Oral language deficits in familial dyslexia: A meta-analysis and review.

Individual Differences in Reading Development

To consider the nature and developmental course of dyslexia, it is important to highlight the resource demands of learning to read. Universally, reading is a process of mapping between the visual symbols on the page (orthography) and the spoken language. However, the nature of the symbols and the units of spoken language to which they connect differ across languages (e.g., Ziegler & Goswami, 2005 ). In an alphabetic language such as English, the foundation of literacy is a system of mappings between letters and phonemes (the smallest speech sounds of words) and a challenge for the child is to abstract the mapping principle ( Byrne, 1998 ). In contrast, Chinese is nonalphabetic; the orthographic symbols are characters comprising semantic and phonetic radicals that correspond to morphemes (units of meaning). The semantic radical provides information about the meaning and the phonetic radical provides a cue to the pronunciation of the word ( Shu & Anderson, 1997 ).

Orthographies also differ in their “transparency”—that is the regularity of the mappings between symbols and sounds. Among European languages, English has the most inconsistent writing system, embodying many exceptions; German and Dutch have relatively few inconsistencies and Finnish is the most consistent. In general, it has been shown that regular languages pose fewer challenges for the beginning learner than irregular languages ( Seymour, 2005 ) but arguably, this is a simplistic view. Languages also differ in the way in which they convey the grammar in writing and differences in morphological structure may moderate the ease of learning ( Seidenberg, 2011 ).

Nonetheless, it is clear that the specific demands of learning to read will differ across languages. Irrespective of this, the child needs to learn the symbol set and how that set maps to the language and this requires explicit awareness of the structure of the language. Once the basic mappings have been established, reading practice (“print exposure”) is required to achieve reading fluency.

Among alphabetic languages it is well established that the predictors of individual differences in decoding (the skill that is impaired in dyslexia) are the same regardless of the transparency of the language: these are, letter knowledge, phoneme awareness, and rapid automatized naming ( Caravolas et al., 2012 ; Ziegler et al., 2010a ). However, reading development proceeds faster in the more transparent orthographies than in opaque orthographies ( Caravolas et al., 2013 ). The process of learning is more protracted in languages that have large symbol sets and in nonalphabetic languages where mappings are to meaning rather than sound ( Nag, Caravolas, & Snowling, 2011 ). Nonetheless, a similar set of skills appear to predict progress in these languages—namely symbol knowledge, metalinguistic awareness, and rapid automatized naming; the corollary is that deficits in these skills compromise decoding in dyslexia.

The Role of Environmental Factors

The majority of research on dyslexia has focused on its proximal cognitive causes; however, within a developmental framework it is important also to consider distal risk factors. Given the differences between orthographies, one extrinsic factor which affects the rate of reading development is the language of learning. Thus, it might be expected that the prevalence of dyslexia will depend on the transparency of the language; indeed it has sometimes been speculated that the difficulty of the English language could be a cause of dyslexia. However, data from more regular orthographies provide little evidence in support of this contention (for Dutch: Blomert, 2005 ; for German: Fischbach et al., 2013 ).

Social-demographic factors can also affect reading attainments. For example, higher rates of reading difficulty have been reported in inner-city samples than in rural areas (except in developing countries where the opposite trend is seen) and it is well-known that there is a social gradient in reading attainment such that reading is poorer in disadvantaged groups (e.g., Rutter & Maughan, 2005 ). Among the factors that could account for demographic variation are differences in parental education level ( Phillips & Lonigan, 2005, for a review ) or in quality of schooling, although whether such differences are truly environmental, rather than reflecting genetic factors, is a moot point ( Friend, DeFries, & Olson, 2008 ). The quality of the home literacy environment (e.g., Frijters, Barron, & Brunello, 2000 ) is one factor that likely reflects the correlated effects of genes and environment ( ge correlation). Moreover, evidence suggests that the home literacy environment may at least partially mediate the influence of socioeconomic status (SES) on children’s literacy outcomes (e.g., Chazan-Cohen et al., 2009 ).

Home literacy environment is defined by a range of factors concerning parents’ and children’s attitudes and dispositions toward reading, including how long parents spend reading with their children and parents’ own reading behavior. An important component is “shared book reading” that has a significant impact on children’s oral language development ( Bus, van Ijzendoorn, & Pellegrini, 1995 ). In addition, direct teaching of literacy concepts is observed in some families and this strengthens the foundational skills for reading. More important, differences in parental attitude to literacy appear to affect how children’s reading skills develop: while direct instructional practices focusing on print concepts facilitate the early development of decoding skills, shared language experiences around books have a greater impact on reading comprehension ( Senechal & LeFevre, 2001 ). An important question that arises from this research is how the literacy environment in families in which there is a parent (or indeed a child) with dyslexia differs from that in a family where family members are free of literacy problems. Family risk studies provide the opportunity to do this and to track possible changes in both child- and parental attitudes as reading develops over time.

Perceptual and Cognitive Deficits in Dyslexia

A major thrust of research on dyslexia has been to specify the underlying deficits that are candidate causes of the condition. Such impairments (that may be cognitive and/or have their origins in basic perceptual processes) mediate the impact of heritable, brain-based differences on behavior ( Morton & Frith, 1995 ; Pennington, 2002 ; Ramus, 2003 , 2004 ).

Within this approach, the predominant view for many years was that dyslexia could be traced to deficits within the phonological system of language ( Melby-Lervåg, Lyster, & Hulme, 2012 ; Vellutino, Fletcher, Snowling, & Scanlon, 2004 ). As we have seen, phonological skills are critical foundations for learning to read in alphabetic systems. More generally, phonological deficits have been reported to characterize dyslexia in logographic Chinese ( Hanley, 2005 ; Ho, Chan, Lee, Tsang, & Luan, 2004 ; Ho, Chan, Tsang, & Lee, 2002 ) and poor readers of alphasyllabic scripts ( Nag & Snowling, 2011 ). However, a problem in assessing the causal status of phonological deficits is that performance on phonological tasks (such as phoneme awareness and nonword repetition) is influenced by reading skill ( Morais & Kolinsky, 2005 for a review ). It follows that deficits in these processes could be correlates (rather than causes) of poor reading. An advantage of family risk studies is that phonological processing can be measured before the onset of literacy. As we shall see, all family risk studies have assessed the development of phonological skills.

At a more fine-grained level, research has pursued the causes of the phonological deficits in dyslexia. In now classic work, Tallal (1980) proposed that dyslexia is caused by a problem with the rapid temporal processing of auditory information needed for the perception of speech sounds, leading to a cascade of difficulty from auditory processing through speech perception to phonological skills. Family risk studies are well placed to test such causal chains from the early stages of language development. Accordingly many of these studies draw on research suggesting deficits in speech perception in dyslexia (e.g., Adlard & Hazan, 1998 ; Nittrouer, 1999 ; Serniclaes, Van Heghe, Mousty, Carré, & Sprenger-Charolles, 2004 ; Ziegler, Pech-Georgel, George, & Lorenzi, 2009 ) or in basic auditory processing ( Hämäläinen, Salminen, & Leppänen, 2012 for a review ).

A separate line of investigation has focused on possible causes of difficulties with the letter-by-letter structure of words (orthographic deficits) in dyslexia. One theory is that spatial coding deficits affect ocular motor control ( Boden & Giaschi, 2007 ; Kevan & Pammer, 2008 ; Vidyasagar & Pammer, 2010 ). Alternatively, problems in the system of visual attention could affect the left-to-right extraction of orthographic information critical for parsing letter strings before decoding ( Facoetti, Paganoni, Turatto, Marzola, & Mascetti, 2000 ; Valdois, Bosse, & Tainturier, 2004 ). In addition, there are modality—general theories that aim not to explain particular features of dyslexia but rather seek overarching explanations (e.g., Ahissar, Lubin, Putter-Katz, & Bani, 2006 ; Nicolson & Fawcett, 1990 ; Vicari, Marotta, Menghini, Molinari, & Petrosini, 2003 ). Whereas such theories may hold promise for understanding how dyslexia relates to other co-occurring disorders (comorbidity, e.g., Rochelle & Talcott, 2006 ) they do not explain why dyslexia can and sometimes does occur in the absence of any other cognitive deficits.

Nevertheless, as a long history of the search for subtypes of dyslexia attests, these causal hypotheses are not mutually exclusive and it is important to recognize that dyslexia is a heterogeneous condition (e.g., Ramus et al., 2003 ). As Pennington (2006) has argued, the etiology of complex disorders like dyslexia is multifactorial and involves the interactions of risk and protective factors. Longitudinal studies of children at family risk of dyslexia that follow children from early childhood to formal schooling can reveal the risk factors associated with a dyslexia outcome. In addition, because dyslexia is a dimensional disorder, the study of unaffected relatives can be informative in highlighting protective or compensatory factors that mitigate familial risks. Such risks (that could be biological processes or cognitive impairments) can be described as “endophenotypes” ( Skuse, 2001 ).

According to Bearden and Freimer (2006) , an endophenotype is a “marker” that is associated with the disorder in the population and expressed at a higher rate in unaffected relatives of probands than in the general population. Put another way, it is intermediate between the genotype and the phenotype and, importantly, the impact of such processes can be moderated or compensated for by areas of skill or through interventions. This particular characteristic of an endophenotype deserves mention in relation to comorbidities. When two neurodevelopmental disorders frequently co-occur it is probable that they have endophenotypes in common ( Thapar & Rutter, 2015 , for a review). Prospective family risk studies can identify putative endophenotypes of dyslexia (or subclinical features) that mark the presence of co-occurring disorders For example, findings of family risk studies can elucidate relationships between difficulties with oral language observed in preschool and later written language disorders, in short, the comorbidity between dyslexia, specific language impairment ( Bishop & Snowling, 2004 ; Pennington & Bishop, 2009, for review s).

Study Rationale and Hypotheses

Although the discussion above highlights the importance of taking a longitudinal perspective to understanding dyslexia, current knowledge of dyslexia draws mainly on cross-sectional studies involving the comparison of individuals with dyslexia and controls at one specific point in time. Such evidence cannot distinguish adequately between causal and noncausal reasons for associations. Following the pioneering work of Scarborough (1989 , 1990 ), who followed a group of 2-year-old English-speaking children deemed to be at high-risk of dyslexia because of an affected parent, many prospective studies of children at family risk of dyslexia have begun in recent years. At the time of writing, 21 independent studies have been completed, resulting in 95 behavioral publications that are the focus of this review. In addition, our review found 23 descriptive studies using neurophysiological measures that are listed in Supplemental Material Table S1 (in the online supplement file). We used the findings of the behavioral studies to test the hypotheses that follow. Hypothesis 1: The prevalence of dyslexia in children at family risk. We predicted that dyslexia would be more common in at-risk than control families (Hypothesis 1a) and more prevalent in English because it has an opaque orthography than in other European languages (Hypothesis 1b); in addition, given the large character set, we expected a high prevalence in Chinese. Irrespective of orthography, because dyslexia is a dimensional disorder with no clear boundaries, we predicted that prevalence would depend upon the cut-off used for “diagnosis” (Hypothesis 1c).

Hypothesis 2: The home and literacy environment of children at family risk of dyslexia. Because of the influence of genetic factors and the environments correlated with them, we hypothesized that the home literacy environment would differ between families in which there is a history of reading difficulty from that in families where parents are free of such problems (Hypothesis 2a). We did not specify the ways in which these differences would be manifest but we anticipated that parents with dyslexia would read less for pleasure than controls (Hypothesis 2b). If the home literacy environment influences children’s reading attainment, we predicted that parental literacy skills (a proxy for gene—environment correlation; van Bergen, van der Leij, & de Jong, 2014 ) would account for independent variance in children’s reading outcomes (Hypothesis 2c).

Hypothesis 3: Endophenotypes of dyslexia. Although endophenotypes can take many forms, we are concerned here with the perceptual and cognitive deficits that are observed among children at family risk of dyslexia. These can be expected to be present in the preschool years before dyslexia is diagnosed (where they can be construed as cognitive risk factors; Hypothesis 3a). Based on the overlap between dyslexia and language impairment ( Bishop & Snowling, 2004 ; Ramus et al., 2012 ), we expected that delays and difficulties with speech and language development would be common (Hypothesis 3b). More specifically, for alphabetic languages we predicted that three of the skills considered foundational for literacy (letter knowledge, phoneme awareness, and rapid automatized naming [RAN]; Caravolas et al., 2013 ) would show developmental delay in the preschool period (Hypothesis 3c) and deficits in each would characterize dyslexia in the school years (Hypothesis 3d). Given the known continuity of risk for dyslexia among family members ( Pennington & Lefly, 2001 ), we predicted that unaffected children at family risk of dyslexia would also have poor literacy skills relative to controls but not severe enough to warrant a diagnosis of dyslexia (Hypothesis 3e) and deficits/endophenotypes should be observed in unaffected children but to a milder degree (Hypothesis 3f). Finally, we aimed to evaluate the causal status of basic deficits in speech perception, visual, and auditory processing, attention and motor skills as additional risk factors (Hypothesis 3a).

Hypothesis 4: Predictive relationships between early cognitive abilities and later reading. Reading builds upon spoken language and there are similar heritable influences on both reading comprehension and language comprehension (e.g., Keenan, Betjemann, Wadsworth, DeFries, & Olson, 2006 ). In this light, we predicted that preschool measures of oral language would predict later reading outcomes, particularly reading comprehension (Hypothesis 4a). Based on previous findings (e.g., Caravolas et al., 2013 ) we also predicted that, for alphabetic languages, there would be three predictors of decoding skills and hence dyslexia: phonological awareness, symbol knowledge (letters in alphabetic languages) and RAN (Hypothesis 4b).

Hypothesis 5: Interventions for dyslexia. Arguably the ultimate aim of research on dyslexia is to identify effective interventions that will ameliorate its impact on educational attainments. In addition, training studies are important theoretically as tests of causal hypotheses ( Snowling & Hulme, 2011 ). We expected that interventions incorporating training in letter knowledge and phoneme awareness would improve decoding skills in children at family risk of dyslexia (Hypothesis 5a). However, we did not expect such interventions to impact reading comprehension beyond gains in decoding unless they incorporated training in broader oral language skills (e.g., vocabulary training; Hypothesis 5b).

Table 1 presents a summary of the hypotheses that guided the review. First, to assess Hypotheses 1(a–c) we summarize data from studies examining prevalence of dyslexia in family risk studies; second, to assess Hypotheses 2 (a and b), we use data on the home and literacy environment and to assess Hypothesis 2c we use data examining parental skills as predictors of outcome. To assess evidence for endophenotypes (Hypotheses 3 a–f), we adopt a developmental perspective, presenting evidence from different dev

Method

The review was designed and is reported in line with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement ( www.prisma-statement.org ). PRISMA is a consensus statement developed by an international group of researchers in health care for the conduct and reporting of systematic reviews and meta-analyses. When there is a sufficient number of studies (two or more), we will do a meta-analysis by calculating a mean effect size. If a summary effect is not possible, we will do a systematic review and report effect sizes for individual studies.

Details concerning the literature search, criteria for inclusion and flow of studies are shown in Figure 1 . The literature search consisted of the following components: Electronic databases (ERIC, Medline, PsychInfo, and all Citation Databases included in ISI web of knowledge from 1980–30 July 2015 with keywords in abstracts at-risk, famil* risk paired with dyslexia, reading disorders, decoding, and decoding problems), citation search on author names, emails to authors and posting at the list server for Scientific Studies of Reading, hand search of journals (Dyslexia, Annuals of Dyslexia, Scientific Studies of Reading), scanning reference lists, and Google Scholar.

Inclusion criteria

To be included a study had to consist of a sample of children with familial risk of dyslexia. Family risk was defined as having a parent and or a sibling (a first-degree relative) with dyslexia. The at-risk status had to be verified by testing or self-report of the relative with reading problems. In addition, the studies had to include a control group of children with no known family risk of dyslexia. Studies vary in the nomenclature they have used. To be consistent, in this review, we use the terms “FR–Dyslexia” to refer to children at family risk (FR) who are identified as having reading problems, “FR–NR” to refer to children at family risk who are identified as typical in their literacy outcome (normal reader) and “control” to refer to children from low-risk groups with no family history, who are considered free of reading difficulties.

In the review we focus on three different study designs: (a) group comparisons, (b) longitudinal prediction studies, and (c) experimental intervention studies. We also include references to neurophysiological studies that are tabulated in Table S1 in the online supplement file. The studies included had to report data so that an effect size could be calculated or significance testing reported for either one of three types of comparisons, based on the different designs we found in the studies in this field: (Comparison Type 1) between children at family risk of dyslexia (FR+) and children without such risk (control). In these studies, reading skills are treated as a continuous variable, and family risk children are not separated into two groups based on whether they have dyslexia or not; (Comparison Type 2) between children at family risk later “diagnosed” with dyslexia (FR–Dyslexia) and control children without family risk (control); and (Comparison Type 3) children at family risk who did not fulfil criteria for dyslexia (FR–NR) and control children without family risk (control).

For the longitudinal prediction studies, in addition to the criteria outlined above, the study must have reported a longitudinal analysis concerning either (a) unique predictors of outcomes treated as continuous variables, or (b) unique predictors of literacy outcome when this is treated as a categorical variable (i.e., presence or absence of reading problem). For a study to be included, the analysis needs to have been conducted so that the predictive patterns pertaining to children at family risk and children without such risk could be compared, or the effect of group membership could be determined. For the experimental intervention studies, some kind of training must have been included with the aim of ameliorating difficulties of reading or literacy in children at family risk.

The neurophysiological studies in Table S1 in the online supplement file used a variety of methodologies and were not amenable to meta-analysis. Most of these studies are of small samples; however, their findings will be used when appropriate to reinforce conclusions.

Coding

Some of the at-risk studies are longitudinal and report data from different stages in development. When coding the studies, it became clear that although there are 95 different publications, many were based on the same study sample (there are 21 independent samples). In Appendices A , B , and C (Column 1 in parentheses after the author name), the sample on which the publication is based is indicated. Special attention had to be taken in the coding to avoid bias related to dependency in the data. To make use of as much information as possible from the different publications, information was coded for different developmental stages: (a) Infants and toddlers (below the age of 3 years), (b) Preschool (below 5.5 years and before formal reading instruction starts), (c) Early Primary school (up to 4th grade), and (d) Late primary school/secondary school (from 5th grade and upward). By doing this, we were able to code information from longitudinal studies twice without merging data from the same study in the same analysis. Therefore, we did not violate the assumption of independence in the data. Nonetheless effect sizes in the different developmental stages will be related because 20 out of 69 publications included in the analyses of group comparisons have data coded from more than one developmental stage.

Furthermore, scrutiny of the studies revealed that several different measures were often reported for the same construct, sometimes in different publications. Table 2 presents the indicators that we selected to represent the higher-order constructs in the review. If there was more than one indicator for a construct from the same developmental stage (e.g., Boston naming test and word definitions for vocabulary knowledge), either in the same or in different publications, the mean of the indicators coded was used in the analysis. An advantage of this procedure is that the mean effect size will be more reliable because it is based on two measures of the same construct. The number of studies and sum of participants reported for the different analyses refers to the number of independent samples and participants that have provided data on a construct and the number of effect sizes contributing to a mean effect size is reported in parentheses. However, if the same test was reported in several different publications on the same sample, the same test was only coded once. In many cases, prevalence of dyslexia is reported for the same sample in several publications. However, such data were only coded once for each independent sample based on one set of reading tests in each study. When it was unclear whether data were based on the same sample or parts of the same sample, authors were contacted by email and asked for clarification. Most authors responded, but in cases where they did not, we have based the overview on information provided in the articles.

A random sample of 80% of the studies was coded by two independent raters. The interrater correlation (Pearson’s) for main outcomes (means, SD , sample size, and age) was r = .98 and agreement rate = 93%. Any disagreements between raters were resolved by consulting the original article or by discussion.

Procedure and Analysis

The coding of studies and analyses were conducted using the “comprehensive meta-analysis” program ( Borenstein, Hedges, Higgins, & Rothstein, 2005 ). Two different effect sizes were used in the meta-analysis. For group differences between the family risk groups and the children from low-risk groups with no family history, we used Cohen’s d . Cohen’s d s were calculated using Hedges’ corrections for small sample sizes ( Hedges & Olkin, 1985 ). When Cohen’s d is negative, children from high-risk groups with a family history have the lowest score. For judging the size of the effect, for group comparisons Cohen’s tentative guidelines were used. According to Cohen, d = 0.2 is a small effect, 0.4 is moderate and 0.8 is large. However, note that according to Cohen (1968) , such guidelines should only be used when no better basis for estimating the effect size is available. In the intervention studies, if Cohen’s d is positive, the group that has received the intervention has the highest gain between pre- and posttest. For the intervention studies, two influential policy organizations (What works clearinghouse [WWC] and Promising Practices Network [PPN]), have set a limit of d = 0.25 for when results of high-quality randomised trials should be taken as having policy implications (see Cooper, 2008 ). In the intervention studies we adopt these guidelines in preference to Cohen. For estimates of prevalence of children in the family risk samples and the control samples affected with dyslexia, we used percentage affected with dyslexia as the effect size.

Mean effect sizes were estimated by calculating a weighted average of individual effect sizes using a random effects model. A mean effect size was calculated if there were two or more studies. A 95% confidence interval (CI) was calculated for each effect size to establish whether it was statistically significantly larger than zero. If confidence intervals cross zero, the result is not significantly different from zero.

To examine the variation in effect sizes between studies, the Tau 2 was used ( Hedges & Olkin, 1985 ). As a rule of thumb, if Tau 2 exceeds 1, the variation between the studies is large. I 2 was also used to determine the degree of true heterogeneity. I 2 assesses the percentage of between-study variance that is attributable to true heterogeneity rather than random error. Notably, I 2 does not say anything about the size of the variation between the studies in general, that is, I 2 can be 100% or 10% but in both cases variation between studies can be small or large. Thus, I 2 can only be used to determine the part of the heterogeneity that is due to true variation between studies rather than sampling error.

For moderator variables, studies were separated into subsets based on the categories in the categorical moderator variable, and a Q -test was used to examine whether the effect sizes differed between subsets. When there were fewer than two studies in a category, this analysis was not conducted. To examine the size of difference between subsets of studies, overlap between CIs was also examined. In cases of multiple significance tests of moderators on the same data set, the results are reported both with and without Bonferroni correction.

When we coded articles, it became clear that there were numerous instances of missing data. If data were critical to calculate an effect size, articles with missing data were excluded if authors did not respond to an email request to provide the data (see inclusion criteria in flowchart). In cases where an effect size could be computed on one outcome but data were missing on other outcomes or moderator variables, the study was included in all the analyses for which sufficient data were provided.