Assessing risk in postmarketing surveillance.

We need to find some way to assess the probability of risk for individuals taking medications for a prolonged period of time either singly or, as is usually more likely, in combinations. It is not possible to perform detailed, controlled studies for all combinations of drugs, not even those most commonly administered. Therefore, we must rely primarily on data from various case studies of attending physicians. Since this data is not collected in a controlled setting, we require a statistical technique that can do the following: be modified to handle data sequentially as information continues to accumulate, be very robust for dependent data, make minimal assumptions concerning the underlying distribution of the data, and be used for multivariate data. A statistical technique--nonparametric density estimation--is available that will satisfy all of the above requirements. We will demonstrate its use in postmarketing surveillance.