The hypothesis of risk compensation implies that persons experiencing a real or perceived change in the riskiness of an activity will alter their consumption of that activity to obtain a preferred combination of risk and reward. In evaluating whether individuals display compensating behavior in response to safety interventions, not all persons subject to the intervention will necessarily display compensating behavior, even if the hypothesis is correct: the hypothesis has testable implications only for the subset of persons subject to the intervention who perceive that their risk has changed. This paper argues that methodologies that include persons for whom the hypothesis has no testable implications (against a null hypothesis of no compensation effect) result in estimates of the compensation effect and test statistics which are biased towards zero. Previously published data on motor-vehicle-related injuries to cyclists and pedestrians in Britain before and after a mandatory safety-belt-use law went into effect were used to infer the size of this bias. In these data, the inclusion of persons for whom the hypothesis of risk compensation has no testable implications appears to have resulted in estimates of a risk-compensation effect which are too small by about half. This work suggests that the British data are consistent with a risk-compensation effect of 7-13 percent, and raises important methodological issues in testing the hypothesis of risk compensation.