Causation

To say that a treatment caused an improvement is to compare what happened with the treatment with what would have happened to the same person without it.

Each person has two potential outcomes, and the causal effect for that person is the difference between them (Rubin, 1974). Only one can ever be observed; the other is the counterfactual. An average effect can nevertheless be estimated when the treated and untreated groups are comparable, which randomisation ensures.

A causal diagram, or directed acyclic graph (DAG), represents each variable as a point and each assumed causal effect as an arrow (Greenland et al., 1999). Association passes along any unblocked path between two variables, whatever the direction of the arrows. Four structures account for most differences between association and causation.

Collider bias

In the simulation, an anxiety score and a memory-problem score are independent, and people are referred to a memory clinic when their sum is high. Among referred patients the two scores are negatively correlated, because those with little anxiety must have had marked memory problems to be referred.

When people choose their own treatment, those who receive it usually differ from those who do not, and these differences can explain a difference in outcomes. Random allocation breaks this link: every characteristic, measured or not, is expected to be balanced between the groups, and any imbalance is due to chance, which the statistical analysis allows for.

A trial also needs a control group, which differs only in not receiving the treatment, and often a placebo, an inactive treatment that looks the same. Expectation alone can have biological effects: in people with Parkinson’s disease, a placebo was followed by a substantial release of dopamine in the striatum (de la Fuente-Fernández et al., 2001). Blinding keeps participants, clinicians and assessors unaware of who received what, and allocation concealment prevents those enrolling participants from predicting the next allocation (Schulz et al., 2010).

Functional MRI and EEG record which brain regions are active during a task, but activity can accompany a behaviour without being needed for it. To test whether a region is necessary or sufficient, neuroscientists change its activity and measure the consequences.

Many causal questions cannot be answered by randomising people, for example questions about head injury or genotype. Quasi-experimental designs use features of the world that act like randomisation; each is persuasive only when its assumptions are plausible (Craig et al., 2017).

Bradford Hill described nine aspects of an association to weigh before judging it causal, and stressed that none is decisive on its own (Hill, 1965). The exception is temporality: a cause must precede its effect (Rothman and Greenland, 2005).

The claim that smoking protects against Parkinson’s disease shows how a causal claim is assessed. Smoking causes many serious diseases; nothing here suggests that anyone should smoke.

1. The association

   A meta-analysis of 44 case–control and 4 cohort studies found a relative risk of 0.59 (95% CI 0.54 to 0.63) for people who had ever smoked and 0.39 (0.32 to 0.47) for current smokers, compared with people who had never smoked, and the risk fell with each additional 10 pack-years (Hernán et al., 2002). The association is strong, consistent and graded, and unlikely to be due to chance.

2. Alternative explanations

   Parkinson’s disease begins years before diagnosis, and its early changes might make people less inclined to smoke (reverse causation). Smokers also die earlier of other causes, before a diagnosis can be made (survivorship bias).

3. Experimental evidence

   In a randomised, placebo-controlled trial, 163 people with recently diagnosed Parkinson’s disease wore nicotine or placebo patches for a year. Nicotine did not slow progression (Oertel et al., 2023).

4. Conclusion

   The trial tested progression, not onset, and one component of tobacco, so whether smoking affects the risk of developing the disease remains open. A fair summary is: “strongly associated; causal explanation not established”.