THE PROMISE OF ELECTROPHYSIOLOGY

Electroencephalography (EEG) and magnetoencephalography (MEG) provide direct noninvasive measures of neural activity, in the sense that they track the voltages (EEG) or the magnetic fields (MEG) generated at the scalp by cortical neuronal currents. When the electrical activity recorded across many related trials is averaged together, the resulting measure of activity that is time locked to a specific event is known as an event-related potential (ERP). The magnetic counterpart is sometimes referred to as an event-related field (ERF). Both techniques are fully noninvasive: scalp voltages are nowadays typically recorded from an array of electrodes that are embedded in a flexible electrode cap worn by the subject; scalp magnetic fields are recorded from an array of sensors (typically 140–250 in current systems) housed in a stationary cryogenic dewar that surrounds the head.

Hemodynamic techniques such as functional magnetic resonance imaging (fMRI) and positron emission tomography (PET) provide excellent spatial resolution, and are hence well suited to address questions about the localization of specific brain functions. These techniques are therefore particularly useful for investigating the possible causes of DLIs from a bottom-up perspective that focuses on the impact on subsequent language development of nonlinguistic precursors to language. To the extent that bottom-up approaches provide detailed localization hypotheses, they can be usefully tested using fMRI or PET, as discussed by Müller (2005). On the other hand, the strength of electrophysiological techniques is their excellent temporal resolution, which is on the order of milliseconds. Distributions of scalp voltages in ERP recordings provide a useful dependent measure, but offer only limited information about brain localization. In contrast, magnetic fields are not attenuated by brain and bone matter in the same way as electrical current, and therefore MEG offers improved localization, with accuracy that is sometimes on the order of a few millimeters. Nevertheless, MEG can only rarely match the spatial precision of fMRI and PET. However, because most current accounts of language functions provide far more detailed hypotheses about the timing of language processes than they do about the localization of language processes, electrophysiological measures are well suited to investigations of DLIs that focus on the details of the language phenotype.

In the ideal case, electrophysiological measures provide an opportunity to pinpoint language processes in the brain with millisecond and (in the case of MEG) millimeter accuracy. In many cases this is possible without the interference of explicit tasks, something that is particularly attractive when working with disordered populations. Furthermore, it is possible to build upon a body of relatively well-established findings from studies of unimpaired language. Therefore, electrophysiological measures offer a great deal of promise for tracking the precise nature of language disorders, for identifying the specific features where they differ, and for understanding why different genetic causes often lead to similar language outcomes. Enthusiasm for these approaches is well motivated. However, it is important to also recognize the challenges that must be addressed if the promise of electrophysiological techniques is to be fully realized. The challenges mostly derive from the fact that although EEG and MEG provide a very rich record of electrical activity in the brain, we are still only at the very early stages of understanding the neural code that they read out. Current techniques for separating signals of interest from irrelevant noise are still relatively rudimentary in most studies. Furthermore, even if it were possible to track cognitive processes with perfect resolution in time and space in the brain, we would face the serious theoretical limitation that current hypotheses are not detailed enough to tell us what to look for in such ideal recordings. In the following, I elaborate on each of these challenges in more detail.

THE DOUBLE-EDGED NATURE OF TEMPORAL PRECISION

The extreme temporal precision of electrophysiological recordings is both a blessing and a curse. It is a blessing because it makes it possible to distinguish a neural response that reliably occurs 100 ms after a particular event from a neural response that reliably occurs 120 ms after the same event (e.g., Obleser, Lahiri, & Eulitz, 2003; Poeppel et al., 1997). This is the kind of time resolution that must be reckoned with if we are to understand linguistic computation in detail. Temporal precision is also a curse, because it can be difficult to isolate neural responses that are not precisely time locked to an eliciting event. The most widespread approach in electrophysiological studies of language averages the responses elicited by hundreds or thousands of presentations of a similar event type, in order to generate a record of the activity that is time locked to the eliciting event. A response that is not time locked to the eliciting event will not survive the averaging process.

The need for responses to be time locked to the eliciting event explains why linguistic anomalies have played a central role in electrophysiological studies of language. Brain activity is most likely to be time locked to an eliciting event when the event is not predictable. Linguistic anomalies and violations are primary examples of unpredictable events. The most famous and extensively studied ERP response component is the N400, a negative voltage shift with a broad central scalp distribution that typically appears 300–500 ms after presentation of a word that is syntactically congruous but semantically anomalous (Example 1a) or unexpected (Example 1b; Kutas & Hillyard, 1980, 1984; for a recent review see Kutas & Federmeier, 2000). A small number of MEG studies have elicited the magnetic counterpart of the N400 and have determined the source to be primarily in the posterior temporal lobe (Halgren et al., 2002; Helenius, Salmelin, Service, & Connolly, 1998).

Studies of morphological and syntactic anomalies have yielded a family of different ERP responses. A wide variety of different syntactic anomalies, including ungrammaticalities of various kinds (Example 2) and garden path sentences (Example 3), have been shown to elicit a positive voltage deflection with a broad posterior scalp distribution and an onset latency of 300–600 ms after the critical word. This response is known as the P600 or the syntactic positive shift (Hagoort, Brown, & Groothusen, 1992; Neville, Nicol, Barss, Forster, & Garrett, 1991; Osterhout & Holcomb, 1992). In a narrower set of studies of syntactic anomaly an anterior negativity has been observed in the 300–500 ms latency range. This response is most commonly known as the left-anterior negativity (LAN), because of the fact that it shows a left hemisphere focus in some studies (Coulson, King, & Kutas, 1998; Friederici, Pfeifer, & Hahne, 1993; Münte, Heinze, & Mangun, 1993; Osterhout & Mobley, 1995). The LAN has mostly been elicited by ungrammatical sequences, but there is at least one recent report of a LAN elicited by well-formed garden path sentences like Example 4 (Kaan & Swaab, 2003).

In an even narrower range of studies, a very early response known as the early left anterior negativity (ELAN) has been observed, with peak latencies in the 150–250 ms range. Although this response has sometimes been taken to indicate a general stage of syntactic structure building that occurs extremely rapidly, it should be borne in mind that most current evidence for the ELAN derives from just two constructions, one each in English (Example 5; Neville et al., 1991) and German (Example 6; Friederici et al., 1993; Hahne & Friederici, 1999). In the only existing study of MEG responses to syntactic anomalies, Friederici, Wang, Herrmann, Maess, and Oertel (2000) report that the magnetic counterpart of the ELAN response is primarily accounted for by activity in the anterior temporal lobe. For recent reviews of ERP responses to syntactic anomalies, see Hagoort, Brown, and Osterhout (1999) and Friederici (2002).

In studies of speech sound processing, the electrophysiological literature is similarly dominated by studies of unpredictable events. Many studies of the mismatch negativity (MMN) elicited by oddball sounds in sequences of similar sounds have shown sensitivity to phonetic and phonological properties (e.g., Dehaene–Lambertz, Dupoux, & Gout, 2000; Näätänen et al., 1997; Phillips et al., 2000; for a recent review, see Phillips, 2001). Even the N100 and its magnetic counterpart M100, which is an early response elicited in auditory cortex by any sharp change in an auditory stimulus, have been found to be affected by predictability (e.g., Sanders, Newport, & Neville, 2002; for a recent review, see Näätänen & Winkler, 1999).

Although much has been learned from the fact that different types of linguistic anomaly elicit different types of electrophysiological response, it remains unclear what the link is between the mechanisms involved in the detection of anomalies and the mechanisms involved in successful processing of well-formed language input. Therefore, in order to realize the promise of the temporal precision of electrophysiological approaches, while also avoiding its pitfalls, it will be important to address two challenges. The first challenge involves the need to understand successful processing of nonanomalous material.

A growing body of research addresses Challenge 1. Although electrophysiological research on language has been dominated by studies of anomalous or unexpected events, there is also a growing set of findings about ERP and MEG responses associated with the processing of congruous stimuli. The N400 response was first observed as a response to semantically anomalous words, but it is now well understood that the N400 is a more general response that is elicited by all content words, and that it varies in amplitude as a function of frequency, priming, and cloze probability (Federmeier & Kutas, 1999; Fischler, Bloom, Childers, Roucos, & Perry, 1983; van Petten, Kutas, Kluender, Mitchiner, & McIsaac, 1991). The responses elicited by semantically odd sentences are merely a special case of this more general phenomenon. A similar conclusion is beginning to emerge from ERP research on syntactic processing. A small number of recent studies have shown that the P600 response, which is normally associated with the detection and repair of syntactic anomalies, is also elicited in the construction of fully congruous long-distance dependencies in English (Kaan, Harris, Gibson, & Holcomb, 2000; Phillips, Kazanina, & Abada, in press) and German (Fiebach, Schlesewsky, & Friederici, 2002), leading to the suggestion that the P600 reflects a more general measure of the costs associated with syntactic structure building. However, the morphosyntactic processes that underlie evoked responses such as the LAN and the P600 remain poorly understood. There are also findings that suggest that the working memory demands of holding an incomplete syntactic dependency in memory elicit a sustained anterior negativity (SAN), which may span many words of a sentence (Fiebach et al., 2002; King & Kutas, 1995). However, work in this area is in its infancy, and there is still a great deal to be learned.

The second challenge involves the need for less dependence on time-locked responses.

ELECTROPHYSIOLOGICAL PROFILES OF NORMALLY DEVELOPING CHILDREN

In electrophysiological studies involving adults there is a relatively large body of findings about the processing of linguistic anomalies, and rather less is known about electrophysiological markers of successful linguistic processing. The situation is more acute in the case of studies of children. There are only a limited number of existing electrophysiological studies of language processing in children, and most of these have focused on the responses elicited by different kinds of anomalies.

A number of studies with children have investigated N400-like responses elicited by words in lists (Coch & Holcomb, 2003; Coch, Maron, Wolf, & Holcomb, 2002), primed words (Dykman, Ackerman, Loizou, & Casey, 2000) or anomalous words in sentence contexts (Atchley et al., in press; Holcomb, Coffey, & Neville, 1992; Mills & Schweisguth, 2001). These studies have demonstrated an N400-like response in children aged between 4 and 12 years that is modulated by expectancy in a similar manner to the adult N400. In some studies it is reported to occur later than its adult counterpart, suggesting slower processing.

Published findings about responses to morphosyntactic anomalies in normally developing children are even more scarce, but a couple of findings suggest an interesting developmental delay. Although two studies report that syntactic anomalies elicit P600-like responses in children, either at the same latency as adults (Atchley et al., in press) or at some delay (Friederici & Hahne, 2001), some other studies report nonadultlike response profiles. One study of anomalous German plural nouns in sentence context found that the morphological anomaly elicited a LAN/P600 response pattern in adults, but that a broadly distributed N400-like response was present in children aged 6–8 years, and that the P600 response first appeared in children aged 11–13 years (Lück, Hahne, & Clahsen, 2001). Similarly, an N400-like response was elicited by word-order violations in a group of English-speaking 4-year-olds (Mills & Schweisguth, 2001). Interestingly, a similar pattern of results was found in a study of syntactic anomalies involving adult agrammatic aphasic patients (Hagoort, Wassenaar, & Brown, 2003). In another study, the N400 response was found to be present by age 5 but the ELAN was not present until substantially later (Hahne & Friederici, 1999, as cited in Koelsch et al., 2003). Note that in each of these cases the children who fail to show an adultlike ERP profile in response to syntactic anomalies nevertheless show adultlike ability to detect the anomalies in a behavioral task. Therefore, an important goal for future studies will be to investigate the reason for the protracted development of responses to syntactic anomalies.

A larger number of studies have investigated the MMN response elicited in infants and children by acoustic or phonetic contrasts. The main finding in this area has been that MMN response profiles reflect the major developmental changes in speech perception abilities as infants change from “universal listeners” around 6 months of age to “native language listeners” by around 12 months of age, and then largely track native language discrimination abilities thereafter (Cheour et al., 1998; for a review, see Cheour, Leppänen, & Kraus, 2000). Nevertheless, there are important developmental changes in the amplitude and timing of the MMN response that occur between infancy and school age (Kurtzberg, Vaughan, Kreuzer, & Fliegler, 1995).

Therefore, in light of the limited current understanding of the ERP profile of normally developing children, particularly with regard to word recognition and sentence processing, it will be important to develop reliable normative benchmarks in order to make effective use of electrophysiological approaches in studying DLIs.

The need for detailed models of the development of normal ERP response profiles is particularly acute in light of claims that at least some language impairments reflect a pattern that is normal but delayed (e.g., Rice, 2002). We have seen that ERP response profiles continue to develop in normally developing children after the time when they have reached ceiling performance on some language tasks. This implies that behavioral performance is a weak predictor of electrophysiological responses in children. This suggests that the normal practice of comparing language impaired children with younger controls that are chosen based on their performance on behavioral tasks may be inadequate in electrophysiological studies.

GRANULARITY OF THEORETICAL MODELS

The fine-grained temporal resolution of electrophysiological measures is most valuable if we have theoretical models of normal and disordered language processes that share a similar granularity. It is only useful to know that ERP response component X reliably precedes ERP response component Y by 50 ms if we can link this finding to a model that accounts for the cognitive processes underlying components X and Y and that makes timing predictions on the order of tens of milliseconds. In other words, detailed measurements require equally detailed linking hypotheses.

Unfortunately, however, most currently available models lack the temporal granularity of electrophysiological recordings. Most models in theoretical linguistics explicitly disavow predictions about the timing of real-time processes, and even those that do make claims about real-time processes tend to have a temporal granularity at the level of individual words (e.g., Kempson, Meyer–Viol, & Gabbay, 2001; Phillips, 2003; Steedman, 2000). At the sentence level, most psycholinguistic models also have a granularity that is close to the level of individual words. There is a small number of computational models whose temporal dynamics make more fine-grained predictions about the time course of sentence processing (e.g., Spivey–Knowlton & Tanenhaus, 1998). However, these models primarily focus on the dynamics of selecting among competing alternative parses in ambiguity resolution, and provide less information on the question of how errors might be diagnosed or how possible structural analyses are generated (but cf. Vosse & Kempen, 2000). The relative scarcity of explicit models of structure generation and anomaly detection has meant that the electrophysiological literature on anomaly detection has been forced to rely on models that were designed to account for quite different types of phenomena.

It is commonly assumed in ERP studies of sentence processing that the timing of a response to a violation at a given linguistic level (e.g., syntax, semantics) reflects the timing of the processes that ensure successful processing at the same level of analysis. For example if violations that can be characterized as “phrase structure violations” elicit an ELAN component, with a characteristic latency of 150–250 ms, then it is assumed that successful syntactic phrase structure building may occur in the same time-window (Friederici, 1995, 2002). In Friederici's model the fact that certain syntactic violations trigger earlier responses than semantic anomalies (i.e., [E]LAN vs. N400) is taken to indicate that successful syntactic processing precedes successful compositional semantic interpretation. Furthermore, these claims about the priority of syntactic over semantic information are also linked to claims about the priority of syntactic over semantic information in models of structural ambiguity resolution, such as Frazier's garden path model (Frazier, 1987; Frazier & Rayner, 1982).

However, notions like “syntax first” entail rather different commitments in each of these three areas. In models of structural ambiguity resolution syntax first amounts to the claim that syntactic simplicity outranks other measures of fitness based on frequency and plausibility in decisions about competing structural analyses of a sentence. Such models typically place less emphasis on how the competing analyses are identified in the first place. This first type of syntax first claim has been highly controversial in psycholinguistic research (Clifton et al., 2003; Frazier, 1987; MacDonald, Pearlmutter, & Seidenberg, 1994; Trueswell, Tanenhaus, & Garnsey, 1994). By contrast, in models of the successful generation of unambiguous structures, syntax first amounts to the claim that compositional semantic interpretation is only possible for word combinations that are sanctioned by the syntax. This is a relatively standard assumption, although it is coming under increasingly close scrutiny (Jackendoff, 2002; Kim, Chen, Rippey, & Osterhout, 2003; Tabor, Galantucci, & Richardson, 2004). Lastly, in models of error detection, syntax first entails the claim that syntactic errors are detected more quickly than semantic errors (Friederici, 1995; McElree & Griffith, 1995), but this does not necessarily entail a commitment about the time course of successful structure generation. For this reason, there is a need for caution when linking findings from violation-based ERP studies to models of the time course of sentence processing.

Although it has often been assumed that the timing of ERP responses to violations at a given level of analysis reflects the timing of successful processing at the same level, this assumption has not been directly tested. Some interesting recent studies suggest that timing estimates derived from ERP studies conform well to timing estimates derived from behavioral speed—accuracy trade-off studies (Bornkessel, McElree, Schlesewsky, & Friederici, in press), but these comparisons involve violation detection paradigms in both cases. In fact, some of the most interesting timing arguments in the ERP literature involve cases where violations at one level of analysis provide evidence for the timing of successful processing at a different level of analysis. For example, Garnsey, Tanenhaus, and Chapman (1989) used the timing of the N400 response elicited by semantic anomaly to draw inferences about the timing of syntactic processing. They reasoned that in filler-gap constructions like Example 7, the fact that the fronted phrase is a semantically anomalous direct object of the verb could only be detected if syntactic processes have first identified it as the object of the verb. Thus, the timing of the onset of the N400 response (which typically occurs less than 400 ms after word onset) provides a good estimate of the time by which syntactic processing of the filler-gap dependency must have occurred.

A related logic has been successfully applied in electrophysiological studies of lexical access, where more explicit models of successful processing are available. For example, Levelt and colleagues (Levelt, Praamstra, Meyer, Helenius, & Salmelin, 1998) use an independently motivated model of picture naming to interpret timing information from an MEG study. According to their model, phonological encoding of a word is the first level of analysis that is sensitive to word frequency. Therefore, the timing of the first MEG response component that shows sensitivity to word frequency in a picture naming task places an upper bound on the timing of word form retrieval. A similar approach has been successfully applied in MEG studies of word recognition (see Pylkkänen & Marantz, 2003, for a review).

Electrophysiological studies have been used with some success in studies of aphasic patients, due to the availability of theoretical claims that link aphasic deficits with disruptions in the timing of syntactic processing operations. For example, in the context of hypotheses that comprehension difficulty in Broca's aphasia reflects delayed lexical access (Milberg, Blumstein, Katz, Gerschberg, & Brown, 1995) or delayed lexical integration processes (Hagoort, 1993; Tyler, Ostrin, Cooke, & Moss, 1995), it is particularly informative to find that N400 responses elicited by semantically anomalous words are delayed by around 100 ms in low-comprehending aphasic patients, relative to high-comprehending aphasics and normal controls (Swaab, Brown, & Hagoort, 1997). Similarly, in the context of the proposal that syntactic processes are slowed in Broca's aphasia (Friederici & Kilborn, 1989; Haarmann & Kolk, 1991), it is interesting to find that syntactic violations like Example 6 elicit the normal pattern of an early ELAN component followed by a P600 in patients with subcortical lesions, but elicit only the P600 in a group of patients with left anterior cortical lesions (Friederici, von Cramon, & Kotz, 1999). It is interesting to find differences between the ERP response profiles of aphasics and normal controls, but the comparison is far more informative when guided by specific time-based hypotheses. For recent reviews see Kotz and Friederici (2003) and Friederici and Kotz (2003).

In the area of DLIs, however, we find mostly static characterizations of the language impairments in terms of knowledge that is present or absent, or skills that are stronger or weaker. We find few accounts that make predictions about the time course of language processing in DLIs. The upshot of this is that extant models of DLIs provide little guidance regarding how to exploit the temporal precision of electrophysiological measures. This provides the fourth challenge for neurocognitive studies of DLIs.

Even in the absence of more detailed time course predictions, electrophysiological approaches can still be useful for the study of DLIs. First, because ERP and MEG methods track automatic brain responses and do not require explicit tasks from the speaker, they may provide an opportunity to demonstrate linguistic sensitivity that is not apparent in standard behavioral measures that require explicit tasks such as button pressing or metalinguistic judgments. Second, electrophysiological studies may reveal differences in the time course of language processing between normal and language impaired individuals, by showing that similar responses are elicited in normal and affected subjects, but with different timing. In this way, electrophysiology could be used to provide an impetus for more fine-grained hypotheses about the time-course of language processing in impaired individuals.

One example of how both of these properties of ERPs might effectively be exploited in research on DLIs can be found in the area of filler-gap dependencies. It is commonly reported that children with DLIs have particular difficulty with sentences that involve filler-gap dependencies, such as object relative clauses (Example 8a) and wh- questions (Example 8b). In a number of instances this is manifested in the form of difficulty with object relative clauses in comprehension tasks (Karmiloff–Smith et al., 1997; Stavrakaki, 2001; Zukowski, 2004). In at least one case, it has been argued that in a special subclass of children with SLI this difficulty extends to grammaticality judgment and production (van der Lely & Battell, 2003), such that children produce sentences that contain fronted wh-phrases but no gap (e.g., Example 9), and judge similar sentences to be well formed.

It would be instructive to follow up on these behavioral findings in DLIs using a version of the ERP study by Garnsey et al. (1989) described above. If children show an N400 response to sentences with implausible filler-gap dependencies, then this would indicate that they are able to construct filler-gap dependencies, even if this is not apparent in behavioral tasks. Furthermore, the results might provide useful information about the time course of constructing such dependencies. In an ERP study conducted with normal adults, Phillips and colleagues (2004) found that the completion of a semantically congruous filler-gap dependency elicited a P600 response, as had been found previously by Kaan et al. (2000), and also found that the P600 response had a significantly later onset for longer filler-gap dependencies that spanned two clauses. In light of the finding that longer filler-gap dependencies are computed more slowly than shorter dependencies, it is possible that a comparison of normal and disordered populations would show a delayed N400 response to semantically anomalous filler-gap dependencies in children with DLIs.

IMPROVING SIGNAL/NOISE RATIOS (SNRS)

An additional challenge for realizing the promise of electrophysiology in research on DLIs derives from technical issues affecting the practical use of these measures with children and affecting their use in individual diagnoses. Solutions to these problems will require investment in the development of infrastructure and improved analytical tools.

Scalp voltages or magnetic fields reflect the combined effect of many different neural sources that are simultaneously active. In ERP studies, the standard approach to improving the SNR is to compute averages over many different presentations (typically 30–100 trials) of the same or equivalent stimuli, and to then create grand averages of recordings from 15 to 30 participants. The effect of this is that ERP studies tend to be rather long. A typical sentence comprehension study with 4 conditions (two target conditions, two control conditions) and a bare minimum of 30 trials per condition requires 120 experimental sentences per subject. When combined with at least twice as many filler items, in order to minimize the development of experiment-specific strategies, this yields at least 360 sentences, requiring over an hour of recording time. Time for electrode application and clean-up can add another hour to the study. The duration of the studies creates obvious challenges for working with normal or disordered children. To make matters worse, ERP recordings are particularly susceptible to motion artifacts and eye movements, and thus studies with children may require even more trials in order to achieve the same SNR attained in adult studies.

MEG recordings add the promise of localization in addition to precise temporal resolution. However, this also introduces additional technical challenges. Since the sensors in whole-head MEG arrays have a fixed position inside the liquid-helium filled dewar, individual sensors are not consistently aligned with specific scalp locations or brain regions. Individual sensors have different positions in a head-defined coordinate space, depending on the size and position of the head. The practical consequence of this is that it is not normally possible with MEG to use the grand-averaging procedures that are standard in ERP studies. Instead, it is necessary to convert recordings from individual participants into a sensor-independent format before data are combined across participants. Localization information must be computed on an individual basis. This means that higher individual SNRs are needed, and thus more trials are needed per participant. For example, whereas ERP studies of syntactic processing typically present each participant with 30–50 tokens of each sentence type, MEG studies of similar phenomena have presented each participant with much larger numbers of tokens per condition, ranging from 100 tokens per condition per participant in studies that have computed more approximate localization information (e.g., Helenius et al., 1998) to as many as 390 tokens per condition per participant in a study that combined localization information from MEG and fMRI recordings (Friederici et al., 2000). In the latter case, this required each subject to undergo multiple recording sessions. Clearly, it would be difficult to undertake studies of this length with normal or disordered children. Furthermore, whereas the use of head-mounted electrode caps in ERP studies allows participants to move their head during the study, the head must remain motionless during an MEG study, due to the fact that the sensors are in a fixed position. Also, the fact that children have smaller heads than adults has implications for the quality of MEG recordings. For the gradiometer sensors that are used in most current MEG systems, the strength of the MEG signal is inversely proportional to the cube of the distance from the source to the sensor. For example, for a child who is lying in a whole-head MEG device, strong signals will be recorded from occipital regions at the rear of the head, where the scalp is touching the MEG dewar, but weaker signals will be recorded from other areas where a few centimeters separate the scalp from the MEG sensors. This latter issue may be addressed by MEG devices that are specially designed for use with smaller heads, or by using older MEG devices that have smaller arrays of sensors that only cover part of the head, and hence can be moved to fit any head size.

Therefore, the fifth challenge facing electrophysiological investigations of DLIs is how to maximize SNRs, such that recording sessions can be kept to a length that is manageable for children. It is probably no accident that the most successful line of electrophysiological language research on children to date has involved MMN studies of phonetic perception. Since the individual experimental trials are shorter in these studies, typically involving just one syllable, it is possible to record many hundreds of trials in a relatively short period of time.

An additional consequence of the signal/noise problem is that electrophysiological studies of language processing have tended to rely on group data in order to achieve reliable results. This has implications for studies of DLIs, where there would be interest in using electrophysiological measures as diagnostic tools for assessing individuals. This is only feasible to the extent that it is possible to achieve a good SNR from the number of trials that can be reasonably recorded from a single person. Until that is possible, electrophysiological measures will continue to be most effective for group studies. One study of syntax-related ERP effects in aphasia used a single case study approach with analyses of individual data (Friederici, Hahne, & von Cramon, 1998), and was able to find significant effects of anomaly detection. However, this study also showed that only very large ERP effects are likely to be statistically reliable in case studies.