Relating Subjective and Objective Pharmacovigilance Association Measures - Clustering Challenges in Biological Network - page 284

threshold for a single adverse event to be significant at the level α is ρ ab <α ,
corresponding to U ab > log 10 α . In plots like Fig. 15.1, however, the question

posed is which of M adverse events is significant, where M is typically much
larger than 1 (in Fig. 15.1, M = 3702). The Bonferroni correction replaces α
with α/M , increasing the significance threshold to U ab >log 10 ( M/α ). While
it is known that the Bonferroni correction is conservative [21], experience has
shown it to be extremely useful in focusing our attention on drug/adverse event
combinations of practical importance [10].

−

15.3. Subjective Associations

The statistical association measures R ab and U ab are objective measures, based
on the ensemble of drug/adverse event records available in spontaneous report-
ing repositories like the FDA's AERS database considered here. This particular
database also includes subjective data on the association between each drug listed
in an ISR and the adverse events listed with it. That is, every drug entry is clas-
sified into one of four categories: primary suspect (PS), secondary suspect (SS),
interacting (I), or concomitant (C). Less than 0 . 1% of the drug records considered
here are classified as interacting, with approximately 27% classified as primary
suspect, approximately 14% as secondary suspect, and the remainder (58 . 3%)
classified as concomitant. This classification is extremely inhomogoenous across
the different drugs, however, suggesting the following numerical measure of sub-
jective association between Drug A and Adverse Event B. For each ISR listing
the drug and the adverse event, assign a suspect classification of 1 if the drug is
classified as PS, SS, or I, and 0 if it is classified as C. The suspect fraction S ab is
then defined as the fraction of ISR's listing both Drug A and Adverse Event B that
have suspect classifications of 1.
Fig. 15.2 presents a modified version of the pharmacovigilance map for the
drug infliximab shown in Fig. 15.1, where the sizes of the points have been made
proportional to S ab . For comparison, the corresponding plot for the drug fluoxe-
tine is shown in Fig. 15.3, where the point sizes are again proportional to S ab .In
both cases, points falling below the Bonferroni-corrected 5% significance thresh-
old have been omitted to make the plots easier to interpret. Comparing these two
plots, it is clear that S ab is large for almost all adverse events associated with
infliximab, even those adverse events exhibiting a negative objective association
with the drug (i.e., adverse events with R ab < 1). In contrast, the S ab values
for the drug fluoxetine exhibit a much wider range of variation across the adverse
events than those for infliximab do. Also, note that the S ab values for adverse
events with R ab < 1 are generally smaller than those for R ab > 1. Both of these