Non-Decision Time Effects in the Lexical Decision Task - eScholarship

Non-Decision Time Effects in the Lexical Decision Task Christopher Donkin (Chris.Donkin@newcastle.edu.au) Andrew Heathcote (Andrew.Heathcote@newcastle.edu.au) Scott Brown (Scott.Brown@newcastle.edu.au) Department of Psychology, University of Newcastle, NSW, 2308, Australia Sally Andrews (sallya@psych.usyd.edu.au) Department of Psychology, University of Sydney, NSW, 2006, Australia Abstract It has been argued that performance in the lexical decision task (LDT) does not provide a direct measure of lexical access because of the effect of decision processes. We re- examine LDT data and fits of the diffusion decision model reported by Ratcliff, Gomez and McKoon (2004) and show that they assumed too little role for non-decision processes in explaining the word frequency effect. Our analysis supports an effect of frequency on decision and non-decision time. Keywords: Lexical decision task; diffusion model Reading is one of the most remarkable abilities achieved by the human mind. One of the key aspects enabling reading is the ability to recognize a string of characters as being a word, a process called “lexical decision”. The lexical decision task (LDT) is a paradigm for studying word identification in which participants are presented with a string of letters and they must quickly decide whether or not the letters form a word. If the letters presented do make a word, then the time taken to make a ‘word’ response is thought to give information about how long it took to retrieve the word from their database of words, a process referred to as lexical access. The word frequency effect is one of the most robust findings from the LDT paradigm: words used less frequently in natural language take longer to indentify than higher frequency words. Historically, the word frequency effect has been reported as a difference in mean reaction time (RT) for correct responses between low and high frequency words. Mean RT from high and low frequency words usually differs by around 60-80ms. However, RT in the LDT is quite variable, typically having a standard deviation of greater than 100ms. Some of this variability is because of differences between words within a frequency class, but variability also occurs between the same word on different occasions. Variability in RT is also positively skewed, with a longer right (slow) than left (fast) tail in RT distribution, and the length of the right tail has been found to vary systematically in LDT experiments. Hence, researchers have begun to investigate differences in the entire RT distribution between high and low frequency words, rather than just the mean RT (Andrews & Heathcote, 2001; Balota & Spieler, 1999; Plourde & Besner, 1997). More recently, there have been lexical theories proposed that account for effects on all aspects of RT distribution (Ratcliff, Gomez and McKoon, 2004; Yap, Balota, Cortese & Watson, 2006). RT distributions have been shown to be well characterized by the ex-Gaussian distribution (Luce, 1986). The ex-Gaussian distribution is produced by convolving (i.e., adding samples from) the Gaussian and Exponential distributions. It has three parameters, the mean (µ) and standard deviation (σ) of the Gaussian component and the mean of the exponential component (τ). These parameters give information about the shape of the RT distribution. In particular, the µ parameter is affected by the speed of the fastest responses made by participants. Similarly, the τ parameter is affected by the length of the right tail of the RT distribution. Differences in parameter estimates from fits of the ex- Gaussian to high and low frequency RT distributions indicate that there are changes in the very fastest and slowest responses made by participants. Changes in µ of approximately 20-30ms have been reported (Andrews & Heathcote, 2001; Balota & Spieler, 1999; Plourde & Besner, 1997). These changes indicate that the entire RT distribution shifts to be slower for less frequent words, independently of any changes in the shape of the distribution. In the same applications of the ex-Gaussian, changes in τ of approximately 35-45ms were observed, suggesting that the right tail is longer when the words to be identified are less frequent. Balota and Chumbly (1984) argued that the data from LDT tasks come from a combination of the lexical process and the decision process. Ratcliff et al. (2004) furthered this line by arguing information about lexical access can only be obtained from RT after accounting for the decision process. In other words, even studying the full range of behavioral data in the LDT (i.e., accuracy and RT distributions for correct and error responses) does not by itself provide clear information about lexical access. To address this issue they fit a model of the decision process, the diffusion model, to their LDT data and used estimates of its parameters, and the parameters of a simple characterization of non-decision processes, to examine lexical access. When Yap et al. (2006) compared the diffusion account with a hybrid two- stage model of the LDT based on Balota and Chumbly’s work, they concluded in favor of the diffusion model. The diffusion model account of RT is composed of two parts – a decision time and a non-decision time. The account of LDT starts by assuming that a stimulus is perceived and encoded. This is followed by lexical access, which gives an estimate of how much evidence the stimulus provides for each response (word and non-word in an LDT). This evidence determines the rate at which information is accumulated, called drift rate, and drives the decision part of the diffusion model. The time taken for the initial