Describe an association without overclaiming causation

An association tells you that two measured variables vary together in the analysed data. A causal effect asks what would change under a defined intervention or comparison.

Those statements are not interchangeable. Yet the difference cannot be managed by replacing “caused” with “was linked to” after the analysis is finished. The question, design, timing, measurement and analysis must support the claim from the beginning.

This guide provides a causal-claim audit. It helps you decide whether to write a descriptive association, a prediction or a causal effect—and which rival explanations must remain visible.

Start with the question, not the verb

Three studies can use the same variables and regression model while answering different questions.

On a narrow screen, scroll the table sideways to read every column.

Objective	Core question	Evidence target	Defensible language
Descriptive or associational	How do measured variables differ or move together in these data?	A clearly defined pattern in a stated sample, population and period	“was associated with,” “differed by,” “was higher among”
Predictive	How well does information available at one point predict an outcome in new cases?	Out-of-sample performance, calibration and a specified prediction setting	“predicted,” “improved prediction,” “classified,” with performance measures
Causal	What would happen to an outcome if a defined exposure, intervention or strategy changed?	An identified effect for a stated contrast and target population	“increased,” “reduced,” “had an effect,” qualified by design and assumptions

A variable may predict an outcome without causing it. An umbrella predicts rain because it is used when rain is expected; opening umbrellas does not cause rainfall. Likewise, a strong association can be useful for description or prediction while remaining unsuitable as an effect estimate.

The reverse mistake also matters. Observational does not automatically mean noncausal. Carefully designed analyses of observational data can target causal effects when their comparison, timing and identification assumptions are explicit and defensible. Hernán’s commentary on causal euphemisms argues that hiding a causal objective behind vague associational language can obscure the methodological scrutiny the study needs.

Name the objective first. Then judge whether the evidence meets it.

Build the effect before interpreting the coefficient

A causal question requires more than two variable names. Write the intended comparison in a complete record:

the target population;
the exposure, intervention or strategy;
the alternative or comparator;
the outcome and its measurement time;
the follow-up period;
the effect measure, such as a mean difference, risk difference or risk ratio;
the analysis population and handling of events after assignment or exposure;
the assumptions that connect the observed data to the effect.

For example, “Does study-room use affect engagement?” is underspecified. Which students, what amount of use, compared with what alternative, over which period, and measured by which engagement outcome?

A more precise causal question might be: among enrolled first-year students, what is the 12-week difference in mean engagement score if access to reservable study rooms is offered from week 1, compared with access beginning in week 7? That statement still does not prove the effect is identifiable, but it makes the proposed intervention, comparator, population and timing inspectable.

The four-targets framework similarly separates the target estimand, target population, target trial and target validity. It also warns that emulating a trial does not remove measurement, selection, confounding or generalization problems.

Run the causal-claim audit

Use the audit before writing an effect verb.

On a narrow screen, scroll the table sideways to read every column.

Audit field	Question to answer	If the answer is weak or unknown
Objective and estimand	Is the study estimating a defined effect, or only describing an association or building a prediction?	Do not let an effect verb invent a causal objective after analysis
Temporal order	Was the exposure or intervention defined before the outcome began to occur?	Retain reverse causation as an explanation
Assignment and exchangeability	Why should the compared groups be alike in causes of the outcome, apart from the exposure?	Confounding remains; adjustment for a short convenience list is not sufficient
Selection and missingness	Could entry, retention, response or complete-case inclusion depend on both exposure and outcome-related factors?	Qualify the association and examine selection mechanisms
Measurement	Are exposure, comparator and outcome measured consistently and with enough validity for the intended effect?	Misclassification or construct mismatch may create, hide or distort an association
Analysis choices	Were models, transformations, outcomes and subgroup analyses justified independently of the observed result?	Report selection and multiplicity; avoid presenting a chosen result as inevitable
Target population	Does the study sample support inference to the people, settings and implementation conditions named in the claim?	Separate internal validity from generalization
Uncertainty	What effect sizes remain compatible with the estimate, interval and assumptions?	A significance label cannot replace magnitude, uncertainty or bias assessment

The CDC Field Epidemiology Manual distinguishes statistical evidence against chance from causal judgment. An observed association can also arise from selection bias, information bias, confounding and errors in design, execution or analysis.

The audit is not a ritual list. Each issue should change the claim it affects.

Worked association: what the data support

The study, institution, sample and numerical output below are wholly fictional. They were constructed to teach interpretation and do not report participant data.

A fictional college administers one survey near the end of term. It records students’ self-reported weekly use of reservable study rooms and a 0–100 engagement score during the same period. A regression compares students reporting at least three hours of weekly room use with students reporting none, adjusting for year level and commuting time.

The fictional adjusted mean difference is +4.2 points, with a 95% confidence interval from +1.5 to +6.9 points.

A suitable results sentence is:

Students reporting at least three hours of weekly study-room use had a mean engagement score 4.2 points higher than students reporting no use, after adjustment for year level and commuting time (95% CI 1.5 to 6.9).

This sentence identifies the groups, scale, estimate, adjustment and interval. It does not say that study-room use increased engagement.

The causal claim is not supported because several alternatives remain unresolved:

Temporal order: room use and engagement were reported for the same period. More-engaged students may have chosen to use rooms.
Confounding: motivation, study workload, timetable gaps, peer networks and access to a quiet home space could influence both room use and engagement.
Selection: students who completed the voluntary survey may differ from nonrespondents in ways related to both variables.
Measurement: self-reported hours may be inaccurate, and the engagement scale may not capture learning or academic performance.
Model dependence: adjustment for year level and commuting time does not establish that all relevant common causes were measured and modelled correctly.
Generalization: the analysed respondents do not automatically represent all students or another institution.

A suitable discussion sentence is:

The cross-sectional association does not establish that study-room use increased engagement. Same-period measurement leaves temporal order unresolved, and self-selection into room use and survey response could partly or wholly explain the difference.

That is more informative than “causation cannot be inferred” on its own because it names the specific routes by which the estimate could be misleading.

Why cautious verbs are not enough

Writers sometimes replace “caused” with “influenced,” “affected,” “led to,” “contributed to” or “was linked to.” The first four still express causal ideas. The last can be so vague that readers cannot tell whether the intended claim is descriptive or causal.

Changing the verb does not change the estimand or remove bias.

If the objective is associational, say what was associated, in whom, when and on what scale. If the objective is causal, state the intended effect and evaluate whether the design and assumptions identify it. The methodological primer on descriptive and causal observational studies stresses that the two objectives face different questions: confounding belongs to a causal contrast, while selection and measurement can distort both causal and descriptive estimates.

A useful wording ladder is:

Observed pattern: “Engagement scores were higher among respondents reporting more room use.”
Quantified association: “Reporting at least three hours of room use was associated with a 4.2-point higher mean score.”
Adjusted association: “The association remained after adjustment for year level and commuting time.”
Causal effect: “Offering early room access increased engagement by…”—only when the comparison and assumptions justify that effect.

Do not climb the ladder merely because the p-value is small or the coefficient changes little after adjustment.

When causal language can be justified

Random assignment can create a defensible causal comparison, but even a randomized study needs an outcome, time point, analysis population and integrity checks.

This second example is also wholly fictional.

A fictional training programme randomly assigns 1,000 consenting participants, 500 per group, to a new reminder system or the standard reminder system. The analysis plan prespecifies 30-day module completion as the outcome and compares groups as assigned. Completion is recorded for everyone: 320 of 500 participants assigned to the new system and 275 of 500 assigned to the standard system.

The estimated intention-to-treat risk difference is +9 percentage points, with a 95% confidence interval from +3 to +15 percentage points.

A bounded causal statement is:

Assignment to the new reminder system increased 30-day completion by an estimated 9 percentage points relative to the standard system in the randomized study population (95% CI 3 to 15).

The claim concerns assignment to the reminder system, not guaranteed exposure to every message. It assumes the randomization and outcome records were implemented as described. It does not automatically generalize beyond the enrolled population, follow-up period or programme conditions.

Observational analyses can also pursue causal questions. A credible analysis may emulate a target trial or use another justified identification strategy, align eligibility and time zero, define treatment strategies, measure sufficient pre-exposure causes, address selection and missingness, and examine sensitivity to assumptions. The label “observational” neither grants nor forbids causal language. The design logic and evidence do the work.

Keep results and discussion sentences separate

The results section should report the estimate produced by the analysis. The discussion should explain the claim it can support and the alternatives that remain.

For an associational analysis:

Results: “Room use was associated with a 4.2-point higher mean engagement score (95% CI 1.5 to 6.9) after the stated adjustment.”
Discussion: “The same-period survey cannot determine whether room use preceded engagement, and unmeasured self-selection may explain the association.”

For the fictional randomized comparison:

Results: “Completion was 64% in the new-reminder group and 55% in the standard group; the risk difference was 9 percentage points (95% CI 3 to 15).”
Discussion: “Random assignment supports an effect of assignment to the new system for this study population, although transfer to other programmes and the effect of actually receiving every reminder remain separate questions.”

Use the results-versus-discussion guide when the two jobs become blurred. Use the confidence-interval guide to preserve uncertainty, and the practical-importance guide before treating a detected effect as consequential.

A practical 12-step workflow

Write the research objective as descriptive, predictive or causal.
Name the target population and analysis sample.
Define the exposure or intervention and its comparator.
Define the outcome, scale, time zero and follow-up.
State the estimand and effect measure.
Explain why the compared groups should support that contrast.
Check temporal order and reverse causation.
Check confounding, selection, missingness and measurement.
Record model, transformation, outcome and subgroup choices.
Report the estimate and interval without upgrading the claim.
Separate internal validity from generalization.
Draft the sentence, then remove any verb whose causal strength exceeds the audit.

Transparent reporting helps readers assess what was planned, done, found and inferred. The STROBE initiative provides reporting checklists for major observational designs, but it explicitly does not present them as design prescriptions or quality scores. Completing a checklist cannot rescue an unidentified causal effect.

Final claim check

Before submitting, ask:

Does the sentence answer the question the study actually set out to answer?
Is the exposure or intervention contrasted with a clear alternative?
Did the proposed cause precede the outcome?
Are the main rival explanations named rather than hidden behind “may”?
Does adjustment have a causal rationale, or is it only a list of available variables?
Is uncertainty visible?
Is the claim about the analysed sample or an unsupported wider population?
Would a reader mistake prediction for intervention effect?
Would the sentence still be defensible if the result were not statistically significant?
Have you stated the assumptions on which any causal wording depends?

The aim is not to make every sentence cautious. It is to make each sentence exact. A well-supported causal analysis should state its causal objective and assumptions plainly. An associational analysis should describe the observed relationship just as plainly, without letting significance, software or vocabulary imply an effect the design did not identify.

Statistics and data-analysis support can help you examine the estimand, design assumptions, alternative explanations and reporting language using your own data while you retain responsibility for the research and its conclusions.