Aggregation Bias and Input-Output Regionalization

Conventional wisdom holds that results from input-output (IO) models with greater sectoral detail are superior to those from models with less detail. However, there is an implicit assumption that the more detailed data are as accurate as their aggregated counterparts. In this paper, we explore the tradeoffs between sectoral detail and model accuracy in the context of IO regionalization, a practical context in which greater sectoral detail is commonly achieved via the imputation of missing values. This reality is especially apparent for increasingly smaller geographical regions where privacy concerns result in more suppressed and undisclosed data. As the number (or share) of disaggregated values that require imputation increases, the disaggregated model results will also deviate further from perfect accuracy. Is there a point at which using an aggregate model with greater certainty – relying on more reported and less imputed data – will provide results that are superior to a disaggregated model with greater potential imputation error and uncertainty? To address these questions, we design and implement simulation experiments founded on the concept of aggregation bias that enable us to evaluate the likelihoods that aggregate models would be superior to their disaggregated counterparts.