and all more or less revolve around the concept of expected proportional reporting, and observed detection of disproportionate reporting applied to large databases. They look at the AEs for a particular drug and compare the count of these same AEs for the remaining drugs in a PV database.
This basic PRR is simple and uses a 2 × 2 table:
Note: PRR = A(A + C)/B(B + C).
Note: PRR = 345/6901 (345 + 6901)/291 (291 + 14556) = 1.67.
In words, the proportion of a particular AE divided by all AEs seen with the drug of interest is divided by the proportion of this AE divided by all AEs seen with all drugs in the database. In the previous example, if 3.28% of all AEs seen with drug X are chest pain and 1.96% of all AEs seen with all the other drugs in the database (excluding drug X and its AEs) are chest pain, then the PRR is 3.278%/1.96% or 1.67. This means that there are (dis)proportionately more chest pain AEs with drug X compared with all the other drugs in the database, and this is noteworthy as a possible signal.
Since it is unusual that the PRR will be exactly 1.0, when the PRR is calculated for all AEs in the database, every PRR will either be below 1.0 or above 1.0. In theory, values above 1.0 would suggest an AE that is more frequently reported to drug X than other drugs in the data. One must be careful not to over interpret the data, especially since there are more than 80,000 terms in MedDRA®, and one could theoretically calculate some 80,000 PRRs. In practice, one sets a threshold above which the PRR is considered noteworthy, such as a value of, for example, 3.0. Any AE that has a PRR above 3.0 will be considered a signal and will be further examined. The higher the PRR, the greater the specificity but the lower the sensitivity. Alternatively, one might simply take the 10 or 20 highest PRRs and evaluate those regardless of how high the PRR is above 1.0.
There are other issues with using this score. When the database is small, problems can occur.
Finally, if the safety database used for the denominator of the PRR is small or has a high proportion of a particular type of patient or disease this may also produce flawed PRRs. Various other statistical methods, or filters, may be added to this calculation to refine the technique to attempt to increase sensitivity. Some have adopted a rule of thumb that signals are worth pursuing if the PRR is more than 3.0, the chi-squared value is more than 4, and that there are at least three of the particular AE in question. Thus, if one has a sufficiently large database, the PRR could be programmed to run periodically (e.g., monthly or quarterly) using the filters as noted to generate possible signals. This method will be less useful if MedDRA coding is not crisp and correct. As always, the issue here is generating too many signals with too many false positives for the personnel available to review the signals.
Although the disproportionality method is commonly used by companies and health agencies to screen and detect potential signals, other methods have been developed to compensate for some of these problems. Some of the other approaches are found in the broad category of “Bayesian approaches”. These methods account for the number of cases (cell sizes) and decrease the sensitivity of the PRR score if the cell sizes are small. One method, the Bayesian confidence propagation neural network (BCPNN), was developed by the Uppsala Monitoring Centre and is used for signal detection in their database. Other methods, such as the gamma poisson shrinker (GPS) and the multi-item gamma poisson shrinker (MGPS), are also used to attempt to make the PRR more useful. These methods have been used by various health agencies, including the Food and Drug Administration (FDA), EMA, and Medicines and Healthcare Products Regulatory Agency (MHRA). Treatment of these methodologies in detail is beyond the scope of this book, and the reader is referred to the standard textbooks of pharmacoepidemiology. A good, approachable summary of the field is available in Module IX Addendum I — Methodological aspects of signal detection from spontaneous reports of suspected adverse reactions (EMA Good PV Practices Web page).14
A brief note on other terms you may run across:
Event rate — The number of subjects experiencing an AE as a proportion of the number of people in the population at risk over a specific period of time. For example, 5 per 1,000 person-days.
Absolute risk — The probability of occurrence of an AE in patients exposed to a drug. For example, one may say the absolute risk of a myocardial infarction with drug X is 5%. Obviously, this value is often hard or impossible to obtain and is one reason that pharmacovigilance exists.
Absolute risk reduction — The arithmetic difference between two absolute risk rates. For example, the absolute risk of a myocardial infarction with drug X is 5% and with drug Y 2%. The risk reduction is 5% − 2% = 3%.
Relative risk (or risk ratio) — In a trial or in an observational cohort, the ratio between the rate of an adverse outcome (e.g., an AE) in a group exposed to a treatment and the rate in a control group. It is a measure of the strength of a cause–effect relationship. For example, the rate of myocardial infarction in the drug X group is 5% and in the control group 2.5%. The ratio (relative risk) is 5%/2% = 2.5%.
Relative risk reduction — The difference in event rates between two groups, expressed as a proportion of the event rate in the untreated group.
Odds ratio — Also known as estimated (or approximate) relative risk. In case-control studies concerning drug safety, the ratio between the rate of exposure to a suspect drug in a group of cases (with the AE) and the rate of exposure in a group of non-cases (i.e., controls without the AE). Like relative risk from cohort studies, the odds ratio is a useful estimation of the strength of cause–effect relationships. This parameter is often used in systematic reviews and meta-analyses.
Risk difference or attributable risk — The difference between the rate of an adverse outcome (e.g., an AE) in a group exposed to an experimental drug and the rate in a control group.
Number needed to harm (NNH) — Also called number needed to harm one. The NNH is the number of patients that must be exposed to a drug to produce an AE/ADR in one patient. Exposure may, for example, be one course or 1 year of treatment. For example, one could calculate based on a study that the number needed to produce one case