2. Methodological aspects

View

1. Causation and correlation

1.1. Causality of an epidemiological association


In contrast to experimental designs as used in clinical trials, pharmacoepidemiology researchers often rely on nonexperimental or observational study designs. Thus, pharmacoepidemiology employs complex study designs and statistical analyses to evaluate rather intricate causal relationships between a therapeutic intervention, including medicine exposure, and an outcome and to find the true causal relationships.

However, before causality can be assessed, each study must be evaluated to determine whether its design is appropriate, the study size is adequate and systematic bias (see below, 2.5) has not influenced the observed association. In addition, the magnitude (both in size and frequency) should be sufficiently large. Causality should ideally be supported by evidence from several epidemiological studies in various geographic regions. Supporting toxicological and pharmacological data are also important. Epidemiological data should be interpreted with caution and in the context of other available scientific information.

Guidelines are available to help assess the possible causality of associations observed in epidemiological studies. The following should be considered (after Bradford Hill, 1965[1] and Rothman 1986[2]):

• Biological plausibility. When the association is supported by evidence from clinical research or toxicology about biological behaviour or mechanisms, an inference of causality is strengthened.

• Temporal association. Exposure must precede the disease, and in most epidemiological studies this can be inferred. When exposure and disease are measured simultaneously, it is possible that exposure has been modified by the presence of disease and it may be especially hard to confirm any or the true causal effect.

• Study precision and validity. Individual studies that provide evidence of an association are well designed with an adequate number of study participants (good precision) and well conducted with valid results (i.e., the association is not likely due to systematic bias).

• Strength of association. The larger the relative risk, the less likely the association is to be spurious or due to unidentified confounding. However, a causal association cannot be ruled out simply because a weak association is observed.

• Consistency. Repeated observation of an association under different study conditions supports an inference of causality, but the absence of consistency does not rule out causality.

Specificity. A putative cause or exposure leads to a specific effect. The presence of specificity argues for causality, but its absence does not rule it out.

Dose–response relationship. A causal interpretation is more plausible when an epidemiological gradient is found (e.g., higher risk is associated with larger exposures).

• Reversibility or preventability. An observed association leads to some preventive action, and removal of the possible cause leads to a reduction of disease or risk of disease.


[2] Rothman KJ. Modern epidemiology. Boston: Little, Brown, 1986:299-304.