Phonological Similarity in Serial Recall: Constraints on Theories of Memory

In short-term serial recall, similar-sounding items are remembered more poorly than items that do not sound alike. When lists mix similar and dissimilar items, performance on the dissimilar items is of considerable theoretical interest. S. Farrell and S. Lewandowsky (2003) recently showed that if guessing strategies are controlled, dissimilar items on mixed lists are recalled more accurately than on pure dissimilar lists, a finding that challenges several current theories of serial recall. This article presents two experiments that extend the generality of the mixed-list advantage for dissimilar items and then applies three theories of memory—the primacy model, SIMPLE, and SOB—to the data. The simulations show that the data are best explained by the SOB theory (S. Farrell, 2006; S. Farrell & S. Lewandowsky, 2002) which, unlike most other current theories, posits that similarity has an effect at the time of encoding. Phonological Similarity 3 Phonological Similarity in Serial Recall: Constraints on Theories of Memory The phonological similarity effect in short-term serial recall refers to the well-replicated finding that lists composed of similar-sounding items are less accurately recalled in the correct order than lists in which items do not sound alike (e.g., Baddeley, 1966, 1968; Conrad, 1964; Henson, Norris, Page, & Baddeley, 1996; Wickelgren, 1965a, b). This effect is of considerable generality, occurring with consonants (Baddeley, 1968) as well as with words (Baddeley, 1966; Coltheart, 1993; Henry, 1991). The phonological similarity effect is typically considered a benchmark result of short-term serial recall that must be accommodated by relevant theories (e.g., Brown, Preece, & Hulme, 2000; Nairne & Kelley, 1999). The phonological similarity effect also occurs when phonologically similar (e.g., B, P, T) and dissimilar (e.g., K, Q, R) items are mixed together on study lists. Mixed lists are particularly diagnostic because the level of recall of the dissimilar items can differentiate between competing theories of memory. Initial reports (e.g., Baddeley, 1968; Bjork & Healy, 1974; Henson et al., 1996) suggested that dissimilar items on mixed lists are recalled with the same accuracy as items on pure dissimilar lists. This apparent absence of an effect of intermixing similar items on the recall of dissimilar items compelled several theorists to propose that serial recall involves two independent stages of processing, with order errors occurring between positional tokens in a primary stage, and with similarity-based confusions occurring in a separate secondary stage that operates only on similar items (e.g., Burgess & Hitch, 1999; Henson, 1998; Page & Norris, 1998a, b). More recently, Farrell and Lewandowsky (2003; see also Farrell, 2006) showed that when guessing strategies are controlled—either by presenting all list items for re-ordering at test or by equalizing total ensemble size across list types—dissimilar items on mixed lists were in fact better recalled than their counterparts on pure dissimilar lists (we refer to this finding as a Phonological Similarity 4 “mixed-list advantage” from here on). Farrell and Lewandowsky (2003) suggested that their results challenged two-stage theories and proposed that models in which similarity affects encoding processes, such as the feature model (e.g., Neath, 2000) or SOB (Farrell & Lewandowsky, 2002), provided a more suitable explanation. Farrell (2006) demonstrated that the SOB model accounted for the mixed-list advantage at a quantitative level, whereas a two-stage model failed to provide an adequate account. However, the work by Farrell and Lewandowsky (2003) and Farrell (2006) was limited in two ways: First, at an empirical level, the mixed-list advantage could be explained by appealing to encoding or retrieval strategies. Second, although the mixed-list advantage promises to differentiate between two classes of theory, its theoretical implications still await thorough examination. This article addresses both of those limitations. We first report two mixed-list experiments that extend the findings of Farrell and Lewandowsky (2003) using a procedure that reduces the possible role of encoding or retrieval strategies. We then compare three theories of memory by simulation and examine their ability to handle the mixed-list advantage for dissimilar items. The simulations show that the two-stage primacy model (Page & Norris, 1998a, b) cannot accommodate the mixed-list advantage. The SIMPLE temporal distinctiveness theory (Brown, Neath, & Chater, 2002) produces a mixed-list advantage with some, but not all, parameter values. The SOB theory of Farrell and Lewandowsky (2002) and Farrell (2006), by contrast, reliably produces a mixed-list advantage and overall provides the best quantitative account of the data. The simulation results suggest that similarity effects are best modelled by assuming that similarity affects encoding into memory rather than contributing only to confusions at retrieval. Phonological Similarity 5 MIXED LIST PHONOLOGICAL SIMILARITY IN SERIAL RECALL: THE EMPIRICAL PATTERN The use of lists that contain a mixture of items of different types has been a highly diagnostic tool in several areas of memory research. For example, in recognition memory, lists containing a mixture of strong and weak items (e.g., items that are presented for 1 s and 2 s, respectively) have been central to arguments about the relative merits of various computational models (e.g., Murdock & Kahana, 1993a, b; Shiffrin, Ratcliff, Murnane, & Nobel, 1993). Although adjudication between those models has been based on a plethora of data, the influence of the— mixed-list-based—list strength effect has been remarkable. In serial recall, recent work on the word length effect has likewise exploited the diagnosticity of mixed lists (Hulme, Neath, Stuart, Shostak, Surprenant, & Brown, 2006). One reason for the widespread utility of mixed lists is that relevant models inevitably handle the benchmark differences between pure lists of the different item types under consideration: For example, models of recognition memory predict performance to be better on pure strong than pure weak lists; models of short-term mmory can accommodate the ubiquitous word length effect and they also universally predict that lists of similar items will be recalled worse than lists of dissimilar items. Models may differ, however, in their ability to handle the results from mixed lists. Concerning phonological similarity, Baddeley (1968, Experiment 5) reported the first experiment in which items within a list were either mutually dissimilar (e.g., M K and R) or highly similar (e.g., V, D, and T), with the two types being interleaved to form a list such as M V K D R T. This mixed list is commonly referred to by its abstract code as DSDSDS, where S and D represent, respectively, a similar and dissimilar item. Baddeley (1968) found that the accuracy with which D items were recalled from pure lists (DDDDDD) was identical to that of D items in the corresponding serial positions on mixed lists (e.g., SDSDSD). Henson et al. (1996) replicated Phonological Similarity 6 the result using a more refined method and analysis. The absence of a mixed-list advantage for dissimilar items was interpreted as evidence against chaining models (e.g., Lewandowsky & Murdock, 1989; see Henson et al., 1996, for further discussion) and became a benchmark result for theorizing during the 1990’s (see Brown et al., 2000; Burgess & Hitch, 1999; Henson, 1998; Henson et al., 1996; Page & Norris, 1998a). Notwithstanding its acceptance for an extended period, the absence of a mixed-list advantage is quite counter-intuitive. Consider the extreme case of a single dissimilar item being embedded in a similar-sounding list (e.g., the letter X in the list B D G X T P); on the basis of the ubiquitous isolation effect—often called the “von Restorff effect” after its initial investigator—the dissimilar item should be recalled more accurately in this case than when it is surrounded by other dissimilar items. Indeed, the presence of other isolation effects in serial memory tasks (e.g., Bone & Goulet, 1968; Cimbalo, Nowak, & Soderstrom, 1981; Kelley & Nairne, 2001; Lippman, 1980) might lead one to question why there was no facilitation for dissimilar items on the mixed lists in the studies by Baddeley (1968) and Henson et al. (1996). Farrell and Lewandowsky (2003) pursued this question and found that recall of dissimilar items is unaffected by the presence of similar items only when people can trade off order and item errors. Hence, even when the pattern of correct responses gave the appearance that recall of dissimilar items did not differ between list types (i.e., pure D vs. mixed), the underlying transposition rates indicated that order memory was better for dissimilar items on mixed lists than on pure lists. Performance appeared equal only because of compensatory differences in intrusions and omissions (Experiment 1). When item errors were controlled by use of a reconstruction task (Experiment 2) or were equalized across list types by equalizing vocabulary size (Experiment 3), a mixed-list advantage for dissimilar items was observed even at the level of correct responses. To give an impression of the consistency of those results, Figure 1 summarizes the data from all Phonological Similarity 7 experiments of Farrell and Lewandowsky (2003) by plotting the transposition error rates for dissimilar items on mixed lists (mixed D) against transposition rates for items in corresponding serial positions on pure lists (pure D). Contrary to the early results by Baddeley (1968) and Henson et al. (1996), nearly all data points deviated systematically and considerably from the diagonal (which would represent equal performance on both list types), such that there were always more transpositions for an item in a pure dissimilar list than in a mixed list. Although these results are consistent and replicable (e.g., Farrell, 2006), two plausible objections can be raised against the procedure of Farrell and Lewandowsky (2003). First, in their studies the various list types were presented contiguously in blocks of 40 trials of the same type. Although this followed precedent (Henson et al., 1996), the blocked presentation regime may have induced people to develop strategies that were specific to each list type, such as encoding of the mixed lists into two separate “streams” of similar and dissimilar items. Farrell and Lewandowsky (2003) sought to address this problem by showing that the mixed-list advantage did not differ between early and late trials within a block; nonetheless, the development of strategies cannot be completely ruled out. In recognition of this issue, Farrell (2006) presented lists in a random order, and still obtained a sizeable mixed-list advantage. However, the mixed lists of Farrell (2006) contained only a single dissimilar item, which gives rise to a second concern. In the studies by Farrell and Lewandowsky (2003) and Farrell (2006) some lists contained a single dissimilar item that was surrounded by similar items (e.g., B D G X T P). In consequence, the overall experimental sequence contained more similar than dissimilar items, thus creating an imbalance in potential proactive inhibition or “token dose” (Bridges & Jones, 1996) between item types. We now report two experiments that resolved both potential objections to the earlier studies. Phonological Similarity 8 MIXED-LIST EXPERIMENTS