Note: This section provides an overview of RGPS computations. For a detailed technical summary, see Regression-Adjusted GPS Algorithm. You must have Adobe® Acrobat® Reader installed to view the technical summary.
When you specify data mining parameters for a two-dimensional MGPS run without subset variables, you can include Regression-Adjusted Gamma Poisson Shrinker Algorithm (RGPS) computations. RGPS is a hybrid algorithm, which combines the methodology of Extended Logistic Regression (ELR) and MGPS.
Note: You must install R as described in the Empirica Signal and Topics Installation Guide to enable RGPS computations.
In general, RGPS differs from MGPS, PRR, and ROR computations in that it adjusts for the effects of polypharmacy and therefore reduces the likelihood of masking biases and "innocent bystander" biases. These biases occur when a drug other than the drug of interest is highly associated with the event of interest.
Masking bias occurs when a highly associated drug other than the drug of interest appears in greater number in reports that do not also include the drug of interest. Masking bias inflates the expected count for a drug-event combination of interest. Consequently, the disproportionality estimate for the drug-event combination of interest is reduced. A lower disproportionality estimate can lead to a missed signal.
"Innocent bystander" bias occurs when a drug with no causal connection to an adverse event is often co-prescribed with a highly associated drug. In this case, both drugs can appear to be associated with the adverse event.
Like MGPS, RGPS calculates a Relative Reporting Ratio (RR) for each itemset in the database. The RR is defined as the observed count (N) for the itemset divided by the expected count (E). However, RGPS computes E using the results of an Extended Logistic Regression (ELR) analysis rather than the Mantel-Haenszel approach.
Note: The Empirica Signal application allows you to run only one RGPS computation at a time.
For a combination of a drug and event, the signal score is the ERAM (Empirical-Bayes Regression-adjusted Arithmetic Mean) value. ERAM is the shrunken observed-to-expected reporting ratio adjusted for covariates and concomitant drugs.
The 5 percent (ER05) and 95 percent (ER95) confidence intervals for the ERAM are also calculated.
For an RGPS run, you have the option of including drug-drug interaction computations. Interaction estimates are calculated for drugs where the number of drug-event reports exceeds the specified minimum interaction count.