Measurement error is usually construed as the deviation from a true value. But measurement in organismal biology typically is a complex, multi-step process, and living systems are dynamic and spatially heterogeneous, often hindering a unique definition of a measurand. Thus, a single true value of an organismal measurement, independent of a specific measurement process, does not exist, which also invalidates the classic realist notion of measurement error. Instead, I advocate an instrumentalist view of measurement that shifts the focus away from an unmeasurable true value toward the practical question of which biological, technical, and experimental factors influence measurement outcomes. I propose to include non-biological factors, such as different measurement environments and devices, observers, and the sequence of measurement, as technical covariates in biometric models. I present a simple least squares implementation for estimating mean effects and differences in variance while correcting for such “attributable” measurement effects. Repeated measurements are only of limited help in this regard because biological measures typically are not completely repeatable. The variance among partly repeated measures neither reflects the actual replicability of measurements, nor is there any benchmark as to how small this variance should be. Often, a large part of “unattributable” measurement effects is due to definitional uncertainty as well as the temporal and spatial heterogeneity of biological systems. I further demonstrate that averaging repeated measures has only negligible effects on the standard error of statistical estimates; increasing sample size (biological replicates) is generally more efficient than repeating measures (technical replicates).
Philipp Mitteroecker (Fri,) studied this question.