Common Analytic Pitfalls In Studies of Autism Risk Factors or Phenotypic Characteristics

Background: Studies of risk factors for, or phenotypic characteristics of, autism have commonly fallen short of rigorous methodology, instead applying inappropriate analytic strategies even when the study design was of high quality. Pitfalls include: incorrect adjustment for intermediate outcomes, and lack of control for case-group in evaluating quantitative traits. Additionally, standard reasoning regarding attribution of causal components has generally not accounted for multifactorial causation within individuals.

Objectives: To apply rigorous methodologic principles to analyses evaluating autism risk factors or quantitative traits and to thereby determine validity of previously reported associations and their interpretations.   

Methods: This project evaluated commonly employed analytic strategies that (a) derived correlations between a continuous exposure measurement and scores on specific behavioral scales; (b) examined the association between a periconceptional exposure and risk for autism; and (c) calculated heritability without accounting for the role of combined effects of genes and environmental factors. The underlying incorrect assumptions will be highlighted.

Results: In the first example, scores on various quantitative scales such as the Aberrant Behavior Checklist (ABC), or Systematizing or Externalizing scales, were correlated with several biomarkers in a case-control design, but without control for case status. Nine possible scenarios are identified that involve the relationships between biomarker and the ABC or other scale score in the two groups. It is shown that when cases and controls differ markedly on the scale scores, only one of these nine possible scenarios would be consistent with the conclusions drawn from an analysis combining cases with controls. Eight scenarios do not permit a correct description of within group associations between the biomarker and the behavioral score. In the second example, the association between fertility treatments and autism were adjusted for multiple births and other conditions such as gestational age that would be downstream of the exposure. This strategy is demonstrated to result in biased estimates of effects. Conditions when it would lead to an unbiased estimate can be demonstrated. Finally, variance components analysis is applied to the problem of genes and environment, and the underlying assumptions behind interpretation of monozygotic vs. dizygotic concordance are critiqued.

Conclusions: The complexity of the etiology and phenotypic heterogeneity in autism provides for multivariate relationships that may require more sensitive methods than are commonly used in this field. Use of Directed Acyclic Graphs can facilitate the choice of variables to be used in multivariable analysis. The study of biomarkers and behaviors demands an appreciation of heterogeneity of effect measures, especially when comparing children with and without autism; thus, caution is advised prior to grouping typically and atypically developing children. Finally, attribution of causes under simplistic assumptions may have led to systematic underestimation of the role of certain types of etiologic factors. A multifactorial model of variance components predicated on interactions can lead to more realistic estimates of etiologic fractions.