Word can be recognized

Three common profiles of poor reading

SWRDs involve core weaknesses in decoding, usually related to phonological weaknesses, coupled with at least average vocabulary knowledge and broad language comprehension (Catts, Adlof, & Weismer, 2006; Leach, Scarborough, & Rescorla, 2003). In addition to their difficulties reading individual words, children with SWRDs also frequently have problems with reading fluency and reading comprehension. However, their difficulties in these latter areas are associated entirely with problems in word reading, not language comprehension. Poor fluency may be due either to inaccurate or nonautomatic reading of individual words. Likewise, for these children, poor reading comprehension is caused entirely by problems in word reading; if children can decode a text accurately and with reasonable fluency, then they can comprehend it. Jamie, the first child described at the outset of this chapter, had a profile of SWRD. SWRD is a common profile in students with dyslexia, which was Jamie’s diagnosis.

SRCDs involve the opposite profile, one that is distinct from dyslexia (Cutting et al., 2013). Poor readers with SRCDs have at least average decoding and phonological skills, but nevertheless, their reading comprehension is impaired (Leach et al., 2003). Usually students’ reading comprehension difficulties are associated with oral language weaknesses in areas such as vocabulary, broad listening comprehension, or higher-level language abilities such as pragmatics (Catts et al., 2006; Norbury & Nation, 2011). Therefore these students tend to display comprehension weaknesses not only during reading but also in listening activities, such as during teacher read-alouds or class discussions. If students with SRCDs have problems with reading fluency, those problems are based in language, not decoding. For example, students with this profile might read text slowly because they are having trouble comprehending it (Valencia, 2011). Eli’s profile exemplified SRCDs. He had a reading disability, but one different from dyslexia.

Children with a mixed profile, or MRDs, have a combination of the above types of difficulties (Catts et al., 2006; Leach et al., 2003). They not only have weaknesses in decoding and phonological skills but also have a core comprehension component to their reading problems. Clinicians might encounter a profile of MRD in students with broad language disabilities or in some students with autism spectrum disorders (Norbury & Nation, 2011). Although students with MRDs often have weaknesses in listening comprehension, their reading comprehension typically is more impaired than their listening due to the additional influence of poor decoding. Fluency problems in students with MRDs may relate to both factors, poor decoding and language weaknesses.

Knowledge about individual students’ profiles can help clinicians understand and integrate a wide array of assessment data, with important implications for instruction, assessment, and accommodations (Catts, Kamhi, & Adlof, 2012; Spear-Swerling, 2015). To be effective, interventions must properly target a child’s component weaknesses. Children with SWRDs and MRDs generally benefit from phonics interventions, whereas those with SRCDs do not, because their difficulties lie outside the domain of decoding (Aaron, Joshi, Gooden, & Bentum, 2008). Students with MRDs need more than phonics intervention to be successful; they also require intervention targeting the source(s) of their comprehension weakness, such as vocabulary, background knowledge, syntax, or comprehension monitoring (Oakhill, Cain, & Elbro, 2015). In assessment, oral reading fluency CBMs may be very useful for monitoring progress in children with SWRDs and MRDs but probably not for students with SRCDs, who will require assessments more focused on comprehension. Students with SWRDs often benefit from accommodations that involve listening to text that is too difficult for them to read, whereas for those with SRCDs, merely hearing a text read aloud is less helpful (Erickson, 2013).

Measurement of underlying cognitive processes involved in reading achievement

The measurement of reading achievement is a complex undertaking because it includes a series of separate but integrated processes. Deficits in word recognition, reading rate, reading fluency, decoding skills, vocabulary, or weaknesses in listening and language comprehension can underlie an impairment in the development of reading comprehension as an academic skill set. To date, extensive research studies have documented that problems in several distinct areas can, and do, contribute to differences between good and poor readers (Swanson & Hsieh, 2009; Wagner & Torgesen, 1987; Wolf & Bowers, 1999). As a result, a number of instruments have been developed to further explicate specific underlying deficits if problems in reading comprehension are found during the measurement of achievement. There are at least eight commonly used measures of phonological awareness itself. A list of these can be found in the Mather and Abu-Hamour (2013). In addition, the various test publishing companies have several others on the market at this time.

The particular measure with which we are most familiar is the Comprehensive Test of Phonological Processing (CTOPP) (Wagner, Torgesen, & Rashotte, 1999). The CTOPP is a measure of phonological awareness, phonological memory, and rapid naming. Difficulty in any of these areas is considered a common cause of reading disability. The first version of the test was developed for children between the ages of 5 and 6, but we are most familiar with the second version, available for individuals between the ages of 7 and 24. The second version consists of seven core subtests and six supplemental subtests. In our experience, it is necessary to administer all 12 subtests with adolescents and young adults as there appears to be a ceiling effect when just the core subtests are given. There is now a CTOPP-2 that contains normative data collected in 2008 and 2009 which was published in 2013 (Wagner, Torgesen, Rashatte, & Person, 2013). However, the basic instrument remains the same and results in five composite scores: Phonological Awareness; Phonological Memory; Rapid Symbolic Naming; Rapid Non-Symbolic Naming; and an Alternate Phonological Awareness Composite Score.

The CTOPP does a remarkable job of sorting out phonological deficits from rapid naming deficits. Rapid naming deficit is a factor important for reading skills that was identified early in the work of Denckla and Rudel (1976). Rapid naming skills have more recently become the focus of research and intervention studies that are concerned with identifying and diagnosing an underlying developmentally based learning disorder involving reading fluency and its impact on comprehension (Wolf & Bowers, 1999; Wolf & Katzir-Cohen, 2001; Wolf, 1986, 2016; Wolf, Miller, & Donnelly, 2000). This research has found that the majority of children with developmental reading disabilities start with weaknesses in naming speed, and then go on to develop problems in reading fluency. Thus the ability of the CTOPP to differentiate these various skills is crucial to better understand an individual’s reading disorder.

The role of fluency is also an important factor in reading disorders. As discussed earlier, reading fluency has been added to several test batteries (e.g., WIAT-III and WJ-IV). This is at least in part due to the work of Wolf. Indeed, in recent years, Wolf and Katzir-Cohen (2001) explicated a new conceptualization for fluency. They propose that reading fluency is the product of a combination of the initial development of accuracy and automaticity in core reading that include the sublexical and lexical processes. These processes are perceptual, phonological, orthographic, and syntactic in nature, and important for integrating single word and connected text. Once fully developed, fluency relies on a combination of accuracy and rate in which the actual decoding is relatively effortless, oral reading is smooth with correct prosody, and attention can be focused on comprehension rather than the basic mechanics of reading fluency. Based on these considerations, various strategies have been formulated to increase reading fluency. There are several excellent papers (Alber-Morgan, 2006; Mastropieri, Leinart, & Scruggs, 1999) detailing research validated strategies, which can be incorporated into achievement findings and recommendations.

Information input channel

The interplay between stimulus-receiving modalities has important implications for learning. Verbally presented information (e.g., classroom instruction and audiobook) gains direct access to the auditory cortex of the brain, whereas reading or pictorial information must be visually inspected and orthographically converted to PH code before it can enter the auditory cortex and be held temporarily in PH STM for additional processing by the CE. This is why some parents report that their children are better able to comprehend and recall information when it is read to them orally rather than having to read and orthographically convert the information themselves. These children are unable to convert the read information efficiently and lose significant information before it becomes available in the PH STM store to be used for passage comprehension-related tasks (i.e., the system becomes bottlenecked).

A recent study of the bottleneck phenomenon involving children with ADHD revealed that two interleaved processes contributed to reading comprehension deficits in ADHD: slowed orthographic conversion and underdeveloped CE processes (Friedman et al., 2017). Slowed orthographic conversion can occur for several reasons; however, one of the most common causes is that automaticity has not been developed for a large number of basic words, math facts (addition, subtraction, and multiplication), and relevant rules (literacy and numeracy) when attempting to activate this information in long-term memory (LTM). The bottleneck occurs because the pursuit of the needed words or math facts slows the speed by which it reaches the limited capacity PH STM store, which can only maintain the information for several seconds unless it is constantly refreshed via covert rehearsal. The resulting consequence is that many children will forget the task instructions—particularly multistep instructions—prior to fully retrieving the needed information from LTM.

Isolated Word Scoring

The error rate of speech recognition is defined as “the average fraction of items incorrectly recognized.” Here, an item can be a word, a subword unit (e.g., a phone), or an entire utterance. For an isolated word recognition system, the error rate is defined as

Here, N is the number of words in the test utterance and N_E the number of words incorrectly recognized. The latter can be subdivided into substitution error, N_S, and deletion (incorrect rejection) error, N_D:

Sometimes the fraction of correctly recognized words, C = 1 – E, called correctness, is used:

These measures do not include so-called insertions, since it is assumed that the beginning and the end of each word can be detected directly from the energy of the signal. However, in real applications in which speech is contaminated by noise, it is not always easy to detect word boundaries, and sometimes noise signals cause insertion errors. Therefore, under these real-world conditions, the same measure as that used in continuous word scoring, which will be described later, is also used in the isolated recognition task.

For isolated word recognizers, a measure more specific than the various contributions to the error rate, a confusion matrix, has also been used, in which the class of substitutions is divided into all possible confusions between words. The confusion C_ijis defined as the probability that word i is recognized as word j. The value C_ii is the fraction of times word i is correctly recognized. These probabilities are estimated by measuring the number of times the confusion took place,

where N_ij is the number of times word j was recognized on the input word i. The confusion matrix gives more detailed information than the error rates. Insertions and deletions can also be included in the matrix by adding a null word i = 0 (nonvocabulary word). Then, the row C_0j contains insertions, the column C_i0 the deletions, and C₀₀ = 0. Using this expanded confusion matrix, the error rate can be calculated from the diagonal elements, i.e., E = 1 – σ_iC_ii = σ_i≠jC_ij. The elements C_ij for i ≠ j are called the off-diagonal elements.

Achievement

Most methods used for the identification of students with LDs include low achievement as a defining characteristic, which is necessary but insufficient for the identification of LDs. Identification procedures based on any model used for the identification of LDs will include measures of a child’s level of achievement. Such information is crucial because it provides an observation of the exact behaviors of interest, the child’s ability to read, write, spell, and solve mathematical problems (Fletcher & Miciak, 2017; Siegel, 1999; Stanovich, 1999). Furthermore, empirical data suggest that there are five different forms of LDs that each impact a different area of achievement: word recognition and spelling, reading comprehension, mathematical computations, mathematical problem-solving, and written expression (Fletcher et al., 2018).

Standardized norm-referenced measures of achievement are used to assess a child’s level of performance in these areas. Common examples of such measures include the Woodcock–Johnson Test of Achievement, currently in the fourth edition (Schrank, Mather, & McGrew, 2014), the Kaufman Test of Educational Achievement, currently in the third edition (Kaufman & Kaufman, 2014), and the Wechsler Individual Achievement Test, currently in the third edition (Psychological Corporation, 2009). Each of these test batteries measures the specific domains impacted by LDs and includes norms for age and grade that allow for standard scores to be computed. The provision of standard scores supports efforts to determine if a child is achieving at a level comparable to his or her peers. This is often operationalized as a cut point. For example, a cut point of the 25th percentile could be adopted. Those children for whom 75% of their age group outperformed them on a given measure of academic achievement (e.g., untimed isolated word reading) would be deemed as exhibiting low achievement in this area. Yet, difficulties arise in determining the exact placement of cut points that impact the provision of services to individual children (Francis et al., 2005). These difficulties emphasize that if they are to be used, the creation of cut points should be made with great care.

Tempo properties of AR

The specific issues of AR with regards to the form of MIR’s relationship to VR and VDR, differ in the fact that an audio data stream is connoted by tempo-related properties, and also properties relating to frequency and sound characteristics such as tone, pitch, timbre, melody and harmony. Again, an audio stream is simply a set of data that must be elaborated and made useable by the processing system by identifying its defining characteristics. The same thing can be said for document treatment, where the theory and technique of AR share things common to all MIR, but specialize in some particular respects.

It is clear that with regard to audio, it is a completely different language and medium from those dealt with in TR, VR and VDR, although the latter has a number of relationships to sound. In the case of content-based treatment of audio objects and audio documents, the same theoretical issues concerning the whole of MIR can be proposed. In fact, it still means working directly with certain content-related elements and concrete and specific objects, as intangible as sounds may seem, without accepting excessive mediation of systems based on the terms and the translation of documented content in text.

Even here it is possible to speak of contentual elements being treated within themselves with consistent tools, and semantic aspects that can be examined later for a total treatment of the sound object. Treatment, in addition to considering the characteristics of sound data as such, and the feelings that they derive from, must also give an account of the complex and abstract interpretation that sounds, and particularly music, may be given. What remains permanent is the non-solvability of issues such as harmonics, where neither contentual nor semantic considerations can come to final conclusions, and where even human emotivity struggles to define their quality – let alone a machine. Content-based treatment of audio documents, based on the same common problems of processing all document types, can boast some real revolutionary developments, effectiveness and limitations.

Among the more problematic points of AR, is the transcription of sound tracks. Music, dialogue or other sounds must be registered in the language of the system with propriety and adherence to the original object in order to be effectively dealt with, whatever the status of the document to which the technique is applied. Nothing is too different, in essence, with problems occurring in other forms of MIR when dealing with a poorly digitized document, which cannot give a correct representation of its content. Digital audio technology does seem, however, to already be at the right level in order to define the best conditions for confronting these issues.

Among other problems is the alignment of different versions of the same sound track, as well as comparing changes between tracks. In such cases the process can be tricked by simple changes in interpretation or performance, and not by recognizing similarities between two objects. This is parallel to the cases of different points of view in VR and VDR, or of homonymy in TR. Specific emphasis will be given to the issues of audio-thumbnailing, or audio abstracts and audio browsing. They are connected, in as much as the ability to draw a complete effective sound synthesis allows the ability to browse and evaluate an audio document in analysis and search.

Finally, there can be different types of audio document processing allowed by AR systems: general analysis and research based on sound data; the comparison between segments of a single sound file; the comparison of sound segments of different types; analysis and recovery of short individual sounds; description and retrieval of sound tracks in audio-visual objects; search by selecting words in songs or dialogues; the processing and indexing of speech; and the treatment of various sounds and noises.

Some modalities of AR already functioning are: speaker identification, based on the ability to recognize human voices regardless of the words spoken; a typical similarity query, which queries a database using sample tracks; and query-by-humming, which has the advantage of allowing search by similarity to an audio model hummed by a user — more or less similar to visual sketches.

George Tzanetakis provides a concise presentation of the extensive problems of AR and, while declaring that attention to audio document content-based search systems are fairly recent, he states how these techniques are the first to be commercially exploited thanks to the increasing dissemination of music files.³⁴

Lately there have been instances of experimental algorithms best capable of processing the sound content of a document, even if such general methods are derived from other methods of MIR or ASR. Music data – with which the author deals exclusively – due to its complexity requires specific systems and techniques which are now on their way to a major development, and AR systems even now can deal with both simple sound files and the most complex musical documents. The ‘feature vectors’ of musical documents represent models of content characteristics, and are treated by MIR systems for the typical content-based operations of ‘indexing’, ‘similarity search’, ‘recognition’ and ‘music browsing’ all taking place via the specific methods of AR.³⁵

The analysis of ‘pitch content’, or ‘harmonics’, has one of its most critical points in ‘polyphonic transcription’, the process of converting a recording into a symbolic representation, similar to that of a musical score containing the notes played by several instruments. Translation and transcription of sounds onto a data table is an essential element for analysis. However, it is very difficult to obtain this automatically. The best tool currently available for this transcription is ‘Pitch Histogram’, consisting of statistical representations of the harmonic development characteristics in a piece of music. The machine cannot give anything more than a statistical representation of the rich harmonies existing in music.

‘similarity matrix’. The diagonal of the matrix has its highest value when it corresponds to the similarity of a segment. The visual transposition of this structure can reveal the hierarchical and regular composition of musical signals, and reveal the repetitions within a piece, their degree of similarity, and in general the structural explanation of the sound of a piece.

The fourth area of interest in AR is that of ‘linking audio to scores’. Sound tracks can be represented both by proper audio signals and by symbolic transcripts, and it is theoretically possible to perform search and confrontational operations through the symbolic tables, presenting results in sound form. The symbols can permit a comparison of concrete sound signals, providing that the transcription is very accurate and allowing for the systematic errors that a machine makes. A practical case for this treatment is ‘query-by-humming’, where the user can hum a melody sample which is then translated by the system into appropriate symbols. Obviously, to associate symbolic and sonorous representations, the system must develop so that it can identify the similarity of audio signals with symbols.³⁸

Fully aligning the basic principles of MIR, Tzanetakis concludes that the ability to ‘automatically comprehend’ audio signals by ‘actual musical means’ is essential to developing and solving one of AR’s key issues: to search and recover a sound track without any knowledge or need for terminological metadata or any kind of external reference.

Methodological challenges and future research directions

One methodological issue present in alcohol administration studies is the timing and rate of alcohol absorption. An examination into the variability in alcohol absorption during a drinking session found that the average peak BAC was 0.073 with a range of 0.047–1.00 g/dL (Winek, Wahba, & Dowdell, 1996). The average peak time, i.e. when the BAC reached its highest was 17.4 minutes, ranging from 0 to 74 minutes. Other factors such as the timing and nature of one’s last meal and participants’ gender can further impact upon absorption rate and target BAC, as seen in Hildebrand Karlén et al. (2015) where female participants reached significantly higher BACs than males.

To complicate things even further, research has shown that memory encoding and retrieval might be affected differently depending on the specific limb of the BAC curve. Söderlund, Parker, Schwartz, and Tulving (2005) showed that alcohol impaired encoding in cued and free recall and recognition of completed word fragments regardless of limb, but word recognition was only impaired during the ascending BAC. Thus, how alcohol affects our memory not only depends on the memory task utilized, but also on the specific timing of when the memory was tested, that is on the ascending or the descending limb of the BAC curve. Future studies should therefore examine the effects of alcohol on eyewitness memory performance on different limbs on the BAC curve to shed more light on when exactly episodic memory might be impaired.

Whilst there are methodological difficulties associated with the timing of the alcohol administration and absorption, there are also challenges associated with assigning appropriate control groups. Studies typically use a control group, in which participants are knowingly not consuming alcohol, and/or a placebo group, in which participants are under the impression that they are consuming alcohol but do actually not receive any. The placebo condition is useful in measuring the behavioral and cognitive effects of expecting alcohol in the absence of pharmacological effects. Eyewitness memory studies have now begun to use fully-balanced placebo designs to control for alcohol expectancies (see Flowe et al., 2019; Gawrylowicz et al., 2019). In addition, to the usual alcohol, placebo and sober control group, a fourth group is included: the reverse placebo group (individuals do not believe that they received alcohol when they actually did). Gawrylowicz et al. (2019) found that their reverse placebo group performed consistently poorer on a cued recall task, they gave fewer correct responses and made more errors compared to the alcohol, control and placebo group. Flowe et al. (2019) found a significant expectancy effect for recall completeness when participants were interviewed with the Self-administered Interview. There was no significant effect of actual alcohol consumption. These findings suggest that pharmacological effects of alcohol might not be solely responsible for differences in memory performance, but that alcohol-related expectancies play a crucial role too.

Including placebo groups can be challenging, as it requires convincing people that they did or did not receive alcohol when in fact they did or did not. In Flowe et al.’s (2019) study 27% of participants who had been told that they received tonic water thought that they drank alcohol, whereas 22% believed they had alcohol when in fact they consumed tonic. Similarly, Schreiber Compo et al. (2017) reported that 16% of placebo participants believed that they had not consumed alcohol. Even if the placebo manipulation is successful, placebo participants often report feeling less intoxicated than their intoxicated counterparts (Kneller & Harvey, 2016).

To summarize, the administration of alcohol in laboratory settings comes with a myriad of methodological challenges. These challenges range from variations in peak BACs to ensuring timing of administration is appropriate and equal. What’s more is that the inclusion of viable placebo groups is often not possible, as the drink deception is often difficult to execute, especially in the reverse placebo condition.

Subtypes of Learning Disabilities

An enormous body of research has accumulated with various approaches for identifying subtypes of LDs. Inquiry into possible subtypes of LDs began with Johnson and Myklebust (1967) and colleagues (e.g., Boshes & Myklebust, 1964; Myklebust & Boshes, 1960). They proposed a nonverbal disability characterized by the absence of serious problems in areas of language, reading, and writing, but with deficiencies in social perception, visual-spatial processing, spatial and right–left orientation, temporal perception, handwriting, mathematics, and executive functions such as disinhibition and perseveration. The neuropsychology literature later characterized a nonverbal LD as reflecting a right hemisphere deficit (Pennington, 1991; Rourke, 1989); however, it remains the least well understood.

Subtyping holds widespread appeal because it offers a way to explain the heterogeneity within the category of LDs, to describe the learning profile of students with LDs more specifically, leading to an individualized approach to intervention. Still, there is considerable debate over the existence of LDs subtypes as distinct categories that can be reliably identified, how best to categorize LDs subtypes, and whether instructional implications differ by subtype.

According to Fletcher et al. (2003), there are three main approaches to subtyping LDs: achievement subtypes, clinical inferential (rational) subtypes, and empirically based subtypes.

The first approach, achievement subtypes, relies on achievement testing profiles. For example, subgroups of reading difficulties are differentiated by performance on measures of word recognition, fluency, and comprehension. Subgroups of reading disability, math disability, and persons with a FSIQ score below 80 exhibit different patterns of cognitive attributes (Fletcher et al., 2003; Grigorenko, 2001). Compton et al. (2011) found distinctive patterns of strengths and weaknesses in abilities for several LD subgroups, in contrast to the flat pattern of cognitive and academic performance manifested by the normally achieving group. For example, students with LDs in reading comprehension showed a strength in math calculation alongside weaknesses in language involving listening comprehension, oral vocabulary, and syntax, whereas students with LDs in word reading showed strengths in math problem-solving and reading comprehension alongside weaknesses in working memory and oral language.

The second approach, clinical inferential, involves rationally defining subgroups based on clinical observations, typically by selecting individuals with similar characteristics. One example is students with a core deficit in phonological processing. The double-deficit model of subtypes distinguishes three subtypes: two subtypes with a single deficit in either phonological processing or rapid automatic naming, and a third subtype with a double-deficit in both areas (Wolf & Bowers, 1999; Wolf, Bowers, & Biddle, 2000). Another subtype model classifies two types of poor decoders: those having phonological dyslexia, orthographic/surface dyslexia, and those having mixed dyslexia, meaning weaknesses in both phonological and orthographic processing based on performance on reading nonwords and irregular exception words (Castles & Coltheart, 1993; Feifer & De Fina, 2000).

The third approach, empirically based subtypes of LDs, is based on multivariate empirical classification. It uses techniques such as Q-factor analysis and cluster analysis, and subsequent measures of external validity. Empirical subtyping models have been criticized for being atheoretical and unreliable; however, these models have provided additional support for rational subtyping methods, including the double-deficit model and differentiating garden-variety from specific reading disabilities (Fletcher et al., 2003). For example, a study by Pieters, Roeyers, Rosseel, Van Waelvelde, and Desoete (2013) used data-driven model-based clustering to identify two clusters of math disorder: one with number fact-retrieval weaknesses, and one with procedural calculation problems. When both motor and mathematical variables were included in the analysis, two clusters were identified: one with weaknesses in number fact retrieval, procedural calculation, as well as motor and visual-motor integration skills; a second with weaknesses in procedural calculation and visual-motor skills.

The identification and classification of a LD relies on either a dimensional or a categorical framework. Subtyping efforts are based on evidence that the heterogeneity within LDs is best represented as distinct subtypes. For example, reading and math LDs can be differentiated because students with reading LDs tend to have a relative strength in mathematics, whereas students with mathematics LDs tend to have a relative strength in reading (Compton et al., 2011). However, some researchers contend that the attributes of reading disability and math disability are dimensional, and efforts to categorize these as distinct subtypes are based on cut scores and correlated assessments (Branum–Martin, Fletcher, & Stuebing, 2013). Continued research is needed to advance our understanding of LD subtypes and their instructional implications for providing tailored intervention to a heterogeneous population of individuals with LDs.

3.3.1 Word-recognition accuracy and speed

Word-recognition accuracy and speed were assessed with the computerized lexical decision subtest of the ProDi-L test battery (Richter et al., 2017; see also Richter, Isberner, Naumann, & Kutzner, 2013). Children were presented with 16 words (e.g., Traktor [tractor]) and pseudowords (e.g., Spinfen) in randomized order. Their task was to decide whether the presented letter string was a real word or not by using two response keys (yes/no). Pseudowords were orthographically and phonologically legal letter strings and varied in their similarity to actual German words. Pseudowords similar to actual words were constructed by changing the first character of an existing word (e.g, Name → Bame). Pseudowords dissimilar to actual words were constructed by combining the syllables of two existing words with irregular spellings. For example, the pseudoword Chilance was constructed by combining the first syllable of the word Chili and the second and third syllables of the word Balance. The pseudowords also included pseudohomophones (1–3 per measurement point), which sound like real words but have a different orthographical form (e.g., Heckse instead of Hexe/witch). These items cannot be solved via the application of phoneme-grapheme translation rules but require direct word recognition via the lexical route. Seven items in the first measurement point, nine items in the second measurement point and eight items in the last two measurement points were regular and irregular real German words. Different but parallel words and pseudowords were used at all four measurement points. Apart from the slight difference in the proportions of words and pseudowords in the first and second item set (which was due to an error), the item sets were strictly parallelized according to the item features, mean accuracy and mean response time of each item set which were obtained in another cross-sectional study (Richter et al., 2013).

The word stimuli were systematically varied in frequency and number of orthographical neighbors. They had an average frequency of 1.25 (SD = .87), retrieved from the CELEX German lemma lexicon (metric: Mannheim written frequency, logarithmic; Baayen, Piepenbrock, & Gulikers, 1995; Baayen, Piepenbrock & van Rijn, 1993), an average length of 5.62 (SD = 1.56) characters and on average 1.75 (SD = 2.46) orthographical neighbours. The pseudowords were matched in length and frequency to the word stimuli. Pseudowords were based on words with an average frequency of 1.03 (SD = .66), they had an average length of 6.31 (SD = 2.16) characters and on average 1.69 (SD = 3.25) orthographical neighbours. In order to examine whether words and pseudowords differed in frequency, length, and orthographical neighbours across the measurement points we ran three separate analyses of variance. The results indicated no significant differences (for all comparisons, p > .17) between words and pseudowords across the measurement points. These results suggest that largely parallel items were used at each measurement point.

Following the ProDi-L manual, two criteria were applied to identify and remove outliers. Logarithmic latencies that were three standard deviations below or above the mean logarithmic latency for the item in the norming sample were coded as missing. The idea behind this criterion is that very short response times are likely to indicate an irregular response, such as clicking through items without reading them, and thus they should not be included in further analyses. Likewise, very long response times are likely due to disturbances, mind wandering, etc. Furthermore, for each child, response times that deviated more than two standard deviations from the average of the individual logarithmic response times were also coded as missing. Further data preparation was performed separately for each measurement point according to the procedure reported by Karageorgos et al. (2019). The sum of correct responses was transformed into proportions representing word recognition accuracy. Furthermore, a words-per-minute score was calculated as an indicator of word-recognition speed. The number of correct and incorrect responses to words and pseudowords was multiplied by 60 000 ms and then divided by the overall latency across all items measured in ms. A child, for example, who responded to 10 items in 10 000 ms received a score of 60 words per minute. Words-per-minute scores were not computed for participants with more than 10 % missing values (due to the outlier removal criteria discussed above) at the relevant measurement point. Thus, word-recognition speed scores were missing for 449 of 4380 data points. The test-retest reliability between measurement points was computed as the intraclass correlation of word-recognition scores at the end of each school year for a total of 692 children (those with complete data sets) using the R-package irr (Gamer, Lemon, Fellows, & Singh, 2019). A two-way mixed-effects model for mean rating and absolute agreement was used for computing the ICC (Koo & Li, 2016; McGraw & Wong, 1996; Price et al., 2015). According to the interpretation guidelines proposed by Ciccheti (1994), the estimated test-retest reliability (i.e., stability) was good for the accuracy score, ρ_I = .624, F(691, 32.9) = 3.48, p < 0.001, 95 % CI [.40, .75], and fair for the words-per-minute score, ρ_I = .50, F(543, 6.7) = 4.46, p = 0.005, 95 % CI [.01, .73].

3.1.1 Logogens and imagens

Morton’s (1979) word-recognition model uses modality-specific visual and auditory input logogens and output logogens to account for recognition performance. In DCT, the variety of modality-specific representations expanded to include auditory, visual, haptic, and motor logogens, as well as separate logogen systems for the different languages of multilingual individuals. Moreover, DCT treats logogens as hierarchical sequential structures of increasing length, from phonemes (or letters) to syllables, conventional words, fixed phrases, idioms, sentences, and longer discourse units – anything learned and remembered as an integrated language sequence. Imagens also are multimodal (visual, auditory, haptic, motor) representations organized hierarchically into spatial nested sets that are most apparent in visual objects – for example, we see pupils within eyes within faces within rooms within houses within larger scenes, and so on. Importantly, the modality–specificity of logogens and imagens excludes abstract mental representations such as propositions. Thus the functional domains associated with stimulus meaning and cognitive abilities are conceptualized entirely in terms of modality specific logogens and imagens.