Abstract
Regional purchase coefficients (RPCs) trade-adjust national input-output accounts for subregional analyses. They measure the share of an area's demand that is fulfilled by production in that area. Herein, I extend RPC-estimation approaches by applying a spatial Durbin estimator in two complementary forms — a static cross-sectional specification, a simple update of the long-standing econometric approach, and a dynamic time-space-recursive panel extension with both spatial lags and a temporal lag of the dependent variable. I also estimate at two complementary geographies over the same span: U.S. state-level and 132 sub-state freight zones. I rationalize the goods into four substantively distinct commodity groups and refit the dynamic specification within each. Five findings follow. First, RPCs are extremely persistent year-over-year, displaying little detectable net drift. Second, because they are seeded with priors, dynamic specifications fit extraordinarily well — at the data-record level, error falls by more than an order of magnitude relative to the static cross-section. Third, the residuals are substantially heteroscedastic, with a gradient that runs primarily by survey coverage and secondarily by transport mode; partition-stratified diagnostics support reading high persistence as an economic feature rather than a data artifact, since thinly-covered (heavily-interpolated) commodities show lower persistence than well-covered ones — the inverse of what data-smoothing would suggest. Fourth, the modifiable areal unit problem appears: RPCs are systematically lower at the zone level than for states, so home bias is meaningfully lower in smaller areas. Fifth, a metropolitan self-supply premium for manufactured goods appears in the raw geography but does not once controlled as legacy information in fixed effects. The above not only yields some insight into how to estimate RPCs more accurately but also into how to produce better multiregional input-output accounts.