Relating Subjective and Objective Pharmacovigilance Association Measures - Clustering Challenges in Biological Network - page 290

of the distinct terms appearing 10 times or less in the database, and almost three
fourths (72 . 3%) appearing 30 times or less. These observations motivate the use
of the top 100 adverse events in the case study presented here, all of which appear
more than 3,000 times in the AERS database.
To obtain a collection of drugs that was representative of those appearing in the
AERS database, a random sample of 50 text strings was drawn from the DRUG-
NAME field of the AERS drug data files. Of these, 2 were exact duplicates of
other drug names in the sample, and 12 others were omitted because they were
either different names for the same drug (e.g., “quinapril” and “accupril”), dietary
supplements (e.g., “multi-vitamins”), or otherwise un-interesting (e.g., “Eckerds
brand lice shampoos”). The remaining 36 drugs were:
accupril, aldioxa, aminocaproic acid, aspirin, atenolol, azasetron,
bactrim, candesartan, carbamazepine, cisplatin, clonidine, clopidro-
gel, clozapine, coumadin, depamide, dipiperon, etanercept, fluoxe-
tine, hydrochlorothiazide, ibuprofen, insulin, levodopa, lorazepam,
luprolide, metoprolol, mianserin, morphine, mycophenolate, omepra-
zole, oxaprozin, oxycodone, paracetamol, pravastatin, prednisone, ro-
fecoxib, and tamsulosin.
In addition to these 36 drugs, the following 15 were included because they were
of independent interest:
atorvastatin, bosentan, ciprofloxacin, cocaine, dutasteride, galan-
tamine, gatifloxacin, infliximab, levofloxacin, lovastatin, moxifloxacin,
simvastatin, tenofovir, valdecoxib, and zolpidem.
The association measures R ab , U ab , S ab , ψ ab and µ ab were computed for each of
these 51 drugs with each of the 100 most frequently ocurring adverse reactions in
the AERS database.
Evaluating the five association measures considered here for each adverse
event yields 500 numbers for each drug. Since the main question of interest is how
these different association measures are related, the primary clustering variables
used are the 10 correlations between each distinct pair of association measures as
they vary over the 100 adverse events. In addition, since the subjective measure
S ab is of particular interest, the mean S ab value over the adverse events is also
included in the clustering variable set. The complete list of the eleven attribute
variables used for clustering drugs is given in Table 15.2.
15.6.2. The Clustering Approach
The 51 drugs considered here were clustered using the eleven variables listed
in Table 15.2 and the Partitioning Around Medoids (PAM) method described