How to read a study about harm or causation
Assess observational associations by examining timing, comparison groups, confounding, measurement and absolute risk.
You have a reason to review the evidence for an exposure when you see it associated with an undesired outcome. This does not establish whether the exposure caused that outcome. Individuals may already differ in ways that affect their health before receiving a treatment, encountering an exposure or choosing a behaviour.
Identify the exposure, outcome, comparison groups and observation period as you read a harm study. Ask how the investigators tried to make the comparison fair. The answers typically provide more information than the strength of the headline wording.
Identify the design
Cohort studies compare outcomes over time among groups with differing histories of exposure. Case-control studies begin with people with and without an outcome of interest, then compare previous exposure. A cross-sectional study measures both the exposure and the outcome at approximately the same point in time.
These designs address different questions and have different weaknesses. Cross-sectional data may leave the order of events unclear. Case-control studies can investigate uncommon outcomes efficiently, but selecting controls and measuring past exposure require close examination.
Determine how the investigator defined exposure. A single prescription, repeated use and measured doses are not synonymous. Check whether the outcome was assessed similarly across groups and whether knowledge of a person’s exposure could have influenced its assessment.
Examine the comparison groups
Confounding occurs when another factor helps explain an observed association. For instance, people referred to a specialist service may be more unwell before they arrive. Worse outcomes in that group would not, by themselves, show that the service caused harm.
Look for a reasoned account of the important confounders and how they were measured. An adjusted result is not automatically a causal result. Adjustment cannot reliably remove the influence of factors that were omitted, poorly measured or modelled inappropriately.
Also consider reverse causation. Early symptoms of an outcome can change a person's behaviour or lead to treatment. The apparent exposure may therefore be a consequence of an emerging problem. A convincing interpretation needs a plausible order in time.
Check when follow-up starts
The comparison groups need a defensible, comparable starting point. If someone must remain alive and event-free for a period before qualifying as exposed, counting that period incorrectly can make the exposure appear protective. This is one form of immortal time bias.
Ask about exclusions, changes in exposure and loss to follow-up. People who stop treatment or leave the dataset may have different risks from those who remain. A large sample does not eliminate these design problems.
Translate the association into counts
Imagine a hypothetical cohort followed for one year. An outcome occurs in 40 of 1,000 exposed people and 20 of 1,000 unexposed people. The observed risks are 4% and 2%, giving a risk ratio of 2 and a risk difference of 2 percentage points. These are risks over one year, not incidence rates per unit of person-time.
The relative comparison sounds large, while the absolute comparison shows an additional 20 events per 1,000 people. Both belong in a fair description. Neither calculation proves that exposure caused those extra events.
The inverse of the observed risk difference is 50. Calling that a causal number needed to harm requires assumptions about causation and comparability that the arithmetic alone cannot supply. For a case-control sample, the sampled case and control counts generally cannot be used to calculate population risks in this way. An odds ratio is not automatically interchangeable with a risk ratio.
Judge the whole explanation
Read the confidence interval alongside the estimate. Examine whether the association changes substantially after adjustment, whether different reasonable analyses agree and whether the authors discuss plausible alternative explanations.
STROBE helps researchers report observational studies clearly. It is not a certificate of methodological quality. Structured risk-of-bias tools can support appraisal, but their scope matters. ROBINS-I concerns non-randomised studies of interventions, rather than every possible exposure question.
A careful conclusion distinguishes an observed association from a defensible causal interpretation. It records the design limitations, the scale of the absolute risk and the population studied. An observational result can be valuable while still leaving uncertainty about why the groups differed.
Methodology references
- STROBE reporting checklist for cohort, case-control and cross-sectional studies
- ROBINS-I: a tool for assessing risk of bias in non-randomised studies of interventions
- Cochrane Handbook, Chapter 25: Assessing risk of bias in a non-randomized study