Uncertainty in inferences from groups to individuals
Epidemiological risk factors do not in fact refer to individuals, but rather to investigated populations. From the paragraph above we can apparently deduce that the situation would be simpler if Eva were a man, since there are studies available on the treatment effect of anti-cholesterol drugs in healthy men (13). In this section we will clarify why epidemiological data cannot directly help reduce strict uncertainty to calculated risk in individual clinical decisions.
Making inferences from groups to individuals is simple when the groups are homogenous, for example black or white balls. People are extremely heterogeneous. Epidemiological studies therefore need to be large to statistically even out all individual idiosyncrasies. To be able to infer the risk for one individual from the risk for a certain group of men, we need to assume that he is representative of the group. He must not possess any medically relevant idiosyncrasies.
But such idiosyncrasies exist. Some of them may be included in the next study, with a finer resolution in the information on physiological and genetic parameters. Other idiosyncrasies are too subtle for statistical epidemiology. One could imagine an overweight, but heavily grief-stricken pastry chef who quit smoking six months ago. He deals with grief by eating his own elaborate patisseries. If he stops comfort-eating there is a high risk that he will start smoking again. In this case, dietary advice will increase his risk of cardiovascular disease, in spite of all convincing epidemiological data stating that overweight is an independent risk factor. After all, smoking entails an even higher risk.
To the extent that the doctor is aware of such idiosyncrasies, she will have to assess their importance. The question will be whether the pastry chef’s representativity is so small that discarding the standard estimates of risk will be reasonable. The answer to this question will often be uncertain. Quantifying this uncertainty in the form of risk is nearly hopeless, since there is no epidemiological knowledge on grief-stricken pastry chefs. The doctor is therefore facing a situation characterised by strict uncertainty.
On other occasions, such issues objectively exist without the doctor being aware of them. From a bird’s eye view, we are apparently observing a paradoxical situation: it seems less problematic to use the risk estimate when we have less information on the patient. Could a lack of knowledge help us avoid the problem of strict uncertainty?
The paradox can be solved by realising that before a decision on medical procedures is made, we are making a choice of methodology. It is always possible to choose to handle strict uncertainty as risk (possibly adjusted with the aid of some medical discretionary judgment). The question is whether this is wise. The doctor who receives the pastry chef as his patient discerns this choice of methodology and will in this case have ample reason to discard the risk estimate. However, an implicit choice of methodology is invariably made.
How large is the medical effect of using risk estimates instead of de-emphasising them and instead spending more time and effort on looking for relevant idiosyncrasies in the patient’s history, life situation and self-concept? Is the effect positive? Can we know anything in general about this at all?
In our opinion, it is self-evident that establishing mathematical control over clinical uncertainty with the aid of risk estimation is an illusion.