The effects of inclusion on academic achievement,socioemotional development and wellbeing of children with special educational needs

One size does not fit all—Inconsistent effects of inclusion on learning and psychosocial adjustment of children with special needs

The effects of placing children with special needs in grades K-12 into inclusive educational settings are inconsistent. Findings from this review suggest that, in general, inclusion neither increases nor decreases learning and psychosocial adjustment of children with special needs in the OECD countries.

What is this review about?

This review and meta-analysis summarises evidence from studies exploring the effects of inclusion on children with special needs in regular educational settings, compared with segregated educational settings.

The review summarises evidence from 15 non-randomised studies in three meta-analyses on language and literacy outcomes, math outcomes and overall psychosocial adjustment of children.

What is the aim of this review?

This review aims to explore the effects of placing children with special needs in inclusive educational settings, compared with segregated placement on measures of academic achievement and psychosocial adjustment. A secondary aim is to explore how different child or setting characteristics (gender, age, type and severity of special need, part- or full-time inclusive education, and co-teaching) might moderate outcomes.

What studies are included?

This review includes studies that evaluate the effects of placing children with special educational needs in mainstream educational settings (inclusion). A total of 94 studies were identified. However, only 15 of these were assessed to be of sufficient methodological quality to be included in the meta-analyses and only between six and eight studies could be synthesised in each of the meta-analyses.

Four studies were from the USA, three from the UK, two from the Netherlands, and one each from Switzerland, Finland, Germany, the Czech Republic, Belgium and Denmark.

The studies included children with multiple types of disabilities such as learning disorders/intellectual disabilities, autism spectrum disorders, ADHD, physical handicaps, visual impairments, and Down syndrome.

The studies all had important methodological weaknesses. None of the included studies used experimental designs with random assignment.

What are the main findings of this review?

Results of the meta-analyses do not suggest any consistent positive or negative effects of inclusion on children's academic achievement as measured by language, literacy and math outcomes, or on the overall psychosocial adjustment of children. The studies in the analysis demonstrated a wide range of both large positive and large negative effect sizes; and although the average effect sizes did favour inclusion, they were small and none were statistically significant.

How has this intervention worked?

Supporters of mainstreaming or inclusion claim that segregated educational placement causes stigmatisation and social isolation, which may have detrimental effects on the self-concept, social identity, and self-confidence of students with special educational needs.

On the other hand, opponents suggest that placement in general education classrooms may have adverse effects for children with special needs, especially if the time and resources allocated for individualisation and differentiation are not aligned with student needs.

In line with these opposing positions, the findings from this review suggest that while some children with special needs may benefit from inclusive educational placement, other children may benefit from traditional special education in a segregated setting. Unfortunately, it was not possible to explore the effects of different kinds of inclusive education for different kinds of children with special needs.

What do the findings of this review mean?

The overall effects of inclusion on the academic achievement and overall psychosocial adjustment of students with special needs are inconsistent. Our findings are very similar to the results of previous systematic reviews and meta-analyses, which include studies published before 2000. It is very unlikely that inclusion in general increases or decreases learning and psychosocial adjustment in children with special needs.

These findings point to the need for an individual assessment of the specific child's educational and psychosocial needs rather than a one-size-fits-all approach to placement in special education.

Research should explore the effects of different kinds of inclusive education for different kinds of children with special needs, to expand the knowledge base on what works for whom.

How up-to-date is this review?

The review authors searched for studies up to 30 September 2021.

Description of the condition

The term children with special educational needs (SEN) refers to highly diverse populations of children with a wide range of physical, cognitive, and socioemotional disabilities or difficulties as well as strengths and resources causing them to require varying degrees of special educational support and assistance (Squires, 2012; Vehmas, 2010; Wilson, 2002).

Several studies document significant gender imbalances in the number of students who receive special educational support, and for most disability categories, the prevalence is higher for boys (Skårbrevik, 2002). The reasons for these imbalances are likely complex, and it is beyond the scope of the present review to account for the hypotheses and findings associated with each disability category. However, a general hypothesis across different categories of SEN is the notion that SEN are more likely to remain undetected in girls as symptoms and difficulties in girls may be less visible to educators (Arms et al., 2008).

Traditional special education consists of segregating students with special needs from mainstream students within separate and typically smaller classrooms or educational settings. However, as early as in the 1930s, a movement originally known as mainstreaming, and in more recent years as inclusion, has sought to bring an end to segregated placement as the preferred educational option for students with special needs (Carlberg & Kavale, 1980). In 1994, the idea of inclusive education became even more consolidated when the Salamanca Statement was adopted by representatives from 92 countries, resulting in an international shift in policy. This meant that far more students with special needs started entering general educational settings around the globe (Murawski & Lee Swanson, 2001; Ruijs & Peetsma, 2009).

The terms inclusion, mainstreaming, integrated placement, and cross-categorical instruction all refer to educational settings with a group composition consisting of a mixture of students with and without SEN. In the present review, we have chosen to use the term inclusion to refer to general educational settings in which there is a mixture of students with and without SEN. Ideally, inclusion and inclusive education should be based on an educational approach in which the goal is to continuously address and respond to the diversity of needs of all learners through increasing participation and reducing exclusion within and from education. Inclusion thus may involve changes and modifications in content, approaches, structures, and strategies, with a common vision which covers all children and a conviction that it is the responsibility of the regular system to educate all children. Inclusion emphasises the provision of opportunities for equal participation of children with disabilities (physical, social, and/or emotional) whenever possible into general education, but leaves open the possibility of personal choice and options for special assistance and facilities for those who need it (UNESCO, 2019). Inclusion as an ideological and theoretical movement was built on a philosophical foundation, but during the last sixty years, the number of empirical studies addressing inclusive education has grown tremendously. However, findings on the efficacy of inclusion on student outcomes are still far from unequivocal (Kavale & Forness, 2000; Lindsay, 2007; Ruijs & Peetsma, 2009). This is where the present review contributes, as the aim of the review was to summarise contemporary evidence on the effects of inclusive education when compared to a traditional segregated approach on students' learning, socioemotional adjustment, and well-being. It is important to consider the types of students who might benefit most from inclusive education. As stated earlier, it is possible that the effects of inclusive education may be different for girls and boys. Further, a child's cognitive and socioemotional skills and needs develop throughout childhood and adolescence (Lightfoot et al., 2009) and thus, it is possible that the potential benefits of inclusive education may vary depending on age and other characteristics of the child.

Description of the intervention

At the core of inclusion is the principle that students with special or additional learning needs or disabilities belong in mainstream education. The fundamental principle of inclusive education is that all children should learn together, regardless of any difficulties or differences (UNESCO, 2019; Winter & O'Raw, 2010). However, operationally within the present review, we define inclusion as an educational setting with a mixture of children with and without SEN. In the present review, the intervention termed inclusion may thus be defined as any given group composition within a general educational setting which contains at least one child with an identified SEN.

Based on the core principles of inclusion, there are many ways in which inclusion may be practised and adjusted, and thus there are numerous characteristics within the inclusive setting, which may vary across the included studies. This review included studies of all kinds of inclusive education meaning that placement in the inclusive setting may be full time or part-time. Special education students are a diverse group, as outlined in the section entitled ‘Types of participants’, and studies were included regardless of the type of SEN of the student population and regardless of the ratio of students with and without special needs within the inclusive setting. We included studies in which the general education teachers are provided with support and continuing professional development aimed at helping the teachers accommodate the needs of special education students and studies of inclusive settings in which no such support was offered to the teachers. It is often referred to as co-teaching when two or more professionals deliver substantive instruction to a diverse or blended group of students within the same physical space (Murawski & Lee Swanson, 2001). In the present review, we also included studies in which special education teachers and/or teaching assistants were present within the general education setting (co-teaching) and studies in which they were not.

This list of possible variations in student and classroom characteristics is not exhaustive, and in summary, within the present review, we included studies of all kinds of inclusive education as long as the studies were aimed at exploring the effects of inclusion in comparison to segregated special educational settings. We planned to conduct moderator analyses to explore the impact of specific characteristics of the inclusive educational setting and the characteristics of SEN on student outcomes.

How the intervention might work

Considering the rapid global development towards inclusive educational placement for students with SEN, there is a rather surprising lack of pedagogical, psychological, or didactic theories regarding the specific ways in which inclusive education may affect students with special needs' academic and socioemotional development. Group composition within the educational setting may play a significant role in determining the academic achievement, socioemotional development, and overall wellbeing of students with special needs. Theoretically and ideologically, scholars favouring mainstreaming or inclusion propose that segregated educational placement causes stigmatisation and social isolation which may have detrimental effects on the self-concept, social identity, and self-confidence of students with SEN (Dyssegaard & Larsen, 2013). Secondly, being placed in a general education classroom along with typically developing peers is proposed to benefit students' academic growth through peer effects (Rea et al., 2002; Tremblay, 2013). Finally, it is hypothesised that social interaction with general education peers may provide developmental opportunities that are not present in smaller, specialised units (Fisher & Meyer, 2002).

On the other hand, opponents of inclusive education suggest that placement in general education classrooms may have adverse effects for students with special needs, especially if the time and resources allocated for individualisation and differentiation are not aligned with student needs. In such cases, students' learning opportunities and wellbeing may also suffer, resulting in damages to self-concept (Zeleke, 2004), social isolation or bullying (Monchy et al., 2004; Pijl et al., 2010), stress (Pitt & Curtin, 2004), negative self-perception, and lower self-confidence (Bakker et al., 2007; Ruijs & Peetsma, 2009).

Hegarty (1993) provides a narrative review of the literature on inclusion and suggests that a number of factors are associated with positive student outcomes in inclusive settings. These are: (1) instruction based on student achievement needs, (2) materials and procedures that allow students to proceed at their own pace, (3) additional time for students who need it, (4) increased student responsibility for their own learning, (5) co-operation among students in achieving goals, (6) support teaching, and (7) collaboration among special and general education teachers.

In sum, the impact of inclusion on the academic achievement, socioemotional development, and wellbeing of students with special needs may be hypothesised to be both positive and negative, and the current knowledge base is rather unclear, leaving special educators and policymakers uncertain when making decisions on special education provision.

Why it is important to do this review

Since the 1980s, a number of reviews on the impact of inclusion on students with special needs have been published (Hegarty, 1993; Madden & Slavin, 1983; Ottenbacher & Cooper, 1982; Wang & Baker, 1985). Results are equivocal, and several reviews point to a number of methodological challenges and weaknesses of the study designs within the included primary studies. In summary, most reviews suggest a neutral or small positive impact of inclusion on most outcomes. However, all reviews also point to the need to study the impact of potential moderators more thoroughly, as there may be several interaction effects between student and classroom characteristics, such as student disability category × proportion of students with disabilities within the classroom and disability category × presence of teaching assistants. Therefore, it was important to conduct the present review to explore the impact of potential moderators associated with student and classroom characteristics.

In the following section, we present the existing reviews and their main findings.

In a systematic review and meta-analysis, which included 50 primary studies exploring the effects of special versus regular class placement for children with special needs, Carlberg and Kavale (1980) concluded that for students with special needs consisting of below average IQs, special classes were significantly inferior to regular classes on all outcome measures (separate analyses were carried out for achievement, social/personal, and other measures). However, for students with behavioural disorders, emotional disturbances, and learning disabilities, special classes were superior to regular classes.

Madden and Slavin (1983) conducted a narrative review of the effects of mainstreaming/inclusion on students with mild academic disabilities. The review does not include a description of the search strategy for identifying records or the criteria used to determine eligibility for inclusion in the review. The review concludes that among methodologically adequate studies, findings indicate few benefits on academic and social outcomes of placement in full-time special education compared with part-time placement with resource support or full time regular class placement for students with mild academic disabilities.

Ottenbacher and Cooper (1982) conducted a systematic review and meta-analysis, which included 43 primary studies exploring the effects of class placement (special class, regular class and resource class defined as placement in regular education classroom with resource support and the possibility for part-time segregated education) on the social adjustment of students with mild cognitive disabilities. The overall results suggest a very small effect in favour of special class placement over regular class placement. However, when special class placement was compared with resource class placement, results were insignificant but favouring resource class placement.

Wang and Baker (1985) conducted a systematic review and meta-analysis, which included 11 primary studies exploring the effects of mainstreaming/inclusion on children with SEN. To be eligible for inclusion in this review, primary studies needed to provide information on the effects of mainstreaming on students with special needs placed in a regular education setting. The studies had to use a control group consisting of special needs students with comparable impairment classifications placed in a segregated learning environment. The final selection of studies included 11 studies published between 1975 and 1984. The included studies used a wide variety of outcomes, but within the meta-analysis, outcomes were synthesised into three categories: performance, attitudinal, and process effects, and separate analyses were carried out for each of the three outcome categories. The study found small-to-moderate beneficial effects of inclusion on all outcomes with an overall mean weighted effect size across all studies and all three categories of outcomes of 0.33.

Hegarty (1993) provides a narrative review of the literature on integration (inclusion) of students with different disabilities. The narrative review is based on a literature review which was commissioned by the Centre for Educational Research and Innovation under the Organisation for Economic Cooperation and Development (OECD) and was conducted by researchers in five different countries. The review does not include a description of the search strategies for identifying records or the criteria used to determine eligibility for the selected primary studies. Within the review, a number of factors which are associated with effective integration programmes are identified.

Baker et al. (1995) describe a review and meta-analysis by Baker et al. (1995), which included 13 primary studies exploring the effects of inclusive placement on academic and social outcomes for students with special needs. We have been unable to retrieve the original publication, but according to Baker et al. (1995), this study found a very small effect in favour of inclusive placement on academic outcomes (0.08) and a small to moderate effect on social outcomes (0.28).

Sebba and Sachdev (1997) provide a review as part of a research report on what works for whom in inclusive education. The review does not include a description of the search strategy or the criteria for inclusion/exclusion of studies for the review. Within the research report, the authors suggest an overall positive impact of inclusive education and list a number of potential moderators such as attitudes of teachers and parents as well as a number of recommendations for the implementation of inclusive education.

McGregor and Vogelsberg (1998) provide a narrative review of studies of both the effects of inclusive schooling on student outcomes and studies focusing on issues related to the implementation of inclusion. The review includes both quantitative and qualitative studies including case studies. Results are difficult to synthesise, but suggest an overall positive impact of inclusion based on the main findings: (1) students with disabilities demonstrate high levels of social interaction in settings with typically developing peers, but placement alone does not guarantee positive social outcomes; (2) interactive small group contexts facilitate skill acquisition and social acceptance; (3) friendships develop between students with disabilities and typically developing peers.

Freeman and Alkin (2000) conducted a systematic narrative review in which it was concluded that on measures of academic achievement and social competence, children with mental retardation placed in general education perform better than children with mental retardation placed in special education classrooms. This review was only about children with mental retardation and did not include meta-analyses.

Murawski and Lee Swanson (2001) conducted a systematic review and meta-analysis which included 6 studies exploring the effectiveness of co-teaching on student outcomes of both general education students and students with SEN. Co-teaching was defined as two or more professionals delivering substantive instruction to a diverse or blended group of students within a shared/common physical space, and thus in this review co-teaching is a form of inclusion. The outcomes within the included studies were grades, achievement scores, social, and attitudinal outcomes. The review found co-teaching to be effective (average total effect size of 0.40). It is unclear what the control conditions within the included studies were and two of the included studies did not have a control group, but used a pre-test/post-test research design.

Lindsay (2007) provides a narrative review of the effectiveness of inclusive education for students with SEN. The review provides a historical overview of the vast literature before 2000 and a search of studies published 2001–2005 in eight journals on special education. The search identified 1373 studies and points to the fact that only 1% of the identified papers were comparative outcome studies. The review concludes that there is a lack of evidence for the effectiveness of inclusion and argues that where evidence does exist, the balance is only marginally positive. Lindsay (2007) thus supports the need for an updated systematic review and meta-analysis on the effectiveness of inclusion for students with special needs, with special attention to the potential impact of student and classroom moderators.

In a systematic narrative review of the effects of inclusion on both learning and socioemotional outcomes of students with and without special needs, Ruijs and Peetsma (2009) point to mixed findings regarding the effects of inclusion on student outcomes and suggest a number of potential moderators. The authors conclude that there is a need for more research. This review has not been updated since publication and doesn't include meta-analyses.

In 2009, a systematic review of evidence comparing the academic performance of students with special needs in different educational settings was carried out by the Canadian Council on Learning. The review included 30 primary studies. The search strategy for identifying studies was not described. The included studies examined students with learning disabilities, intellectual disabilities, language impairments, and mixed disabilities. The quality of each study was rated as either ‘high’, ‘medium’ or ‘low’ based on criteria related to transparency and research design, and effect sizes were retrieved. No meta-analyses were carried out, but the authors provide tables illustrating the number of effect sizes for each disability category favouring either inclusive or segregated settings along with the quality ratings of the studies from which they were retrieved. The authors point to mixed findings, but conclude that the balance of evidence shows favourable academic outcomes for students with SEN educated in inclusive settings, however they also note that results are not homogenous, and that effects are generally small in magnitude.

Dyssegaard and Larsen (2013) provide a systematic review and narrative synthesis on the effects of including children with special needs in mainstream teaching in primary and lower secondary school, and on which of the applied educational methods have proven to have a positive effect. The narrative synthesis is based on 43 studies of which 16 studies were deemed to have a ‘high level of evidence’. The included studies consist of randomised controlled trials, non-randomised controlled trials, systematic reviews, cohort studies, longitudinal studies, and studies using a pre-test/post-test design. The systematic review included studies focusing on outcomes for both mainstream and special needs students and does not include a meta-analysis. The conclusion points to mixed findings regarding the overall effectiveness of inclusion on the academic achievement and psychosocial adjustment of special needs students and suggests that the effects may vary depending on the age of the child and the overall school and teacher attitudes towards inclusion. Furthermore, the review suggests that the effectiveness of co-teaching may depend on the educational background and continuous professional development of both special and general education teachers and of teaching assistants.

Carrol et al. (2017) provide a rapid evidence assessment of studies focused on approaches, strategies, and interventions supporting children and young people with SEN in mainstream schools. The rapid evidence assessment is based on a systematic search in a single database (ERIC) as well as a strategy of consulting experts within the relevant fields. The initial search identified 1046 papers of which 505 were later excluded due to low quality of evidence. The rapid evidence assessment points to a number of implementation strategies, pedagogical, and didactic approaches which have shown positive results. Furthermore, the study points to evidence gaps and suggests the need for further research. The rapid evidence assessment does not include a meta-analysis.

In the present review, besides being up-to-date, we conducted an extensive risk of bias assessment of all included studies, and we provide separate meta-analyses for each conceptual outcome (academic achievement: language/literacy outcomes and math outcomes, and psychosocial adjustment). Furthermore, we hoped to be able to conduct moderator analyses based on the children's specific disability categories and the specific type of inclusion setting.

Traditional segregated special education is costly and in a time of increased interaction between special and general education systems and constraints on educational spending, policymakers must consider the cost-efficiency of different special needs provisions.

As more students with SEN enter general education settings, educators and policymakers must consider how the needs of these students are met in different settings and on what grounds placement in general or special educational settings should be determined. As previously noted, the current knowledge base is ambiguous with many findings suggesting a complex interplay between student and classroom characteristics (Carlberg & Kavale, 1980; Mesibov & Shea, 1996; Peetsma et al., 2001), leaving special educators and policymakers uncertain when making decisions on special education provision and highlighting the need for a comprehensive review of the effectiveness of inclusion on student outcomes.

Criteria for considering studies for this review

Search methods for identification of studies

Data collection and analysis

The total number of potential relevant studies constituted 20,183 hits after duplicates were removed. A total of 94 studies met the inclusion criteria and were critically appraised by the review authors. The 94 studies analysed data from 87 different samples. Only 15 studies (analysing 15 different samples) could be used in the data synthesis. 79 studies could not be used in the data synthesis, as they were judged to have too high risk of bias and, in accordance with the protocol, were excluded from the meta-analysis on the basis that they would be more likely to mislead than inform. One study (Rangvid, 2021) did not have any outcomes in common with the other 14 studies and could therefore not be synthesised in a meta-analysis but was reported in an additional table.

Meta-analyses of three child outcomes (overall psychosocial adjustment, language/literacy learning outcomes, and math learning outcomes) were conducted on each metric separately. All analyses were inverse variance weighted using random effects statistical models that incorporate both the sampling variance and between study variance components into the study level weights. Random effects weighted mean effect sizes were calculated using 95% confidence intervals (CIs).

A number of other relevant outcomes were reported in a single study only (measures on academic tests other than language/literacy and math, learning motivation, parent and teacher reported SDQ on friendship and bullying, socialisation skills and school completion rates) and could therefore not be synthesised in the meta-analyses. The effect sizes and 95% CIs were calculated and reported in an additional table.

Description of studies

See supplementary descriptive table.

Risk of bias in included studies

The detailed risk of bias coding for each of the 94 studies is available in a supplementary document.

Summary scores from the risk of bias assessment are shown in Table 6. All 94 studies were non-randomised studies and were rated using the ROBINS-I tool. If a study reported on several outcomes, a risk of bias assessment was carried out for each of those outcomes. If ratings differed, only the most favourable assessment is reported in Table 6, since this assessment may have been a reason to include the study in the meta-analysis. As stated in the protocol, we stopped the assessment of a non-randomised study outcome when it was rated ‘Critical’ on any of the items, therefore not all studies are rated on all domains.

Seventy-nine studies were rated with an overall critical risk of bias, corresponding to a risk of bias so high that the findings should not be considered in the data synthesis. The overall Critical risk of bias rating was mainly due to issues on the Confounding bias item; 69 were rated Critical risk of bias on this item; that is, they failed to establish a comparison group that was balanced on important confounders and further either did not control for any confounders. One study had a critical risk of selection bias, three studies had a critical risk of classification bias (of which two also were rated Critical on confounding) and seven studies were judged to be in critical risk of bias due to ratings of serious risk of bias in multiple domains (six on confounding and selection bias and one on confounding, selection, measurement and reporting bias).

Fourteen studies were rated with an overall serious risk of bias, and one study with an overall moderate risk of bias.

Of the 25 studies not rated Critical risk on the Confounding item, 20 had Serious issues on this item, three were rated Moderate risk of bias and two were not rated on this item. Besides the one study rated Critical risk of bias on the Selection bias item, one study was rated low risk of bias, nine were rated Moderate risk of bias and 13 were rated Serious risk of bias. Besides the three studies rated Critical risk of bias on the Classification bias item, eight studies were rated Low risk of bias, five were rated Moderate risk of bias and three were rated Serious risk of bias. Six studies were rated Low risk of bias on the Deviation item, two were rated Moderate and one was rated Serious risk of bias. On the Missing data item 12 were rated Low risk of bias and one was rated Moderate risk of bias. On the Measurement item, five were rated Low risk of bias, nine were rated Moderate and two were rated Serious risk of bias. Fourteen studies were rated Moderate risk of bias on the Selection of Reported Results mainly because there was no a priori analysis plan and a further two were rated Serious risk of bias as they had other issues in addition to no a priori analysis plan.

The risk of bias assessment thus left us with only 15 studies which could be used in the data-synthesis, 14 with an overall Serious risk of bias rating and one study which was rated as in overall Moderate risk of bias.

Effects of interventions

Fifteen studies provided standardised mean differences and variances or data that enabled calculation of standardised mean differences and variance effect estimates that could be used in the data synthesis. No adjustments were necessary for clustering as the studies did not analyse whole classes (only one or a few students in a class).

One study (Cosier, 2013) analysed inclusion as a continuous variable, that is, time spent in general education. Thus, the reported coefficients reflect the effect of a 1-h increase in time spent in general education on student achievement. Mean time spent in general education as well as the standard deviation of the time spent in general education was reported. We will use the effect of a standard deviation increase in the time spent in general education in the data synthesis and investigate the robustness of results in the sensitivity analysis. Thus, the results of the study Cosier (2013) will reflect an increase in 11.6 h spent in general education.

OPEN IN VIEWER

Figure 1

Flowchart

Meta-analysis

Child outcomes

Overall psychosocial adjustment

Eight studies analysed the effect of inclusion on the overall psychosocial adjustment of children with special needs as reported by themselves, by their parents or by teachers. Measures used include: Vineland Adaptive Behaviour Scale, Socialisation Subscale; Teacher Report Form; Self-Perception Profile for Adolescents; The Personality Assessment Questionnaire; The Strengths and Difficulties Questionnaire; ADOS-G, module 3: Child self-report; Learning motivation based on the Finnish LTL framework subscale for Avoidance Orientation. The random effects weighted standardised mean difference favouring the intervention group was 0.20 (95% CI: −0.01 to 0.42) and not statistically significant. The forest plot is displayed in Figure 2. There is some heterogeneity between the studies; the estimated τ ² is 0.05, Q = 18.64, df = 7 and I ² is 62% as displayed in Figure 2. Although the heterogeneity as indicated by I ² was moderate, the estimate of τ ² indicates that the evidence for this outcome is inconsistent, in the sense that the distribution of effect sizes spans both positive and negative effect sizes. To ease interpretation concerning the implication of heterogeneity, we calculated a 95% prediction interval. A prediction interval reflects the variation in treatment effects over different settings, including what effect is to be expected in future primary studies in similar settings. The range was (95% PI: −0.41 to 0.81) and based on this prediction interval we calculated the probability that inclusion in a future setting will have a true-negative effect to be 22%.

OPEN IN VIEWER

Figure 2

Meta-analyses: Overall psychosocial adjustment

Language and literacy

Six studies analysed the effect of inclusion on language and literacy learning outcomes as measured by standardised and curriculum based tests. The random effects weighted standardised mean difference was 0.04 favouring the intervention (95% CI: −0.27 to 0.35) and not statistically significant. The forest plot is displayed in Figure 3. There is some heterogeneity between the studies; the estimated τ ² is 0.11, Q = 39.63, df = 5 and I ² is 87% as displayed in Figure 3. Although the average effect of inclusion was close to a null value it reflects that there were both large negative and large positive effect sizes in the primary studies. The prediction interval (95% PI: −0.98 to 1.06) also reflects that the study effects were dispersed over a wide range. The probability that inclusion in a future setting will have a true-negative effect was estimated to be 46%.

OPEN IN VIEWER

Figure 3

Meta-analyses: Language and Literacy

Math

Eight studies analysed the effect of inclusion on math learning outcomes as measured by standardised and curriculum-based tests. The random effects weighted standardised mean difference was 0.05 favouring the intervention (95% CI: −0.16 to 0.26) and not statistically significant. The forest plot is displayed in Figure 4. There is some heterogeneity between the studies; the estimated τ ² is 0.06, Q = 28.72, df = 7 and I² is 76% as displayed in Figure 4. The average effect of inclusion was close to zero, but the primary study effects were dispersed over a wide range, giving a prediction interval of (95% PI: −0.60 to 0.70) and the probability that inclusion in a future setting will have a true-negative effect was estimated to be 43%.

OPEN IN VIEWER

Figure 4

Meta-analyses: Math

Other outcomes

A number of relevant outcomes which we could not synthesise in the meta-analyses, as they were only reported in a single study each, were extracted. The outcomes were measures on academic tests other than language/literacy and math, learning motivation, parent and teacher reported SDQ on friendship and bullying and socialisation skills. The effect sizes and 95% CIs are reported in Table 7.

OPEN IN VIEWER

Table 7

Additional outcomes

Study	Measure	Outcome	SMD [95% CI]
Hienonen et al., 2021	Average of Finnish, mathematics, foreign language, and science	GPA	−0.42 [−0.66, −0.18]
Hienonen et al., 2021	Ninth grade test	Foreign language	−0.26 [−0.50, −0.02]
Hienonen et al., 2021	Ninth grade test	Science	−0.33 [−0.57, −0.09]
Hienonen et al., 2021	Learning motivation scales	Mastery extrensic orientation	−0.31 [−0.55, −0.07]
Hienonen et al., 2021	Learning motivation scales	Perfomance approach orientation	−0.28 [−0.52, −0.04]
Hienonen et al., 2021	Learning motivation scales	Mastery intrinsic orientation	0.22 [−0.02, 0.46]
Hienonen et al., 2021	Learning motivation scales	Performance avoidance orientation	0.07 [−0.17, 0.31]
Rowley, 2012	SDQ-P	Friendships	0.02 [−0.61, 0.65]
Rowley, 2012	SDQ-P	Fighting/bullying	0.84 [0.21, 1.47]
Rowley, 2012	SDQ-P	Victimisation (being bullied)	0.05 [−0.62, 0.72]
Rowley, 2012	SDQ-T	Friendships	0.04 [−0.65, 0.73]
Rowley, 2012	SDQ-T	Fighting/bullying	0.31 [−0.42, 1.04]
Rowley, 2012	SDQ-T	Victimisation (being bullied)	−0.76 [−1.41, −0.11]
Rowley, 2012	ADOS-G, module 3: Child self-report	Friendships	−0.69 [−2.34, 0.96]
Reed, 2012	Vineland Adaptive Behaviour Scale (VABS)	Socialisation skills	0.85 [0.44, 1.26]

			Risk difference [95% CI]
Rangvid, 2021	.	Enrolment in a qualifying upper secondary education programme year 1 after secondary education	0.21 [0.16, 0.25]
Rangvid, 2021	.	Enrolment in a qualifying upper secondary education programme year 2 after secondary education	0.15 [0.11, 0.20]
Rangvid, 2021	.	Enrolment in a qualifying upper secondary education programme year 3 after secondary education	0.14 [0.07, 0.20]
Rangvid, 2021	.	Take school leaving exam	0.32 [0.30, 0.34]
Rangvid, 2021	.	Pass school leaving exam	0.29 [0.26, 0.32]
Rangvid, 2021	.	Pass school leaving exam and specific requirement for vocational education	0.24 [0.21, 0.27]
Rangvid, 2021	.	Pass ‘minimum competency threshold’ in core subjects (grade 4)	0.09 [0.05, 0.12]

Abbreviations: CI, confidence interval; SMD, standardised mean difference.

Secondary outcomes

One study (Rangvid, 2021) reported school completion rates (take school-leaving exam and pass school-leaving exam). The effect sizes and 95% CIs are reported in Table 7.

Child overall psychosocial adjustment

Co-teaching

Due to invariability between studies on this moderator it was not possible to conduct a subgroup analysis.

Autism spectrum disorder/other types of special needs

Of the eight studies providing effect estimates of overall child psychosocial adjustment, two studies used a sample of children with autism spectrum disorder only and six used samples of children with other types of special needs. The forest plot for the eight effect estimates is displayed in Figure 5. Pooled results for the two subgroups showed a statistically non-significant negative effect of −0.12 (95% CI: −0.46 to 0.22) for children with autism spectrum disorder and a statistically significant positive effect of 0.32 (95% CI: 0.05 to 0.59) for children with other types of disabilities. There was no heterogeneity of effects among studies in the subgroup of children with autism spectrum disorder (τ ¹ = 0.00, I ² = 0%) while there was some heterogeneity for children with other types of special needs (τ ² = 0.06, I ² = 70%). The CIs of the subgroups overlapped. There was no evidence to support the hypothesis that the effect differs for children with or without autism spectrum disorders.

OPEN IN VIEWER

Figure 5

Subgroup analyses, autism spectrum disorder/other types of special needs

Children with physical disabilities (including severe visual impairment/blindness)/all other types of disabilities

Of the eight studies providing effect estimates of the overall child psychosocial adjustment, two studies used a sample of children with physical disabilities and six used samples of children with other types of special needs. The forest plot for the eight effect estimates is displayed in Figure 6. Pooled results for the two subgroups showed a statistically non-significant negative effect of −0.06 (95% CI: −0.16 to 0.28) for children with physical disabilities and a statistically non-significant positive effect of 0.26 (95% CI: −0.06 to 0.57) for children with other types of disabilities. There was limited heterogeneity of effects among studies in the subgroup of children with physical disabilities (τ ² = 0.01, I ² = 20%) while there was some heterogeneity for children with other types of special needs (τ ² = 0.08, I ² = 61%). The CIs of the subgroups overlapped. There was no evidence to support the hypothesis that the effect differs for children with or without physical disabilities.

OPEN IN VIEWER

Figure 6

Subgroup analyses, children with physical disabilities (including severe visual impairment/blindness)/all other types of disabilities

Child mean age 15 or higher/child mean age less than 15

Of the eight studies providing effect estimates of the overall child psychosocial adjustment, three studies used a sample of children with a mean age of 15 or higher and five used samples of children with a mean age less than 15. The forest plot for the eight effect estimates is displayed in Figure 7. Pooled results for the two subgroups showed a statistically significant positive effect of 0.47 (95% CI: 0.08 to 0.86) for children aged 15 or higher and a statistically non-significant positive effect of 0.02 (95% CI: −0.10 to 0.15) for children younger than 15. There was some heterogeneity of effects among studies in the subgroup of children aged 15 or older (τ ² = 0.07, I ² = 59%) while there was very little heterogeneity for children younger than 15 (τ ² = 0.00, I ² = 6%). The CIs of the subgroups overlapped. There was no evidence to support the hypothesis that the effect differs for children aged 15 or higher versus children younger than 15.

OPEN IN VIEWER

Figure 7

Subgroup analyses, child mean age 15 or higher/child mean age less than 15

Child mean age 10 or higher/child mean age less than 10

Of the eight studies providing effect estimates of the overall child psychosocial adjustment, six studies used a sample of children with a mean age of 10 or higher and two studies used samples of children with a mean age less than 10. The forest plot for the eight effect estimates is displayed in Figure 8. Pooled results for the two subgroups showed a statistically non-significant positive effect of 0.27 (95% CI: −0.02 to 0.56) for children aged 10 or higher and a statistically non-significant positive effect of 0.05 (95% CI: −0.26 to 0.36) for children younger than 10. There was some heterogeneity of effects among studies in the subgroup of children aged 10 or older (τ ² = 0.07, I ² = 72%) while there was no heterogeneity for children younger than 15 (τ ² = 0.00, I ² = 0%). The CIs of the subgroups overlapped. There was no evidence to support the hypothesis that the effect differs for children aged 10 or higher versus children younger than 10.

OPEN IN VIEWER

Figure 8

Subgroup analyses, child mean age 10 or higher/child mean age less than 10

Math

Co-teaching

Of the eight studies providing effect estimates of math outcomes, two studies reported on an inclusive setting which was co-taught by teachers with qualifications in both general and special education and six studies had no mentioning of co-teaching in the inclusive setting or were unclear as to whether co-teaching occurred. It should be noted that it is possible that in some of the former studies, there was some form of co-teaching in the inclusive setting, which was just not specified in the published studies. The forest plot for the eight effect estimates is displayed in Figure 9. Pooled results for the two subgroups showed a statistically significant positive effect of 0.35 (95% CI: 0.07 to 0.63) for children in inclusive settings with co-teaching and a statistically non-significant negative effect of −0.04 (95% CI: −0.28 to 0.20) for children in inclusive settings in which there was no description of co-teaching. There was no heterogeneity of effects among studies in the subgroup of children in inclusive settings with co-teaching (τ ² = 0.00, I ² = 0%) while there was some heterogeneity for children in inclusive settings in which there was no description of co-teaching (τ ² = 0.06, I ² = 79%). The CIs of the subgroups overlapped. There was no evidence to support the hypothesis that the effect differs based on whether there is co-teaching in the inclusive setting.

OPEN IN VIEWER

Figure 9

Subgroup analyses, co-teaching

Autism spectrum disorder/other types of special needs

Due to invariability between studies on this moderator it was not possible to conduct a subgroup analysis.

Children with physical disabilities (including severe visual impairment/blindness)/all other types of disabilities

Due to invariability between studies on this moderator it was not possible to conduct a subgroup analysis.

Child mean age 15 or higher/child mean age less than 15

Of the eight studies providing effect estimates of math outcomes, two studies used a sample of children with a mean age of 15 or higher and six used samples of children with a mean age less than 15. The forest plot for the eight effect estimates is displayed in Figure 10. Pooled results for the two subgroups showed a statistically non-significant negative effect of −0.10 (95% CI: −0.88 to 0.68) for children aged 15 or higher and a statistically significant positive effect of 0.11 (95% CI: 0.02 to 0.20) for children younger than 15. There was a substantial heterogeneity of effects among studies in the subgroup of children aged 15 or older (τ ² = 0.30, I ² = 93%) while there was no heterogeneity for children younger than 15 (τ ² = 0.00, I ² = 0%). The CIs of the subgroups overlapped. There was no evidence to support the hypothesis that the effect differs for children aged 15 or higher versus children younger than 15.

OPEN IN VIEWER

Figure 10

Subgroup analyses, child mean age 15 or higher/child mean age less than 15

Child mean age 10 or higher/child mean age less than 10

Of the eight studies providing effect estimates of math outcomes, two studies used a sample of children with a mean age of 10 or higher and six used samples of children with a mean age less than 10. The studies with children with a mean age of 10 and older are the same as the studies with a mean age of 15 and older. The subgroup effects are thus identical to the subgroups analysed above. There was no evidence to support the hypothesis that the effect differs for children aged 10 or higher versus children younger than 10.

Sensitivity analysis

Sensitivity analyses were planned to evaluate whether the pooled effect sizes were robust across study design and components of methodological quality.

No randomised controlled trials were included in the meta-analyses, so the impact of study design could not be evaluated. For methodological quality, we carried out sensitivity analyses for the confounding, selection, classification, measurement and reporting risk of bias items of the risk of bias checklists. We examined the robustness of our conclusions when we excluded studies with a serious risk of bias assessment.

The analyses are performed separately by outcome, essentially replicating the meta-analyses conducted in Analysis 1.1, Analysis 1.2 and Analysis 1.3. We further examined the robustness of our conclusions when we increased the reported effects from a one standard deviation increase in time spent in general education to an increase of two standard deviations increase in time spent in general education in the study using time spent in general education as a continuous variable.

The results of the sensitivity analyses are provided in Tables 8–11 and displayed in Figures 11–14.

OPEN IN VIEWER

Table 8

Sensitivity analysis: Child overall psychosocial

	Number of studies	Mean	95% CI
Studies excluded	k	SMD	Lower	Upper
All	8	0.20	−0.01	0.42
Overall: All studies were rated serious	-	-	-
Confounding: All studies were rated serious	-	-	-
Selection: Serious risk of bias removed	6	0.17	−0.11	0.44
Classification: Serious risk of bias removed	6	0.04	−0.10	0.18
Deviation: No study was rated serious		-	-	-
Missing data: No study was rated serious		-	-	-
Measurement: Serious risk of bias removed	7	0.19	−0.04	0.43
Reporting: Serious risk of bias removed	7	0.19	−0.04	0.43

Abbreviations: CI, confidence interval; SMD, standardised mean difference.

OPEN IN VIEWER

Table 9

Sensitivity analysis: Language and literacy

	Number of studies	Mean	95% CI
Studies excluded	k	SMD	Lower	Upper
All	6	0.04	−0.27	0.35
Overall: all except one study were rated serious	-	-	-	-
Confounding: Serious risk of bias removed	2	0.31	−0.35	0.97
Selection: Serious risk of bias removed	2	−0.22	−0.77	0.33
Classification: Serious risk of bias removed	4	0.15	−0.31	0.60
Deviation: Serious risk of bias removed	5	−0.03	−0.36	0.31
Missing data: No study was rated serious	-	-	-	-
Measurement: No study was rated serious	-	-	-	-
Reporting: No study was rated serious	-	-	-	-

Abbreviations: CI, confidence interval; SMD, standardised mean difference.

OPEN IN VIEWER

Table 10

Sensitivity analysis: Math

	Number of studies	Mean	95% CI
Studies excluded	k	SMD	Lower	Upper
All	8	0.05	−0.16	0.26
Overall: all except one study were rated serious	-	-	-	-
Confounding: Serious risk of bias removed	2	−0.01	−0.19	0.17
Selection: Serious risk of bias removed	4	0.05	−0.11	0.22
Classification: Serious risk of bias removed	6	0.10	−0.05	0.25
Deviation: Serious risk of bias removed	7	0.01	−0.21	0.23
Missing data: No study was rated serious	-	-	-	-
Measurement: No study was rated serious	-	-	-	-
Reporting: No study was rated serious	-	-	-	-

Abbreviations: CI, confidence interval; SMD, standardised mean difference.

OPEN IN VIEWER

Table 11

Sensitivity analysis time spend in general education

	SMD	Lower	Upper
Sensitivity: Language and literacy
Cosier, 2013 used with 1 SD change in general education	0.04	−0.27	0.35
Cosier, 2013 used with 2 SDs change in general education	0.09	−0.29	0.46
Sensitivity: Math
Cosier, 2013 used with 1 SD change in general education	0.05	−0.16	0.26
Cosier, 2013 used with 2 SDs change in general education	0.08	−0.16	0.32

Abbreviation: CI, confidence interval.

OPEN IN VIEWER

Figure 11

Sensitivity analysis

OPEN IN VIEWER

Figure 12

Sensitivity analysis

OPEN IN VIEWER

Figure 13

Sensitivity analysis

OPEN IN VIEWER

Figure 14

Sensitivity analysis

There were no appreciable changes in the results following removal of any of the studies nor were there any appreciable changes of doubling the effect from the study using time spent in general education as a continuous variable.

OPEN IN VIEWER

Figure 15

Child overall psychosocial adjustment

OPEN IN VIEWER

Figure 16

Language and literacy

OPEN IN VIEWER

Figure 17

Math

In summary, the conclusions of the main syntheses do not change.

Summary of main results

Based on the 15 studies, which could be used in the data synthesis, three meta-analyses were conducted on the outcomes overall psychosocial adjustment, language/literacy, and math. Overall, there were too few studies included in any of the meta-analyses and the degree of heterogeneity was too high in order for us to draw any conclusion concerning the effectiveness of inclusion. At most, the results from eight studies could be pooled in a single meta-analysis. All the meta-analyses showed a positive weighted average and none of them were statistically significant. Heterogeneity was present in all analyses and 95% prediction intervals were wide reflecting that the distribution of effect sizes spanned both large positive and large negative effect sizes.

Overall completeness and applicability of evidence

In this review, we included a total of 15 studies in the data-synthesis. This number is relatively low compared to the large number of studies (94) meeting the inclusion criteria. The reduction was mainly due to the remaining studies being judged to be in critically high risk of bias (mainly confounding bias) and, in accordance with the protocol, we excluded these from the data synthesis on the basis that they would be more likely to mislead than inform on the size of the effect of the intervention. If all the 94 studies had provided an effect estimate with a lower risk of bias, the final list of useable studies in the data synthesis would have been larger which, in turn, would have provided a more robust literature on which to base conclusions.

In total, samples from 19 countries were represented by the 94 studies meeting the inclusion criteria. The 15 studies used in the data synthesis covered only nine countries (Belgium, Czech Republic, Denmark, Finland, Germany, the Netherlands, Switzerland, UK, and USA). The more narrow geographical coverage may limit the applicability of the evidence.

We analysed all available child outcomes. However, the small number of studies reporting on the same outcomes makes us reluctant to draw any firm conclusions, as only between six and eight studies could be pooled in a meta-analysis. Heterogeneity was present in all analyses and 95% prediction intervals were wide, reflecting that the distribution of effect sizes spanned both large positive and large negative effect sizes. The small number of studies unfortunately precluded a thorough investigation of heterogeneity and the impact of potential moderators, and thus more research is needed.

As can be seen from Table 5, a high proportion of the studies used in the meta-analyses are based on populations of children in which children with more severe disabilities or special needs were excluded from the analysis. This selection was typically done to ensure comparability between children placed in inclusive settings (treatment) and control children placed in segregated special education. This selection practice reflects the fact that in many school systems, placement of children with special needs is dependent on assessments of symptom severity, with an increased tendency for children with severe and pervasive difficulties to be placed in segregated special education and children with mild or moderate difficulties to be placed in inclusive settings. Therefore, estimating the effects of inclusion based on comparisons between children placed in inclusion vis a vis children in segregated special education is all but straightforward; and while excluding special education children with the most severe disabilities from the analytical sample may be a necessary step to ensure comparability in individual studies, it means that the current review has limited applicability when it comes to the placement conditions of children with the most severe types of special needs or disabilities. This is an unfortunate limitation given the need for more knowledge regarding the effects of educational placement for all children with SEN, regardless of disability type or symptom severity.

We found no indication of publication bias.

Findings from the present review are relevant to policy and decision-makers as well as to educational placement agencies, as they emphasise the need to base a child's educational placement on an individual assessment of the child's needs. The fact that there were no studies reporting on the long term effects of inclusion is perhaps not surprising as longitudinal research is costly and difficult to carry out. However, based on the theoretical hypothesis that inclusive placement decreases stigma, social isolation and marginalisation of children with special needs, it is possible that these effects do not manifest themselves while the child is still in school, and thus more research should explore this issue further, by including follow-up measurements of outcomes into adulthood.

Quality of the evidence

The overall methodological quality of the included studies was very low, no experimental studies in which children were randomly assigned to intervention and control conditions were identified. Overall the risk of bias in the included non-randomised studies was high. We examined the risk of bias using the model ROBINS–I, developed by members of the Cochrane Bias Methods Group and the Cochrane Non-Randomised Studies Methods Group (Sterne et al., 2016) for non-randomised studies. Among the included non-randomised studies, only 15 of 94 studies were not rated critical risk of bias, and of these 15 studies not rated critical risk of bias, only one study had an overall moderate risk of bias, while the remaining 14 studies were judged to be in serious risk of bias. The level ‘critical’ means: the study (outcome) is too problematic in this domain to provide any useful evidence on the effects of intervention, and it is excluded from the data synthesis.

The quality of the evidence in this review was enhanced by excluding studies assessed to be at critical risk of bias using the ROBINS–I tool from the data synthesis. We believe this process excluded those studies that are more likely to mislead than inform.

Furthermore, we performed a number of sensitivity analyses for each outcome to check whether the obtained results were robust across methodological quality, carrying out sensitivity analyses for the confounding, selection, classification, measurement, and reporting risk of bias items of the risk of bias checklists. We examined the robustness of our conclusions when we excluded studies with a serious risk of bias assessment. We further examined the robustness of our conclusions when we increased the reported effects from a one standard deviation increase in time spent in general education to an increase of two standard deviations increase in time spent in general education in the study using time spent in general education as a continuous variable. The overall conclusions did not change.

Overall, in all meta-analyses, there was a moderate amount of heterogeneity between studies and inconsistency in the direction and magnitude of effects. Some effects favoured inclusion and some effects favoured segregated placements.

Potential biases in the review process

We performed a comprehensive electronic database search, combined with grey literature searching, and hand searching of key journals. All citations were screened by two independent screeners from the review team, and one review author (NTD) assessed all included studies against inclusion criteria.

We believe that all the publicly available studies on the effects of inclusion for children with special needs in the OECD countries published after 2000 were identified during the review process. However, 23 references were not obtained in full text.

We were unable to comment on the possibility of publication bias as at most eight studies were included in the same meta-analysis. Thus, we cannot rule out that there are still some missing studies.

We believe that there are no other potential biases in the review process as two members of the review team (NTD, ANBO, TF, MWK, MHC) independently coded the included studies. Any disagreements were resolved by discussion. Further, decisions about the inclusion of studies were made by two members of the review team (MWK, MHC) and at least one review author (ANBO, TF, NTD). Assessment of study quality and numeric data extraction was made by one review author (TF/NTD/ANBO) and was checked by a second review author.

Members of the review team at VIVE Campbell were: Research assistants: Malene Wallach Kildemoes and Maluhs Haulund Christensen.

Agreements and disagreements with other studies or reviews

Lindsay (2007) provides a narrative review of the effectiveness of inclusive education for students with SEN. The review provides a historical overview of the vast literature before 2000 and a search of studies published 2001-2005 in eight journals on special education. The search identified 1373 studies and points to the fact that only 1% of the identified papers were comparative outcome studies. The review concludes that there is a lack of evidence for the effectiveness of inclusion and argues that where evidence does exist, the balance is only marginally positive (Lindsay, 2007). Our findings are very similar to the findings from Lindsay (2007) and to all previous systematic reviews and meta-analyses, suggesting that even though the number of published studies has grown a lot in the past two decades, the proportion of comparative outcome studies in this field is extremely low, and even when a control group is used, the vast majority of studies are still in serious risk of confounding bias due to the fact that educational placement is not random. To the contrary, educational placement is most often based on an assessment of the individual child's needs, and at the moment this seems like the most ethically sound choice since the evidence base on the effects of inclusion to this day provides little guidance.

Implications for practice

Based on the findings from the present review, it is clear that the quantitative evidence on the effects of inclusion does not suggest that placing children with special needs in inclusive settings is by default superior to placement in traditional segregated special educational settings on measures of children's psychosocial adjustment and learning, rather results should be interpreted with caution. However, we must emphasise that this absence of evidence is not evidence of absence. Heterogeneity was present in all analyses and 95% prediction intervals were wide, reflecting that the distribution of effect sizes spanned both large positive and large negative effect sizes. Further, the possibilities of investigating the impact of potential moderators was limited, and the moderator analyses were inconclusive due to the low number of studies included in each meta-analysis and due to limited reporting of and variability in the potential moderators.

It is possible that future focused primary studies or systematic reviews may be able to identify significant positive or negative effects for children with certain types of special needs or that specific models of inclusion may be more effective than others at increasing academic achievement, psychosocial adjustment and wellbeing in children with special needs. At the moment, the current evidence base suggests that one size does not fit all, and there is only limited contemporary evidence on what works for whom in special education when it comes to comparing inclusion to segregated placement for subgroups of children with special needs. Furthermore, as stated in the description of included studies in the review, 11 studies were based on child populations in which children with more severe disabilities were excluded from the analysis. This challenges the generalisability of the findings from the present review, and thus there is limited evidence on the effects of inclusion for children with more severe disabilities. Similarly, very few studies contained detailed descriptions of the resources available to children within both inclusive and segregated settings. This constitutes a serious flaw in the evidence base, as it is theoretically possible that the effects of inclusion are dependent upon the resources available to the children within the educational setting.

Therefore, the choice of educational placement should continue to be based on an individual assessment of the child and detailed knowledge of the specific local inclusion model available. As is evident from the included studies in this review as well as in findings from previous reviews, the scientific literature on the effects of inclusion is vast and comes from numerous countries (Lindsay, 2007; Madden, 1983; Ruijs, 2009), and while it is not possible to draw any definitive conclusions regarding the international overall effects of inclusion, practitioners may find useful locally relevant contemporary knowledge by reading some of the included studies within the present review or by exploring the qualitative research literature.

Implications for research

The overall methodological quality of the included studies was very low, no experimental studies in which children were randomly assigned to intervention and control conditions were found. The 15 studies which could be used in the meta-analyses were all except for one judged to be in serious risk of bias. The risk of bias assessment of the included studies suggests that the risk of confounding bias is a serious problem in almost all available studies. This finding is perhaps not surprising and reflects the fact that educational placement is very seldom random and most often reflects an assessment of the severity of the SEN. Based on this finding, it could be argued that the overall direction of bias favours inclusion, as most of the included studies within the review described school placement procedures in which children with more severe special needs were more likely to be placed in segregated settings.

Unfortunately, within the present review, it was not possible to test the potential moderating effects of severity of SEN across included studies. The participating children had very diverse difficulties and diagnoses ranging from children who only had a specific learning disability to children with multiple severe physical and mental disabilities. However, based on the descriptions of educational placement procedures in the included primary studies, it is clear that in many cases inclusive settings were not considered appropriate for children with the most severe disabilities, and in other cases children with more severe disabilities were excluded from the analyses to ensure balance on potential confounders between intervention and control groups. Thus, findings from the present review might have been different if all children from across the ranges of severity of disability had been included. Within the included studies in the review, the descriptions of the resources available to the children in the included and segregated settings were limited, meaning that in many studies it was unclear if there was co-teaching or additional personnel within the inclusive classrooms or if children with specific disabilities had access to technology that might enable them to compensate for their disabilities. This needs to be explored further in future studies, as it is theoretically possible that the effects of inclusion are dependent upon the resources available within both the segregated and the inclusive setting.

From a scientific perspective, the evidence base would be strengthened by carrying out randomised studies. However, there are a number of ethical concerns, which need to be addressed. First, it would be important to ensure that the inclusive setting is fitted with sufficient resources and personnel to meet the needs of children with special needs, and thus inclusive education should not become simply a way of reducing the costs of special education. Second, it would be important to conduct an individual assessment of eligibility of the child before randomisation, to ensure that there are no individual needs of the specific child which can only be met in either an inclusive or a segregated educational placement. Third, participation in a randomised trial should be based on informed consent from the parents, and thus it may not be feasible to conduct randomised studies within this field.

Results of the meta-analyses do not suggest on average any sizeable positive or negative effects of inclusion on children's academic achievement in language, literacy, and math, or in the overall psychosocial adjustment of children. The average point estimates did favour inclusion, though small and not statistically significant, and heterogeneity was present in all analyses and there was inconsistency in direction and magnitude of the effect sizes.

This finding is very similar to the results of previous systematic reviews and meta-analyses, which include studies published before 2000, and thus although the number of studies in the current meta-analyses is rather limited, it can be concluded that it is very unlikely that inclusion in general increases learning and psychosocial adjustment for children with special needs. Rather, the results may be seen to suggest that for some children with special needs, segregated placement may be the better educational choice—at least when compared with the current local models of inclusion. Future research should thus aim to explore the effects of different kinds of inclusive education for different kinds of children with special needs, to expand the knowledge base on what works for whom.

In this review, none of the studies which could be used in the meta-analyses reported usable follow-up outcomes and this finding is also similar to reviews published before 2000. Theoretically, it is possible that potential positive effects of inclusion emerge at a later stage in the child's life, and thus it would be very useful for future research to include time points past the end of the intervention. More specifically, it would be useful to explore outcomes such as social skills, social integration, self-perception, marginalisation and employment to fully test the ideological and theoretical hypotheses that inclusive education may protect the child from the effects of stigma and lead to more social integration and less marginalisation or social isolation in adulthood.