Event definitions shape surrogate-risk inference: Measurement sensitivity in lane-changing safety analysis

Surrogate safety analysis turns trajectories into events and models the high-risk ones. What counts as an event is rarely reported as a modeling decision, yet it shapes the population and every downstream inference. Using 27,219 lane-change markers from the exiD drone dataset, we apply three semantic filters (lenient, baseline, strict) that reduce the modeling-ready set from 12,925 events to 2,609, an 80% drop. A composite surrogate-risk label built from seven indicators supports two models: a full model with access to those indicators, and a reduced model without them, evaluated under location-transfer validation with calibration. The full model reconstructs the label (AUC 0.980–0.989) under both definitions; the reduced model drops from AUC 0.709 to 0.630, with a three-way control assigning about 42% of the drop to sample size and class imbalance, about 2% to within-set relabeling, and 56% to the analyzed-population change. The direction holds in all twelve cells of a grid of alternative strict-filter thresholds. Calibration improves probability reliability but does not recover the lost discrimination: under the strict definition fixed-threshold classification collapses while ranking weakens. Event construction is part of the measurement model, and reporting accuracy without event-definition sensitivity overstates what trajectory-based models can claim.