specific node. From this love node, activation will spread to the positive valence node and, from there, to all other nodes representing positive concepts. This spreading activation may preactivate potential targets with a positive connotation (e.g., sun) but not negatively connoted targets (e.g., death). This account is ruled out by fanning (see, e.g., J. R. Anderson, 1974): The amount of activation spreading from the source node to connected nodes must be limited, which is why the amount of activation arriving at the goal node depends on the number of links emanating from the source node. In the case of valence nodes, the number of connected nodes is too large for the spreading activation to produce observable evaluative priming effects.
An elegant alternative is provided by distributed memory models of priming (see Masson, 1991, 1995; McRae, de Sa, & Seidenberg, 1997). In these models, each semantic concept corresponds to a specific pattern of activated processing units within semantic space. The semantic relatedness between concepts is determined by the number of shared activated units. Thus, priming effects are explained by a faster transition from the pattern corresponding to the prime concept to the pattern corresponding to the target concept if the two are semantically related, in which case there are shared units that are already in the appropriate mode of activation. This model is prima facie perfectly suited to account for S–S-based evaluative priming, given the assumption that a considerable part of the activation pattern represents the evaluative features of a concept (see Spruyt et al., 2002; Wentura, 1999, 2000). Note, however, that this account suggests that at any specific point in time, the complete pattern of only one single concept can be activated in semantic space.
In contrast to this, the three-process model claims parallel activation of several concepts that (1) might mutually facilitate each other, while (2) they might oppose each other with regard to response determination. To further corroborate this claim, Schmitz and Wentura (2012) applied a variation of the Eriksen flanker task (Eriksen & Eriksen, 1974), which resembled a priming task except for the temporal position of the prime. (The prime appeared after target onset. We will refer to this “prime” as a flanker in the following to avoid an intermixture of task terminologies.) In the flanker task, a target word presented in the center of the screen had to be categorized as denoting either a person (e.g., brother) or an animal (e.g., snake). A flanker word appeared both above and below the target (i.e., the same word flanked the target on both sides) and was also a person or an animal word. Furthermore, both targets and flankers had either a positive or a negative valence. Thus, both targets and flankers varied in terms of (task-irrelevant) evaluative categories, as well as (task-relevant) semantic categories. While semantic compatibility was confounded with response compatibility, evaluative congruency was not.
The typical flanker effect is characterized by faster target categorization given target–flanker response compatibility, as compared with incompatibility. We aimed to examine whether the occurrence of this effect was influenced by the evaluative congruency between the target and flanker. Because we hypothesized that evaluative congruency would strengthen target-to-flanker activation (due to process 2 of the three-process model), we created presentation conditions under which the flanker effect should be diminished without this additional activation boost (or in other words, conditions that avoided a ceiling in the flanker effect). To this end, we used a target–flanker onset delay of 50 ms, with a 100-ms flanker presentation interval (in contrast, the target remained on screen until the response). We predicted increased flanker activation in the case of evaluative congruency between the flanker and target, leading to a larger potential of the flanker to facilitate (in the case of semantic compatibility) or disturb (in the case of incompatibility) the target response. As predicted, the flanker effect was moderated by evaluative congruency and arose only for evaluatively congruent target–flanker pairs. This pattern of results corroborates the assumption of the three-process model that the flanker is able to interfere with the target response only if its activation is supported by the evaluatively congruent target (due to process 2). Since the target onset preceded the flanker onset, this result supports the notion of parallel activation.
Corroboration of the three-process model assumptions by ERP evidence
However, there is a caveat. One might argue that, in fact, only a single distributed pattern is active at any point in time. An evaluatively congruent flanker could strengthen those parts of the target pattern that overlap with the flanker representation, while simultaneously weakening the activation of nonoverlapping parts. Thus, according to this argument, in cases of evaluative congruency and semantic incompatibility, target response slowing is caused not by response conflict between concurrently activated target and flanker responses but, instead, by a short temporary dilution of the response part of the target pattern, which occurs when a response-incompatible flanker representation is active. Once the flanker is offset (while the target remains on-screen), the target pattern’s response part is fully reactivated to determine the response. Thus, parallel activation of a response-incompatible target and flanker representation accompanied by an explicit response conflict might not be necessary in order to explain target response slowing due to evaluatively congruent and semantically incompatible flanker activation.
The end result—that is, longer RTs in the case of semantic incompatibility, as compared with compatibility, given that flanker and target are evaluatively congruent—does not distinguish between the rival hypotheses derived from the three-process model and this type of distributed memory model. An independently established indicator of response conflict might, hence, be informative as an additional dependent measure for testing the two rival hypotheses. In particular, the N2 component of the event-related potential (ERP) has been identified in the electrophysiological literature as a marker of response conflict (see Van Veen & Carter, 2002). The N2 component arises around 250–380 ms after stimulus onset, with a maximum at central or frontocentral electrodes, and has been found in different response conflict paradigms (Bartholow et al., 2005; Van Veen & Carter, 2002; see Folstein & Van Petten, 2008, for a review). N2 amplitude differences have been directly associated with the standard flanker effect, with larger negativities in cases of response-incompatible flankers, as compared with compatible ones (e.g., Kopp, Rist, & Mattler, 1996; Van Veen & Carter, 2002).1
As in previous studies, we analyzed N2 mean amplitudes with respect to semantic categorization conflicts between target and flanker. However, we also manipulated evaluative congruency (as in Schmitz & Wentura, 2012) and expected this factor to moderate the semantic response compatibility effect. The three-process model suggests that the activation of the briefly presented flanker should be maintained by an evaluatively congruent target. Only in that case should we then expect response conflict between a target and flanker, leading to the associated flanker effect in N2 mean amplitudes (i.e., larger negativities given semantic incompatibility, in comparison with compatibility). In the case of evaluative incongruency, the flanker effect is expected to be (almost) absent.
In addition to the N2 component, we also assessed two later components that were typically analyzed in response interference paradigms, the P3 component and the lateralized readiness potential (LRP). The P3 is a ubiquitous component prominent in all kinds of decision-making tasks, typically arising around 300–500 ms after stimulus onset at parietal electrodes (see Duncan-Johnson & Donchin, 1982; Pritchard, 1981). Of special interest for our purpose is the finding that the P3 latency increases when stimulus categorization becomes more difficult (see Kutas, McCarthy, & Donchin, 1977; Liu, Xin, Jin, Hu, & Li, 2010; Magliero, Bashore, Coles, & Donchin, 1984; McCarthy & Donchin, 1981). Accordingly, in the flanker task, longer P3 latencies have been reported in incompatible, as compared with compatible, conditions (see Coles, Gratton, Bashore, Eriksen, & Donchin, 1985; Gehring, Gratton, Coles, & Donchin, 1992; Smid, Mulder, & Mulder, 1990). Hence, in the present experiment, P3 peak latencies were analyzed with regard to target–flanker response compatibility. Target categorization was predicted to be more difficult if the flanker calls for a conflicting semantic categorization. Again, we expected this effect to occur prominently in cases of evaluative congruency, when the activation of a semantically incompatible flanker is maintained by the target. In cases of evaluative incongruency, however, the flanker was expected to be too weakly activated to interfere with target categorization.
The LRP represents the lateralized part of the readiness potential (see Vaughan, Costa, & Ritter, 1968) that is at least partly generated in the primary motor cortex (see Coles, 1989; Miller & Hackley, 1992). The LRP is seen as an index of selective response preparation (see Coles, 1989; Gratton, Coles, Sirevaag, Eriksen, & Donchin, 1988; Miller & Hackley, 1992) and maximally arises over central scalp locations contralateral to the hand that is responsible for the movement (thereby reflecting the contralateral organization of the motor cortex; Brunia, 1988). The LRP onset indicates the beginning of side-specific response preparation (e.g., Coles, 1989). In the flanker task, shorter LRP onset latencies arose for response-compatible, as compared with incompatible, flankers (e.g., Carrillo-de-la-Peña, Lastra-Barreira, & Galdo-Álvarez, 2006). One might also expect this effect in our variant of the flanker task. However, obtaining LRP onset latency differences between semantically compatible and incompatible conditions that are constrained to evaluative congruency does not resolve the conflict between the two rival hypotheses—that is, the conditioned response interference hypothesis and the “diluted response representation” hypothesis (see above). Instead, the LRP data would be meaningful with regard to our rationale only if we obtained evidence for an LRP deflection associated with a flanker-related response tendency—that is, an LRP deflection in the direction of the wrong response for semantically incompatible flankers (in case of evaluative congruency). Such an LRP deflection in response to response-incompatible flankers has been reported in several studies (see Bartholow & Dickter, 2008; Kopp et al., 1996; Van ’t Ent 2002); it was, however, often constrained to trials with slow responses (Eder, Leuthold, Rothermund, & Schweinberger, 2012; e.g., Gratton et al., 1988; Heil, Osman, Wiegelmann, Rolke, & Hennighausen, 2000). In our experimental setting, it was, a priori, unlikely that we would find it: Since the flanker appeared with an onset delay of 50 ms in comparison with the target onset, the preparation of both responses (i.e., the response associated with the target and the response associated with the semantically incompatible flanker) should temporally overlap to a greater extent than in flanker studies with flanker onset preceding target onset. Therefore, we did not expect to observe a clear positive LRP deflection in response to the flanker; nevertheless, we analyzed the LRP data with regard to this issue.
Finally, we analyzed the N400 component with regard to evaluative congruency. The N400 is a broad negative-going ERP component (e.g., Heil & Rolke, 2004; Kiefer & Martens, 2010) that is sensitive to semantic context (for reviews, see Brown & Hagoort, 2000; Kutas & Federmeier, 2000; Kutas & Van Petten, 1994; Osterhout & Holcomb, 1995; Pritchard, Shappell, & Brandt, 1991). Most important, it has been widely explored in semantic priming tasks, with the main finding that targets accompanied by semantically incongruent primes elicit a relatively more negative deflection than do targets in the congruent condition (e.g., Bermeitinger, Frings, & Wentura, 2008, 2010; Heil & Rolke, 2004; Kiefer & Martens, 2010). Thus, the N400 was related to the processing required to activate a word’s meaning (e.g., Kandhadai & Federmeier, 2008). It has been found in evaluative priming tasks as well (i.e., more negative deflection in cases of evaluative incongruency, as compared with congruency), corroborating the claim of S–S-based processes in evaluative priming contexts (see Eder et al., 2012; Zhang, Lawson, Guo, & Jiang, 2006). Since these S–S-based processes reflect the suggested process 2 of the three-process model (i.e., mutual facilitation, with varying involvement of prime-to-target and target-to-prime facilitation, depending on the specific experimental parameters), larger N400 mean amplitudes in cases of evaluative incongruency, as compared with congruency, would additionally corroborate our theoretical assumptions.
Note, however, that the analyses of the P3 and N400 components can only be an aside to our main focus (i.e., exploring the N2), because it can be expected that both components overlap, which is typically the case whenever the experimental paradigm involves an explicit decision (see, e.g., Heil & Rolke, 2004; Kiefer & Martens, 2010; Kounios & Holcomb, 1994; Kutas & Van Petten, 1994). Therefore, unambiguously identifying both components may be more difficult in our experiment, as compared with experiments that are exclusively designed to find effects in either the P3 or the N400 component. In order to separate the analyses as much as possible, we focus on N400 mean amplitudes as a function of evaluative congruency, because the variation of evaluatively congruent versus incongruent conditions is orthogonal to (and therefore independent of) the binary decision categories. In addition, one might argue that N400 mean amplitudes are a function of semantic compatibility (see, e.g., Bermeitinger et al., 2010). We will report the N400 results with this factor as well.
We employed the same task, design, and materials as in Schmitz and Wentura (2012, Experiment 3). That is, we examined the moderation of semantic categorization conflicts by the evaluative congruency of target and flanker, using positively and negatively connoted exemplar names from the categories persons and animals as targets and flankers. We employed a 2 (target semantic category) × 2 (flanker semantic category) × 2 (target evaluative category) × 2 (flanker evaluative category) within-subjects design.
Since an evaluatively congruent target was assumed to maintain the flanker activation, a semantically incompatible flanker should strongly interfere with the target response if the flanker and target are evaluatively congruent, whereas a compatible flanker should facilitate the target response in this case. This should result in a significant semantic flanker effect given evaluative congruency, as compared with no (or a significantly smaller) flanker effect given evaluative incongruency. Such an interaction of the semantic and evaluative factors was hypothesized to emerge in mean RTs and N2 mean amplitudes. We expected longer mean RTs and larger N2 mean amplitudes in the case of semantic target–flanker incompatibility, as compared with compatibility, if the target and flanker were evaluatively congruent. Given evaluative incongruency, we did not expect differences in mean RTs and N2 mean amplitudes for semantically compatible and incompatible target–flanker pairs.
In the same vein, P3 peak latencies and the LRP were analyzed with regard to correlates of response conflict that is influenced by evaluative congruency. This should manifest itself in longer P3 peak latencies for semantically incompatible, as compared with compatible, target–flanker pairs if the target and flanker were evaluatively congruent. With regard to the LRP, we explored possible deflections in the direction of the wrong response for semantically incompatible flankers in the case of evaluative congruency.
Finally, we analyzed the N400 component, hypothesizing relatively more negative mean amplitudes in cases of evaluative incongruency, as compared with congruency.
Thirty participants (15 male, 15 female) completed the experiment. The median age was 25 years (range from 19 to 32 years). All participants were native German speakers and had normal or corrected-to-normal vision. None of them reported any neurological impairment. They gave written informed consent prior to their inclusion in the study and were paid 8 €/h for their participation. All participants were debriefed after the experiment.
We employed a 2 (target semantic category) × 2 (flanker semantic category) × 2 (target evaluative category) × 2 (flanker evaluative category) within-subjects design. With this design, we can assess semantic flanker effects (i.e., the RT difference between semantic compatible flanker–target pairs and semantic incompatible flanker–target pairs), given evaluative congruency and incongruency of the pairs, respectively.
As outlined, first, we expect a clear semantic flanker effect given evaluative congruency. Schmitz and Wentura (2012, Experiment 3) found an effect of dz = 0.75. Given a sample size of N = 30 and an α-value of .05, an effect of size dz = 0.75 could be detected with a probability of 1 − β = .99. Second, the semantic flanker effect should be reduced given evaluative incongruency. In the ideal case, this flanker effect is not significant and is significantly reduced, as compared with the evaluative congruency condition. With regard to the latter test (i.e., the difference of the two flanker effects), Schmitz and Wentura found an effect of dz = 0.38. This effect (given N and α as above) could be detected with a probability of 1 − β = .65. To account for this unsatisfactory outcome, we increased the trial number from 160 to 240 trials to increase the effect sizes. With a power of 1 − β = .80, we were able to detect an effect of dz = 0.46. Power calculations were done with GPower (Faul, Erdfelder, Lang, & Buchner, 2007).
All the stimuli were German words. We selected 15 positive and 15 negative exemplar names from the categories persons and animals (most of which were also used by De Houwer, Hermans, Rothermund, & Wentura, 2002, as well as Schmitz & Wentura, 2012). Mean valences on a scale from 1 (very negative) to 9 (very positive)—as rated by the participants after finishing the experiment—were Mpositivepersons = 7.57 (SDpositivepersons = 0.73), Mnegative persons = 2.69 (SDnegative persons = 1.11), Mpositive animals = 6.76 (SDpositive animals = 0.75), and Mnegative animals = 3.47 (SDnegative animals = 1.18). Ratings for positive and negative words differed significantly [t(14) = 24.43, p < .001, for person words, and t(14) = 17.45, p < .001, for animal words]. The sets of person and animal words did not differ with regard to mean valence, t(29) = 0.07, p = .95. Word length varied from four to eight letters and was matched as closely as possible for the four categories. Mean numbers of letters were Mpositivepersons = 6.27 (SDpositivepersons = 1.44), Mnegativepersons = 6.27 (SDnegativepersons = 1.44), Mpositiveanimals = 6.40 (SDpositiveanimals = 1.35), and Mnegativeanimals = 6.07 (SDnegativeanimals = 1.22). The complete list of experimental stimuli can be found in the Appendix.
Participants were individually tested in an electrically shielded and sound-attenuated chamber. The experiment was run using E-Prime (Psychology Software Tools, Version 2.0) with a standard PC and a 22-in. LCD monitor; viewing distance was ca. 0.8 m. The experimental procedure consisted of a practice phase, the flanker task, and the final valence rating. Instructions were given on the screen. For both the practice phase and the actual flanker task, participants were told to categorize every centrally presented word as quickly and as accurately as possible as a person or animal, using the “c” and “m” keys of a standard keyboard with their left and right index fingers, respectively. The assignment of semantic categories to keys was counterbalanced across participants.
In the practice phase, the sequence of each trial was as follows. A fixation cross (height/width: 1 cm) appeared centrally for a jittered time interval of 250, 500, or 750 ms. It was followed by a blank screen of 200 ms and a word (black 18-point Courier New font, written in capital letters), presented centrally until keypress. In the case of an incorrect response, written feedback (“Wrong! Continue with the correct key!”) was given on-screen. The next trial started after a blank screen for a jittered intertrial interval of 1,250, 1,500, or 1,750 ms. The practice phase comprised 60 trials in a random sequence (i.e., each word was displayed once). If more than 15 errors (i.e., 25 % of the trials) occurred, the whole phase was repeated.
We created four conditions by crossing the factors semantic compatibility and evaluative congruency with regard to the target–flanker relation. Each of the 60 words was presented in each of these four conditions once as a target and once as a flanker, so the whole task comprised 240 trials. The trial sequence was randomized, with the constraint that the target and flanker were never the same word on any given trial and neither the target nor the flanker was repeated on immediately successive trials. Sixteen warm-up trials (i.e., 4 trials per condition) preceded the experimental trials, and there were self-paced breaks after every 60 experimental trials.
EEG recording and analyses
The electroencephalogram (EEG) was recorded continuously from 60 active Ag/AgCl scalp electrodes, mounted in a preconfigured elastic cap (Brain Products, Munich) and labeled according to the extended 10–20 system (see Sharbrough et al., 1991). Vertical eye movements were monitored using an electrode placed below the left eye and the Fp1 electrode (located above the left eye). Electrodes placed at the outer canthi of the eyes measured horizontal eye movements. Recordings and offline data processing were performed using BrainVision Recorder and BrainVision Analyzer 2.0.1 software, respectively (Brain Products, Munich). All channels were amplified with BrainAmp DC amplifiers. EEG signals were sampled at a rate of 500 Hz and online band-pass filtered (0.1–250 Hz). Signals were referenced online to the left-mastoid electrode and were rereferenced offline to averaged mastoids. Impedances for all electrodes were kept below 20 kΩ. Data were filtered offline with a low-pass filter of 40 Hz (slope 24 dB), and eye movements were corrected using the independent-component analysis. ERPs were calculated from individual epochs of 1,700 ms (including a baseline of 200 ms before target onset); epochs still containing artifacts in any EEG channel were rejected2 (in total, 3.33 % of trials were rejected due to artifacts). Data were baseline corrected using the 200 ms before target onset. For each participant, ERPs were averaged separately for each condition.
On the basis of previous research and visual inspection, P3 peak latencies were analyzed at the Pz electrode. Since peak latencies largely varied between participants and conditions, we chose a wide peak-detection search window from 380 to 680 ms after target onset.
Unless otherwise noted, all effects referred to as statistically significant throughout the article are associated with p-values of <.05, two-tailed.
Mean RTs were derived from correct responses. The mean error rate was 3.7 %. Furthermore, trials with RTs that were 1.5 interquartile ranges below the first or above the third quartile with respect to the individual distribution (see Tukey, 1977), were below 200 ms, or were above 1,500 ms were discarded (1.7 % of all trials).
Mean response times (RTs; in milliseconds) as a function of semantic condition and evaluative condition (errors in percentages in parentheses), with flanker effects for RTs (in milliseconds; standard errors in brackets)
An analogous ANOVA on error rates yielded a significant main effect of semantic condition, F(1, 29) = 5.51, p < .05, η2 = .16, which corresponded to a positive flanker effect of M = 1.7 % (SD = 3.9 %), t(29) = 2.35, p < .05, dz = 0.43. Neither the main effect of evaluative condition, F(1, 29) = 1.75, p = .20, η2 = .06, nor the interaction, F < 1, reached significance.
After the exclusion of trials due to artifact rejection and application of the same exclusion criteria as for the behavioral data, the mean numbers of included trials were M = 54 (SD = 4) in the semantically compatible and evaluatively congruent condition, M = 53 (SD = 3) in the compatible and incongruent condition, M = 53 (SD = 4) in the incompatible and congruent condition, and M = 51 (SD = 3) in the incompatible and incongruent condition.
N2 mean amplitudes
Analogous to the behavioral data, the analysis of N2 mean amplitudes focused on the interaction in the semantic condition (compatible vs. incompatible) × evaluative condition (congruent vs. incongruent) design. Furthermore, the topographic information was included by the within-subjects factors of caudality (anterior, central, posterior) and laterality (left, middle, right). For data inspection, we calculated for each region of interest the decisive interaction term: flanker difference (i.e., RTincompatible − RTcompatible) for evaluatively congruent trials minus flanker difference for evaluatively incongruent trials. There were 2 participants who had outlying values (one at each tail of the distributions) on most of these terms. Since parametric analyses are heavily burdened by outliers, these 2 participants were excluded for parametric analyses. However, to corroborate results, we added nonparametric analyses including the outliers wherever possible. That is, whenever the corresponding F-value of the parametric analyses has only one numerator df, a Wilcoxon signed-rank test can be calculated. Given conditions of nonnormality, the Wilcoxon test had better power characteristics than did the parametric counterparts (see Van den Brink & Van den Brink, 1989). To keep the exposition concise, in the following report, only ERP results referring to the factors of semantic condition and/or evaluative condition are reported.
Most important, a caudality (anterior, central, posterior) × laterality (left, middle, right) × semantic condition (compatible vs. incompatible) × evaluative condition (congruent vs. incongruent) repeated measures multivariate analysis of variance (MANOVA) yielded a significant two-way semantic condition × evaluative condition interaction, F(1, 27) = 5.73, p < .05, η2 = .18 (z = −1.99, p < .05, for the corresponding Wilcoxon signed-rank test), which was further qualified by laterality, F(2, 26) = 4.24, p < .05, η2 = .25. All remaining effects (referring to the factors semantic condition and/or evaluative condition) failed the conventional level of significance, all Fs < 2.16, all ps > .10.
Flanker effects in N2 mean amplitudes (in μV) of the grand-average ERPs at averaged left, middle, and right regions of interest within the time interval 300–380 ms after target onset as a function of evaluative condition (standard errors in parentheses)
For evaluatively incongruent conditions, a laterality (left, middle, right) × semantic condition (compatible vs. incompatible) MANOVA yielded a clearly insignificant flanker effect, F(1, 27) < 1 (z = −0.30, n.s.), as well as an insignificant interaction of the flanker effect with laterality, F(2, 26) = 1.73, p > .19, η2 = .12. For the sake of completeness, this null result held as well for the electrodes FCz, M = 0.24 μV (SD = 1.92 μV), z = .01, n.s., and Cz, M = 0.11 μV (SD = 2.11 μV), z = 0.24, n.s.
P3 peak latencies
P3 peak latencies (in milliseconds) of the grand-average ERPs at the Pz electrode within the time interval 380–680 ms after target onset as a function of semantic condition and evaluative condition (standard deviations in parentheses), with flanker effects (in milliseconds; standard errors in brackets)
The interaction reached significance, z = −1.64, p = .05 (one-tailed), while both main effects were nonsignificant (z = −0.69, p = .49, for the main effect of semantic condition, and z = −1.66, p = .10, for the main effect of evaluative condition). In simple-effect tests, the flanker effect was not significant, given neither evaluative congruency, M = 10 ms (SD = 63 ms), z = −0.98, p = .33, φ = 0.18, nor incongruency, M = −18 ms (SD = 60 ms), z = −1.86, p = .06, φ = 0.34. In anticipation of the discussion, we would like to emphasize that the P3 peak latencies in semantically incompatible conditions were longer in cases of evaluative congruency than of incongruency, resulting in a significant difference of M = 27 ms (SD = 74 ms), z = 2.10, p < .05, φ = 0.38. In contrast, P3 peak latencies in compatible conditions did not differ significantly between the evaluatively congruent and incongruent conditions, M = 1 ms (SD = 64 ms), z = −0.22, p = .83, φ = 0.04.
As is clearly visible in Fig. 3a, there was no observable LRP deflection into the range of positive values for the semantically incompatible condition in cases of evaluative congruency (blue line). Thus, there seems to be no preactivation of the wrong (i.e., the flanker-associated) response.
However, some authors recommend separately analyzing trials with fast and slow responses (see, e.g., Eder et al., 2012; Miller, 1998). The logic behind this separation is that the exclusion of trials with incorrect responses may complicate the search for incorrect response preparation by response-incompatible flankers, since the activation of the incorrect response was presumably strongest in these excluded trials, leading to fast incorrect responses (see Gratton et al., 1988). Thus, fast correct trials are possibly trials with only weak initial activation of incorrect responses. The LRP pattern of slow correct responses, however, might better capture the initially incorrect response tendency that is finally overcome by the target-related response.
Mean amplitudes in grand-average lateralized readiness potentials (in μV) for slow responses within the time interval 340–390 ms after target onset as a function of semantic condition and evaluative condition (standard deviations in parentheses), with flanker effects (in μV; standard errors in brackets)
N400 mean amplitudes
To test the hypothesis of an evaluative congruency effect, we conducted a caudality (anterior, central, posterior) × laterality (left, middle, right) × semantic condition (compatible vs. incompatible) × evaluative condition (congruent vs. incongruent) repeated measures MANOVA on N400 mean amplitudes. Again, we report only ERP results referring to the factors semantic condition and/or evaluative condition. Most important, we found a main effect of evaluative congruency, F(1, 29) = 4.60, p < .05, η2 = .137 (z = −2.19, p < .05), with relatively more negative amplitudes given evaluative incongruency, as compared with congruency. It was not significantly moderated by semantic condition, F(1, 29) = 1.09, n.s., η2 = .036 (z = −1.55, p = .12), caudality, F(2, 28) = 2.14, p = .14, η2 = .132, or laterality, F(2, 28) = 1.78, p = .19, η2 = .113 [F < 1.34, n.s., for all higher interaction terms involving evaluative condition except for the three-way interaction of laterality, semantic condition, and evaluative condition, F(2, 28) = 3.02, p = .07, η2 = .1786].
In analogy to the factor evaluative condition, we report the analyses for the factor semantic condition, since we expected a more negative-going wave in case of semantic incompatibility, as compared with compatibility. While the main effect of semantic condition failed to be significant, F(1, 29) = 3.41, p = .08, η2 = .105 (z = −1.59, p = .11), the caudality × laterality × semantic condition interaction just missed the conventional criterion of significance, F(4, 26) = 2.64, p = .06, η2 = .289. Numerically, the main effect of semantic condition had the same sign for all nine regions of interest and corresponded to more negative amplitudes given semantic incompatibility, as compared with compatibility. At midline central and posterior regions of interest combined, the main effect of semantic condition was significant, t(29) = −2.10, p < .05, dz = 0.38 (z = −2.09, p < .05), whereas it was not at the remaining regions of interest combined, t(29) = −1.72, p = .10, dz = 0.31 (z = −1.55, p = .12); t(29) = 2.08, p < .05, dz = 0.55 for the difference (z = 2.58, p < .01). In addition, it should be noted that a significant semantic compatibility effect could be observed at the single electrodes CPz (M = −0.73 μV, SD = 1.92 μV), z = −2.01, p < .05, Pz (M = −0.77 μV, SD = 1.89 μV), z = −2.71, p < .01, and POz (M = −0.60 μV, SD = 1.74 μV), z = −2.07, p < .05, which have been reported in other priming studies exploring the N400 (Eder et al., 2012; Kiefer, Weisbrod, Kern, Maier, & Spitzer, 1998).
All remaining interactions involving the factor semantic condition were insignificant, Fs < 1.87, n.s., while the three-way interaction of laterality, semantic condition, and evaluative condition (as reported above) just missed the conventional level of significance.
The present experiment investigated the assumption that in a semantic variant of the flanker task with evaluatively connoted target and flanker words, the semantic flanker effect is moderated by the evaluative congruency of target and flanker. In mean RTs, we found a significantly positive flanker effect in cases of evaluative congruency but no effect in cases of incongruency. This pattern of flanker effects in RTs replicated the findings by Schmitz and Wentura (2012, Experiment 3) and corroborated the idea that an evaluatively congruent target may maintain the activation of a response-irrelevant stimulus (i.e., flanker or prime), as postulated in the three-process model of evaluative priming.
Most important, we observed a clear result for N2 mean amplitudes. Again, the flanker effect depended on evaluative congruency: There were larger N2 mean amplitudes given semantic incompatibility, as compared with compatibility, if the flanker and target were evaluatively congruent. In contrast, if the flanker and target were evaluatively incongruent, the flanker effect clearly did not differ from zero. Since the N2 component has typically been associated with response conflict (see Folstein & Van Petten, 2008; Van Veen & Carter, 2002; but see below), the significant influence of evaluative congruency on the N2 flanker effect corroborates the assumption that the flanker’s potential to interfere with the target response depended on its evaluative congruency with the target: If an evaluatively congruent target helps maintain the activation of the flanker, the flanker may facilitate the target response or interfere with it.
We should briefly mention that Wendt and colleagues (2007) questioned the close correspondence of the N2 component and response conflict. In a flanker design, they used stimulus-compatible (SC; i.e., the target and flanker were identical), stimulus-incompatible but response-compatible (SIC; i.e., the target and flanker were different stimuli that were assigned to the same response), and response-incompatible (RIC; i.e., the target and flanker were different stimuli that were assigned to different responses) conditions. The authors found larger N2 mean amplitudes for the RIC and SIC conditions, in comparison with the SC condition, and concluded that the N2 flanker effect is sensitive to response conflict, as well as to stimulus incompatibility (but see Van Veen & Carter, 2002; Van Veen, Cohen, Botvinick, Stenger, & Carter, 2001). In the present experiment, the target and flanker from a single trial were always different stimuli; that is, the relation of both stimuli varied with regard to response-relevant categories, but not with regard to stimulus identity. Thus, we can safely conclude that our result reflects conflicts that are related to a response-relevant categorization process.
As was discussed above, distributed memory models would offer the alternative explanation that any RT delay associated with an evaluatively congruent and semantically incompatible flanker arises not from response conflict between concurrently activated responses, but from a temporary dilution of the response part of the target’s activation pattern, which occurs when a response-incongruent flanker representation—whose response part does not overlap with the target pattern—is active. However, the conditional flanker effect in N2 mean amplitudes indicates an involvement of response conflict and, hence, supports the notion of the three-process model that evaluatively congruent target and flanker representations are simultaneously activated and interact with regard to response execution, while an evaluatively incongruent flanker is not sufficiently activated to influence the target response. The interpretation of our results in terms of response conflict is somewhat corroborated by the P3 and LRP results, although these findings should be taken with some caution, since the effects in these two components are not very strong.
In P3 peak latencies, the interaction of evaluative congruency and semantic compatibility was significant, but only in a one-tailed test. Both single flanker effects failed to be significant, as well. Interestingly, P3 peak latencies clearly differed between evaluatively congruent and incongruent conditions given semantic incompatibility, but not given compatibility (see Table 3). Since semantic incompatibility (in contrast to compatibility) is associated with response conflict and, thereby, with larger effort needed for target categorization, the kind of evaluative overlap of target and flanker may moderate the solution of the response conflict: While evaluative congruency may increase response conflict, incongruency may help resolve it. In contrast, since the target and flanker elicit the same categorization response if they are semantically compatible, target categorization is neither facilitated by evaluative congruency nor hampered by incongruency. Of course, due to its post hoc character, this interpretation needs further examination. Bartholow (2010; see also Bartholow & Dickter, 2008) argued that P3 peak latencies reflect categorization effort independently of response processes. Thus, the weak effects in P3 peak latencies may indicate that the experimental manipulation in the present study mainly influenced processes at the response level.
For the LRP, we found some tentative evidence for the activation of the wrong response by semantically incompatible flankers in cases of evaluative congruency, but not in cases of incongruency. However, possibly due to our specific presentation parameters, the pattern was not as clear-cut as it would be in the ideal case. Restricted to slow responses (for this rationale, see, e.g., Eder et al., 2012; Miller, 1998), there was a positive deflection in the LRP (i.e., a tendency toward the wrong response) for semantic incompatibility given evaluative congruency that significantly deviated (in a sign test) from the corresponding LRP for semantic compatibility given evaluative congruency (i.e., there was a flanker effect with regard to the activation of a flanker-related response). This effect was clearly missing in the evaluatively incongruent condition. However, since the mean of the positive dip missed the criterion of significance and the interaction in the semantic compatibility × evaluative congruency analysis clearly failed to be significant, we should take this evidence with some caution.
Finally, we found evidence for a moderation of the N400 mean amplitudes by evaluative congruency. Note that—besides the special design to test for conditioned flanker effects—our experiment had all the ingredients of a semantic priming experiment, which is the usual paradigm to establish N400 effects, albeit with some unusual features: There is a target stimulus accompanied by a prime stimulus (here, the flanker) that is either related (here, evaluatively congruent) or unrelated (here, evaluatively incongruent). The task is a binary decision task with response categories that are orthogonal to the relatedness factor (like a lexical decision task, which is the typical task in semantic priming research). Thus, finding a N400 effect in evaluative priming tasks with positive as well as negative SOAs perfectly matches our claim that evaluatively congruent stimuli mutually facilitate each other. As an aside, we found some evidence for larger N400 mean amplitudes in cases of semantic incompatibility, as compared with compatibility. The variation of N400 mean amplitudes with regard to semantic compatibility in a semantic categorization task replicates comparable findings for evaluative congruency in the evaluation task (see Eder et al., 2012; Zhang et al., 2006).
Our results corroborate the assumption that the evaluative features of semantic concepts are processed without any task requirement and provide further evidence for a privileged processing of the evaluative connotations (see, e.g., Bargh et al., 1996; Murphy & Zajonc, 1993). Schmitz and Wentura (2012) could even show that a similarly preferred processing is not true for the semantic categories persons and animals. In the last section, we will consider the constraints for a model of the memory representations of the evaluative connotations. Such a model must allow for a significant influence of evaluative congruency on nonevaluative categorization processes and should provide an interpretation of S–S-based evaluative priming effects.
Memory representation of evaluative connotations
Our conditional flanker effects are, just like S–S-based evaluative priming effects, explainable with the idea that evaluative congruency between semantic concepts is accompanied by the parallel activation and mutual facilitation of these concepts. This raises the question of how evaluatively congruent concepts may interact. As was described above, the associative network model by Bower (1991) suggests that the activation of an evaluatively connoted concept spreads to all evaluatively congruent memory representations, such that in a sequential evaluative priming task, the encoding of the target is facilitated by a preceding, evaluatively congruent prime (see, e.g., Spruyt et al., 2002). As was discussed above, the fanning effect, however, speaks against the plausibility of spreading activation within an associative network as the process underlying S–S-based evaluative priming effects (see J. R. Anderson, 1974).
Parallel distributed models (e.g., Masson, 1991, 1995; McRae et al., 1997) provide a reasonable alternative mechanism for facilitated activation of evaluatively congruent concepts: Transition from a prime to a target concept is facilitated when the evaluative part of the target pattern is already preactivated. However, with regard to the response explanation of evaluative priming effects in the evaluation task and the explanation of the present results, as well as those by Schmitz and Wentura (2012), an additional mechanism is necessary that allows for the parallel activation of prime and target responses. Not only does our model propose such a mechanism, there is also a biologically plausible mechanism of parallel activation that is relatively well-understood.
In the neuroscience literature, parallel activation of representations in both perception and memory has been linked to phase synchronization of neural firing (see Gray, 1994; Klimesch, 1996; e.g., Raffone & Van Leeuwen, 2001; Raffone & Wolters, 2001; Vogel, Woodman, & Luck, 2001; Wolters & Raffone, 2008). Synchronous firing means that activity of all neural populations—which can be thought to represent activation units in a cognitive model—that participate in the representation of an object or concept in perception or memory oscillates synchronously. This mechanism enables the complete and unambiguous activation of a specific semantic concept, as well as the concurrent parallel activation of different semantic concepts without confusion, due to different oscillation frequencies. That is, a specific activation unit can be activated by different concept patterns at the same time by participating in distinct oscillations (see Raffone & Van Leeuwen, 2003). What, however, is missing in these accounts is an implementation of mutual facilitation of evaluatively congruent concepts, as postulated by the three-process model.
Criteria for the inclusion of trials were the following: maximum amplitude in the recording epoch of ±200 μV, maximum difference between two successive sampling points of 50 μV, maximum difference of 150 μV in successive intervals of 200 ms, and lowest allowed activity change of 0.5 μV in successive intervals of 100 ms.
The activity in each of the nine regions was computed by averaging the activities at four different electrodes: anterior-left, F3, F5, FC3, FC5; anterior-mid, Fz, FCz, FC1, FC2; anterior-right, F4, F6, FC4, FC6; central-left, C3, C5, CP3, CP5; central-mid, Cz, CPz, CP1, CP2; central-right, C4, C6, CP4, CP6; posterior-left, P3, P5, PO7, PO9; posterior-mid, Pz, POz, PO3, PO4; posterior-right, P4, P6, PO8, PO10.
The flanker effect given evaluative incongruency was associated not only with a lower mean, but also with a largely increased standard deviation (32 vs. 18 ms for evaluative congruency). This has made the interaction test quite powerless, since the difference variable of the two flanker effects (which is tested against zero in the interaction test) becomes very noisy. Note that testing the flanker effect given evaluative congruency (M = 20 ms) against a random variable with M = 10 ms (i.e., the mean of the flanker effect given evaluative incongruency) and SD = 18 ms (i.e., the SD of the flanker effect given evaluative congruency) yielded t-values between t = 1.97 and t = 2.66 (median = 2.14), associated with p-values between p = .013 and p = .058 (median = 0.042) in 10 drawings of the random variable.
The reader might note that for the fast responses (see Fig. 3b), there was an unexpected positive dip in the semantically compatible condition, given evaluative congruency. On the basis of visual inspection, we determined the LRP mean amplitude within the time interval 190–240 ms after target onset for all conditions. Neither the positive dip of M = 0.17 (SD = 1.15), t(29) = 0.79, p = .43, nor the corresponding flanker effect (i.e., given evaluative congruency), t(29) = 1.47, p = .15, dz = 0.27, nor the interaction in a semantic condition × evaluative condition repeated measures ANOVA, F(1, 29) = 1.19, p = .29, η2 = .04, was significant.
Since there was no indication of a semantic condition ×x evaluative condition interaction, F(1, 29) = 1.25, p = .27, η2 = .041, at midline electrodes, at which evaluative congruency effects in N400 mean amplitudes have previously been reported (see, e.g., Eder et al., 2012), we will not further elaborate on this three-way interaction.
This research was funded by the International Research Training Group (IRTG) "Adaptive Minds" of the Deutsche Forschungsgesellschaft (DFG).
The authors thank Dr. Ullrich Ecker for helpful comments on an earlier version of the manuscript.
- Bermeitinger, C., Frings, C., & Wentura, D. (2010). N400 effects for cateogy exemplars primed by category labels. Scientific Report, PsyDok, 1–10.Google Scholar
- Bower, G. H. (1991). Mood congruity of social judgments. In J. P. Forgas (Ed.), Emotion and social judgments. International series in experimental social psychology (pp. 31–53). Elmsford: Pergamon Press.Google Scholar
- Brown, C., & Hagoort, P. (2000). On the electrophysiology of language comprehension: Implications for the human language system. In M. W. Crocker, M. Pickering, & C. Clifton (Eds.), Architectures and mechanisms for language processing (pp. 213–237). New York: Cambridge University Press.Google Scholar
- Carrillo-de-la-Peña, M. T., Lastra-Barreira, C., & Galdo-Álvarez, S. (2006). Limb (hand vs. foot) and response conflict have similar effects on event-related potentials (ERPs) recorded during motor imagery and overt execution. European Journal of Neuroscience, 24, 635–643.Google Scholar
- Klauer, K. C., & Musch, J. (2003). Affective priming: Findings and theories. In J. Musch & K. C. Klauer (Eds.), The psychology of evaluation: Affective processes in cognition and emotion (pp. 7–49). Mahwah: Lawrence Erlbaum Associates Publishers.Google Scholar
- Kutas, M., & Van Petten, C. K. (1994). Psycholinguistics electrified: Event-related brain potential investigations. In M. A. Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 83–143). San Diego: Academic Press.Google Scholar
- Masson, M. E. J. (1991). A distributed memory model of context effects in word identification. In D. Besner & G. W. Humphreys (Eds.), Basic processes in reading: Visual word recognition (pp. 233–263). Hillsdale: Lawrence Erlbaum Associates.Google Scholar
- Masson, M. E. J. (1995). A distributed memory model of semantic priming. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 3–23.Google Scholar
- Osterhout, L., & Holcomb, P. J. (1995). Event related potentials and language comprehension. In M. D. Rugg & M. G. H. Coles (Eds.), Electrophysiology of mind: Event-related brain potentials and cognition (Oxford psychology series, Vol. 25, pp. 171–215). New York: Oxford University Press.Google Scholar
- Pritchard, W. S., Shappell, S. A., & Brandt, M. E. (1991). Psychophysiology of N200/N400: A review and classification scheme. In J. R. Jennings, P. K. Ackles, & M. G. H. Coles (Eds.), Advances in psychophysiology: A research annual (Vol. 4, pp. 43–106). London England: Jessica Kingsley Publishers.Google Scholar
- Tukey, J. W. (1977). Exploratory data analysis. Reading: Addison-Wesley.Google Scholar
- Wentura, D., & Rothermund, K. (2003). The “meddling-in” of affective information: A general model of automatic evaluation effects. In J. Musch & K. C. Klauer (Eds.), The psychology of evaluation: Affective processes in cognition and emotion (pp. 51–86). Mahwah: Lawrence Erlbaum Associates.Google Scholar