Autism is a complex neurodevelopmental disorder in which the interactions of genetic, epigenetic and environmental influences play a causal role. Despite the compelling evidence for a strong heritability, the etiology and molecular mechanisms underlying autism remain unclear. High phenotypic variability and genetic heterogeneity confounds the identification of susceptibility genes. The lack of robust indicators to tackle this complexity in autism has led researchers to seek for novel diagnostic tools to create homogenous subgroups. Several studies have indicated that patients with autism have higher rates of minor physical anomalies (MPAs) and that MPAs may serve as a diagnostic tool; however, the results have been inconsistent. Using the cumulative data from seven studies on MPAs in autism, this meta-analysis seeks to examine whether the aggregate data provide evidence of a large mean effect size and statistical significance for MPAs in autism. It covers the studies using multiple research methods till June 2007. The current results from seven studies suggested a significant association of MPAs in autism with a robust pooled effect size (d=0.84), and thereby provide the strongest evidence to date about the close association between MPAs and autism. Our results emphasize the importance of MPAs in the identification of heterogeneity in autism and suggest that the success of future autism genetics research will be exploited by the use of MPAs. Implications for the design of future studies on MPAs in autism are discussed and suggestions for further investigation of these important markers are proposed. Clarifying this relation might improve understanding of risk factors and molecular mechanisms in autism.
Autism (OMIM 209850) is a severe neurodevelopmental disorder, characterized by qualitative impairments in social interaction and communication, accompanied by repetitive and stereotyped behaviors and interests. These symptoms manifest in the first 3 years of age with a lifelong persistence.1 The prevalence is estimated to be approximately 1 in 150, making it one of the most prevalent medical conditions of childhood.2, 3 Boys are affected approximately four times more than girls, with an even higher ratio in milder forms of the broad spectrum.4
The vulnerability for developing autism is highly genetically determined. Twin studies indicated substantially greater concordance for autism in monozygotic than in dizygotic twins.5, 6, 7, 8, 9, 10 Moreover, family studies revealed a recurrence risk of 5–6% among siblings of affected individuals, which is much higher than the prevalence rate in the general population.11, 12 Together, these data show that autism is a strongly genetically influenced multifactorial childhood psychiatric disorder.13, 14, 15
The prevailing view is that its etiology involves a complex interaction between multiple genes and possibly environmental insults, leading to an aberrant neurodevelopment.16, 17, 18 Researchers have attempted to overcome the challenges posed by such a complexity with reliable diagnostic tools, including the study of head circumference and other morphological characteristics.19, 20, 21, 22 Excessive head growth found in the first year of life, in children later diagnosed with autism, has been one of the most promising quantitative traits.23, 24, 25, 26, 27
As to other morphological characteristics, an excess of minor physical anomalies (MPAs) in autistic individuals received specific attention. MPAs are defined as slight morphological deviations that have no serious medical or cosmetic significance to an individual. However, they are of great value to the clinician because they can be utilized as indicators of underlying disease susceptibility or disturbed development (for example, they are found to be more common in individuals with an obvious major embryonic defect).28, 29, 30, 31 The presence of such MPAs in autism has been suggested to be related to the shared genetic risk of developing autism.32, 33
Although in the psychiatric literature ‘MPAs’ is generally accepted, many different terms are used to describe them, including dysmorphic features, minor congenital anomalies or minor malformations. The majority of the studies that have assessed the incidence of MPAs in autism used the Waldrop scale (vide infra), with occasional modifications and omissions of items.34, 35, 36 MPAs in that list originated from an unpublished study by Goldfarb and Botstein37 to classify schizophrenic patients. However, although the Waldrop scale is able to dissociate patients from controls, it has been criticized for inherent limitations regarding both content and form, its restricted range of 18 items, low sensitivity, subjective nature, ethnic effects and inter-rater reliability.38, 39, 40, 41
To date, there are a number of studies that have examined the prevalence of MPAs in children with autism.32, 42 Although most of these studies showed excess MPAs in autistic individuals as compared to healthy controls, the findings are inconsistent regarding the magnitude and extent of the case–control differences. There is substantial variation in MPA scores across studies, for autistics as well as for the control groups. Moreover, effect sizes in the individual studies have not been quantitatively reviewed and integrated in a meta-analytical way.
The aim of the present meta-analysis was to produce a synthesized effect-size estimate that has considerably more power than the individual studies.43 In addition, in order to identify the sources of variation across the studies, effects of the study characteristics on the findings were analyzed.
Materials and methods
An extensive bibliographic search was conducted to identify relevant articles that examined the incidence of MPAs in autism. Pubmed, Cinahl, PsycINFO and the Cochrane library were searched from inception to June 2007. For the query translation, Medical Subject Headings terms were used where they were available. The thesaurus for index terms was also checked to identify possible synonyms. The keywords used in the computerized search were ‘clinical morphology’, ‘minor physical anomalies’, ‘dysmorphology’ and ‘autism’. The reference lists cited in these studies and published reviews were examined to identify additional studies. In addition, individuals with expertise in the area of dysmorphology were asked for studies which were in press or any other papers of particular interest. Abstracts of studies identified by the search strategies were then scrutinized to determine whether they could be included or not. This search returned 78 potential hits of which the abstracts were evaluated and 16 articles were found to be relevant to the study.
Studies were considered eligible for inclusion if: (1) they were designed as a case–control study where the controls were healthy children; (2) they had used the Waldrop scale or some variant of it in the MPA assessment; (3) they had presented sufficient data to compute effect size in the form of a standardized difference between means (that is, means, standard deviations, exact P, t or F values); (4) they were written in English. As the primary focus of this study was on MPAs that are listed on a standard scale (Waldrop),35, 36 reports that examined only head circumference or major abnormalities were all excluded. Studies that reported previously published data were also excluded. Selection of the studies for inclusion was completed by two authors (HMO and JWH). Authors of the identified studies were contacted if there were queries regarding their studies.
Data were independently extracted by two authors (HMO and JWH), using a structured pro forma. For each eligible study, recorded data variables were authors, year and country of publication; demographic variables (mean age, male/female ratio and ethnicity); diagnostic criteria (if applicable); study size (number of participants and controls) and rating methodology (whether raters were trained to recognize dysmorphic features, blinding of the authors and inter-rater reliability). The assessment scale used to identify and quantify MPAs was classified as ‘Waldrop scale’ or ‘Waldrop scale modified’. The number of MPA scale items used in each study was also recorded. Any discrepancy between ratings was discussed and resolved by consensus.
The key to meta-analysis is defining an effect size capable of representing the quantitative findings of a set of research studies in a standardized form that permits meaningful comparison and analyses across the studies.43 Therefore, for each individual study, an unbiased standardized mean difference (d) was calculated. This effect size was computed as the difference between the mean of the autistic group and that of the control group, divided by the pooled standard deviation. The resulting effect size was corrected for upwardly biased estimation due to small sample size by using Hedges' formula.43, 44 When means and standard deviations were not available, d values were calculated from the reported t or F values. After computing individual effect sizes for each study, a weighted mean effect size (g) was obtained which indicated the magnitude of the association across all studies.44 Each effect size was weighted by the inverse of its sampling variance when calculating the pooled effect size, to account for the different sample sizes on which each effect size was based.45 An effect size between 0.2 and 0.5 is considered to be weak, between 0.4 and 0.6 moderate, and greater than 0.8 is considered to be a large effect size.46
To investigate the significance level of the effect, a 95% confidence interval (CI) and z-value was calculated. Once the effect size for each study was obtained, the variance across effect sizes was assessed. The homogeneity statistic Q was calculated to test whether the observed variability in the distribution of effect size estimates is greater than that expected from sampling error.43 As published research findings suggest a number of variables that may influence effect size, a weighted regression analysis was performed to determine the extent to which selected study characteristics might explain between-study variations in effect size. Potential moderator variables for analysis were year of publication, the number of Waldrop scale items used and the use of siblings as control group. Other variables with a potential influence on effect size such as IQ, case–control sex ratio, autistic symptoms and types of autism could not be analyzed because of the small number of studies reporting results for these parameters. Data analyses were performed using random-effects framework. A random-effects model assumes that each observed effect size differs from the population mean by differences in sample sizes plus a value that represents other sources of variability assumed to be randomly distributed. To obtain a good estimate of the random effects variance component, we chose the noniterative method based on the method of moments.43 Subsequently, a sensitivity analysis was performed to explore the influence of each study's effect size on the overall effect size, by deleting each study sequentially. The pooled effect size for the remaining studies is recomputed each time with the removal of each study, along with their 95% CI. This analysis allowed testing the overall robustness of the meta-analysis as well as detection of the most influential studies.
To investigate the possibility of publication bias, Rosenthal's fail-safe N statistic was computed.48 Publication bias implies that studies with no effect may not be published, posing a threat to the stability of the obtained effect size. This method determines the number of unpublished studies with null results that would be required to reduce the overall effect size to a nonsignificant level. A large fail-safe number makes the ‘file-drawer’ problem negligible. Furthermore, publication bias was also assessed using Egger's test.49
Description of studies
The combined literature search yielded 78 references. After eliminating overlapping references and those that clearly did not meet the criteria, 16 studies were identified and retrieved for further scrutiny. Of those, one was excluded because it was investigating major congenital anomalies, rather than MPAs,50 two because controls were not included33, 51 and two because the Waldrop scale or a variant of it in their assessment was not used.52, 53 Two more studies were excluded because of absence of sufficient data to compute a mean effect size even after contacting the investigators.54, 55 Two final studies were excluded because of the absence of relevant data in the published articles and no response from the investigators.56, 57
In the end, a total of seven studies, published between 1975 and 2005, met our inclusion criteria and contributed to the meta-analysis.5, 58, 59, 60, 61, 62, 63 Each sample was included independently into analysis. These studies included 330 patients with autism (mean age: 9.75, 80% male) and 382 healthy controls (mean age: 10.3, 70% male). In two studies only boys were included,60, 62 and three studies had mixed ethnicities.58, 59, 60 Three studies used siblings as their control group;5, 58, 60 four of the seven selected studies were conducted in the United States,58, 59, 62, 63 two were carried out in Canada60, 61 and one in the United Kingdom.5 The main characteristics of the studies are listed in Table 1.
Effect sizes were calculated of all studies that provided mean MPA scores on the basis of the Waldrop scale. As graphically presented in Figure 1, the results of our meta-analysis indicate that mean MPA scores in patients with autism differ from those of healthy controls.
In all seven studies the direction for the effect size indicates that autism patients show higher MPA scores than controls. None of the seven studies had a negative effect size or included the value zero, indicating a statistically significant effect for each study. Finally, a single pooled effect size of 0.84 (P<0.001) was found, with 95% CI ranging from 0.47 to 1.21. The results of our meta-analysis indicate a significant difference in the mean number of MPA scores between patients with autism and healthy controls. The pooled effect size is in the range of a robust effect.43, 46 There was considerable heterogeneity across studies (Q=36.90, d.f.=6, P<0.01) and distributional analysis of effect size estimates indicated one positive outlier. Without this potential outlying case, the effect size distribution was no longer heterogeneous. However, the new combined effect size was still statistically significant. We were reluctant to remove this study from the analysis because removal may lead to an underestimation of the estimated mean effect size. In addition, the study completely fit our inclusion criteria, and closer examination revealed nothing unusual about the outlying study. Thus, the size and statistical significance of the pooled effect size is not strongly influenced by the outlying study. Results of the sensitivity analysis revealed that removing any single study failed to result in a significant shift in the pooled effect size estimate (Figure 2).
The largest negative shift occurred following removal of the study by Gualtieri,59 which was expected based on the large effect size reported in this study. The largest positive shift occurred following removal of Soper et al.61 In both cases, however, there was a negligible net effect on the overall pooled estimate. This indicates that all seven studies were similarly influential and that the meta-analysis is generally robust. Because of the large CI (0.47–1.21; see Table 1) and the small number of studies, there was sufficient variability to warrant further analysis. Therefore, we performed weighted regression analysis where the relationship between quantitative study characteristics and effect size was explored. Neither the number of MPA scale items nor the use of siblings as controls, or gender rates accounted for a significant proportion of the between-study variance in effect size. None of the regression models were statistically significant (P>0.5). However, the failure to find a moderator is not surprising given the small number of studies included.
Publication-bias analysis indicated a fail-safe number of 43 that means that at least 43 studies reporting no effect need to be found before the mean results are no more significant, large enough to credence to our findings. The estimated bias with the Egger's test was 0.22 (95% CI 14.2–14.64), P=0.97, which also indicates an absence of publication bias.43
This meta-analysis integrated the results of seven studies that compared MPAs of 330 patients with autism with those of 388 healthy controls. Each original study consisted of a relatively small sample of autistic cases ranging from 20 to 74. The results indicate that the mean total MPA scores in children with autism differ from those of healthy controls. In each of the seven studies the autistic sample had significantly higher rates of MPAs than those of the controls. In none of the individual studies, the magnitude of the case–control difference was in the small range. Moreover, the pooled effect size calculated from all seven studies was 0.84 (P<0.001), which is considered to be a large effect. This compelling, robust effect showed strong consistency by sensitivity analysis, regardless of the dataset removed. The meta-analysis supports the conclusion that patients with autism have significantly more MPAs than those of normal controls; this finding is consistent with the findings of the individual studies.53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63
Despite this finding, a high degree of variability in the magnitude of the effect size magnitude was observed among these seven studies. In order to identify potential factors that could be responsible for constraining this variation, we performed regression analyses. None of the earlier mentioned moderator factors (that is, year of publication, number of Waldrop scale items used and the use of siblings as control group) were able to account for a statistically significant relationship to the observed between-study variation in effect sizes (P>0.05). This failure to identify a moderating factor is not surprising given the small number of studies included in this meta-analysis. The effect of other potentially important moderator variables (for example, sex, IQ, familiarity) could not be analyzed because of a lack of information on these characteristics in most of the studies as well as the small number of studies.
Although this meta-analysis clearly shows a robust effect for significant excess of MPAs in the autistic population, the results are complicated by a number of methodological issues. Although having the advantage of being short, there have been concerns about Waldrop scale's limitations. A major methodological issue is the unblended nature of the measurement. Although a good inter-rater reliability can be achieved, one should be cautious that patients with autism might be different during even the minimal interaction needed to measure some of the Waldrop scale items (for example, head circumference). Although few studies have been carried out blinded, complete blinding is very difficult given the close personal interaction involved when assessing MPAs and facial measurements. In addition, one should keep in mind the subjectivity regarding the validity of certain Waldrop items (for example, hair quality, hypertelorism, clinodactyly). In addition, we believe that other methodological variability among the studies may explain some of the variation in MPA scores: inconsistencies in sample size and composition, differences in MPA items, lack of consensus in the terminology, and ethnic diversity. For instance, the majority of cases were Caucasian, however, three studies used mixed ethnicities in their population samples.58, 59, 60
As we confirmed the higher rate of MPAs in autistic patients as compared to controls, we are faced with new challenging questions. First of all, why do autistic patients have higher rates of MPAs? Apparently, a common genetic vulnerability for developing autism is reflected in MPAs. Several developmental genes have recently been identified that play a paramount role in shaping body structures.64 Moreover, new insights into craniofacial morphogenesis have indicated that a rapidly increasing number of genes are known to regulate cerebro-craniofacial development.65, 66 It can be speculated that the genes that determine the craniofacial morphology overlap with candidate genes for autistic disorders. MPAs could also be related to prenatal infection, or other environmental exposures which are associated with autism. Prenatal or postnatal exposure to infections, such as rubella, herpes simplex virus and cytomegalovirus, has been reported in several patients with autism. Thalidomide exposure during pregnancy is consensually associated with autism.67 In addition, both de novo and familial cytogenetic abnormalities may be associated with an increased number of MPAs.68 Copy number variation (CNV), including deletion and duplication, translocation, inversion of chromosomes, has been identified in some individuals with autism.14, 69 In fact, Engels et al.70 showed a direct association between the severity of physical anomalies and the chance to find mutant CNVs. The results of this meta-analysis suggest that MPAs in autism are, at least in part, related to the risk of developing the disease and that these MPAs may therefore precede the clinical onset of the disorder.
Another critical question is whether these physical anomalies in autism are broad population characteristic of all patients in the spectrum, or whether patient–control differences derive from overrepresentation of those abnormalities among only a specific subgroup of patients. Some evidence for the latter comes from Miles et al.33 who hypothesized that autistic patients with high MPA scores represent ‘nonfamilial or sporadic’ autism due to single environmental insults or nontransmitted genetic events, whereas autistic patients with low MPA scores represent ‘familial’ autism (where genetic clustering of psychiatric disturbance reflects variable expression of the underlying genotype).33, 42 These findings were confirmed by Links et al.60 and replicated in a later study by Miles et al.33
Another fundamental issue to be addressed is whether sets of certain physical anomalies are related to specific phenotypic behavioral characteristics in children with autism, and whether clustering of certain anomalies to groups of patients would yield homogenous subgroups. Except one early study by Walker,63 no other study in the literature dealt with clustering. However, he found random association and heterogeneity in distribution with few exceptions (for example, high palate and low setting ears).
A final, more speculative, question is about the specificity of MPAs. Are the MPAs seen in autism different from those in other disorders? In a recent meta-analysis, a higher prevalence of MPAs was also established in schizophrenia.31 Do MPAs seen in autism have a different etiology than those in schizophrenia, or do disorders associated with MPAs share a common etiological basis with schizophrenia and autism? Findings indicating overlapping markers could provide important clues regarding the underlying genetic bases of these disorders. Some evidence for such an overlap comes from the observation that individuals with autism spectrum disorders may also be at greater risk for schizophrenia.71 And, recent findings indicate that most complex disorders are probably rooted in genetic variation that is significantly shared by multiple disease phenotypes.72 Robust diagnostic specificity is often lacking for several other disease markers as well as MPAs and reflects the fact that different disorders may share genes, and also share partially overlapping neural substrate dysfunction and clinical features.73
Limitations and strengths of this meta-analysis
There are certain limitations that should be borne in mind when interpreting the results.
First, as with all meta-analysis studies, the results depend on the quality of the individual studies. Although we used well-defined inclusion criteria, we had to accept some methodological diversity among the studies in order to compare a sufficient number of studies. We should also mention that the inclusion or exclusion of a given study in this analysis was not based on the scientific value of the publication. We had to exclude some valuable publications, as they did not meet the specific goal of the present study.
Second, the diagnosis of autism was occasionally problematic in the early studies, written before the introduction of the fourth edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM) classification. Nevertheless, those studies were included because they met all our inclusion criteria. Sensitivity analysis confirms that the change in the type of criteria used for autism diagnosis does not appear to influence the effect size. Moreover, perhaps, in part, due to a true increase in the prevalence rate or, in part, due to the introduction of DSM 4 or greater awareness of the syndrome, the prevalence of autism has significantly increased over the last decades. Therefore, we cannot exclude that the effect size of the association today may somewhat differ from that of the earlier studies included in the meta-analysis. However, the effect size reported by the most recent study (Soper et al.60) is not different from that of the older studies. And again, sensitivity analyses show that the publication year does not influence the effect size.
Third, three studies were included that measured MPAs in sibling controls, which may introduce a confounding factor. However, because of the small number of studies in the meta-analysis, these studies, which met all criteria, were included. Interestingly, there were no significant differences in MPA scores when compared to either a sibling or nonfamily control group.5, 58, 60
Fourth, we were unable to examine topography of dysmorphic features in autism, because we had too little information comparing these across studies. Yet, such information is fundamental to understanding the timing and nature of dysmorphic events. Interestingly, increased head circumference, which is a well-documented finding in autistic children, was not consistently reported in these seven MPA studies.
Fifth, although the age of the participants has been thought to facilitate differences in effect size among the studies, the results of moderator variable analysis, possibly, in part, due to the small number of studies failed to confirm this hypothesis. Additionally, although gender is known to affect the incidence of autism, the studies included in this meta-analysis did not provide enough data to examine gender effects. Thus the possibility that some of the effects found in this meta-analysis study were caused by confounding factors such as age and sex cannot be ruled out.
Despite these limitations, the present results offer several methodological advantages for future inquiry. This is the first report studying the association between MPAs and autism in a meta-analytic way. This study provides evidence that MPAs are significantly increased in the autistic population and that some, unknown biological mechanism is likely responsible for producing these anomalies which may yield further knowledge about the developmental origins of the disease.
Recommendations for future research
It is obvious that the assessment of MPAs in autism require further study. With the aforementioned caveats in mind, we have the following recommendations.
Although MPA measurement is considered simple, noninvasive and inexpensive, we should caution that their genetic architecture may be as complex as that of autism itself. This does not mean there is no advantage to use them for genetic studies.74, 75 More and larger studies in ethnically homogenous populations are needed to search for a possible correlation between MPAs and family history as well as to achieve sufficient power to search the potential role of moderating variables such as gender, autism symptoms and IQ.
Our results provide strong support for the association between MPAs and autism. This meta-analysis emphasizes the importance of MPAs in the identification of heterogeneity in autism and suggests that the success of future autism genetics research will be exploited by the use of MPAs. Although these findings reflect a vulnerability to developing autism, it is still unclear how and to what extent genes and/or environment are involved. Future studies should focus on the search for susceptibility genes, chromosomal alterations (for example, mutations, duplications, deletions or CNVs) as well as different environmental factors in relation to morphological characteristics by using detailed definitions of the phenotype and an internationally accepted classifying list to enable comparison of the results. MPAs might serve as a helpful instrument in autism research, delineating subgroups which provide a more homogenous basis for unraveling the etiology and predicting prognosis.
We thank Dr Marcus Munafo for critically reading and helpful discussion of the article. We are grateful to Dr Henry Soper, who provided data from his studies for the meta-analysis and Dr Judith Rapoport and Dr Tom Gualtieri for answering specific questions regarding their studies.
About this article
Journal of Autism and Developmental Disorders (2013)