The three main challenges of pharmacovigilance, i.e. to detect, to assess and to prevent risks associated with medicines (Begaud, 2000), may concern both the patient level and the popula-tional level. Similarly, the latter may rely on classical epidemiological studies, e.g. cohort or case-control, or on cases-only analyses, which is the scope of spontaneous reporting (SR).
In cohort studies (Kramer, 1988), subjects are followed in a forward direction from exposure to outcome (e.g. the occurrence of a given disease), and inferential reasoning is from cause to effect. For example, in the case of a cohort study with a reference group, the subjects can be split, at the end of the follow-up, among the four cells of the following classical two-by-two table:
Diseased Not diseased
t0 follow-up t1
In case-control studies, subjects are investigated in a backward direction, from outcome (disease) to exposure and inference is from effect to cause:
Exposed exposed a b N1 diseased (cases)
c d N2 not diseased (controls)
In both designs, the compared groups are generally drawn from a larger source population, which raises the problem of possible selection biases; however, the subjects are generally exhaustively classified according to a binary variable: to present or not to present the considered disease in cohort studies, or to have been or not to have been exposed to the studied factor in case-control studies. SR, per se, is a passive surveillance method involving the whole source-population, e.g. all subjects of a given country treated with a given medicine; however, SR
suffers two major limitations (Begaud, 2000):
• it does not provide any direct and reliable information on the size, characteristics and exposure patterns of the source population;
• the term spontaneous refers to the random character of the case collection from the exposed population; indeed, reporting assumes that the observer (i) identifies the adverse event,
(ii) imputes its occurrence to a drug exposure,
(iii) is aware of the existence of a pharmaco-vigilance system, and (iv) is convinced of the need to report the case if relevant, e.g. new and/or serious adverse drug reactions (ADRs).
This results in the major plague of this surveillance method: an inescapable under-reporting, the magnitude and selectivity of which are unknown and extremely difficult to assess. Indeed, if a number a of cases of a given event have occurred in a population during the "follow-up" period, then it is likely that only a part k = a/U of these cases will be reported, U being the underreporting coefficient varying from 1 to infinity, e.g. U = 4 if 25% of cases have been reported.
Moreover, it is hard to believe that each of the a cases that have occurred have an identical probability 1/U to be reported. Many factors have been shown to influence reporting (Pierfitte et al., 1999) such as the age of the patient, the seriousness of the event and its onset delay. Thus, because of a selection bias, k could be a non-representative sample of the source population of cases.
From a biostatistical point of view, the rather bizzare design of SR could be compared to a cohort study without reference group in which:
• the "followed" population is extremely large, i.e. the whole population of the surveyed territory treated with drugs;
• the characteristics of this population, e.g. age and gender distributions, concomitant diseases, are unknown as are its characteristics of exposure (indications, dose, duration, co-medications, etc.);
• the number of "investigators" is extremely large, i.e., all health professionals in the territory;
• the case collection does not rely on a precise protocol and is thus non-systematic and may be subjective.
Moreover, because of the open character of this method of surveillance (any type of drug, any type of event), there is in fact a quasi infinite number of sub-cohorts, one for each type of drug exposure:
Was this article helpful?