Observational Studies
Dr. Nadim Mahmud
Study Design · Research CurriculumMost clinical research you will read and conduct as a trainee uses an observational design. This module walks through the key design types, when to use each, how to interpret their results, and the biases you need to watch for.
Introduction
Clinical research can be broadly divided into two categories: experimental studies, in which the investigator assigns an exposure or intervention; and observational studies, in which the investigator simply observes. In observational research, participants are not randomized - they are studied as they naturally exist in the world, with their real-life exposures, comorbidities, and treatment histories intact.
The vast majority of research you will encounter as a trainee - and likely conduct yourself - will be observational. Understanding the design, strengths, and limitations of the major observational study types is therefore essential, both for critically appraising the literature and for designing your own projects.
There are three main types of observational study design: cross-sectional studies, cohort studies, and case-control studies. Each has a distinct structure, a specific temporal logic, and a different set of trade-offs. Before diving into each, it helps to see how they fit into the broader landscape of research study designs.
Overview of Research Study Designs
The diagram below maps out the major categories of research study design. Observational studies sit within the analytic branch - meaning they are designed to evaluate associations between an exposure and an outcome. Click any study type in the Observational node to learn more.
Click a study type to learn more
Cross-Sectional Studies
A cross-sectional study analyzes data from a population at a single point in time. Both the exposure and the outcome are measured simultaneously - there is no follow-up. Cross-sectional studies can serve two purposes: a descriptive function (estimating the prevalence of an exposure or outcome in a population) or an analytic function (estimating the association between an exposure and an outcome within the same timeframe).
- Descriptive: Estimating the prevalence of undiagnosed diabetes among adult patients in the Penn system.
- Analytic: Administering a survey to assess the association between screen time and sleep quality at a single clinic visit.
Cohort Studies
In a cohort study, investigators identify a group of participants based on their exposure status (exposed vs. unexposed) and then follow them forward in time to compare the incidence of outcomes between the two groups. Because exposure is assessed before the outcome develops, cohort studies can establish temporal sequence - a prerequisite for causal inference.
The landmark example of a cohort study. Beginning in 1948, investigators enrolled residents of Framingham, MA and followed them longitudinally to identify risk factors for cardiovascular disease. From this single cohort, investigators identified hypertension, hyperlipidemia, diabetes, smoking, obesity, and physical inactivity as major CVD risk factors.
Notably, the same Framingham dataset was later used to conduct retrospective cohort studies, cross-sectional analyses, and nested case-control studies - illustrating how a single rich dataset can support multiple study designs.
Prospective vs. Retrospective Cohort Studies
Cohort studies can be either prospective or retrospective, and the distinction is practical rather than conceptual.
Participants are enrolled and followed forward in real time. Exposure and covariate data are collected as the study proceeds, often with greater control over data quality and completeness.
- Data quality is higher (prospectively collected)
- Can capture exposures and outcomes more precisely
- Time-consuming and expensive
- Loss to follow-up is a concern
The exposure and outcome have already occurred before the study begins. Investigators use existing data (e.g., medical records, administrative databases) to reconstruct the cohort historically.
- Faster and less expensive
- Uses existing electronic health record data
- Limited by what was captured in the data source
- More susceptible to information bias
Case-Control Studies
In a case-control study, investigators start with the outcome and work backwards. They identify participants with the outcome of interest (cases) and participants without the outcome (controls), then compare the prevalence of prior exposures between the two groups.
The case-control design is particularly efficient for studying rare outcomes or diseases with long latency periods. Instead of following tens of thousands of people and waiting years for a rare outcome to develop (as in a cohort study), investigators select a smaller group of cases and match them to controls, making the study far less expensive and faster to execute.
To study the association between second-hand smoke and lung cancer: rather than enrolling 20,000 people and waiting years for lung cancers to develop (a massive cohort study), investigators recruit 100 patients with lung cancer (cases) and 100 patients without lung cancer (controls), then ask both groups about their history of second-hand smoke exposure.
RR vs. OR: Measures of Association
One of the most commonly confused concepts in clinical research statistics is the difference between the Risk Ratio (RR) and the Odds Ratio (OR). The choice between them is not arbitrary - it is determined by study design.
The ratio of two probabilities (risks). Risk = proportion of people in a group who develop the outcome.
RR = Risk(exposed) / Risk(unexposed)
Requires knowing the total number of exposed and unexposed individuals. Valid in cohort studies (both prospective and retrospective), but cannot be calculated from a case-control study.
The ratio of two odds. Odds = probability of the event / probability of the non-event.
OR = Odds(exposed) / Odds(unexposed)
Can be calculated from any 2×2 table, including case-control studies (using the cross-product: ad/bc). Valid in all observational designs.
Why Are the RR and OR Often Different?
When an outcome is common (say, 85% in one group), the RR and OR can diverge dramatically. When the outcome is rare (less than ~10% in both groups), the OR approximates the RR closely. This is called the rare disease assumption, and it is why ORs from case-control studies of rare diseases are often interpreted as if they were risk ratios.
Interactive 2×2 Table
Use the calculator below to compute the RR and OR from a 2×2 table. Try the pre-loaded examples or enter your own values. Notice how the RR and OR diverge when the outcome is common.
e.g., daily fast food and GI symptoms (85% vs 20%)
| Outcome + | Outcome − | Total | |
|---|---|---|---|
| Exposed | 100 | ||
| Unexposed | 100 |
Risk(exposed) = 85.0% | Risk(unexposed) = 20.0%
"4.25-fold higher risk in exposed vs. unexposed"
Odds(exposed) = 5.67 | Odds(unexposed) = 0.25
Edit any cell to explore how values affect the results. The callout above updates automatically.
Cohort vs. Case-Control: When to Use Which?
Both cohort and case-control studies aim to evaluate associations between an exposure and an outcome. The choice of design depends primarily on the rarity of the outcome, the length of the time horizon, and practical resource constraints.
| Feature | Cohort Study | Case-Control Study |
|---|---|---|
| Starting point | Exposure status | Outcome status |
| Direction | Forward (exposure → outcome) | Backward (outcome → exposure) |
| Best for | Rare exposures, multiple outcomes | Rare outcomes, long latency periods |
| Speed & cost | Slow and expensive (especially prospective) | Fast and inexpensive |
| Measures available | RR, HR, OR, incidence | OR only - cannot calculate incidence |
| Can study multiple outcomes? | Yes | No (one outcome per study) |
| Loss to follow-up bias? | Yes - concern if non-random | Less of a concern |
| Evidence strength | Generally stronger (temporality is clearer) | Weaker (retrospective exposure data) |
Observational Studies vs. Randomized Controlled Trials
If observational studies are susceptible to confounding and bias, why do we conduct them at all? The answer is that RCTs, while methodologically superior, are not always feasible, ethical, or appropriate.
- + Exposure is randomized → eliminates confounding
- + Higher internal validity (less systematic bias)
- − Strict enrollment criteria → limited generalizability
- − Expensive, time-consuming
- − Ethically impossible for harmful exposures
- − Not feasible for rare outcomes or long time horizons
- + Real-world populations → higher external validity
- + Can study harmful or unethical exposures
- + Feasible for rare outcomes and long time horizons
- + Fast and inexpensive (especially retrospective)
- − Exposure not randomized → risk of confounding
- − Lower internal validity
Key Biases in Observational Studies
Because observational studies do not randomize, they are susceptible to biases that can distort the estimated association between an exposure and an outcome. Being able to identify these biases - both when reading papers and when designing your own study - is an essential skill in clinical research.
Occurs when the process of enrolling participants is systematically influenced by exposure or outcome status, causing the study sample to be unrepresentative of the intended population. Example: a cohort of hospitalized patients may not reflect community disease burden.
Arises when exposures are inaccurately measured or misclassified. Can lead to over- or underestimation of the true association. Example: self-reported dietary intake is less accurate than biomarker measurement.
A subtype of information bias common in case-control studies. Cases (who have the disease) may remember and report past exposures differently than controls, inflating the apparent association.
Occurs when a period of follow-up during which the outcome cannot occur (by definition) is incorrectly attributed to the exposed group. This can make an exposure look falsely protective. Example: a drug study where patients must survive long enough to receive the drug are counted as "exposed" from enrollment, not from first prescription.
Participants who drop out of a cohort study may differ systematically from those who remain - especially if dropout is related to exposure or outcome status.
Occurs when one exposure group is surveilled more intensively for the outcome than another. A more-surveilled group will appear to have a higher outcome rate simply due to greater detection.
Early detection of disease in one group is mistaken for improved survival, when in fact the group simply had their diagnosis advanced in time. Common in screening studies.
A confounder is a variable that is associated with both the exposure and the outcome, and can distort the apparent relationship between them. Example: in a study of coffee and lung cancer, smoking is a confounder (smokers drink more coffee AND are more likely to develop lung cancer). Confounding is addressed through multivariable adjustment, matching, or propensity score methods.
Interactive: Identify the Study Design
Read each vignette and try to identify the study design before revealing the answer. Pay attention to two key questions: What is defined first - the exposure or the outcome? And: Does the study look forward, backward, or at a single point in time?
Patients in a community are enrolled in a study to determine the risk of coronary artery disease. On the date of enrollment, each participant reports their smoking status. Over the next ten years, they have yearly study visits to ascertain changes in smoking status and occurrence of CAD events.
Investigators wish to examine the association between alcohol intake and pancreatic cancer. They select a group of patients with pancreatic cancer and a group without pancreatic cancer, then solicit detailed prior alcohol histories from both groups.
Investigators wish to study the association between TV watching and obesity. They enroll consecutive patients presenting to a primary care clinic and, at that single visit, ask about hours per week of TV watched and measure current weight and height.
A study aims to determine the association of psoriasis with development of heart disease. The Nurses' Health Study began enrolling participants in 1976 and follows them every 2 years. In 2008, a research team begins a new analysis examining the relative incidence of heart disease among NHS participants with vs. without psoriasis, using data already collected since 1976.
Continue Learning
Observational studies are the foundation of most clinical research. As you build on this module, the next step is understanding how randomized controlled trials differ in design and interpretation - and why they sit at the top of the evidence hierarchy.