In this issue of Pediatrics, Dr Shattuck presents findings from his intensive statistical analysis of diagnostic substitution in US special education data.1 Diagnostic substitution is one factor potentially contributing to the large observed increase in autism prevalence and, unlike many other possible factors, empirical evaluation of this factor can actually be approached by using existing administrative data. However, analyses of this sort are not without their challenges and complexities. When considering diagnostic substitution in US special education data, we are limited to group-level comparisons. We do not know whether individual children have switched classifications, and of course we can never know whether a given child in a particular birth cohort would have been classified differently had they been born either earlier or later. At best, analyses of this type are merely trying to determine if trends in one classification have the potential to offset those in another.
Although Shattuck's models incorporate a number of advantageous features, they essentially distill the assessment of this potential for diagnostic substitution to the sign and statistical significance of logistic-regression–model coefficients. However, assessment of potential diagnostic replacement should be rooted in a comparison of the magnitude, not just the direction, of classification prevalences. It is quite possible that small-magnitude negative associations will be deemed statistically significant when denominators are large, as with the US census counts used here. In addition, because the association-model coefficients are estimates of percent changes in autism classification prevalence associated with absolute change in the prevalences of other classifications, it is difficult to determine by examining these coefficients whether there is, overall, the potential for offset as demonstrated by equal-magnitude, opposite-direction changes. Finally, the models also assume homogenous effects for different birth cohorts, which, as described below, might not be the case.
In a descriptive analysis of US special education data published in March 2005 in Pediatrics Electronic Pages,2 my colleagues and I plotted special education classification prevalence birth-cohort curves and compared trends visually (the vertical distance between birth cohorts capturing the secular trend). I realize now that our choice of a logarithmic scale, although well-suited to our overarching descriptive goals, actually hampered our ability to determine if classification trends offset. Figure 1 shows the original plots for autism and mental retardation (MR) classification prevalences as well as plots rescaled to an arithmetic y-axis. Rescaled plots use the same 6- to 11-year-old age range used by Shattuck. The original plots emphasize the large annual percent increase in the low-prevalence autism classification while obscuring the smaller (in percentage terms) changes in the higher-prevalence MR classification. On the rescaled plots, we now can see that as autism prevalence increases with most successive birth cohort for ages 6 through 9, MR prevalence is decreasing (corresponds to the negative sign in Shattuck's models). However, the graphs also demonstrate that, at least in the US aggregate data, the increases in autism classification prevalence are greater in absolute terms (the distance between lines is farther) than the MR decreases, suggesting that any offset by MR alone would not be complete. In addition, at ages 10 and 11, autism prevalence still increases with successive birth cohorts but MR prevalence no longer decreases consistently.
In preparing this commentary, I have come to appreciate that even the “approachable” analysis of diagnostic substitution is a complex proposition. For example, Shattuck begins to address the notion of simultaneous substitution across several categories, a quite plausible scenario, but this is even more difficult to consider fully than the simple 1-for-1 substitution scenario I have been entertaining here. In the end, we may have to live with the fact that although administrative data suggest the potential for some diagnostic substitution, it remains difficult to confidently ascribe all of the observed autism prevalence increase to this particular phenomenon.
I also believe that the time has come to accept that, given the behavioral basis of the autism diagnosis, the lack of knowledge about autism's underlying etiology, and the limitations of retrospective analyses, we are not likely to develop a conclusive body of evidence to either fully support or fully refute the notion that there has been some real increase in autism risk over the past 2 decades. Accepting this, acknowledging that putative autism-risk genes have yet to emerge from gene-finding studies, and understanding that when genetic susceptibility and environmental triggers work together disease heritability is easily overestimated, the stage now should be set for well-reasoned, carefully designed research exploring the potential role of environmental exposures as well as genetic liability in autism etiology.
Editor's note: I am always hoping that we can find an explanation for the rising incidence of “autism.”
- Accepted November 16, 2005.
- Address correspondence to Craig J. Newschaffer, PhD, Department of Epidemiology, Center for Autism and Developmental Disabilities Epidemiology, Johns Hopkins Bloomberg School of Public Health, 615 N Wolfe St, Room E6030, Baltimore, MD 21205. E-mail:
The author has indicated he has no financial relationships relevant to this article to disclose.
- ↵Shattuck PT. The contribution of diagnostic substitution to the growing administrative prevalence of autism in US Special Education data. Pediatrics.2006;117 :1028– 1037
- ↵Newschaffer CJ, Falb MD, Gurney JG. National autism prevalence trends from United States special education data. Pediatrics.2005;115(3) . Available at: www.pediatrics.org/cgi/content/full/115/3/e277
- Copyright © 2006 by the American Academy of Pediatrics