When “data” are not data: the pitfalls of post hoc analyses that use stock assessment model output

Abstract

The practice of treating stock assessment model output as data in subsequent modeling efforts is becoming more common, aided in part by the growing availability of online repositories of assessment results (misleadingly referred to as “data” bases). Such modeling exercises frequently overlook the uncertainty in the assessment output, the potential bias in estimates and correlation between estimates, and the structural assumptions of the original assessment model. We provide examples of post hoc analyses and discuss the problems in each case. We suggest alternative approaches that could have avoided using assessment model output altogether or suggest analyses that may have exposed the pitfalls of such methods. Whenever possible, we suggest not using stock assessment model output as data in post hoc analyses. If using assessment model output as data is unavoidable, then to address some aspects of the uncertainties associated with using assessment model estimates, we suggest collaborating with lead assessment scientists, sensitivity analyses, errors-in-variables methods, and cross-validation methods. Such additional work is imperative if research that uses stock assessment output as data is to make robust and meaningful contributions to stock assessment methodology and management decisions.