4  The measurement floor

Longitudinal change in regional morphometry is frequently smaller than the measurement error of the pipeline that produces it. Establishing the error floor is a prerequisite for any method comparison, because regions in which the expected change is below the floor cannot discriminate between methods and will dilute any aggregate result.

4.1 Establishing the floor

Estimate, for each region and each measure:

  • Scan–rescan reliability. From back-to-back acquisitions where available, or from published intraclass correlation coefficients for the same pipeline, parcellation, field strength, and sequence. Values are not transferable across pipelines.
  • Expected change magnitude. The population mean change over the study interval, from prior literature or from the data itself.

The ratio of these quantities defines a per-region signal-to-noise ratio. Results should be reported stratified by this ratio, and regions below a stated threshold should be excluded from primary analysis or reported separately. This single analysis preempts the most common reviewer objection to sparse longitudinal modeling papers.

4.2 Longitudinal processing

Processing repeated scans independently with a cross-sectional pipeline introduces variance that is frequently larger than the biological effect. Unbiased within-subject template approaches reduce this substantially and are standard (Reuter et al. 2012). Cross-sectional processing of longitudinal data should be treated as a design flaw, not a simplification.

Note also that longitudinal pipelines induce correlation between timepoints by construction. This is desirable for sensitivity but must be acknowledged when modeling residual covariance, and it can produce regression toward the within-subject template if the template is built from all timepoints including held-out ones. Where a later timepoint is held out for evaluation, the template must be built without it.

4.3 Site, scanner, and protocol effects

In multi-site or long-running studies, scanner and protocol changes are confounded with time. Software version upgrades within a site produce step changes in derived measures that are indistinguishable from accelerated atrophy. Mitigations are longitudinal harmonization procedures, explicit site and software-version terms, and sensitivity analysis restricted to subjects with no scanner change.

4.4 Attrition

Follow-up samples are typically much smaller than baseline samples, and attrition is rarely random with respect to the outcome. Reported results should state the retained sample at each visit. Where attrition is plausibly informative, either a joint model of the outcome and dropout process or inverse probability weighting is required, and the sensitivity of conclusions to this choice should be reported.

4.5 Registration and normalization choices

Normalization by total intracranial volume, by total brain volume, or not at all produces different answers to the question of which regions change fastest. The choice interacts with the compositional constraint discussed in Chapter 2 and should be fixed in advance and reported, with at least one sensitivity analysis under an alternative.