A model is described for determining the seed or propensity matrix from which turning flows at an intersection can be estimated to match given inflow and outflow volumes, using the biproportional method when intersection-specific counts are unavailable. Normalising counts to standard inflow and outflow totals is demonstrated to reveal a striking similarity in underlying propensities across intersections. Functional class is shown to be a poor explanatory of different propensities. Propensity is modelled as a function of angle, competing short cuts, presence of deadend approaches, and grid density. Estimates made using the propensity model seed are found to have prediction errors of about 6% of inflow volume, compared to errors of 4% for estimates that use an intersection-specific count as seed.
Abstract