Comment on Asbury and Wai (2019), “Viewing education policy through a genetic lens,” Journal of School Choice

Abstract

Asbury and Wai (Journal of School Choice, 2019) perform a valuable service by summarizing much available behavior--genetic research on academic achievement. However they consider that no specific policies stem from the research body at this time. Here we do propose a policy based on some of our research using twins, namely that available funding for students struggling with learning to read be targeted to them individually rather than allocated to schools per se. We briefly canvass some practical issues, such as the variety of funding mechanisms, best-practice intervention techniques, and identification of struggling readers. We also outline a general research strategy for uncovering factors contributing to educational attainment that takes behavior-genetic research as its starting point and drills down from there, and advocate including genetically-sensitive methods in a growing list of quantitative research techniques in education.

In their article in this journal, Viewing education policy through a genetic lens, Asbury and Wai (2019) summarized several decades of research into educational attainment that has employed genetically-sensitive methods such as twin studies. Their goal was to bring this body of findings to the attention of educational professionals and to explore its implications for policy. Theirs is one of several recent publications with broadly similar goals (e.g., Asbury & Plomin, 2013; Byrne, Olson, & Samuelsson, 2019).

Data from twin studies, the most frequently-used method, are typically used to parse variation in a characteristic into genetic (heritability, or “nature”) and non-genetic sources (the environment, or “nurture”). Non-genetic influences can be further parsed into factors that twins in a family share, such as family wealth, school attended, and (sometimes) shared teacher, and those that twins do not share, such as individual health issues, separate friends, and (sometimes) different teachers. In many of the research papers that Asbury and Wai (2019) summarize, the heritability of school achievement in literacy, numeracy and other domains is estimated at around 60% of total variance, sometimes higher. They report a substantial range in estimates of shared environment influence, from zero to around 40%, depending on the sample and the academic domain being studied. Non-shared environment influence, which includes measurement error, contributes the remainder of total variance in each study.

With an eye on the substantial heritability estimates, Asbury and Wai (2019) express the following aspiration: “…accepting the importance of genetic influences on educational outcomes, and working to better understand the interface between genes and experiences, should have a profound impact on policy discussions…” (p. 9). Nevertheless, they offer this disclaimer: “…although this body of research is highly robust and replicated, no necessary policy implications follow from it, and, indeed, it raises more questions than solutions at this point” (p. 10).

A suggested policy implication and its empirical basis

We agree that caution is in order when moving from research findings to implementation, but in this commentary we do offer a specific policy that we consider does flow from parts of the research oeuvre. To foreshadow, our recommendation concerns educational funding for students at risk for or experiencing reading disability (RD), and we summarize it here: The funding distribution model to support RD students should reflect the sources of their disability. Available funding to support students struggling with learning to read should be tagged for particular students rather than allocated to schools or classrooms per se.

In developing this policy recommendation, our focus has been on the contribution of the shared environment component of variance to literacy development. It happens that many of the factors commonly thought to play important roles in a child’s literacy growth are ones that would mostly show up as part of shared environment in a twin study. The home literacy climate is often cited as one such source, both in the research literature and public discourse (see Scarborough and Dobrich, 1994, for a review, with evidence of unexpectedly modest effects of preschoolers’ home literacy environment on later reading). The classroom, which twins in a pair often share, particularly early in schooling, is also seen as an influential source of variance (see Byrne et al, 2010, for examples of both research and press commentary, and Taylor, Roehrig, Soden Hensler, Connor, & Schatschneider, 2010, for an example of research using twins to document a classroom effect on literacy growth). In countries with an established system of fee-paying private schools it is often assumed that these schools “value-add” compared to state-funded schools (see Smith-Woolley et al., 2018, for a discussion, and evidence to the contrary). School type effects, if they existed, would typically show up as a source of shared environment because twins mostly attend the same school as each other. Whether or not a child attends preschool and the duration of attendance along with a measure of staff “quality” has often been cited as contributing to variability in subsequent academic (including literacy) development; two or more years of attendance in a preschool with well-trained staff is seen as optimal (see Whitehurst, 2013, for a discussion). Typically, if one twin in a pair attends, the other does too, hence any variability stemming from preschool will be shared.

Despite the wide acceptance that these factors play substantial roles in affecting school-based academic development, the actual quantitative estimates of shared environment effects on literacy in some research is vanishingly small. This is certainly the case in much of our own research, based primarily in Australia but including data from Scandinavia and the USA as well. For example, in a project we dubbed the International Longitudinal Twin Study (ILTS) and which comprised approximately 1000 pairs of twins we estimated the shared environment influence on word and nonword reading and on spelling among Australian kindergartners at exactly zero (Samuelsson et al., 2007). In Grades 1 and 2, using combined results from Australia, Norway, Sweden and the US, shared environment effects for word identification and reading comprehension were also close to zero (maximum, .07), and non-significant (see Byrne, Khlentzos, Olson, & Samuelsson, 2010, and references therein). The exception to this pattern was kindergarten reading and spelling in the US, with values of .33 and .46 respectively. We attribute this to the fact that in Colorado, the site of data collection, attendance is restricted to half days and there is no uniform instructional regimen for literacy. In contrast, Australian children attend full-time, with literacy “milestones” mandated to be met throughout the year. In other words, environmental variability is likely higher in US kindergartens than Australian ones, and as Asbury and Wai (2019) point out, the less uniform the environment (US versus Australian literacy instruction in this case), the higher will be environmental influence. In contrast to the overall low shared environment effects, heritability was uniformly high in the ILTS, ranging from .67 to .91 (again, with US kindergarten values lower at .61 and .34 for reading and spelling respectively). Nonshared environment effects were always significant and ranged from .09 to .26.

In a second project, we used data from the Australia-wide National Assessment Program Literacy and Numeracy (NAPLAN), with a total of 2538 pairs of twins. The literacy tests are of reading comprehension and spelling. A detailed description of the tests, which are administered in Grades 3, 5, 7, and 9, is presented in Grasby, Coventry, Byrne, Olson, and Medland (2016). In Grade 3, shared environment influence on reading was a non-significant .06. It rises in later grades, ranging from .09 to .14, averaging .12. Genetic influence was, as in the ILTS, substantial (.59 – .70), and non-shared environment effects comprised around 25% of total variance throughout (see Tables 4 – 7, Grasby et al., 2016)

These results suggest that, in the context of consistent universal education and relatively high levels of social support (Grasby, Coventry, Byrne, & Olson, 2019), the bulk of the variability in early reading is due to factors that are unique to each child; genes and non-genetic processes that influence a child individually. It is not substantially due to factors that twins would share, anchored in the home or in school-wide or classroom-wide processes.

Recall that the shared environment value in Grade 3 NAPLAN reading comprehension was non-significant (value of .06) but it was significant in each of the three higher grades, 5, 7, and 9 (range .09 ’ .14). It is worth noting that educationists canvass the idea that up to about the end of Grade 3 children are learning to read, beyond that, reading to learn (Loveless, 2020). The NAPLAN tests are conducted about three months into the third-grade school year, so even though the test focuses not on the mechanics of reading (largely captured by the speed and accuracy of word recognition) but on comprehension of text, a child’s Grade 3 score may continue to be substantially influenced by variation in these mechanical processes, which are the ones also largely determining scores in the assessment up to Grade 2 (ILTS data) Thus, the evidence for very little shared environmental influence on early reading remains in place into Grade 3 among Australian students.

As a reminder, our recommendation is that Available funding to support students struggling with learning to read should be tagged for particular students rather than allocated to schools or classrooms per se. This recommendation follows from our observations that factors specific to individual students, namely genes and unique environmental experiences, are the major drivers of variation in reading skill in our samples, particularly in the earliest school grades. With few exceptions, these two sources account for 80% or more of the total variation in our literacy and numeracy measures, on some occasions all of the significant variation. Shared environmental factors, such as school wealth or class size or family-based processes such as the home literacy environment, do not appear to be major sources of average variation within our samples. So, while school principals and/or classroom teachers may choose to support the neediest individual students with any additional funds that are directed to the school itself, they may instead allocate them to more diffuse purposes, like the provision of more books or electronic devices. In our opinion, this is likely to be a less efficient use of financial support for struggling students than funding them directly for remedial activities.

The data on which we have based our recommendations are primarily from Australia, with additional data from Colorado and Scandinavia, and as such describe what is the case and not what could be in other places and at other times. There are indications that the low shared environment influence seen in our samples in fact does not always hold elsewhere. For example, Christopher et al. (2016) report a shared environment value of .10 for a latent variable derived from word identification over multiple grades in Colorado, close to the .12 we noted for Grades 5 – 9 in the NAPLAN data but higher than non-significant estimates for most of the younger samples in the ILTS (range zero to .07). A meta-analysis of twin studies across a range of countries including Australia, China, Scandinavia, The Netherlands, the UK and the US identified an average shared environment value of .10 for reading (de Zeeuw, de Geus, & Boomsma, 2015), with studies in Florida and Ohio reporting values as high as .2 to .3 (Byrne et al., 2019). In addition, it appears that in the US at lower socioeconomic levels shared environmental factors exert more influence on variability in literacy skills than among children in wealthier families (though this does not hold in Australia, see Grasby, Coventry, et al., 2019). Thus, making funds available to ameliorate negative effects of shared environmental factors (family literacy habits, school climate, class size, and so on) may have more utility in societies with reduced environmental support systems than in societies where shared environment influences are as low as in our samples.

Drilling down into the shared environment effect

Having said that, what is needed to guide funding policy is research to drill down to identify the factors that are contributing to the shared environment effect. If classroom-level processes are not among them, our recommendation to provide additional funds for named students rather than schools and classrooms remains unaltered, whatever the size of the shared environment influence. Recent analyses we have conducted indicate that averaged classroom-level processes, a potential source of shared environment estimates when twins share a class, are a negligible factor in our database. This is the case even when the shared environment influence is significant, as in reading at higher grades in the NAPLAN project. Twins who are placed together in a class are barely more similar than ones who are separated into different classes, a comparison that is a sensitive test of the average classroom-level effect (Grasby, Little, Byrne, Coventry, Olson, & Larsen, 2019). Individual grade and test domain results were mostly non-significant, and averaging across all results irrespective of statistical significance yielded an estimate of 2–3% of total variance as being due to classroom placement. An international study with twin samples from the UK and Canada also indicated very weak classroom effects (White et al., 2018). Even in the case of US kindergarten spelling with a shared environment value of .46, the correlation for MZ twins sharing a class was .78 and for those with separate teachers almost as high at .75. For DZ twins, the values were .59 and .52, respectively. Thus, it is likely that factors outside of the classroom are driving the influence of the environment that twins share on early spelling skill.

The school is a possible source of shared environment influence in the NAPLAN results, but we cannot directly test this hypothesis in our samples as we could with classroom, where twins either share or do not share a classroom, because almost all twins in the sample attend the same school as each other. However, we do know from a very large-scale study in the UK that school type (state non-selective, state selective, and private) exerts an almost imperceptible independent influence on academic outcomes; once the academic ability and genetic profiles of students entering the three systems are taken into account, the substantial differences in exam results at age 16 all but disappear (Smith-Woolley et al., 2018). There is no similar research in Australia of which we are aware, so we need to keep an open mind on possible school effects, but the UK results, where the school system has a high degree of similarity to Australia’s, are provocative.

Other sources of any shared environment effect on literacy and numeracy are certainly worth investigating, and several candidates, such as levels of disruption in the home (Johnson, Martin, Brooks-Gunn, & Petrill, 2008) and neighbourhood characteristics (Little, Hart, Phillips, Schatschneider, & Taylor, 2019) are under consideration. However, these factors, which are outside the school environment, are not obviously germane to school funding decisions. Thus, as indicated earlier, our recommendation of student-targeted funding rather than allocation of funds to the school or classroom per se can stand, supported by findings of negligible direct classroom and school effects on variance in literacy.

Some examples of funding policies and delivery

If all school districts in the catchment areas for our own research (Australia, Colorado, Norway and Sweden) and those of others we have cited (Quebec and the UK) did in fact fund individual students for RD, our recommendation would be merely reinforcing the status quo. A survey of funding practices in these jurisdictions is beyond the scope of this commentary, but we can report on the policy in New South Wales, Australia. In NSW there are no dedicated funds to support specific children with RD, nor are any ring-fenced for reading support per se. There do exist student-specific funds for those deemed to suffer from moderate intellectual disability and for other conditions such as Autism Spectrum Disorder, mental health disabilities, language delay if it is comorbid with those disorders, and very high levels of aberrant and anti-social behaviors. In addition, there are extra funds under a scheme known as Every Student, Every School (ESES). Schools with particular demographics suggesting social disadvantage attract a higher rate of ESES funding. The classifications are based on parental occupations, and numbers of indigenous students and those of non-English speaking backgrounds. ESES funding can be used as schools see fit, and experience indicates that they are often employed to support early literacy, though there is no requirement to report back on how funds are used nor on any particular students’ progress.

In our view, schemes such as this Australian state’s are suboptimal precisely because they do not target individual RD students, although the practice of targeting other disability categories indicates that there is no in-principle problem in doing so. They are also deficient in that additional funds under ESES are not likely to be as available to schools with clear social advantage, as defined by high parental occupations and low numbers of indigenous and non-English-speaking background (NESB) students. This could have the effect of leaving unsupported RD students in such schools because principals do not have the opportunity to funnel ESES support to them. Thus, we recommend bringing RD students into the same fold as others who receive dedicated support.

Turning to actual funding mechanisms, here are some examples of those that could be or are employed to deliver support to students identified as needing support. In the case of NSW, funds for eligible individuals (not reading-disabled, as noted above) go to the child’s school, and are typically used to employ an additional teacher or teacher aide which in turn allows the class teacher to devote extra time to the target student. The funds are seen as “belonging” to that student, and could in principle be made available for RD students, as we recommend. In the US, a few states have established Education Savings Accounts (ESA), which divert funds that would normally go to public schools to allow families of children with disabilities to access private schools, individual tutors, and/or to save for college expenses. ESAs operate differently in different states, but may include RD, as in Arizona, for example (see https://www.greatschools.org/gk/articles/section-504–2/). One US state, Mississippi, offers Dyslexia Therapy Scholarships (in addition to ESA) that support parents to enrol a child in a private school with dedicated dyslexia intervention programs, which in turn illustrates another delivery mechanism, enrolment in schools specializing in treatment for RD but with students selected on an individual basis from feeder schools. (We add here that we know of no scientific comparisons of the efficacy of individualized versus whole-of-class RD treatment regimens, though this would be a worthwhile project.)

These schemes illustrate what we imagine as the two kinds of possible funding, allocation to the school but dedicated to particular students, or allocated to families in the form of vouchers (or debit card funds). The common theme is that identified students are the focus, whatever the mechanism of delivery. This feature is the one that we are recommending.

More generally, note that our recommendation of individually-targeted funding to support struggling readers is independent of the overall level of funding for a school system. The recommendation would be weakened if it were the case that overall funding increases went hand-in-hand with increases in literacy performance, including that of the lowest-achieving students. But the evidence for such a relationship is weak, at least in the US. Hanushek, Ruhose and Woessmann (2016) refer to “the very weak correlation between increased spending on schools and higher levels of student achievement” (p. 21). Having said that, however, it may not be overall funding levels that matter but how funds are spent. There is some evidence from the well-known Tennessee class size experiment that smaller class sizes, kindergarten to Grade 3, predict a higher likelihood of attending college later in life (Chetty, Friedman, Hilger, Saez, Schanzenbach, & Yaga, 2011). Thus, hiring more teachers may be worthwhile at these grades, although this evidence does not speak to whether RD children in smaller classes benefit as much as they would from more individualized attention. We know of no evidence on this point.

Hanushek and colleagues do have a series of suggestions for improving schools, as outlined in a forum in Education Next, Fall 2009, pp. 49–56, but they involve a focus on outcomes rather than increased funding. They recommend, for example, more school district flexibility in rewarding quality teaching, greater parental choice in where to send children, and, pertinently, “reasonable levels of funding based on the needs of particular student enrollments…” (p. 51). Based on our evidence and arguments, we would refine the last of these to reasonable levels of funding based on the needs of particular students.

Effectiveness of Intervention

An important consideration is whether intervention at the individual child level actually works. If it does not, if for example only an improvement in classroom-wide instructional techniques has been found to be effective, channelling funds to individuals would be pointless. But here the evidence in favour of individual intervention is solid, as can be seen in the recent review by Kilpatrick and O’Brien (2019). Typical of the studies reviewed is Torgesen, Alexander, Wagner, Rashotte, Voeller, and Conway (2001), who showed both short- and longer-term advances among severely disabled readers, averaging around 14 −18 standard score points on tests of word reading accuracy after intensive, one-on-one intervention. The intervention used what has become a well-accepted menu for intensive intervention (Snowling & Hulme, 2012), a combination of phonemic awareness training, phonemic decoding, and reading practice. It continued for 67.5 hours over eight weeks, and hence must be regarded as a relatively expensive exercise when implemented in schools.

It also seems to be the case that small-group intervention, perhaps up to as many as six children, can be equally effective (Elbaum, Vaughn, Hughes, & Moody, 2000). The Response to Intervention (RtI) framework generally utilizes small-group instruction at its Tier 2 level. Tier 2 is recommended for students who fail to respond to Tier 1, which comprises whole-of-class instruction using “evidence-based” methods. The Tier 2 groups typically meet up to 5 times per week for up to 40 minutes each time, likely to incur substantial costs as well if additional staff are required. The Institute of Educational Sciences classifies the evidence of efficacy for Tier 2 as strong (Gersten et al., 2008).

Evidence of the efficacy of individual (or small-group) intervention is important because it bolsters our case for allocating funds to named students. But it is important too because it shows that even though RD is in part genetically grounded at the population level, it can be treated with some success. There is a tendency, documented by Haslam and Kvaale (2015), for people, when they discover a biogenetic explanation for a disorder, (a) to be relieved (they are absolved from blame to a degree) but (b) to be more pessimistic about successful intervention (because genes cannot be altered). The authors refer to this as the “mixed blessings model.” Thus, if the funding distribution scheme we recommend is implemented, based as it is in part on a documented role for genes in RD, informed parents and indeed teachers may well be susceptible to this mixed-blessing effect and feel less optimistic about a child’s improvement. Evidence that improvement through intervention is possible can likely short-circuit this unhelpful tendency.

Identification of RD

A second question amenable to scientific research findings is who should be selected to be a named recipient of funding. RD is sometimes referred to as dyslexia, which has overtones of a categorical condition, like measles or Down syndrome. But available evidence likens it more to obesity, where the underlying variable, BMI in that case, is continuous and the cut-off is somewhat arbitrary (Miciak & Fletcher, 2019; Peters & Ansari, 2019). So, RD is best seen as the low end of a normal distribution; very poor readers are so not because of defective processes unique to them but because of lower levels of one or more of the processes that are needed to master reading. This makes over- or under-identification of individuals deserving additional support an important consideration when planning how to allocate funds for RD.

Miciak and Fletcher (2019) make recommendations for best-practice identification of RD students, including steps like using a combination of tests to increase reliability and therefore validity, employing confidence intervals to better represent the confidence of identification in individual cases, and avoiding rigid cut points, instead tailoring decisions to the educational needs of individuals. The evidence that small-group intervention may work as effectively as individual treatment will bolster our recommendations and provide comfort to educators, parents and stakeholders as they grapple with the arbitrariness of the inclusion decisions—be as generous as you can be, a stance that may help moderate any feelings of unfairness on the part of parents whose child does not qualify for individualized intervention.

A generalized research strategy

If we now step back from our specific set of policy recommendations, we can outline a generalized strategy that takes behavior-genetic studies of academic performance as a starting point. The first stage is the standard one of quantifying the sources of variability in the academic domain of interest; behavior-genetic research is better-placed to do exactly that than other methods which are not genetically sensitive, as Asbury and Wai (2019) point out. In the case of data from our samples, this research immediately rules out the shared environment as a major source for early reading differences when the educational environment is uniform, with the implications for funding that we have advocated. Stage 2 is to drill down into processes underlying the variance components, as we have illustrated for classroom effects in the knowledge that in some samples shared environment effects are higher than in ours. Our finding that twins in separate classes are almost as similar on literacy measures as those sharing a classroom (Grasby, Little et al., 2019) suggests that the classroom is not the site of whatever shared environment effects are in place.

Stage 2 also applies to unique environment influences and to genetic ones. To drill below the surface of unique environment influences on a phenotype, a valuable resource is the discordant monozygotic twin design. When identical twins in a pair are substantially different for the phenotype (i.e., discordant), the reason cannot reside in genes nor, by definition, in shared environment influence. Larsen et al. (2019) exploited this design to explore sources of unique environment for reading, writing, and numeracy equivalent of two years or more. A variety of biological conditions (such as early exposure to anaesthesia in one twin but not the other), personality factors (such as a difference between the twins in persistence in the face of challenge), and school events (such as the placement of one twin into a higher academic track which discouraged the twin left in the regular classroom) presented as candidates worth further investigation. Asbury, Moran and Plomin (2016) conducted a study with the same design and with some observations that aligned with those of Larsen et al. Some of these factors may have policy implications for educators, such as reconsidering the merits of academic tracking or implementing strategies to help students who give up in the face of challenges to become more persistent. In addition, the existence of these substantial differential effects even in identical twins reinforces the value of individually-targeted funding.

In the case of genetic influence, a twin study restricted to analysing the phenotype cannot drill down into the molecular bases of heritability but can explore hypotheses about the underlying processes intervening between genes and the phenotype of interest. These processes are often referred to as endophenotypes. An example of such a research strategy can be seen in Byrne et al. (2013), where the ability to form print-speech bonds, considered more basic that reading itself, was shown to be substantially heritable and genetically highly correlated with real word identification. Byrne at al. argue that their results mean that students struggling with reading will need many more exposures to printed words than other students will to fix them in memory, with implications for inclusion of repeated reading of appealing materials in remedial practices.

Behavior-genetics in the educational research landscape.

In an important recent article, Singer (2019) has urged educational researchers to adopt quantitative methods beyond those typically favoured, such as randomized control trials, to include longitudinal analyses, studies that refine assessment and measurement, and ones that employ the techniques grouped together as data science. In a sense, in this contribution we are joining with Asbury and Wai (2019) in advocating the further expansion of any list to include genetically-sensitive methods. One straightforward reason for doing this is simply to further clarify the factors that influence students’ academic performance, particularly genetic ones. But an additional reason is that quantifying the influence of genetics on academic performance and growth places some kind of floor under how influential purely environmental factors can be. This floor is not a fixed one; we know that estimates of heritability can vary with social, educational and geographic circumstances (Grasby et al, 2017; Samuelsson et al., 2008), and we know that extraordinary interventions can ameliorate the effects of compromised genetic endowment, as with the metabolic disorder, phenylketonuria. But in educational jurisdictions where genetic influence on literacy is known to be substantial (e.g., in the range of .60 to .75 for reading and spelling in Australia, Grasby et al., 2016), the notion that there exist simple but highly effective environmental “fixes,” such as reducing class sizes or supplying all students with laptops, is likely to be over-optimistic.

The degree to which methods advocated by Singer (2019) can be combined with genetically-sensitive methods, the field will advance more assuredly. Already, there is a number of studies that do this; longitudinal twin studies of literacy growth that have asked whether the already-known stability of individual differences in literacy growth is due to genes that are shared over development or to aspects of the environment that are shared (e.g., Grasby & Coventry, 2016); in the Smith-Woolley et al. (2018) paper cited earlier, “big data” in the form of prior genome-wide screening of 293,723 individuals was harnessed to explore whether private and selective state schools in the UK “value add” to academic achievement compared with non-selective state schools. These are welcome developments in exploiting the merits of distinct quantitative methods.

Conclusion

Thus, we applaud Asbury and Wai (2019) for bringing behavior-genetic perspectives to the attention of educators through the pages of this journal. As they point out, educational research that does not control for genes is at risk of overinterpreting environmental influence on educational outcomes. In this commentary we have attempted to show that, in contrast to their more conservative approach, there already exist sufficiently robust results to make specific policy recommendations, in particular about the distribution of funds to support students with or at risk for RD. We have also attempted to codify the progression that research can take; drilling down into variance components once they have been quantified for particular samples at particular times. And we have suggested that adding behavior-genetic methodology to the armory of available or emerging quantitative techniques in educational research will enrich the field. More policy recommendations will surely follow as psychologists, geneticists and educators continue to cooperate in a joint research endeavour.