Reflections on Attribution and Decisions in Pharmacovigilance
- PDF / 116,448 Bytes
- 5 Pages / 504.57 x 720 pts Page_size
- 46 Downloads / 173 Views
EDITORIAL
ª 2010 Adis Data Information BV. All rights reserved.
Reflections on Attribution and Decisions in Pharmacovigilance Ola Caster and I. Ralph Edwards Uppsala Monitoring Centre, Uppsala, Sweden
1. Causal Relationships In general, a causal relationship depends upon the nature and amount of evidence supporting an attribution hypothesis, such that ‘A causes B’. A can be a sufficient cause of B, meaning that A is always followed by B; or a necessary cause, meaning that B cannot occur without being preceded by A; or both. These deterministic concepts are not relevant to pharmacovigilance. No drug is a sufficient cause of an adverse effect, and there are no examples of necessary causes either. The trigger to modern pharmacovigilance, thalidomide, has come close to being considered a necessary cause of phocomelia, but this is not true. Phocomelia is very rare in the absence of thalidomide but not vanishingly so; x-irradiation is one example of another cause.[1] This example is illuminating also because it shows that the definition of the hypothesis is critical: one must be sure one knows what A and B are. In this case we have used the terms phocomelia, ectromelia, amelia and limb-reduction disorder loosely as synonyms when in reality they are overlapping hierarchical entities. Lacking deterministic causality, we must rely on probabilistic evidence for our attribution hypotheses. Such evidence is present when the probability of B given A is ‘large’, and the probability of B given not-A is ‘small’. Whereas this holds true for thalidomide and phocomelia, where the likelihood of a causal relationship is indeed very high, not even then is causality certain, and we must be aware of that. This requires those of us in pharmacovigilance to always consider the prob-
ability of one causality hypothesis against others. The real challenge is to prove a hypothesis wrong, which does not automatically succeed if there is support for a competing hypothesis because there can be multiple causes to an effect. Confounding is one very common example: if a competing hypothesis regarding the potential confounder C and its relation to B is accepted, for example by showing a statistically strong relationship between C and B, this does not automatically exclude the possibility of a causal link between A and B, though it could be that the probability of the causal link from C to B is greater. It may also be that A and C together have a higher likelihood of causing B via a synergistic or an additive effect, possibly even when A seems to have no effect on B in the absence of C. 2. Sources of Evidence Pharmacovigilance in action is essentially Bayesian: a tentative prior probability for a hypothesis becomes modified up or down as further evidence is obtained, to become a posterior probability for new consideration. Such evidence may come from different sources and also the overall view of the data may use different logic. Individual case reports can provide excellent evidence on attribution in one or more specific patients. For example, posi
Data Loading...
