Much of what is known about the safety effect of various measures must be extracted from implementations in real life rather than from experiments that are staged to meet the dicta of rigorous scientific experimental design. The tools for extracting usable knowledge from data must be tailored to suit this reality.
METHODS of estimation that appear well suited for this task are reported here. First it is shown that what is commonly done is incorrect; it is incorrect to compare the count of "before" accidents with the count of "after" accidents and from this to draw conclusions about the safety effect of a measure. A simple method is provided for the correct analysis of "before" and "after" data. Next the likelihood function is introduced; it serves a dual purpose: First, it allows the assessment of the accuracy with which the safety effect is known. Second, it is a coherent formal device by which results from diverse studies can be accumulated. The ability to accumulate empirical evidence from many small studies is the key to progress in research on safety. The test of the advocated methods is in application; in this case, the examination of the effect on intersection safety of a change from two-way to multiway stop control. Details are given in two companion papers appearing elsewhere in this Record.