If increasing an exposure produces a correspondingly larger effect, the pattern is difficult to explain by confounding alone, because a confounder would have to track the exposure across its whole range.
If increasing an exposure produces a correspondingly larger effect, the pattern is difficult to explain by confounding alone, because a confounder would have to track the exposure across its whole range. This is why dose-response is one of the classical criteria for inferring causation from observation.
A confounder that itself varies with the exposure will produce the same gradient. People taking more of something frequently differ systematically from people taking less, and the gradient then reflects that difference.
Some relationships are threshold effects: nothing below a level, full effect above it. Others are U-shaped, where too little and too much are both worse. The absence of a straight line is not evidence against causation.
Where people adjust their exposure in response to their condition, the sickest may take the most, producing a gradient that points the wrong way. Longitudinal data separate the two and cross-sectional data cannot.
Imprecisely measured exposure pulls an observed gradient toward flatness. A weak-looking dose-response may be a strong one measured badly, which is the opposite of the usual direction of bias and worth remembering.
Studies often split a continuous exposure into groups. Where the cut points were chosen after seeing the data, the resulting pattern can be an artefact of the splitting rather than of the relationship.
A gradient reproduced in independent populations, with different measurement methods, and a mechanism that predicts the shape. Those together move an association a long way toward a causal reading.
Ask whether the gradient was predicted in advance, whether the exposure was measured more than once, and whether the extreme groups differ in anything else obvious.
Not every genuine relationship is a straight line. Some appear only above a level, and some worsen at both extremes. Absence of a clean gradient is therefore not evidence against causation, though it is often reported that way.
Imprecisely measured exposure pulls an observed gradient toward flat, so a weak-looking dose-response may be a strong one measured badly. This runs opposite to most intuitions about how error behaves.
The claims on this page about more produces more is one of the stronger signals a causal claim can carry can be checked against sources outside this site: