BIO 550 Week 6: Does Association Imply Causation?

The answer is no, every student knows it, and the phrase correlation does not imply causation is among the most repeated sentences in the discipline. That means the answer itself earns nothing and the whole post rests on the why. What is being tested is whether you can name the specific alternatives to a causal explanation, and whether you know what would license a causal claim rather than only what refuses one. Plan accordingly: state the answer in a sentence, spend two thirds of the post on the alternatives, and reserve the last third for the harder and more interesting question of how epidemiology ever concludes that something does cause something. That allocation is itself a signal to the marker that you understood which part of the question was actually open. Most of the board will invert that allocation, spending five paragraphs establishing something nobody disputes and one sentence on the part that required judgement.

There are four things other than causation that can produce an observed association, and naming all four is the core of a competent answer. Chance: with enough comparisons some will reach significance without anything real behind them. Bias, which subdivides into selection bias, where the way participants entered the study created the association, and information bias, where the way exposure or outcome was measured differs between groups. Confounding, where a third factor causes both the exposure and the outcome and the association between them is a shadow of that. And reverse causation, where the outcome influenced the exposure rather than the other way round. Give a one-line example for each rather than a definition, because the examples are what show the concepts are usable. A definition can be copied; an apt example cannot, which is why the examples carry more weight here than the taxonomy does.

Confounding deserves the most space, since it is the alternative that most often produces a persuasive and wrong result. The classic structure is easy to state: a factor associated with the exposure and independently causing the outcome will generate an apparent relationship where none exists. Coffee drinking and lung cancer is the standard illustration, because coffee drinkers historically smoked more, and the coffee association disappears once smoking is accounted for. What makes confounding hard in practice is that adjustment only works for confounders you measured. Residual confounding by unmeasured or imperfectly measured factors is the standing limitation of all observational work, and acknowledging it is what separates a sophisticated answer from a confident one. It is also the reason randomised allocation is valuable, since randomisation balances unmeasured factors as well as measured ones, which no amount of statistical adjustment can do.

Then turn the question around, because the interesting half of the topic is not that association fails to prove causation but that it is nonetheless the only evidence available for most public health questions. You cannot randomise people to smoke. So the discipline developed a framework for reasoning from association to causal inference: the viewpoints Bradford Hill set out — strength, consistency, specificity, temporality, biological gradient, plausibility, coherence, experiment, and analogy. Temporality is the only one that is genuinely necessary, since a cause must precede its effect. Hill himself was explicit that these were considerations rather than a checklist to be scored, and repeating that caution is a good sign you read about them rather than absorbing a summary. Specificity in particular has aged poorly, since most exposures of public health interest cause several outcomes and most outcomes have several causes.

A useful addition, and one that will distinguish your post, is the converse claim. If association does not imply causation, it is equally true that absence of association does not imply absence of causation. A real causal effect can be hidden by measurement error, by a confounder acting in the opposite direction, by a study too small to detect it, or by a mixture of subgroups in which the effect runs in opposite directions. Noting this prevents the post from landing on a purely sceptical conclusion, which is the register most answers on this question end in, and it is more useful professionally: the practitioner's problem is usually deciding how much evidence is enough, not proving that certainty is unavailable. Public health decisions are made under exactly that uncertainty, and a practitioner who can only say that nothing is proven has not helped anyone decide anything.

Ground the post in one real example carried through, rather than several mentioned. Smoking and lung cancer is the canonical case for good reason — it was resolved without a randomised trial, through consistency across designs and populations, a dose-response relationship, temporality, and eventually a plausible mechanism. Walking through how that association came to be accepted as causal answers the prompt far better than a list of criteria, because it shows the criteria being applied to something. Cite your sources for the framework rather than reciting it from memory, since the viewpoints are frequently misstated as necessary conditions, and getting that wrong undermines the very point you are making about careful inference. A marker reading twenty posts will notice which ones checked the framework and which ones remembered it. The irony is worth avoiding on a question that is entirely about the discipline of not overclaiming from evidence.