Observational Study

Observational Study

An observational study measures exposures, characteristics and outcomes without a researcher assigning the exposure under investigation. It can describe disease burden, investigate associations, assess safety and, with a suitable design and assumptions, estimate effects relevant to health economic decisions. This page distinguishes the principal designs, follows a concrete comparison from question to result, and explains why an observed association needs more scrutiny before it informs a causal claim.

Choose the design to match the question

The timing of selection, exposure measurement and outcome measurement determines what a study can estimate. “Observational” describes the absence of investigator assignment of the exposure of interest; it does not mean that measurement or analysis is unplanned or that all observational designs are interchangeable.

Design	How participants enter	Typical strength	Important constraint
Cohort	Select people by exposure or eligibility and compare subsequent outcomes, prospectively or from existing records.	Can estimate incidence and directly compare outcome risks over a defined follow-up.	Loss to follow-up, time-related bias and confounding need attention.
Case-control	Select people by outcome status and reconstruct prior exposure.	Efficient for relatively rare outcomes or long latency.	Selection and recall can distort exposure comparisons; sampled case fractions are not population risks.
Cross-sectional	Measure exposure and outcome at or around one point in time.	Describes prevalence and current associations.	Temporal order and incident risk usually cannot be established.

A registry or electronic-health-record study describes a data source, not a design by itself. The same source may support a cohort, nested case-control or cross-sectional analysis. A natural experiment or quasi-experimental analysis can also use observational data, but must specify the assignment mechanism and its own identification assumptions.

Specify the comparison before calculating

Define a target population, eligibility, treatment or exposure strategies, time zero, follow-up, outcome, summary measure and handling of switching or competing events. Align eligibility, exposure classification and follow-up at the same starting point. Otherwise, a patient may need to survive long enough to be classified as treated, creating immortal-time bias.

For an intervention question, a useful discipline is to describe the hypothetical randomized trial one would ideally run and then identify which parts can be emulated with observed data. This does not turn an observational study into a randomized trial: it makes the estimand and assumptions clearer. For descriptive questions such as prevalence, define sampling and measurement instead of forcing a causal comparison.

Calculate and interpret a cohort result

Suppose a fictional cohort contains 1,000 patients beginning treatment A and 1,000 beginning treatment B, with complete 12-month follow-up for a specified adverse event. The observed counts are 80 events under A and 120 under B. The 12-month observed risks are $80/1{,}000=0.08$ and $120/1{,}000=0.12$; the risk difference for A minus B is $0.08-0.12=-0.04$, or four fewer events per 100 treated people. The risk ratio is $0.08/0.12\approx0.67$.

Quantity	Illustrative spreadsheet expression	Result
Risk under A	=80/1000	0.08 or 8%.
Risk under B	=120/1000	0.12 or 12%.
Risk difference, A minus B	=80/1000-120/1000	-0.04, or -4 percentage points.
Risk ratio, A divided by B	=(80/1000)/(120/1000)	Approximately 0.67.

These are crude associations within the selected cohort, not estimates of the effect of assigning A rather than B unless assumptions about comparability and bias are justified. If patients prescribed A were younger or less ill at baseline, some of the difference might predate treatment. If follow-up time differs or death prevents the event, a simple count divided by baseline enrollment may no longer estimate the intended 12-month risk; choose an appropriate time-to-event or competing-risk analysis. The example is invented and has no clinical interpretation.

Examine bias and the assumptions for causal interpretation

Confounding arises when causes of treatment choice also predict outcomes. Selection bias can arise from inclusion, retention or conditioning on variables affected by treatment; information bias can arise from inaccurate exposure or outcome measurement. These threats need design-specific investigation, not a generic adjustment checklist.

Threat	Concrete example	Possible response and remaining limit
Confounding by indication	Sicker patients receive a more intensive medicine.	Measure baseline severity and compare clinically eligible alternatives; residual unmeasured severity may remain.
Selection	A database captures only insured patients who stay enrolled.	Describe coverage, compare inclusion and attrition, and assess generalizability; missing people may differ.
Misclassification	Adherence is inferred from a filled prescription.	Validate or vary exposure definitions; dispensing does not prove ingestion.
Time-related bias	Treated status is assigned after a qualifying survival period.	Set a common time zero and define strategies and follow-up prospectively.

Regression adjustment, matching, weighting and instrumental-variable approaches answer different questions under different assumptions; no technique automatically removes unmeasured confounding. Check positivity or overlap when comparing groups, specify whether effects are intended for all eligible patients or a subgroup, and examine missingness and sensitivity to plausible unmeasured factors. A reporting checklist such as STROBE improves transparency but does not certify freedom from bias. For intervention-effect evidence, structured risk-of-bias assessment such as Cochrane's ROBINS-I is distinct from reporting completeness.

Connect the evidence to health economic evaluation

Observed utilization, costs, quality-of-life measurements and event rates can inform a model when their population, perspective, time period and measurement match the decision problem. A baseline event risk and a causal relative treatment effect have different roles: substituting a confounded crude association for a treatment effect may bias projected costs and QALYs. Examine whether costs include all relevant care, whether outcomes reflect the intended population, and how uncertainty and heterogeneity will be represented.

When comparing economic outcomes in an observational cohort, define the estimand and account for skewed costs, differential follow-up, censoring, confounding and uncertainty. Report absolute as well as relative effects where possible, because decision models often require baseline risks and absolute event counts. A precise estimate can still be systematically wrong; uncertainty intervals describe sampling variability under a model, not all sources of bias.

Report what the study can support

Provide setting, recruitment, dates, eligibility, exposure and outcome definitions, denominators, follow-up, missing data, analytic choices and material limitations. Separate an observed association from an effect attributable to an intervention. Discuss transportability to the policy population, including differences in disease severity, access and usual care, before applying the estimate elsewhere.

Sources and further reading

• STROBE Initiative, checklists for transparent reporting of cohort, case-control and cross-sectional studies: https://www.strobe-statement.org/checklists/
• Cochrane Handbook, Chapter 25, assessment of bias in non-randomized intervention studies, including confounding: https://www.cochrane.org/authors/handbooks-and-manuals/handbook/current/chapter-25
• Cochrane Methods, ROBINS-I, for risk of bias in results from non-randomized studies of interventions: https://methods.cochrane.org/bias/risk-bias-non-randomized-studies-interventions
• Hernán MA and Robins JM, “Using big data to emulate a target trial when a randomized trial is not available” (2016): https://pubmed.ncbi.nlm.nih.gov/26994063/