Annals of Dyslexia

, Volume 57, Issue 1, pp 3–32

Reading development subtypes and their early characteristics

Authors

    • Department of PsychologyUniversity of Jyväskylä
  • Asko Tolvanen
    • Department of PsychologyUniversity of Jyväskylä
  • Anna-Maija Poikkeus
    • Department of Teacher EducationUniversity of Jyväskylä
  • Kenneth Eklund
    • Department of PsychologyUniversity of Jyväskylä
  • Marja-Kristiina Lerkkanen
    • Department of Teacher EducationUniversity of Jyväskylä
  • Esko Leskinen
    • Department of Mathematics and StatisticsUniversity of Jyväskylä
  • Heikki Lyytinen
    • Department of PsychologyUniversity of Jyväskylä
Article

DOI: 10.1007/s11881-007-0003-0

Cite this article as:
Torppa, M., Tolvanen, A., Poikkeus, A. et al. Ann. of Dyslexia (2007) 57: 3. doi:10.1007/s11881-007-0003-0

Abstract

The present findings are drawn from the Jyväskylä Longitudinal Study of Dyslexia (JLD), in which approximately 100 children with familial risk of dyslexia and 100 control children have been followed from birth. In this paper we report data on the reading development of the JLD children and their classmates, a total of 1,750 children from four measurement points during the first two school years. In the total sample, we examined whether heterogeneous developmental paths can be identified based on profiles of word recognition and reading comprehension. Secondly, we studied what kind of early language and literacy skill profiles and reading experiences characterize the children with differing reading development in the follow-up sample. The mixture modeling procedure resulted in five subtypes: (1) poor readers, (2) slow decoders, (3) poor comprehenders, (4) average readers, and (5) good readers. The children with familial risk for dyslexia performed on average at a lower level in all reading tasks than both their classmates and the controls, and they were overrepresented in slow decoders subtype. Differences between the subtypes were found in the early language and literacy skill development, as well as in the reading experiences of the reading subtypes.

Keywords

Reading subtypesFamilial dyslexia riskLanguage developmentSimple view of reading

Introduction

Learning to read is one of the most important goals of the first school years. The simple view of reading (Gough & Tunmer, 1986; Tunmer & Hoover, 1992) defines reading ability as a function of decoding and comprehension skills, but at the early phases of learning to read, word recognition and reading comprehension are difficult to separate. After children become more fluent readers, reading comprehension emerges more clearly as a closely related but separate skill from basic word recognition ability (e.g., Catts, Hogan, & Fey, 2003; Nation, 2005; Storch & Whitehurst, 2002). In terms of the profiles of children’s reading skill, the relationship between word recognition and reading comprehension manifests as four reading subtypes: children without reading difficulties, children with primary problems in word recognition, children with primary problems in comprehension, and children with both word recognition and comprehension difficulties (often regarded as garden-variety poor readers) (e.g., Catts et al., 2003; Leach, Scarborough, & Rescorla, 2003; Nation, Clarke, Marshal, & Durand, 2004; Shankweiler et al., 1999). The prevalence of reading difficulty subtypes varies greatly depending on the age and characteristics of the sample. Children with reading difficulties most typically show difficulties in word recognition, and those children for whom poor comprehension skills combine with average word recognition comprise only about 6 to 15% of the poor readers (Catts et al., 2003; Leach et al., 2003; Nation et al., 2004; Shankweiler et al., 1999).

Empirical work on reading difficulty subtypes deriving from the simple view of reading model has typically employed cut-off scores of word recognition and reading comprehension (Catts et al., 2003; Leach et al., 2003; Nation et al., 2004; Shankweiler et al., 1999). Because word recognition skills and reading comprehension correlate highly, particularly in the early school years (e.g., Catts et al., 2003; Shankweiler et al., 1999), the stability and generality of the reading difficulty subtypes suggested thus far is unclear. In two recent reading type classifications on random samples of Finnish second graders, only three reading subtypes were identified (Lerkkanen, Rasku-Puttonen, Aunola, & Nurmi, 2004; Poskiparta, Niemi, Lepola, Ahtola, & Laine, 2003): one with reading comprehension difficulties, a second with both word recognition and reading comprehension difficulties, and a third with no reading difficulties. In contrast to previous studies, these data did not support a reading subtype with only word recognition difficulties. This finding may, however, be due to the close link between word recognition and reading comprehension after only 2 years of formal school instruction, to sample size, or to statistical power issues regarding the traditional cluster analysis employed by the researchers. Due to the relatively low incidence of dyslexia, it is also difficult to identify children with only word recognition difficulties from a small random sample.

The first aim of the present study is to identify reading difficulty subtypes by using a design and methodology that enables us to overcome some of the limitations of earlier studies. First, small sample sizes have been an issue in many of the previous studies. In this study, we examine the development of word recognition and reading comprehension in a large sample (n = 1,750) of Finnish first and second graders, a sample that comprises a total of 93 school classes. Second, most of the previous reading subtype classifications have been based on single time points, and longitudinal follow-up of the reading subtype stability has rarely been carried out (e.g., Lerkkanen et al., 2004). In this study, we identify children based on the analysis of heterogeneity in their developmental paths in word recognition and reading comprehension during the first two grades and across several times of measurement. Third, the effect of classroom membership has not previously been addressed in studies of reading difficulty subtypes although the relevance of its effect on children’s learning has been demonstrated (McCoach, O’Connell, Reis, & Levitt, 2006; Rowan, Correnti, & Miller, 2002). Because our data consist of whole classrooms, we were able to assess and control for the classroom effect in the design. Fourth, we were able to avoid the limitations of the more traditional cut-off score-type reading analysis by using a more advanced method of subgroup identification (mixture modeling, Muthén, 2001). The advantages of mixture modeling over the use of cut-off criteria, traditional cluster analysis, and I-states as objects analysis (ISOA) analysis (used in Lerkkanen et al., 2004; see Bergman & El-Khouri, 1999) are that it allows for the isolation of measurement error, basing the subgrouping on several measurements instead of relying on one measurement occasion. It also provides statistical tests for selection of the best solution. Thus, mixture modeling affords a much more powerful tool for the estimation of whether different reading profiles are apparent in the present data. In addition, the explicit inclusion of children with high familial risk for dyslexia raises the probability of identifying children with word recognition difficulties.

The second aim of the present study is to examine the extent to which children following different reading development profiles manifest differences in early skill development and reading experiences. These analyses are based on the detailed longitudinal information available for the 200 children with and without familial risk for dyslexia (included within the total sample of 1,750 children) who have been followed since birth in the Jyväskylä Longitudinal Study of Dyslexia (JLD). Some earlier reading difficulty subtyping studies have included language and literacy predictors from Kindergarten-age onwards (Catts et al., 2003; Leach et al., 2003; Poskiparta et al., 2003), and others have shown that there are different language and cognitive profiles behind reading disability at school age (e.g., Morris et al., 1998). Less is known about the earlier language development of children with varying reading profiles, and in particular, little is known about the early developmental paths leading to poor comprehension (Nation et al., 2004).

Reading difficulty subtype studies and those focusing on predictors of continuous reading outcomes both suggest that the early risk factors for difficulties in word identification and/or reading comprehension difficulties are partially different (e.g., Bishop & Snowling, 2004; Catts et al., 2003, 2005; Leach et al., 2003; Muter, Hulme, Snowling, & Stevenson, 2004; Nation & Nordbury, 2005; Oakhill, Cain, & Bryant, 2003; Phillips & Lonigan, 2005; Tunmer & Hoover, 1992). Letter knowledge and phonological awareness have consistently been identified as the best proximal predictors of future word recognition skills (e.g., Adams, 1990; Byrne, 1998; Elbro, Borstrom, & Petersen, 1998; Gallagher, Frith, & Snowling, 2000; Lonigan, Burgess, & Anthony, 2000; Pennington & Lefly, 2001; Scarborough, 2001; Snow, Burns, & Griffin, 1998; Vellutino, Fletcher, Snowling, & Scanlon, 2004). In addition, especially in highly consistent orthographies such as Finnish or German, relatively strong predictive associations have been observed also between word reading and naming speed (e.g., Holopainen, Ahonen, & Lyytinen, 2001; Wimmer, Mayringer, & Landerl, 1998, 2000; Wimmer & Mayringer, 2002). The close association between rapid naming and reading skill in these languages can be understood through the transparency of the writing system. In Finnish, for example, the writing system consists of only 24 grapheme–phoneme combinations and every word can be read through reliance on this highly bidirectionally consistent phonological strategy. This makes the acquisition of basic reading accuracy a quick and easy process for beginning readers (e.g., Seymour, Aro, & Erskine, 2003). The ease with which reading accuracy can be attained within a transparent language has turned the focus towards identification of reading difficulties with measures of reading fluency.

Individuals with poor reading comprehension skills (in the absence of word recognition difficulties) do not typically have difficulties with phonological awareness tasks but they perform below average on a wide range of oral language measures, especially in tasks that tap vocabulary, listening comprehension, semantics, and morphosyntax (e.g., Cain et al., 2004; Nation et al., 2004; Roth, Speece, & Cooper, 2002; Stothard & Hulme, 1995; Sénéchal & LeFevre, 2002; Storch & Whitehurst, 2002). In addition, poor comprehension has often been associated with deficits in short-term memory and problems in “higher-level” skills such as poor inference making and comprehension monitoring (e.g., Cain et al., 2004). Some poor comprehenders also perform below average in general cognitive ability (Nation et al., 2004).

Although the language problems of children with word recognition difficulties predominantly reside within the domain of phonological language skills, variation in word recognition has been found to be predicted also by oral language abilities, such as vocabulary knowledge (e.g., Catts, Fey, Zhang, & Tomblin, 1999; Nation & Snowling, 2004). Prospective follow-up studies of children with familial risk for dyslexia support this finding; in addition to the most obvious measure, phonological awareness, children at risk for dyslexia typically perform on average at a lower level than control children in letter knowledge, rapid serial naming (RAN), and tasks measuring vocabulary (e.g., Elbro et al., 1998; Gallagher et al., 2000; Lyytinen et al., 2004; Lyytinen et al., 2006a, 2006b; Scarborough, 1990; Snowling, Gallagher, & Frith, 2003). Interestingly, Catts et al. (2003) reported that, in their sample, poor comprehenders performed as poorly as children with word recognition difficulties in phonological awareness and rapid naming. This finding may, however, be explained by the poor comprehenders’ difficulties in understanding the task demands (Nation, 1999).

Previously, we (Lyytinen et al., 2006b) have examined the heterogeneous profiles of the JLD children’s language and literacy skills and their association with early reading and spelling ability. The results suggested that there are at least three routes to difficulties in reading acquisition, with the most explicit routes being characterized by problems in phonological awareness, naming speed, and letter knowledge. In the analyses by Lyytinen et al. (2006b), the reading outcome was a composite of fluent word recognition and spelling measures, and the focus was on heterogeneous paths of early language and literacy development. In the present analyses, however, we focus on the reading development profiles in word recognition and reading comprehension and examine retrospectively the early language and literacy development. The seven skill domains selected to represent relevant early predictors of reading are: (1) receptive and (2) expressive language skills, (3) inflectional morphology skills, (4) memory, (5) retrieving words efficiently from memory (naming speed), (6) letter knowledge, and (7) phonological awareness skills. Furthermore, we will examine the effects of familial risk for dyslexia and the amount of reading experiences on variation in reading development. The amount of reading experiences has previously been found to be higher among children with better reading skills (e.g., Leach et al., 2003; Leppänen, Aunola, & Nurmi, 2005; Scarborough, Dobrich, & Hager, 1991; Sénéchal & Lefevre, 2002).

More specifically, we ask: (1) can heterogeneous developmental paths of reading development during the first two school years (early reading subtypes) be identified in a longitudinal sample of children with familial risk for dyslexia, their controls, and classmates? (2) What kind of early language and cognitive skill profiles and reading experiences characterize children in the JLD follow-up with differing reading development?

Based on the evidence that word recognition and reading comprehension skills already diverge in children after a few years of formal school attendance (e.g., Catts et al., 2003; Leach et al., 2003; Lerkkanen et al., 2004; Nation et al., 2004, Poskiparta et al., 2003; Shankweiler et al., 1999), we expected to find groups of children with differing development. Children with familial risk for dyslexia were expected to be overrepresented among children with reading problems (Elbro et al., 1998; Finucci, Guthrie, A.L. Childs, Abbey, & B. Childs, 1976; Gilger, Pennington, & DeFries, 1991; Hallgren, 1950; Lyytinen et al., 2004; Olson, Datta, Gayán, & DeFries, 1999; Pennington & Lefly, 2001; Scarborough, 1990). By definition, children at risk for dyslexia have a high propensity for difficulties in word recognition in particular, but difficulties in decoding may also be reflected in reading comprehension, especially during the early school years. Difficulties in word recognition were expected to be related most strongly with slow early development of phonological awareness, letter knowledge, and rapid naming. Reading comprehension difficulties, on the other hand, were expected to be related with slow early development, particularly in vocabulary, morphological awareness, verbal IQ, and short-term memory (Cain et al., 2004; Nation et al., 2004; Nation, Clarke, & Snowling, 2002). Finally, the amount of reading practice was expected to be higher among proficient readers than children with any type of reading difficulties (Leach et al., 2003; Sénéchal & LeFevre, 2002).

Method

Participants

Data for this study were drawn from the JLD, a prospective follow-up of children from birth to school age. The JLD seeks to identify early language development and precursors of dyslexia (for the most recent review of results, see Lyytinen et al., 2004). From four successive age cohorts of families invited for screening, a total of 214 families from the city of Jyväskylä and its surrounding communities in the province of Central Finland joined the study prior to the birth of their children. Half of the participating families include a parent who has been diagnosed with dyslexia and who also reports similar problems among immediate relatives. The children from these families are referred to as the at-risk group. The control group comprises children from families whose parents gave no personal or familial report of reading or spelling difficulties. Parents also underwent extensive cognitive and literacy-based assessment (see Leinonen et al., 2001 for full details). In terms of distribution, the level of parental education is representative of the Finnish population and did not differentiate the at-risk and control groups. All the children are native Finnish speakers and have no mental, physical, or sensory difficulties. At the present stage of the JLD study, the youngest of the four cohorts has now completed the third-grade assessments, The attrition rate has remained low, with 199 of 214 families continuing to participate in the project from the beginning to the end of third grade.

When the JLD children entered school (in the year they turned seven years of age), group assessments of the first two of the four age cohorts were conducted in schools for the whole classrooms. The third and fourth JLD age cohorts were assessed individually (n = 82). Altogether 1,803 children participated in the group tests. Group tests were administered twice during the first grade (November and April) and twice during the second grade (November and April). Children who participated in only one of the four group assessment sessions were excluded from the data. The data used in the present analyses comprised 1,750 children from 93 classes. Of these children, 191 were JLD follow-up participants.

Among the classmates’ data, there were missing data as we controlled for the presence of the JLD follow-up participants but not for the other children on the group assessment day. Complete data from all four assessment times were available for 1,147 children (997 classmates, 85 at-risk, 65 controls). The most typical missing value pattern resulted from either a JLD follow-up participant’s change of classroom after the first grade or new children entering the classroom at the beginning of the second grade (for whom there were no data on the first-year assessments). Because of these two reasons, information was available only for the first-grade assessments (T1 and T2) for 100 children and for the second-grade assessments (T3 and T4) for only 152 children. It should be noted that stability was high in the classroom membership: 1,448 children remained classmates through the first two grades. Other types of missing value patterns concerned less than 3% of the children. There were no teacher changes within a school year in any of the classrooms, but the teacher did change between the first and second grade in the case of 114 children. In the mixture modeling, missing data were handled through missing value implementation in Mplus.

Measures and procedure

Reading skills

Fluent word recognition was assessed four times and reading comprehension three times (i.e., all other times except for the fall semester of first grade). The tester was either a trained graduate student or a classroom teacher. Detailed information and consultation was provided for the teachers. Reading skills were assessed in a group test format during normal classroom time. Scoring was conducted by the research team and classroom teachers received feedback of the test results concerning their own classroom. Descriptive statistics and correlations of the reading skills are reported in Table 2.

Fluent word recognition

The child’s task was to match a picture with a correct word from among four alternative words. The test consisted of 80 items. In the first grade, children had 5 min, and in the second grade, 2 min, to complete the task. Items were identical at time points 1 and 4, and at time points 2 and 3. This task forms part of a standardized national reading achievement test battery (Lindeman, 2000).

Reading comprehension

In April of the first grade, a 12-item sentence comprehension task was used (Lerkkanen, Poikkeus, & Ketonen, 2006). The test comprised 12 pictures, each with a set of four sentences describing the events depicted in the picture. Within each sentence set, there was always one sentence with a word that did not match with the picture. The child’s task was to identify the word that did not fit. Two points were awarded for each correct identification with one point granted if the child had marked several words but included the correct identification. The task had a time limit of 10 min.

Text comprehension tasks from a national reading achievement test battery (Lindeman, 2000) were used for the second-grade assessments. The text for the fall semester of second grade was a nonfiction text concerning judo (146 words in length). After reading the text, the child was asked to answer 11 multiple choice items and one item that required the child to arrange seven statements in the correct sequence based on the information gathered from the text. The text for the spring semester of the second grade was a nonfiction text concerning gymnastics (114 words in length). There were 11 multiple choice items and one item that required the child to arrange five statements in the correct sequence based on the information gathered from the text. One point was awarded for each correct answer, resulting in a maximum of 12 points. Children completed the task at their own pace, but the maximum time allotted was one lesson (45 min).

Early skills

A theoretical basis underpinned the selection of the core skill domains and the identification of the key measures in each skill area evolved from empirical analysis of the data. To ensure high reliability of the measures, composite means of separate measures were formed where possible. The composites for each domain were formed by calculating a grand mean of the standardized scores. The standardization of each measure was based on the JLD follow-up control group children’s mean and standard deviation. In Table 1, we report the descriptive statistics in each domain of early skills and reading experiences.
Table 1

Descriptive statistics of standardized values in early skills and reading experiences (available only for the JLD follow-up sample)

 

n

Min

Max

M

SD

Skew.

Kurt.

Receptive language

1–1.5 years

190

−1.55

2.50

−0.08

0.81

0.59

−0.03

2–2.5 years

186

−2.88

3.03

−0.16

1.08

0.46

0.84

3.5 years

171

−1.87

2.32

−0.10

0.98

0.35

−0.35

5 years

187

−2.59

2.17

−0.20

1.12

−0.12

−0.81

Expressive language

1–1.5 years

192

−1.08

3.74

−0.06

0.78

1.87

4.58

2–2.5 years

191

−2.90

2.11

−0.04

0.84

−0.51

0.67

3.5 years

190

−2.27

2.94

−0.24

1.01

0.61

0.17

5.5 years

186

−3.98

2.19

−0.24

1.15

−0.67

0.39

Morphology

3.5 years

162

−2.02

3.02

−0.20

1.04

0.51

−0.31

5 years

191

−3.29

1.74

−0.25

1.05

−0.53

−0.09

Phonological awareness

3.5 years

191

−2.21

1.06

−0.22

0.70

−0.43

−0.32

4.5 years

191

−2.15

1.55

−0.17

0.73

−0.11

−0.17

5.5 years

191

−3.71

2.12

−0.25

0.93

−0.09

0.76

6.5 years

192

−2.79

1.32

−0.19

0.93

−0.29

−0.57

Letter knowledge

3.5 years

182

−0.61

2.83

−0.10

0.89

2.18

3.86

4.5 years

190

−1.20

1.80

−0.20

0.98

0.63

−1.04

5.0 years

192

−1.59

1.50

−0.22

1.04

0.24

−1.42

5.5 years

191

−2.25

1.19

−0.29

1.14

−0.25

−1.37

6.5 years

192

−3.60

0.86

−0.28

1.17

−0.95

−0.14

Memory

3.5 years

184

−2.83

1.96

−0.12

0.89

−0.06

−0.43

5.0 years

192

−2.03

2.19

−0.12

0.77

0.12

−0.21

6.5 years

188

−2.67

2.92

−0.02

1.07

0.23

−0.02

RANa

3.5 years

167

−3.05

1.91

0.07

0.96

1.07

0.90

5.5 years

189

−7.13

2.09

0.36

1.47

1.90

5.28

6.5 years

192

−6.70

1.57

0.47

1.45

1.91

4.58

Verbal IQ

8.0 years

190

−2.38

2.64

−0.18

0.89

0.19

−0.11

Performance IQ

8.0 years

190

−3.17

2.88

−0.03

1.12

−0.12

−0.23

Shared reading

2 years

174

−3.28

1.62

−0.05

1.14

−0.46

−0.32

4 years

180

−2.67

2.91

−0.08

1.04

0.10

0.39

5 years

180

−2.61

3.48

0.13

1.13

0.06

0.23

6 years

180

−2.42

2.98

0.01

1.00

0.06

0.23

7 years

162

−2.29

3.02

0.02

1.04

0.00

0.13

8 years

156

−1.63

3.13

0.03

0.98

0.02

−0.31

Reading alone

4 years

191

−2.40

2.89

0.17

1.17

0.18

−0.25

5 years

191

−2.29

2.73

0.15

1.11

0.18

−0.25

6 years

174

−1.93

3.09

0.10

0.99

0.05

−0.59

7 years

171

−2.01

1.53

0.04

0.94

−0.26

−0.96

8 years

162

−2.42

1.79

0.02

1.03

−0.10

−0.49

Values are standardized (to the JLD control group’s distribution) because most of them are composites of several measures.

aCoding of RAN: the faster the time, the higher z value.

It should be noted that measures that were based on international tests were administered using a Finnish translation of the test with language-specific modifications where necessary. In most cases, we used tests with existing Finnish norms. Examples of these include the Wechsler Intelligence Scale for Children third edition (WISC-III) (Wechsler, 1991, 1999), the MacArthur Communicative Development Inventory (MCDI) by Fenson, Dale, Reznick, Bates, Thal, and Pethick (1994) with Finnish norms reported in Lyytinen (1999), and the Boston Naming Test (BNT) by Kaplan, Goodglass, and Weintraub (1983) with Finnish norms reported in Laine et al. (1993) and Laine, Koivuselkä-Sallinen, Hänninen, and Niemi (1997).

Receptive language

In 12 to 18 months: A composite mean score of three separate measures was used: vocabulary comprehension at 12 and 14 months using the Finnish adaptation (Lyytinen, 1999) of the infant and toddler forms of the MCDI (Fenson et al., 1994) and verbal comprehension at 18 months [Reynell Developmental Language Scales (RDLS; Reynell & Huntley, 1987)] (Cronbach standardized α = 0.78). The MCDI provides the total number of words comprehended by the child as identified from a checklist by the parents. The RDLS provides a score for the child’s ability to take appropriate action in response to verbal instructions from the tester concerning test materials. In 2.5 years: the RDLS was administered to the children for a second time. In 3.5 and 5.0 years: measures were derived from the Peabody Picture Vocabulary Test—Revised (PPVT; Dunn & Dunn, 1981). The PPVT requires the child to select from four alternatives the picture that correctly depicts a spoken word.

Expressive language

In 12 to 18 months: A composite mean score (Cronbach standardized α = 0.93) based on six separate measures was used: vocabulary production at 12, 14, and 18 months; mastery of inflections, maximum sentence length, and expressive language at 18 months. The first five measures were derived from parental reports obtained using the toddler form of the MCDI. The vocabulary production score was derived from the total number of words checked by the parent as being both comprehended and produced by the child. Mastery of inflections was based on the 16-item grammar section of the Finnish toddler MCDI in which the parents reported on their child’s use of suffixes (e.g., “Have you heard your child use the plural form?”). This provides an early measure of the child’s command of morphological rules as observed by the parents but differs from the later tasks of morphological knowledge elicited directly from the child that either require identification of a target picture representing a certain morphological concept or repetition of the target word with the correct ending and correct modifications to the stem of the word. Maximum sentence length was derived by asking parents to write down three of the longest utterances made recently by the child. The mean number of morphemes used by each child in these utterances was computed. The sixth measure was derived from administration of the RDLS expressive language items (vocabulary, content, and structure) at 18 months of age.

At 2.0 and 2.5 years: This composite mean (Cronbach standardized α = 0.92) consisted of seven scores: vocabulary production, mastery of inflections, and maximum sentence length at both 2.0 and 2.5 years, and expressive language at 2.5 years. The first six measures were obtained from parental reports using the toddler form of the MCDI. The seventh measure was obtained by repeating the administration of the RDLS at 2.5 years of age.

At 3.5 and 5.6 years: The BNT (Kaplan et al., 1983) was used to obtain a measure of expressive language at 3.5 and 5.6 years of age. The BNT entails confrontational naming of 60 pictured items. The score is deduced from the correct number of items, either named spontaneously by the child or following receipt of a semantic stimulus cue.

Morphological knowledge

At 2.5 years: This measure was derived from the administration of a 12-item morphology test that taps comprehension of adjectival, noun, and verb inflections. Each item includes a set of three colored pictures: one introduces the target word in its basic form, one depicts an object to which the inflected target word applies, and one is a distracter (e.g., “This house is this high, show me the picture where the house is even higher.”). One point for each correctly comprehended item was allocated. At 3.5 and 5.0 years: Mastery of inflectional morphology was assessed using a Berko-type elicitation test (Lyytinen et al., 2001; P. Lyytinen & Lyytinen, 2004) covering inflections of adjectives, verbs, and nouns. The target words were ancient, now obsolete Finnish words (and thus are unknown to children) of two to four syllables. The words were presented orally together with a colorful drawing and the child is instructed to generate the inflection of the target word. The sum score of correct inflections was used in the current analyses (Cronbach standardized α = 0.79).

Phonological awareness

The composites of phonological awareness were derived from measures of word and/or segment identification, synthesis, continuation, phonological processing, and initial phoneme identification and production. The word and segment identification, synthesis, and continuation tasks were computer-based: the child’s responses were recorded automatically via touch screen or the child’s oral responses were coded online by the experimenter and recorded in digital sound files (for details, see Puolakanaho, Poikkeus, Ahonen, Tolvanen, & Lyytinen, 2003). In the word-level identification task, the child was presented with three pictures of objects on a screen, immediately followed by the name of each object [all compound words, e.g., lentokone (airplane), soutuvene (row boat); polkupyörä (bicycle)]. The child was then asked to identify the picture containing a specified part of the compound (e.g., “In which picture can you hear the sound ‘kone’ (plane)?”). At the segment identification level, the task was the same but with the requirement to identify subword level units (syllables or phonemes) within the target [e.g., “koi” in the word “koira” (dog)]. In the synthesis task, the child was presented with segments (syllables or phonemes) each separated by 750 ms. The task requirement was to blend the segments to produce the resulting word [e.g., per-ho-nen (butterfly) or m-u-n-a (egg)]. In the continuation task, the child was presented with the onset of a “secret” word and asked to guess how the word would continue (e.g., “mu-?”). Only continuations that were meaningful words were coded as correct. Phonological processing comprised part A of the NEPSY (Korkman, Kirk, & Kemp, 1998), whereby the child was required to select one picture from three on the basis of its containing an orally presented segment (e.g., “-ap-”). The initial phoneme identification task entailed the child being shown four pictures of objects with simultaneous presentation of the object name. The child was then required to select the correct picture on the basis of the oral presentation of a subsequent initial phoneme relating to one target (e.g., “In the beginning of which word do you hear ____?”). In the initial phoneme production task, the child was first required to name a picture of an object and then to articulate the first sound (phoneme or letter name) of the object. At 3.5 years, the composite (Cronbach standardized α = 0.59) included five measures, namely, word identification, segment identification, synthesis, continuation, and phonological processing. At 4.5 years, the composite (Cronbach standardized α = 0.71) included five measures, namely, segment identification, synthesis, continuation, initial phoneme identification, and initial phoneme production. At 5.5 and 6.0 years, the composites for each age (Cronbach standardized αs were 0.77 and 82, respectively) were derived from four tasks: segment identification, synthesis, initial phoneme identification, and initial phoneme production.

Letter knowledge

At 3.5 years of age, the child was presented 16 uppercase letters; at ages 4.5, 5, and 5.5 years, 23 uppercase letters; and at age 6.5 years, 29 upper- and lowercase letters, but only the latter were used in the analyses because uppercase letters no longer produced much variance. The child received one point for each correct response (use of a phoneme or a letter name were both coded as correct responses). Letters were presented in four sets (6 + 6 + 6 + 5 letters) and testing was discontinued if the child failed to correctly name any items in a set of letters, except for at age 6.5 years when all 29 letters were presented. The testing always began by presenting the child the letter that was expected to be most familiar to him or her, i.e., the first letter of his or her own first name. The remaining letters were then presented in a fixed order, following the order in which letters are typically taught to Finnish first graders.

Memory

At age 3.5 years, memory was assessed with forward digit span (see Gathercole & Adams, 1994) and with sentence repetition (NEPSY; Korkman et al., 1998). Cronbach standardized α was 0.67. The memory assessments at ages 5 and 5.5 years were combined and included a digit span task (computer presented), a syllable span task (computer presented), a sentence repetition task (NEPSY), and a nonword repetition task (NEPSY). Cronbach standardized α was 0.75. At age 6.5 years, memory was assessed with forward digit span.

Rapid serial naming

Traditional RAN paradigms (Denckla & Rudel, 1974, 1976) were used (except for age 3.5 years, when a matrix of only 30 items was used): naming objects at age 3.5 years; naming objects and colors at age 5.5 years; and naming objects, colors, numbers, and letters at age 6.5 years. Cronbach standardized α was 0.80 at age 5.5 years and 0.89 at age 6.5 years.

Intelligence

Verbal and performance intelligence quotients were obtained at age 8 years with a short-form administration of WISC-III (Wechsler, 1991).

Reading experience

When the children were 2, 4, 5, 6, 7, and 8 years old, parents completed a reading models questionnaire whereby the amount of reading experience received by a child was estimated.

Shared reading

To form a composite score of shared reading, we employed parental reports both on the frequency and on the amount of time of the children’s home reading activities. At the age of 2 years, the shared reading measure consisted of a mean composite score of four items: (1) mother reads to the child, (2) father reads to the child, (3) amount of picture book reading, and (4) the typical duration of a reading episode when the child is reading with an adult. Parents responded to the first three items using a four-point scale (1 = not at all/seldom...4 = daily) and to the fourth item with a three-point scale (1 = less than 5 min/day...3 = longer than 15 min/day).

A mean composite score of the following four items covered shared reading at the ages of 4, 5, 6, 7, and 8 years: (1) mother reads to the child, (2) father reads to the child, (3) the typical duration of a reading episode when the child is reading with an adult, and (4) the total time per day that the child spends reading a book with an adult. Parents responded to the first two items using a five-point scale (1 = not at all/seldom...5 = several times a day) and to the third and fourth items using a three-point Likert scale (1 = less than 15 min/day...3 = longer than 45 min/day).

Reading alone

The amount of children’s solitary reading activities was calculated from three questions at the ages of 4, 5, 6, 7, and 8 years: (1) how often the child reads alone, (2) the typical duration of a reading episode when the child is reading alone, and (3) the total time per day that the child spends reading a book alone. Parents responded to the first item using a five-point scale (1 = not at all/seldom...5 = several times a day) and to the second and third items using a three-point Likert scale (1 = less than 15 min/day...3 = longer than 45 min/day).

Results

Table 2 presents the correlation coefficients, descriptive statistics, and intraclass correlations of fluent word recognition and reading comprehension at each time point. As shown in Table 2, the mean performance in fluent word recognition (number of words decoded per minute) increased rapidly during the 5 months between assessments waves at T1 and T2, but not so rapidly between the last two assessments during the second grade. The stability of fluent word recognition was already very high from the first measurement occasion onwards, but the stability of reading comprehension was not as high as shown by the correlations in Table 2. Fluent word recognition was closely linked with reading comprehension and its correlations with reading comprehension were similar to those among reading comprehension measures across time. Note that the reading comprehension measure at T2 was based on an illustrated sentence comprehension task, whereas at T3 and T4 it was based on reading text and responding to multiple-choice items. As Table 2 shows, the correlations between the sentence level comprehension task and the text level comprehension tasks were almost identical to those between the two text comprehension tasks. These three tasks also loaded into a single comprehension factor in confirmatory factor analysis.
Table 2

Correlation matrix, descriptive statistics, and intraclass correlations for the measures of fluent word recognition and reading comprehension among JLD follow-up sample and their classmates

 

Fluent word recognition

Reading comprehension

T1

T2

T3

T4

T2

T3

T4

Fluent word recognition

       

 T1: 1st grade fall

       

 T2: 1st grade spring

0.82

      

 T3: 2nd grade fall

0.68

0.78

     

 T4: 2nd grade spring

0.67

0.78

0.81

    

Reading comprehension

       

 T2: 1st grade spring

0.57

0.64

0.55

0.52

   

 T3: 2nd grade fall

0.48

0.52

0.47

0.46

0.53

  

 T4: 2nd grade spring

0.43

0.44

0.40

0.42

0.52

0.56

 

Range

0–16

0.80–16

0.50–33

4.50–31

0–24

0–12

0–12

Mean

5.22

8.80

13.90

14.91

14.33

7.83

8.79

SD

3.44

3.54

4.51

4.38

7.30

2.63

2.55

n

1,520

1,520

1,562

1,526

1,518

1,555

1,503

Skewness

1.014

0.340

0.328

0.338

−0.346

−0.582

−0.848

Kurtosis

0.510

−0.674

0.180

−0.145

−1.110

0.016

0.149

Intraclass correlation

0.05

0.07

0.10

0.09

0.04

0.06

0.05

For all correlation coefficients p value < 0.001

Reading skill comparisons of study groups

The present data sample included three subsamples: the JLD follow-up at-risk and control children and their classmates. Table 3 provides the means, standard deviations, sample sizes, and group comparison results of the subsamples. On average, the JLD at-risk children performed at a lower level in fluent word recognition tasks than the JLD control children or their classmates. The JLD control children outperformed the JLD at-risk children and the classmates at the first-grade assessment points. However, at the second-grade assessments, they performed on average at the same level as classmates in fluent word recognition. In reading comprehension, the at-risk children performed at a lower level than the control children and their classmates at the end of the first grade, but had reached the level of their classmates by the fall of the second grade and the level of the control children by the spring of the second grade. The initial advantage displayed by the control children in reading comprehension faded away, and by the spring of the second grade, they were at the same level as the classmates.
Table 3

Descriptive statistics of the observed variables and their mean comparisons

 

AR

CO

CM

Group comparisons

n

M

SD

n

M

SD

n

M

SD

Fluent word recognition

T1: 1st grade fall

100

4.33

3.31

84

6.71

4.17

1,336

5.19

3.39

AR < CM < CO

T2: 1st grade spring

101

7.62

3.62

87

9.91

3.96

1,332

8.82

3.48

AR < CM < CO

T3: 2nd grade fall

98

12.68

4.71

80

14.83

4.62

1,384

13.94

4.48

AR < CM = CO

T4: 2nd grade spring

99

13.57

4.51

76

15.51

4.74

1,351

14.97

4.33

AR < CM = CO

Reading comprehension

T2: Sentences 1st grade spring

101

12.21

8.05

86

16.47

6.79

1,331

14.35

7.24

AR < CM < CO

T3: Text 2nd grade fall

93

7.20

3.15

78

8.74

2.22

1,284

7.82

2.61

AR = CM < CO

T4: Text 2nd grade spring

97

8.67

3.11

73

9.29

2.20

1,333

8.77

2.52

AR = CM = CO

In group comparisons, pairwise Bonferroni corrected values were used, p < 0.05.

AR at-risk, CO control, CM classmates

Examination of heterogeneity in reading profiles: two-level mixture modeling

Next, we examined the heterogeneity in children’s developmental profiles of fluent word recognition and reading comprehension. For this purpose, we utilized the mixture modeling feature of the Mplus v. 4.2 program (L.K Muthén & B.O. Muthén, 1998–2006). Mixture modeling identifies mixtures of subpopulations (latent classes) from the observed data and provides statistical tests to aid in the evaluation of the existence and number of the subpopulations. Because the measures of reading skill were not the same across ages, the reading scores were standardized according to the classmates’ distribution (classmates’ mean = 0 and standard deviation = 1). In learning to read a language with highly regular orthography such as Finnish, the development of reading skills is very rapid after the beginning of reading instruction, and qualitative shifts from being a nonreader to an accurate decoder take place within short time periods. Because the early measures unavoidably hit ceiling after a few months, the use of different measures at different ages is necessary. Due to the large sample size, the standardized scores can be considered as comparable to the development of Finnish first and second graders in the general population. In this type of standardized score mixture modeling, we did not search for heterogeneity in growth in itself, but for heterogeneity in reading profiles.

The modeling was started by searching for the best model to represent the covariance structure of the reading measures in the whole data. The observed variables were found to be best represented by a model with the two latent factors: one for fluent word recognition and one for reading comprehension and their covariance [correlation between the latent factors = 0.81, model fit: χ2(11) = 23.37, p = 0.02, comparative fit index = 1, Tucker Lewis index = 1, root mean square error of approximation = 0.025, standardized root mean square residual = 0.012]. The good fit of the model reflects the structure validity. The four assessments of fluent word recognition loaded strongly onto one latent factor (standardized loadings were 0.85 at assessment 1, 0.96 at assessment 2, 0.81 at assessment 3, and 0.80 at assessment 4). This underscores the strong stability of fluent word recognition through the first two school years. Reading comprehension assessments loaded also onto one single latent factor (standardized loadings were 0.83 at assessment 1, 0.67 at assessment 2, and 0.62 at assessment 3. The connection between latent fluent word recognition and reading comprehension factors was very strong (r = 0.81). Squared standardized loadings serve as reliability estimates and tell how much of the variation in the measured task performance was explained by the factor structure. Here, the latent fluent word recognition factor explained, depending on assessment time, 64 to 92%, and the latent reading comprehension factor explained 38 to 69% of the task performance depending on the assessment time and task.

Next, a two-level model based on the two-factor model was estimated because of the complex data structure. The intraclass correlations (see Table 2) indicated that there was indeed a small but statistically significant classroom membership effect. Of the variance in fluent word recognition performance, 5 to 10% was shared by classroom members, while the rest was due to individual differences. In reading comprehension, classroom effects were between 4 and 6%. This classroom effect on children’s reading skills was controlled for in the subsequent two-level mixture analysis.

The following two-level mixture analysis was based on the above models. Several types of data models may serve as a basis for mixture analysis. In the present analyses, the mixture procedure was based on a model with two continuous factors (i.e., one for reading comprehension and the other for fluent word reading), and one categorical latent variable, denoted by C (see Fig. 1). In the within-level part of the model in Fig. 1, the filled circles at the end of the arrows from the two within factors Wfwr and Wrc to the observed variables fwr1, fwr2, fwr3, fwr4, rc1, rc2, and rc3 represent random intercepts that vary across clusters (i.e., school classes). The random intercept variation between the clusters was modeled with latent variables in the between part of the model and are shown in circles named fwr1, fwr2, fwr3, and fwr4 and rc1, rc2, and rc3. Comparable to the two within-level factors Wfr and Wrc, the model had two between-level factors, Bfwr and Brc. The categorical latent variable, C, represents latent classes or subtypes. The within-level arrows from C to observed variables represent varying intercepts between the subtypes. The filled circle on circle C represents the random mean (representing class or subtype proportion) that varies across clusters. This variation is expressed as C in between levels and it was specified to be uncorrelated with the between-level factors. Note that we refer to the latent classes as reading subtypes hereafter to avoid confusion with school classes.
https://static-content.springer.com/image/art%3A10.1007%2Fs11881-007-0003-0/MediaObjects/11881_2007_3_Fig1_HTML.gif
Fig. 1

The two-level FMA for the reading measures. C latent categorical variable (subtyping), fwr fluent word recognition, and rc reading comprehension. W refers to within level and B to between level. Numbers denote assessment time point

In the present two-level mixture analysis, the focus was not to examine the small between-level variation, but only to omit the classroom effect from the subtype identification. That is, the subtypes were searched based on the within-level variation. The latent subtypes were estimated to differ in their mean values in observed variables, instead of their mean values in factor means. One refers to this type of mixture model where the subtypes are formed based on the indicator values and the continuous factor represents the within-subtype variation as factor mixture model (FMA; see, e.g., Muthén, 2006a, 2006b; Muthén & Asparouhov, 2006). The inclusion of continuous factors allows the indicators to be correlated with each other within a class. This feature was important for the present analyses because the observed variables of fluent word recognition and of reading comprehension across time (even though not exactly the same at each time point) correlated strongly with each other. An alternative to the FMA, the basic latent class analysis (LCA) (where the continuous factor variance is set at zero), allows no such within-subtype variation and aims to identify subtypes of subjects defined as having independence of items (e.g., Muthén & Asparouhov, 2006). One also has the alternative to use the means of the factors in the identification of subtypes [mixture factor analysis (MFA), see Muthén, 2006a, 2006b]. We examined these three alternatives (FMA, LCA, and MFA) but the FMA model was clearly the best solution. The FMA models showed the best model fit indices and, because they allowed the differences between the subtype reading profiles to change over time, they also offered the best theoretical solution. In the FMA, it is possible to set the variances of the continuous factors and the factor loading (slopes) as subtype-variant. Also, these alternatives were examined but FMA with subtype-invariant factor variances, covariance, and loadings was found to fit the data better and this model is therefore reported here. The model was estimated using the MLR estimation method. To avoid local maxima, we increased the amount of the random starting values and the number of iteration rounds according to the Mplus manual (L.K. Muthén & B.O. Muthén, 1998–2006). It should be noted that estimations of these models are computationally heavy and very time-consuming. All children with group tests were included into the two-level FMA analysis (individually tested 82 JLD follow-up children were omitted because they did not have reference classroom information available).

In the mixture modeling procedure, we fitted models with different numbers of latent subtypes. To evaluate the appropriate number of latent subtypes, we used three criteria: (a) the fit of the model, (b) mean probabilities and number of children to be situated into a latent subtype, and (c) the usability and interpretability of the latent subtypes in practice. The fit of the model was evaluated by four criteria: Akaike’s information criteria (AIC), Bayesian information criteria (BIC), adjusted Bayesian information criteria (ABIC), and the Vuong–Lo–Mendel–Rubin test. The lower AIC, BIC, and ABIC values indicate a better model, and significant Vuong–Lo–Mendel–Rubin test results indicate a higher number of subtypes. The five-class solution was confirmed as the best according to three of these criteria (see Table 4 for indices of mixture models with differing numbers of latent subtypes). It also produced the highest log likelihood value, indicating the best fit of the model. A model with seven subtypes was also tested, but the model failed to produce trustworthy estimates despite our persistent efforts, probably due to the representation of an overly complicated model for the data.
Table 4

Indices for mixture models with different numbers of latent classes (subtypes)

Number of classes

log L

BIC

ABIC

AIC

VLMR

n (class 1)

n (class 2)

n (class 3)

n (class 4)

n (class 5)

n (class 6)

1

−11,122.216

22,481.851

22,380.192

22,308.431

1,668

     

2

−10,943.433

22,183.641

22,056.567

21,966.866

0.004

1,441

227

    

3

−10,799.791

21,955.711

21,803.222

21,695.581

0.068

629

185

854

   

4

−10,719.407

21,854.299

21,676.395

21,550.814

0.046

285

182

238

863

  

5

−10,661.590

21,798.020

21,594.701

21,451.180

0.123

174

691

409

212

182

 

6

−10,677.503

21,889.201

21,660.467

21,499.005

0.513

354

296

136

91

7

784

log L log likelihood; VLMR Vuong–Lo–Mendell–Rubin test, p value

Table 5 and Fig. 2 present the descriptive statistics of the final mixture solution of five latent subtypes. The first reading subtype, referred to as good readers, comprised 176 (10.6%) children, and their performance in all reading assessments was at a very high level. The second reading subtype, referred to as average readers, was the largest group and comprised 692 (41.5%) children with average-level reading skills. The third reading subtype, referred to as slow decoders, comprised 411 (24.6%) children whose reading development was somewhat below the average level performance in fluent word recognition tasks but reached the level of average readers in reading comprehension by the end of second grade. The fourth reading subtype, referred to as poor comprehenders, comprised 179 (10.7%) children with average performance in fluent word recognition tasks but whose difference from average and good readers’ performance in reading comprehension increased in time. The fifth and final reading subtype, referred to as poor readers, comprised 210 (12.6%) children with poor performance in all reading assessments. Note that the values of reading skills in Fig. 2 have been standardized according to the classmates’ distribution (classmates’ mean = 0 and standard deviation = 1) to transform the variable scales to be identical. Therefore, the profile changes in Fig. 2, such that those of reading comprehension in slow decoders and poor comprehenders are not absolute but relational to the skill levels of other children at each time point.
Table 5

Descriptive statistics of reading development in the five reading subtypes

 

Good readers

Average readers

Slow decoders

Poor comprehenders

Poor readers

N

M (SD)

n

M (SD)

n

M (SD)

n

M (SD)

n

M (SD)

Fluent word recognition

T1: 1st grade fall

165

12.26 (1.87)a

604

5.52 (2.07)b

359

3.18 (1.98)c

150

4.08 (2.08)d

162

2.46 (1.52)e

T2: 1st grade spring

160

14.04 (2.10)a

599

9.78 (2.68)b

362

7.72 (2.65)c

153

8.74 (2.46)d

162

5.14 (1.87)e

T3: 2nd grade fall

160

20.16 (3.59)a

634

14.91 (3.48)b

347

11.48 (3.56)c

169

14.71 (3.11)b

181

9.27 (2.79)d

T4: 2nd grade spring

149

21.18 (3.42)a

622

15.86 (3.41)b

337

12.80 (3.46)c

167

15.25 (3.29)b

179

10.56 (2.78)d

Reading comprehension

T2: Sentences 1st grade spring

158

20.85 (3.90)a

601

19.60 (3.30)b

362

6.85 (3.49)c

153

14.97 (2.97)d

162

4.44 (3.19)e

T3: Text 2nd grade fall

158

10.19 (1.71)a

637

8.82 (1.95)b

343

6.82 (2.40)c

167

6.84 (2.40)c

185

5.12 (2.34)d

T4: Text 2nd grade spring

146

10.73 (1.36)a

606

10.18 (1.29)b

336

9.08 (1.41)c

163

6.26(1.20)d

185

4.30 (1.51)e

The pairs with same subscript letters do not differ significantly (p > 0.05) based on ANOVA post hoc (Bonferroni corrected) paired comparisons.

https://static-content.springer.com/image/art%3A10.1007%2Fs11881-007-0003-0/MediaObjects/11881_2007_3_Fig2_HTML.gif
Fig. 2

Mean reading skill profiles of the reading subtypes identified with mixture modeling. The available sample size within reading subtypes varied in time points and for good readers was 146–165, for average readers 599–637, for slow decoders 336–362, for poor comprehenders 150–169, and for poor readers 162–185

The mean comparisons between the reading subtypes showed (see Table 5) that the subtypes were distinct at most time points and in both reading measures, except for three occasions: poor comprehenders performed as well as average readers in fluent word recognition at the second grade fall and spring assessments (T3 and T4) and slow decoders and poor comprehenders performed equally well in reading comprehension at the T3 assessment in the fall of the second grade.

Early developmental paths and characteristics of the reading subtypes

Next, we examined what early characteristics or skills may differentiate the reading subtypes that were identified above. For this purpose, we had data available from 191 JLD follow-up participant children, 104 of whom were children with familial risk for dyslexia and 87 control children. Because the reading skills of 82 of the 191 JLD follow-up participants were tested individually, we had no data available for their classmates and consequently, they were omitted from the previous two-level mixture analyses. To obtain subtyping information for these children also, we ran the two-level FMA again with data where the 82 individually tested JLD children were added and each had an individual artificial classroom code with the mean and variance estimates from the previous FMA model set as starting values. The concluding estimates of this model were almost equal to those of the model presented above and the results below are based on this model of all 1,750 children.

Reading subtypes and familial risk for dyslexia

Table 6 presents the frequencies of the five reading subtypes for each study group. The chi-square test statistics showed group differences [χ2(8) = 24.05, p = 0.002]. Analysis of the residuals revealed that the at-risk children were overrepresented among the slow decoders (adjusted residual = 2.6), whereas they were underrepresented among the average readers (adjusted residual = −2.3). For the control group children, on the other hand, membership of the good readers subtype was more common than expected (adjusted residual = 3.1), whereas they were underrepresented among the slow decoders (adjusted residual = −2.2).
Table 6

Cross-tabulation of the reading subtypes by study groups (frequencies in parentheses)

 

Classmates

JLD at-risk group

JLD control group

Good readers (n = 167)

10.3% (160)

9.6% (10)

20.7% (18)

Average readers (n = 643)

42.0% (655)

30.8% (32)

43.7.1% (38)

Slow decoders (n = 478)

24.6% (383)

35.6% (37)

14.9% (13)

Poor comprehenders (n = 241)

10.6% (166)

6.7% (7)

11.5% (10)

Poor readers (n = 221)

12.5% (195)

17.3% (18)

9.2% (8)

Total n

1,559

104

87

Reading subtypes and early language and literacy development

Figure 3 shows the early developmental profiles (prior to school-age) of the five reading subtypes in the seven domains of language and literacy skills: receptive and expressive vocabulary, morphological awareness, phonological awareness, letter knowledge, short-term memory, and rapid naming. See Appendix for the subtype mean comparisons.
https://static-content.springer.com/image/art%3A10.1007%2Fs11881-007-0003-0/MediaObjects/11881_2007_3_Fig3_HTML.gif
Fig. 3

Early developmental skill profiles of the reading subtypes among the JLD participants (n = 191). The sample size within reading types varied in time points and for good readers was 24–28, for average readers 63–70, for slow decoders 40–50, for poor comprehenders 12–17, and for poor readers 18–26

Overall, the profile comparisons showed that the children of the good readers subtype showed the highest performance on all included early tasks, beginning with the early vocabulary, and this difference in comparison to others increased by age. The performance of the average readers was close to that of the good readers in most skill areas. The good readers performed significantly better (pairwise comparisons with Bonferroni correction) than the average readers in phonological awareness after age 5.5 years, in letter knowledge after age 4.5 years, and in verbal IQ. The performance of the poor readers was poorest on all early tasks, and the mean difference between this group and the average readers and good readers reached significance on most tasks with the exception of the very early vocabulary. Furthermore, phonological awareness, and letter knowledge at the age of 6.5 years differentiated poor readers from all the other reading subtypes. Slow decoders also scored below the good and average readers on several measures, particularly on phonological awareness, letter knowledge, and RAN. RAN at the age of 6.5 years differentiated slow decoders from all the other reading types apart from poor readers. A particular dip in the early skill profile of poor comprehenders was apparent in expressive vocabulary development. This group also performed below the good readers on phonological awareness at 5.5 and 6.5 years, letter knowledge after age 4.5 years, verbal IQ, and verbal short-term memory after age 5 years.

Reading subtypes and amount of reading experiences

Figure 4 illustrates the across age profiles of the amount of reading experience (child by him/herself or reading together with a parent) among the five reading subtypes. Good readers, average readers, and slow decoders did not differ in the amount of shared reading experiences at any time point. Poor readers, however, had significantly less shared reading experiences with parents than children who were average readers and good readers in the early years (see Appendix). Good readers engaged more in solitary reading than both poor readers and poor comprehenders at ages 6 and 7 years. At age 7 years, good readers were reading significantly more than all other subtypes.
https://static-content.springer.com/image/art%3A10.1007%2Fs11881-007-0003-0/MediaObjects/11881_2007_3_Fig4_HTML.gif
Fig. 4

Reading experience profiles of the reading subtypes among the JLD participants (n = 191). The sample size within reading types varied in time points and for good readers was 24–27, for average readers 58–70, for slow decoders 39–50, for poor comprehenders 14–17, and for poor readers 19–26

Discussion

This study aimed to substantiate the claims, based on the simple view of reading, that even though word recognition and reading comprehension are strongly related skills, distinct subtypes of reading difficulties with a discrepancy in word recognition and comprehension performance can be identified (e.g., Catts et al., 2003; Leach et al., 2003; Lerkkanen et al., 2004; Nation et al., 2004; Poskiparta et al., 2003; Shankweiler et al., 1999). This study adds to previous findings because it conducted analyses on a large data set within the context of a consistent language (Finnish), included classrooms of students with the whole range of variation in reading skills, controlled for the effects of classroom membership and used several time-points for identification of the reading subtypes. We also analyzed the early language and literacy development, reading experience, and the status concerning familial dyslexia risk of the children belonging to the reading types within the JLD follow-up sample. Using an advanced mixture modeling procedure, we demonstrated five distinct subtypes. These reading subtypes nicely complement the simple view of reading (Gough & Tunmer, 1986; Tunmer & Hoover, 1992). The five reading subtypes differed with respect to early language and literacy development, amount of reading experience, and familial risk status, differences that further strengthen the validity of the classification.

Instead of the four different reading subtypes predicted from the simple view of reading and the types specified in some of the previous studies (e.g., Catts et al., 2003; Leach et al., 2003; Nation et al., 2004; Shankweiler et al., 1999), the mixture analysis of the present study identified five reading types: (1) poor readers (poor word recognition and reading comprehension skills), (2) slow decoders (poor word recognition fluency combined with reading comprehension that reached average levels over time), (3) poor comprehenders (average word recognition combined with a lag in reading comprehension in time), (4) average readers (average word recognition and reading comprehension), and (5) good readers (above-average word recognition and reading comprehension). The findings of the five distinct reading subtypes showed that even though word recognition and reading comprehension were highly correlated (r = 0.81 between across-age factors of fluent word recognition and reading comprehension), children with a discrepancy in reading skills can be identified in our data. The close association of word recognition and reading comprehension was expected (e.g., Shankweiler et al., 1999), and for most of the children, these reading skills did go together. Children who did not show a word reading and reading comprehension discrepancy (poor, average, and good readers) amounted to 64.6% of the children. The slow decoders and the poor comprehenders, on the other hand, showed a developmental pattern whereby their word reading level and reading comprehension separated over time and already showed a discrepancy by the end of second grade. The reading skill pattern of the slow decoders illustrates the intertwining development of fluent word recognition and reading comprehension in that their reading comprehension seemed to begin to develop rapidly only after their word recognition skills had reached the necessary level for text comprehension.

Direct comparison between the present findings and those of previous reading subtyping studies is difficult because of the variability in measures, classification strategies, and ages of the children. In addition, our large sample appears to capture a wider-than-usual variation in reading skills, whereas many previous studies have focused on differentiation at the poor end of reading variation or used smaller samples. At least two additional differences also need to be taken into account. First, our subtype identification was based on several time points, and second, the word recognition test we used was a speeded test, which taps fluency along with accuracy. The use of a speeded measure was a natural choice in the context of the Finnish language as reading difficulties are predominantly manifested in the speed of reading in consistent languages (e.g., Aro & Wimmer, 2003; Porpodas, 1999). To underline this point, we used the label slow decoders to refer to the subtype with below-average fluency in word recognition combined with average reading comprehension at second grade. Despite these differences, our finding of five reading subtypes does not contradict the previous classifications of four reading types drawn from the simple view of reading because our result of five subtypes differentiates the average and good readers, which has not been the focus of the previous studies (e.g., Catts et al., 2003; Leach et al., 2003; Nation et al., 2004; Shankweiler et al., 1999).

It should be noted that there are also subtyping studies that have used a very different type of approach (i.e., identifying variation in a range of languages and cognitive skills and comparing the resultant subtypes of reading skills) and have identified a varying number of subtypes at risk for reading difficulties (e.g., Lyytinen et al., 2006b; Morris et al., 1998). These studies suggest combinations of potential underlying factors or profiles of poor readers instead of searching for heterogeneity in reading development itself and were therefore not cited more thoroughly here.

In two recent Finnish studies from another data set but with children of the same age, only three reading subtypes were identified (Lerkkanen et al., 2004; Poskiparta et al., 2003). These studies failed to obtain two of the present subtypes: (those of our slow decoders and good readers). Perhaps this was due to their smaller sample sizes, problems of statistical power in the traditional cluster analysis, and different sets of variables. Poskiparta et al. (2003) based their cluster analysis on measures of both reading skills and spelling. They explicitly looked for three clusters of children and did not report whether any other solutions were explored. Lerkkanen et al. (2004) used a longitudinal clustering analysis approach (ISOA analysis) with two reading comprehension measures and one word recognition measure in a sample of 90 children. As they speculated themselves, because of their smaller sample size, a further division of different poor reading subtypes was not possible.

The above-mentioned factors, which relate to differences in samples and the method of subtyping, also naturally affect the proportions of children representing different reading subtypes reported in the different studies. To provide some comparisons with previous studies, we interpreted the poor readers’, slow decoders’, and poor comprehenders’ subtypes of the present study as loosely corresponding to the samples of children with reading difficulties used in previous studies (e.g., Catts et al., 2003; Leach et al., 2003; Nation et al., 2004; Shankweiler et al., 1999). In the present study of the children with any reading difficulties, 51.4% displayed the slow decoders’ reading pattern, a pattern characterized by slow word recognition but fast relative growth in reading comprehension. This corresponds well with the finding by Leach et al. (2003), who reported that 49% of their poor readers of similar age had word recognition difficulties without reading comprehension difficulties. However, in their samples of poor readers, Catts et al. (2003) reported that 35.5%, and Shankweiler et al. (1999) reported that only 18%, showed the pattern of solely word recognition difficulties. The proportion of children with both word recognition and reading comprehension difficulties, on the other hand, has varied between 35.7% (Catts et al., 2003) and 72% of the poor readers (Shankweiler et al., 1999). In the present study, 26.3% of the children with reading difficulties displayed both word recognition and reading comprehension difficulties (i.e., belonged to the poor readers subtype), whereas 22.4% were poor comprehenders. The proportion of poor comprehenders in previous studies conducted among samples of poor readers has varied between 6 and 15% (Catts et al., 2003; Leach et al., 2003; Shankweiler et al., 1999). Nation et al. (2004), however, suggested that poor readers amount to about 10% of all children, which corresponds well with the findings of the present study (10.7% were classified as poor comprehenders).

The identification of the separate reading types raises the question of what differentiates these groups of children from each other in earlier skill development or experiences. The variance derived from differences in the quality of the reading environment at school was not the only answer. Based on the present analyses, we can conclude that, in the Finnish schooling system, where nearly all children are enrolled in public schools with a national curriculum and participate in reading instruction that emphasizes systematic use of phonics with a strong focus on letter sound relations, this source of variability was not very strong. However, the effect was statistically significant: 4 to 6% of the variance in reading comprehension and 5 to 10% in fluent word recognition depending on the assessment time point was due to classroom membership effect. The focus of the present study was not a very detailed examination of the determinants of classroom membership effect, and the analyses of the present study merely controlled for the classroom effect by using the within-level variance in the subtyping procedure.

The comparisons of the reading subtypes proceeded with the analyses within our JLD follow-up participants (in this study n = 191), for whom we have detailed data regarding their risk status for dyslexia (based on presence of dyslexia in the family), their early language and literacy development, and their amount of home literacy experience. In the comparison of the JLD children and their classmates, we found that the JLD control children showed significantly higher-than-average reading skills in the first grade, but this early advantage faded in the second grade. This early advantage may be in part associated with their participation in the intensive follow-up study, but also with the fact that the JLD control children were selected as controls based on the normal reading level of their parents and the absence of reading difficulties among close relatives. Among the classmates, however, normal variation of risk for difficulties in reading or other kinds of difficulties exists.

The familial risk status for dyslexia was found to have a significant role in reading development. The children with familial risk for dyslexia performed on average at a lower level in reading than their peers, as expected (Elbro et al., 1998; Hallgren, 1950; Lyytinen et al., 2004; Olson et al., 1999; Pennington & Lefly, 2001; Scarborough, 1990) and the at-risk children were overrepresented in the reading subtype with difficulties in fluent word recognition (particularly in the slow decoders subtype). This finding is in line with the expectation that these children would show reading difficulties that fit the definition of dyslexia, which is based on word recognition/decoding and not on reading comprehension (Lyon, S.E. Shaywitz, & B.A. Shaywitz, 2003). The proportion of children identified either as poor readers or slow decoders was high in the present study (52.9% in the at-risk group, 24.1% in the control group, and 37.1% among classmates) and, as such, does not represent a true reading deficit group. It should be noted that the relative incidence of word recognition difficulties among the at-risk children as compared to controls is higher in the JLD data when a greater number of reading and spelling measures and strict criteria for reading difficulties are applied (Lyytinen et al., 2004).

Comparison of early skill development showed that children with average or good level reading profiles had a higher level of skill in several early language and literacy tasks from early on than children of the other three reading subtypes. The differences between the subtypes were most evident in the tasks of phonological awareness, letter knowledge, rapid naming, and vocabulary. These differences increased over time and became clearly noticeable after age 5 years. The poor readers lagged behind the other subtypes and good readers outperformed the other reading subtypes, particularly in these tasks. These findings are compatible with the existing understanding of the origins of word recognition difficulties (Vellutino et al., 2004), and show that the variation in these skills can also separate children at the higher end of the distribution.

The slow decoders displayed a level of rapid naming performance similar to the poor readers. This finding points to speed as a significant differentiating factor, a result that was expected due to links reported between fast retrieval of symbol names from memory (as in RAN) and reading fluency in particular (e.g., Holopainen et al., 2001; Wimmer & Mayringer, 2002; Wimmer et al., 1998, 2000). The slow decoders significantly outperformed the poor readers in phonological awareness and letter knowledge at age 6.5 years. In addition, the expressive vocabulary and morphology of the slow decoders appeared to be at a higher level than that of the poor readers, but these differences did not reach the level of significance in the paired comparisons. These strengths may support the development of reading in slow decoders in that they performed on a somewhat higher level than poor readers in the fluent word recognition task and their reading comprehension progressed up to an average level.

Vocabulary skills were the most compromised early skill of the poor comprehenders. This trend is in line with previous findings of the predictors of reading comprehension (e.g., Nation et al., 2004; Roth et al., 2002; Stothard & Hulme, 1995; Sénéchal & LeFevre, 2002; Storch & Whitehurst, 2002). Poor comprehenders also showed somewhat below-average skill level in several tasks. It has been suggested that the difficulties of poor comprehenders in understanding the demands of a task explain their wide range of difficulties in task performance (Nation, 1999). Their rapid naming skills after age 3.5 years were, however, at an average level, showing a clear distinction from the slow decoders. It should be noted that reading comprehension is still at a relatively early stage of its development at the end of the second grade, and all of the poor comprehenders may not manifest compromised skills later on. A clearer pattern of deficient oral language and memory may be true for those children whose comprehension skills still lag behind at the later grades. Inclusion of other early skills related to reading comprehension, such as children’s inference-making skill or listening comprehension, might have complemented the characterization of these children (e.g., Cain et al., 2004).

Unlike many previous studies of reading subtypes, we were able to examine the continuum of reading experience among children with differing reading types. The comparisons indicated that children with the strongest reading skills were reading for fun more often than others, in particular poor readers and poor comprehenders, as could be expected (e.g., Leach et al., 2003; Leppänen et al., 2005; Scarborough et al., 1991; Sénéchal & LeFevre, 2002). At this early stage, differences between children in reading skills may be huge; the poorest readers still struggle with basic decoding whereas some of the best readers read short stories already in first grade. Interestingly, the poor readers had also had significantly less shared reading experiences with parents before school age than the good readers. The mechanism behind this association goes potentially through oral language development. The early shared reading experiences have been found to support children’s oral language, particularly vocabulary growth (e.g., Scarborough & Dobrich, 1994; Sénéchal & LeFevre, 2002), which, in turn, is associated with better ability to comprehend text. The direction of causality is, however, unclear. In our previous analyses, for example, large vocabulary at age 3.5 years predicted the amount of shared reading at subsequent years in the at-risk group, whereas the amount of shared reading at age 2 years did not predict vocabulary at age 3.5 years (Torppa et al., 2007).

After school entry, the amount of shared reading with parents increased among poor readers, whereas it was slightly declining in other subtypes, particularly in good readers. The reason for this pattern may be that school entry alerts parents of the poor readers that their children need help with reading. At the same time, the children with poor reading skills engaged less in solitary reading than other children did. It may be that their basic reading skills were not yet at a level that would make reading enjoyable or that their interest in reading is not as high as that of others. These differences in reading experiences have a potentially important role in all academic learning tasks, as a child who reads frequently will further develop his/her reading fluency, learn efficient strategies for gleaning meaning from written text, and gain knowledge about a wide range of topics.

Overall, the mixture analyses of the present study support the previous findings of heterogeneous reading development and also show that, in a highly transparent writing system, a substantial heterogeneity is already present after only 2 years at school. The end of second grade is an interesting stage to assess reading comprehension because, from that point on, Finnish children are challenged with a clearly increasing amount of text reading at school. Because of their slower-than-average reading comprehension development, the children belonging to the poor comprehenders and poor readers subtypes can be expected to have difficulties in several school subjects. The fact that poor readers have difficulties in both fluent word recognition and in reading comprehension tasks is particularly alarming. Even if their fluent word recognition were to progress to a level needed for efficient reading comprehension, their very slow start may hinder them from keeping up with the pace of the increasing demands of normal classroom instruction. Their lower level in several literacy and language skills and potential lack of interest in reading indicate that these children are in need of extra support. In our sample, support seemed to be provided at least by parents in the form of increased level of shared reading. The problems of the poor comprehenders are challenging also because they can easily be left unnoticed (e.g., Leach et al., 2003). Slow decoders were able to comprehend text if given sufficient time, and an important strength of these children is their good reading comprehension. The increasing length and amount of reading materials after second grade may, however, become a stumbling block for these children if they are unable to compensate their slow reading speed with other strengths, such as strong vocabulary skills, good inference skills, or a wide knowledge base.

Acknowledgements

The JLD belonged to the Finnish Center of Excellence Program (2000–2005) and was supported by the Academy of Finland (#213486), the Niilo Mäki Foundation and the University of Jyväskylä, and the Finnish National Graduate School of Psychology. We would like to thank the families who participated in the study. We also thank Matthew Wuethrich and Jane Erskine for polishing the language.

Copyright information

© The International Dyslexia Association 2007