Accident migration: a statistical explanation?

It has been noted, in before-and-after studies of the effectiveness of remedial treatment on accident frequency, that the accident rates apparently rise at untreated sites adjacent to treated sites. This phenomenon has become known as 'accident migration'. This paper proposes a new and purely probabilistic hypothesis, to explain this apparent 'migration' effect. It is argued that the true accident rates are spatially correlated, and it is then shown, by means of simulation results, that the selection process of 'neighboring' sites (which are untreated and adjacent to a treated site) leads to bias in the comparison of their before and after accident frequencies. The magnitude of the bias, it is shown, can match that of the apparent 'migration' effect noted by previous researchers.