Bias and Inference in Spatial Interaction Models with Suppressed Origin–Destination Data
Аннотация
Maximum entropy methods are used to infer the true trip-distribution matrix in cases where parts of the data are suppressed due to privacy concerns. Large proportions of the suppressed data are found to be inferred correctly when the marginal totals in the trip-distribution are known. Entropy-based approaches are further found to outperform a strategy of ignoring suppressed information in cases with suppressed marginal totals and/or a higher cut-off value for suppressing cell information. We show that standard suppression rules induce systematic bias in estimated distance-deterrence parameters, and that entropy-based reconstruction restores macro-consistent origin–destination flows sufficient for correct inference. Our methods are demonstrated to reduce the systematic bias in estimates of the distance deterrence parameter to such small numbers that it is effectively zero, preventing potentially serious bias in estimates and predictions resulting from standard spatial interaction models. We also identify the conditions under which time-series data and inter-zonal distances most improve reconstruction accuracy.