\(H_0^H\): is the era drift an artefact of measurement timing?#

Definition

Era axis only, and conditional on rejecting \(H_0^A\). Null: the era drift is an artefact of measurement timing; the class-parameter-by-year association is nullified after adjustment for the measurement lag (the gap between diagnosis and parent report) and age at evaluation, and does not hold in a small-lag subsample. Alternative: the era drift persists net of the lag and age at evaluation, and within the small-lag subsample, implicating a change in the diagnosed population rather than in the timing of measurement. Estimand: the partial association of the class parameters with year of diagnosis, conditional on the lag and age at evaluation.

Status

Partially written. The test is specified but not yet run as a standalone stage. One preview exists from the neighbouring prevalence read (\(H_0^B\)): the era rise of the Social or behavioral class is nullified once sex, the diagnosis-to-measurement lag, and age at evaluation are netted out, while the era decline of the developmental class and the rise of Broadly affected survive that adjustment. So the prevalence evidence points toward a mixed picture, with part of the era signal explained by measurement timing and part not. The profile-level test below is still to come.

Method#

The planned test reads the direct effect of year of diagnosis on each class parameter, holding the measurement lag and age at evaluation fixed. In measurement terms this is a differential item functioning model: does a feature shift with year of diagnosis once the person’s class and evaluation age are held constant? A parameter whose year association survives the adjustment is drift in the diagnosed population; one that vanishes is a timing artefact. A small-lag subsample re-test then checks that the surviving association is not a residual of the adjustment model, by restricting to probands measured close to their diagnosis, where the lag cannot act.

This article will be completed once the direct-effect stage is implemented and run. Until then, the prevalence adjustment in H_0^A and H_0^D: are the class profiles invariant, and is any drift small? and the prevalence read are the closest evidence on the record.

See also#