receive a high proportion of vaccines. If the database is restricted to reports involving infants, the associations disappear. Another example is the association between Viagra use and various cardiac events—due to the association of both of these items with age and gender, each being most common among elderly men. Similarly in a database spanning many years of reports, any drug very new to the market is more likely to show an association with an adverse event that is newly defined in MedDRA (Medical Dictionary for Regulatory Activities), the coding dictionary for adverse events used in most such databases. To combat these common confounding biases, it is recommended to stratify the database reports by gender, age of patient and year of report (and possibly other variables, if available) and to compute a version of the disproportionality measure that adjusts for stratum effects. This correction for demographic and secular trend confounding seems to be most often used with the measure RR. The correction in this case, originally due to Mantel and Haenszel (1959) consists of computing e separately for the reports within each stratum, and then summing the stratum-specific values to get a total e to use in the ratio n/e. There are also analogous ways to adjust ROR and PRR for stratum effects.
A trickier type of confounding is with the presence of an indication for taking the drug. For example, a drug used to combat cancer might have a reported AE that is a symptom of the cancer for which the drug is prescribed. It would be very difficult to automatically eliminate all such confounding from a computer database analysis. So far, the only feasible approach is to rely upon the medical knowledge of the analyst to recognize and discount such computed associations.
Another problem with disproportionality measures is their often extremely large variance. For example, a database analysis may find tens of thousands of drug–event combinations in which n = 1 and e < 0.001 so that n/e > 1000. In contrast, a value of n = 20, e = 2, n/e = 10, is usually going to be a much more “interesting” drug–event combination on which to follow up. There are two statistical approaches to controlling false positives due to the high variance of n/e. The first is to restrict computation of the disproportionality ratio to combinations in which some measure of statistical significance meets a threshold. For example, Evans and colleagues (2001) recommend restricting PRR to combinations in which n > 2 and the chi-squared statistic for association in the 2 × 2 table is at least 4. The second strategy is to use a Bayesian or empirical Bayesian analysis to produce “shrinkage estimates” that stabilize the ratios by applying a prior distribution that reduces (“shrinks”) the ratios n/e when n and/or e are small. Bate and colleagues (2002 and references therein) describe a “Bayesian confidence propagation neural network” (BCPNN) method that has been used to analyze World Health Organization databases. The development and use of an empirical Bayes model, the “multi-item