Second, when estimating latent scores for each individual, the estimates of the θjs, just like sum scores in the CTT tradition, are not simply observations but estimates with error variance. When computing the confidence intervals for the heritability estimates in the second phase, this uncertainty on the latent scores is not taken into account and the heritability confidence intervals are consequently too narrow and the estimates biased downwards. Moreover, in an IRT framework, the confidence intervals for estimates of individual latent scores are dependent on their location on the scale (actually, the number of items with β-parameters that are similar in magnitude to the person score θ and the items’ discriminatory power, α), whereas in the variance decomposition, it is assumed that measurement error (as included in the non-shared environmental variance component) is independent of location (cf. CTT). For example, many psychopathology scales have only items that refer to relatively rare symptoms. As a consequence, many individuals in the general population score 0, which does not necessarily imply that all actually have the trait to the exact same degree. In
