Finding that two things move together is how most causal questions begin. The observation is real and the inference is the hard part, because several quite different situations produce identical correlations.
Finding that two things move together is how most causal questions begin. The observation is real and the inference is the hard part, because several quite different situations produce identical correlations.
The arrow may point the other way. People who are already unwell may change behaviour in response, producing an association between the behaviour and the illness that is entirely backwards. Longitudinal data help and do not settle it.
A third factor can drive both. Where that factor is unmeasured, no statistical adjustment can remove it, which is the central limitation of observational work and the reason randomisation exists.
How a sample was assembled can create associations that do not exist in the wider population. Studying only people who sought treatment, or only those who stayed in a study, builds in a relationship that has nothing to do with the question.
Testing enough pairs of variables guarantees strong correlations by chance alone. A correlation discovered by searching a large dataset is weaker evidence than one predicted in advance, even when the numbers are identical.
A dose-response pattern, an effect appearing after the exposure rather than before, consistency across different populations and designs, and a plausible mechanism. None is sufficient alone; together they raise confidence substantially.
Observational evidence established several causal relationships that could never have been randomised. Treating all correlation as uninformative discards the only evidence available for many important questions.
Ask what else could produce this pattern, and whether the authors addressed it. A paper that names its most likely confounders and explains what it did about them is doing the work; one that does not is reporting a pattern.
Conditioning on a variable that is caused by both the exposure and the outcome creates an association where none exists. This is why restricting a sample to people who sought care, or who met some entry criterion affected by both variables, can manufacture a relationship from nothing. It is less familiar than confounding and harder to spot, because the adjustment looks like good practice.
A variable on the causal path between exposure and outcome should usually not be adjusted for, because doing so removes part of the effect being measured. Treating a mediator as a confounder is one of the more common ways a real effect is analysed away, and the distinction depends on subject knowledge rather than on the data.