physical attacks on other persons." Participant observers also recorded the timing, location, antecedents, choice of victim (staff or patient), objects used, immediate consequences, and staff observations regarding the presence or absence of external provocation. Antecedents were "external provocation, no external provocation, and unable to report." Victim status was "staff, peer, and combination."
Logistic regression was used to analyze the data. From this limited variable set, the author reports statistical significance among the following variables: antecedent, age group, and victim status. It is unclear why victim status was included as an independent variable, because this would seem to be a variable of choice. Garrison (1984:233) reports that the model's fit was "rather low" and that "a prediction table which employed these 1038 incidents to assess the utility of the model showed overall prediction to be only marginally better than that expected by chance."
Although we could apply formal statistics, such as RIOC (Farrington and Loeber, 1989) to summarize the predictive power of the regressions that were reported above, there is little reason to do so. It is apparent that violence can be "predicted," but that for even the most recidivistic offenders, violence is less likely to occur than to not occur.
Moreover, RIOC tells only part of the story. The important question is what level of accuracy can be achieved when predicting who will and who will not be violent. However, statistical analyses generally involve officially recorded violence such as arrests and hospitalizations. Typically, we might find that for a designated class of subjects, the probability of violence, as so measured, is .3-.5. Whether such predictions are accurate or inaccurate cannot be determined from calculating a statistic such as RIOC. Instead, when interpreting accuracy, we are required to make many "leaps of faith." How likely is it that an act of violence will lead to an official record of its occurrence? How likely is it that the probability of detection is uniform across a population of subjects? Because we are interested in actual violence, not violence as revealed through officially recorded incidents, answering these questions is inescapable when judging predictive accuracy. At this time, only subjective impressions and some imprecise statistics (Blumstein et al., 1985) provide guidance. Although RIOC