1. Association Is a Starting Point, Not a Causal Explanation
Correlation means that two measured variables vary together. Causation means that changing one variable would produce a change in another, under specified conditions. Suppose a study finds that students who use a learning app earn higher test scores. That association does not, by itself, show that the app improves learning. More motivated students might be more likely to use it, or better-resourced schools might provide both the app and additional instruction.
A causal question asks what would happen under an intervention: would otherwise comparable students perform differently if they used the app rather than not using it? Researchers cannot observe the same student both using and not using the app during the same learning period. Study designs therefore seek credible comparisons that approximate this missing alternative. Correlation can point toward a causal relationship, but the explanation requires additional evidence and assumptions.
2. Confounding and Other Reasons Associations Can Mislead
A confounding factor is a shared cause of the exposure being studied and the outcome. In the app example, prior motivation could influence both app use and test performance. Unless researchers address that difference, they may attribute motivation's effect to the app. Reverse causation creates another difficulty: students struggling with coursework might start using the app because their scores are low, rather than the app causing lower scores.
Selection and measurement can also distort results. A survey limited to enthusiastic app users may not represent everyone who tried it. Self-reported usage may differ from actual usage. These problems are distinct from confounding, but all can undermine a causal interpretation.
Statistical adjustment can help address measured confounders, yet controlling for more variables is not automatically better. Adjusting for a mediator—something through which an intervention works—can remove part of the effect researchers want to estimate. Conditioning on a shared consequence of two variables, called a collider, can create a misleading association. Choosing adjustments requires knowledge of how the variables relate, not just access to a large dataset.
3. What Observational Studies Can Show
Observational studies measure exposures and outcomes without assigning the exposure. Cross-sectional studies capture a snapshot and can identify patterns, but often cannot establish which variable came first. Longitudinal cohort studies follow participants over time, helping establish that an exposure preceded an outcome. Case-control studies compare people with and without an outcome and examine earlier exposures; they can be useful for rare outcomes but require careful comparison-group selection.
These designs can support causal conclusions when their methods and assumptions credibly address alternative explanations. Researchers may adjust for baseline differences, compare similar participants, or examine whether findings persist under different analytical choices. However, unmeasured or poorly measured confounders may remain. Following people over time establishes temporal order more clearly; it does not automatically remove confounding.
4. Why Randomized Experiments Strengthen Causal Inference
Randomized experiments assign participants to an intervention or comparison condition by chance. This makes assignment independent of participants' existing characteristics, balancing known and unknown confounders on average. With an appropriate comparison, consistent outcome measurement, and sound analysis, differences between groups can provide strong evidence that the assigned intervention caused an effect.
Randomization is not a guarantee of flawless evidence. Small groups can differ by chance, participants may drop out unevenly, and expectations can affect behavior or assessment. Blinding can reduce some of these problems when feasible. Nonadherence also matters: an analysis based on original assignment estimates the effect of being assigned the intervention, which may differ from the effect of actually receiving it. Finally, an effect established in one setting may not transfer unchanged to other populations or conditions.
5. Natural Experiments and Evidence Across Studies
When randomization is impractical or unethical, researchers sometimes use policy changes, eligibility thresholds, or other external events to construct comparisons. These natural experiments and quasi-experimental designs can support causal inference when their assumptions are credible. For example, comparing changes across affected and unaffected groups requires a convincing case that their outcomes would otherwise have followed comparable trends.
Systematic reviews and meta-analyses assess evidence across studies. They can reveal consistency, differences between settings, and uncertainty more clearly than a single result. But combining studies does not automatically eliminate their biases. Several studies using the same flawed comparison may reproduce the same misleading conclusion. Confidence grows when different methods, with different vulnerabilities, converge on a compatible explanation.
6. Read the Claim Alongside the Design
When evaluating a scientific claim, ask what was compared, whether the exposure preceded the outcome, and why the groups differed in exposure. Look for plausible confounders, missing participants, measurement limitations, and assumptions behind statistical adjustments. Then consider the effect's size and uncertainty—not merely whether a statistical test crossed a significance threshold.
Match the wording to the evidence. “Associated with” describes an observed relationship; “caused” requires a defensible causal argument. Neither a large sample nor an impressive model substitutes for a credible comparison. The most useful conclusion states what the study supports, which alternatives remain plausible, and where its findings may apply.
