Two-phase designs in epidemiological studies of autism prevalence introduce methodological complications that can severely limit the precision of resulting estimates. If the assumptions used to derive the prevalence estimate are invalid or if the uncertainty surrounding these assumptions is not properly accounted for in the statistical inference procedure, then the point estimate may be inaccurate and the confidence interval may not be a true reflection of the precision of the estimate. We examine these potential pitfalls in the context of a recent high-profile finding by Kim et al. (2011, Prevalence of autism spectrum disorders in a total population sample. American Journal of Psychiatry 168: 904-912), who estimated that autism spectrum disorder affects 2.64% of children in a South Korean community. We reconstructed the study's methodology and used Monte Carlo simulations to analyze whether their point estimate and 95% confidence interval (1.91%, 3.37%) were reasonable, given what was known about their screening instrument and sample. We find the original point estimate to be highly assumption-dependent, and after accounting for sources of uncertainty unaccounted for in the original article, we demonstrate that a more reasonable confidence interval would be approximately twice as large as originally reported. We argue that future studies should give serious consideration to the additional sources of uncertainty introduced by a two-phase design, which may easily outstrip any expected gains in efficiency.
Keywords: autism spectrum disorders; epidemiology; prevalence; two-phase screening.
© The Author(s) 2015.