Consider an epidemiological study of EMFs and childhood cancer.

The study looks at a certain number of specific children (for example it might look at 200 children with cancer and compare them to 200 children without cancer). Suppose it finds a statistical association between exposure to EMFs and cancer:

So we know there is a statistical association in the study.

One reason for this might just be chance. Perhaps the 200 children, just by chance, included more of the children exposed to EMFs. If we chose a different 200 children, we wouldn’t find the association. The association in our study does not reflect a true association in the population.

Another reason might be bias. Maybe the way we chose the children meant we got more children with cancer exposed to EMFs, or we got fewer comparison children exposed to EMFs. Or maybe there was some systematic bias in the way we measured exposure. So the children in our study are not representative of the whole population, and the association in our study does not exist in the population. More on bias in studies of EMFs.

But if neither of these applies, then:

We know there is a statistical association in the population.

This still might not mean that EMFs cause cancer. There might be something else in the environment – a factor X. Suppose X causes childhood cancer. But suppose that the children who are exposed to X are also exposed to EMFs. Then it would look as if cancer was associated with EMFs, but this would be a side effect of X. This is called “counfounding” and X is a “confounding factor”.  More on confounding factors.

But if there is no confounding, then:

We know there is a causal effect

Epidemiologists tend to approach a result in the order we’ve laid it out here. They tend to look at a study which has found an association, and accept it as establishing causation only if the alternatives – chance, bias, confounding – don’t seem likely.  See below for the factors epidemiologists take into account.

Two things to note: some of these effects can work in either direction. A bias in the measurements, for example, could often mean that the study finds a smaller association than there really is. And there doesn’t have to be just one of these in any one study, you could have some of the association explained by chance and some by bias, for example.

In our introduction to epidemiology we explain how an observation of a statistical association between an exposure and a disease may be evidence of causation, or it may have other explanations, such as chance, bias or confounding.

Deciding whether to deduce causation or not is a judgement.  But there are yardsticks to help with that judgement.  A statistical association observed in an epidemiological study is more likely to be causal if:

• it is strong (the relative risk is reasonably large)
• it is statistically significant
• there is a dose-response relationship - higher exposure seems to produce more disease
• the scope for bias or confounding seems limited
• the evidence comes from more than one study, preferably studies of different design
• there is biological support from laboratory experiments
• it is physically plausible

The Bradford-Hill criteria

One often-quoted attempt to set out these factors was by Sir Austin Bradford Hill.  His list was:

• Strength
• Consistency
• Specificity
• Temporal relationship
• Biological gradient
• Plausibility 
• Coherence
• Experiment
• Analogy

People sometimes try to use these criteria as a ticklist - tick them all and you've proved causation.  But Hill himself was clear these are just a guide and cannot take away the need for an informed judgement in each specific case.

The IARC rules

One organisation which has developed rules for assessing the strength of evidence and whether it amounts to evidence of causation or not is the International Agency for Research on Cancer.  See more on their rules.