Single versus Dual Process Models of Lexical Decision Performance: Insights from RT Distributional Analysis

This paper evaluates two competing models that address the decision-making processes mediating word recognition and lexical decision performance: a hybrid twostage model of lexical decision performance and a random-walk model. In two experiments, nonword type and word frequency were manipulated across two contrasts (pseudohomophone-legal nonword and legal-illegal nonword). When nonwords became more wordlike (i.e., BRNTA vs. BRANT vs. BRANE), response latencies to the nonwords were slowed and the word-frequency effect increased. More importantly, distributional analyses revealed that the nonword type x word frequency interaction was modulated by different components of the RT distribution, depending on the specific nonword contrast. A single-process random-walk model was able to account for this particular set of findings more successfully than the hybrid two-stage model. Single vs. dual process models 3 Single versus Dual Process Models of Lexical Decision Performance: Insights from RT Distributional Analysis The study of the processes underlying isolated visual word recognition is a major endeavor in experimental psychology, and has provided insights in domains as diverse as psycholinguistics, pattern recognition, computational modeling, attention, and neuroscience. Although many procedures have been developed for studying word recognition, the speeded lexical decision task (LDT) (Rubenstein, Garfield, & Millikan, 1970) remains one of the most widely used tasks (e.g., Murray & Forster, 2004; Ratcliff, Gomez & McKoon, 2004). In this task, participants are presented with a letter string and are required to decide as quickly as possible whether the string forms a word or nonword, most typically with a keypress response. Findings obtained in the LDT have been very influential in informing models of word recognition (e.g., Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001). A number of models have been proposed to accommodate lexical decision performance (see Ratcliff et al., 2004, for a recent review). For example, the classic logogen model (Morton, 1969) posits word detectors (logogens) for every lexical entry. When a word (e.g., DOG) is presented, the logogen for DOG accumulates evidence until some threshold is reached, and word identification takes place. The original logogen model could not handle nonwords, but recent extensions to the model can carry out lexical decision. For example, Grainger and Jacob’s (1996) multiple read-out model (MROM) implements three processes that drive a lexical decision response. Word responses are produced when either the activation level of a single lexical representation (local activity) or the summed activation levels of all lexical representations (global Single vs. dual process models 4 activity) exceed their respective thresholds. Nonword responses are produced when lexical activity does not reach threshold after some (variable) time deadline. The dualroute cascaded (DRC) model (Coltheart et al., 2001) adopts essentially the same principles to accommodate performance in the LDT. Importantly, for these two examples, the speed and accuracy of lexical decision responses are yoked to the activity of the word representations contained in the models’ lexicon. Lexical decisions have also been instantiated in parallel distributed processing (PDP) reading models, which contain distributed orthographic and semantic representations (Plaut, 1997). In the latter framework, words or nonwords that are presented to the model generate varying degrees of stress values, which reflect how semantically familiar they are. A decision criterion can then be adopted that allows the model to discriminate between words (stress values higher than criterion) and nonwords (stress values lower than criterion). This approach assumes that lexical decision performance is driven by the activity of (distributed) representations. In this article, we will focus on two different approaches to lexical decision that emphasize the decision processes tied to lexical decision. Two models which specifically address the decision-making processes that mediate word identification and behavioral responses are the two-stage model of lexical decision performance (Balota & Chumbley, 1984; Balota & Spieler, 1999), and the diffusion model (Ratcliff et al., 2004). It is important to note that both frameworks have been used extensively in accommodating data in binary decision tasks, ranging from memory scanning (Atkinson & Juola, 1974) to episodic memory retrieval (Ratcliff, 1978). Moreover, both the two-stage model and the diffusion model have been shown to successfully accommodate basic lexical decision Single vs. dual process models 5 phenomena. Interestingly, though, these two models are built on very different premises. The diffusion model assumes that a single process can drive lexical decision, whereas the two-stage model suggests that there are two qualitatively distinct processes. Whether lexical decision is better accommodated by a single or a dual process model is another instance of a broader distinction across a wide variety of domains (see, for example, Yonelinas, 2002). Just as the debate between proponents of these two theoretical approaches has ramifications beyond psycholinguistics, the answer to the proposed question will give us greater leverage in understanding how binary decisions are carried out in general. In this paper, we will be exploring this issue systematically. We will begin by examining some interesting constraints in lexical decision performance, then describe the two classes of models, and finally quantitatively evaluate which approach accommodates the results from two experiments more successfully. Interaction between Nonword Type and Word Frequency A critical variable that has been investigated in lexical decision performance is the similarity of the nonwords to real words. Nonwords can be pronounceable and orthographically legal (e.g., FLIRP), unpronounceable and orthographically illegal (e.g., RPFLI), or homophonous with real words (e.g., BRANE). As one might expect, nonword type powerfully modulates lexical decision latencies for word trials, and also produces interactive effects with other variables which influence lexical decision performance. For example, the word-frequency effect (faster lexical decision latencies for frequently encountered words) is strongly modulated by the type of nonword context. Stone and Van Orden (1993) systematically manipulated nonword type and word frequency in lexical decision, and observed the pattern presented in Table 1. As Single vs. dual process models 6 nonwords become more similar to words, two trends are apparent. Lexical decision word latencies become slower, and more intriguingly, the word-frequency effect becomes larger. Stone and Van Orden interpreted these results as consistent with both a pathway selection framework and a random-walk framework. The pathway selection framework proposes that the lexical processing system consists of independent processing modules which are interconnected by pathways. Manipulating the nature of the nonwords alters the task context, and the system strategically selects the pathways that optimize task performance. More relevantly for this paper, the results were also accommodated within a random-walk framework. The random-walk model has been useful for describing various aspects of binary decisions (Ratcliff & Rouder, 1998) and is a member of a more general class of sequential-sampling models. The random-walk perspective conceptualizes lexical decision as an evidence-accumulating process. When a stimulus is presented, noisy information is accumulated over time towards one of two possible decision boundaries, word or nonword in the case of LDT (see Figure 1). A word response is produced when the accumulation process reaches the word boundary; a nonword response is produced when the accumulation process reaches the nonword boundary. For the simplest random-walk model, two parameters are of interest: the signal strength and the response criterion. Signal strength refers to the rate of evidence accumulation, and is greater for stimuli which are processed more efficiently (e.g., high frequency words). The response criterion refers to the distance of the boundaries from the start point; increasing the response boundaries reflects more conservative response criteria. Using this simple random-walk model, Stone and Van Orden (1993) argued that there is a linearly decreasing concave function between signal strength and the amount of Single vs. dual process models 7 time needed to reach criterion (Figure 2), that is, the same change in signal strength has a greater impact on response times when signal strength is lower compared to when signal strength is higher. The interaction between nonword type and word frequency is predicted by this function. Low frequency words have lower signal strengths than high frequency words. When nonwords become more wordlike (e.g., from BRONE to BRANE), word-nonword discrimination becomes more difficult. The signal strengths of both low and high frequency words decrease, leading to longer decision times. Importantly, because of the concave function, word-frequency effects are larger in the pseudohomophone condition than in the legal nonword condition, mimicking the pattern presented in Table 1. Putatively, this account also explains why word-frequency effects are larger in the legal nonword condition than in the illegal nonword condition. Importantly, though, Stone and Van Orden’s data were examined at the level of the mean, and there was no explicit implementation of this model. Hence, it was a descriptive account of the pattern observed in the means. As shown below, analyzing the same data at the level of distributional characteristics may yield further insights that are neither apparent nor intuitive. An alternative account of the nonword type by word frequency interaction was provided by Balota and Chumbley’s (1984) two-process model, which is based on Atkinson and Juola’s two-stage model of memory search. The application of this framework to lexical decision performance is displayed in Figure 2. This model was originally advanced as an account of task-specific effects in lexical decision (Balota & Chumbley, 1984), and was motivated by the observation that frequency effects are different in size across lexical decision, naming, and category classification, three tasks Single vs. dual process models 8 that presumably tap the same word identification process. Balota and Chumbley found that frequency effects were largest in lexical decision, and argued that this pattern was likely due to the fact that the frequency effect reflects both word identification processes and the word/nonword discrimination process that is specific to that task. Balota and Chumbley suggested that words and nonwords could be conceived as reflecting two underlying distributions that vary along a familiarity/meaningfulness (FM) dimension. Participants can use two types of information to make lexical decisions. The first is a relatively fast-acting familiarity based signal and the second is a slower more attention-demanding response, which may involve explicitly checking the spelling of the stimulus. Low frequency words are particularly sensitive to variables that modulate the checking process since low frequency words are more likely to overlap with the nonwords on the FM dimension. Hence, as one increases the overlap between the two distributions by making the nonwords more similar to the words, this further increases the checking process for the low frequency words, thereby slowing these items. Therefore, greater checking will occur for low frequency words when these items are embedded in lists with pseudohomophones, compared to legal nonwords. Moreover, the smallest amount of checking will occur for low frequency words when these items are embedded in lists with illegal nonwords, since the nonword distribution will overlap very little with the word distribution. Hence, the two-stage model also accommodates the Stone and Van Orden lexicality by word-frequency interaction by assuming two distinct processes instead of a single random-walk process. In addition, the framework was able to qualitatively account for various lexical decision effects (e.g., blocking effects and repetition effects) that were troublesome for certain extant word recognition theories. Single vs. dual process models 9 Beyond measures of central tendency In standard chronometric studies, a set of response times (RTs) for a particular experimental condition is collected for each participant. Typically, the mean of those response times (MRT) is then computed, with MRT providing an estimate of the central tendency for that condition. Of course, it is possible that variables do not simply shift the RT distribution, as implicitly assumed by analyses based on means; variables may also change the shape of the distribution. Hence, when possible, it is also useful to also investigate the influence of a variable on the shape (e.g., variance, skew) of a distribution (Heathcote, Popiel, & Mewhort, 1991). For example, fitting the ex-Gaussian function to data (Hohle, 1965; Luce, 1986; Ratcliff, 1979) allows researchers to estimate how different variables shift, skew, or shift and skew RT distributions. Ex-Gaussian analysis characterizes an RT distribution by assuming an explicit model for the shape of the distribution. This model is a convolution of the normal (gaussian) and exponential distributions, and has three parameters: μ, the mean of the normal distribution; σ 2 , the variance of the normal distribution; and τ, a reflection of the mean and standard deviation of the exponential distribution. In addition to providing unusually good fits to positively skewed empirical RT distributions (Luce, 1986, p. 439), one useful consequence of exGaussian analysis is that the algebraic sum of μ and τ is approximately equivalent to the mean when one estimates parameters from empirical data (μ and τ are exactly equal to mean in the theoretical ex-Gaussian model). Briefly, this property allows differences in means to be conveniently partitioned into two components: a component which is associated with distributional shifting (μ) and a component which is associated with Single vs. dual process models 10 distributional skewing (τ). There are at least two other reasons why such a distributional analysis might be valuable. First, Heathcote et al. (1991) pointed out that analyzing mean response times can often be inadequate and misleading because such an analysis does not consider the shape of the RT distributions. For example, they examined Stroop color-naming performance with both traditional and ex-Gaussian analyses. Based on mean response latencies, there was no difference between the congruent (RED displayed in red) and baseline (XXX displayed in red) conditions. This suggests that congruency has no effect on color naming, relative to the baseline. However, ex-Gaussian analyses revealed that naming response times in the congruent condition were facilitated (faster than baseline) in μ, but inhibited (slower than baseline) in τ. In this instance, congruency shifted the RT distribution leftwards while increasing its skew. These two effects cancelled each other out, spuriously producing null effects of congruency (see Spieler, Balota, & Faust, 1996, for a replication of this tradeoff). Second, by exploiting more of the information available in a RT distribution, one can make increasingly sophisticated predictions about how a variable might modulate the shape of a distribution, rather than just asking whether a variable has an effect in mean response times. This is useful when one is trying to adjudicate between two models. Models may be indistinguishable at the level of the mean, but make different predictions at the level of the RT distribution (see Hockley, 1984; Mewhort, Braun, & Heathcote, 1992). The two experiments reported in this paper will be an extension of Stone and Van Orden’s (1993) Experiment 1, with nonword type (legal nonwords, illegal nonwords, & Single vs. dual process models 11 pseudohomophones) and word frequency (high & low) factorially manipulated in a lexical decision task. In order to obtain a sufficient number of observations to provide adequate estimates of RT distributional characteristics, we will collect 100 observations for each of the cells for each participant. Experiment 1 will examine the contrast between legal nonwords (FLIRP) and pseudohomophones (BRANE), while Experiment 2 will examine the contrast between illegal nonwords (RPFLI) and legal nonwords. Importantly, we will examine the joint effects of the two variables on RT distributional properties, using both ex-Gaussian analysis and a non-parametric technique called vincentizing, described in the Results section of Experiment 1. Following the empirical section, we will then describe and implement the two modeling frameworks, and then test which framework better accommodates the observed effects, both at the level of the mean and at the level of distributional characteristics. Experiment 1