Epidemiological Study Designs

Introduction and context

Epidemiology is the study of the distribution and determinants of health-related states or events in specified populations, and the application of this knowledge to control health problems. For health-care managers, epidemiology is not only a research discipline. It supports service planning, quality improvement, resource allocation, outbreak response, prevention policy and evaluation of programs.

A study design is the overall plan used to answer a research question. It determines how participants are selected, how exposure and outcome are measured, when observations are made, whether an intervention is assigned, and how results are interpreted. The choice of design should follow the question rather than convenience alone. A design that is efficient for estimating prevalence may be unsuitable for establishing temporality; a design that provides strong causal evidence may be infeasible or unethical in routine service settings.

The central distinction is between descriptive and analytic epidemiology. Descriptive studies characterize the occurrence of disease by person, place and time. Analytic studies compare groups to assess associations or causal effects. Designs may also be classified as observational or experimental, and as individual-level or population-level.

Learning objectives

By the end of this chapter, the learner should be able to:

1. Distinguish major descriptive, observational and experimental study designs.
2. Select a suitable design for questions about frequency, determinants, prognosis, prevention and service performance.
3. Interpret measures of occurrence and association used in different designs.
4. Evaluate internal validity, external validity, feasibility, ethics and managerial usefulness.
5. Recognize confounding, selection bias, information bias and threats to causal inference.
6. Compare explanatory randomized trials with pragmatic trials and quasi-experimental evaluations.
7. Develop a defensible design strategy for a health-care management problem.

A framework for choosing a design

A useful starting point is to specify the question using the PICOT structure:

• Population: Who is being studied?
• Intervention or exposure: What factor, treatment or policy is of interest?
• Comparator: Against what alternative is it assessed?
• Outcome: What health or service result is measured?
• Time: Over what period is the outcome assessed?

The design should also reflect whether the outcome is common or rare, whether the exposure is modifiable, whether randomization is ethical and feasible, and whether the primary purpose is description, prediction, explanation or evaluation.

No design automatically eliminates bias. Validity depends on how the design is implemented, how variables are defined, how missing data are handled and whether the analysis corresponds to the design.

Descriptive epidemiology

Descriptive studies answer “what, where, when and in whom?” They may use routinely collected records, registries, surveillance systems, clinical databases, census information or specially collected data. They commonly report counts, proportions, rates and distributions by demographic, geographic and temporal characteristics.

A case definition is essential. It specifies the clinical, laboratory or administrative criteria used to classify a person as a case. Changes in the case definition can create apparent changes in disease frequency even when the underlying risk is stable.

For a population in which a disease is measured at a particular time, prevalence is:

Prevalence is a proportion and has no time unit, although it refers to a defined point or period. Incidence concerns new cases. Cumulative incidence over a specified period is:

The denominator must exclude people who already have the outcome and should correspond to the population capable of developing it. When individuals contribute different amounts of observation time, incidence rate is more appropriate:

The unit is, for example, cases per 1,000 person-years. Prevalence is influenced by both incidence and duration. In a relatively stable population with a rare disease and approximately constant rates, prevalence is sometimes approximated by:

This approximation is not appropriate when disease is common, the population is changing rapidly, or incidence and duration vary substantially.

Descriptive studies are valuable for identifying service needs, generating hypotheses and detecting clusters. However, they generally cannot distinguish whether an observed pattern is caused by an exposure, by confounding, by differences in diagnosis or by differential access to care.

Ecological studies

An ecological study uses groups rather than individuals as the unit of analysis. Groups may be countries, districts, hospitals, schools or calendar periods. For example, a manager might compare district-level vaccination coverage with district-level hospitalization rates.

Ecological designs are efficient when exposure data are available only at population level, when the policy operates at group level, or when rapid comparisons across regions are needed. They can also examine contextual factors such as air pollution, health expenditure or facility density.

The major limitation is the ecological fallacy: an association observed between group averages does not necessarily apply to individuals within those groups. A district with higher average pollution and higher mortality does not prove that the individuals who died were more exposed than survivors. Ecological comparisons may also be distorted by differences in age structure, socioeconomic conditions, diagnostic practices and reporting.

Time-series ecological studies examine repeated observations in a population. They are useful for monitoring seasonal disease, assessing abrupt policy changes and identifying secular trends. Autocorrelation, seasonality, concurrent interventions and changes in measurement must be considered.

Cross-sectional surveys

A cross-sectional study measures exposure and outcome at approximately the same time. It can estimate the prevalence of disease, risk factors, service use, unmet need, satisfaction or health behaviors. A representative health survey, for example, may estimate the prevalence of hypertension and the proportion of affected adults receiving treatment.

Cross-sectional surveys may be census-based or sample-based. Probability sampling improves representativeness, while weighting may be needed when selection probabilities differ. A basic prevalence ratio is:

A prevalence odds ratio may also be estimated, particularly using logistic regression, but it should not be interpreted as a risk ratio when the outcome is common.

The major weakness is uncertain temporality. If obesity and depression are measured together, it may be unclear whether obesity preceded depression, followed it, or is associated with a third factor. Non-response can reduce validity if responders differ systematically from non-responders. Recall bias and socially desirable reporting may affect self-reported behaviors.

Cross-sectional surveys are often highly useful to managers because they provide a rapid service profile and can support budgeting and workforce planning. They are less suitable for studying short-lived exposures, rare diseases or causal pathways requiring long follow-up.

Case-control studies

A case-control study begins with outcome status. Investigators select people with the disease or outcome (cases) and comparable people without it (controls), then determine previous exposure histories. The design is especially efficient for rare diseases, diseases with long latency and outcomes requiring extensive diagnostic investigation.

The usual measure is the odds ratio. If exposure status is arranged in a 2 × 2 table:

then:

The odds ratio compares the odds of previous exposure among cases with the odds among controls. Under appropriate conditions, especially when the outcome is rare and controls represent the exposure distribution in the source population, it can approximate the risk ratio. This approximation should not be assumed for common outcomes.

Control selection is the central methodological challenge. Controls should arise from the same source population as cases and should have been eligible to become cases if they had developed the disease. Hospital controls may be convenient but can be inappropriate if their illnesses are related to the exposure under study. Population controls may improve representativeness but be more expensive.

Important biases include:

• Recall bias: cases and controls remember or report exposure differently.
• Selection bias: inclusion depends on exposure and outcome-related factors.
• Interviewer bias: knowledge of case status influences questioning or classification.
• Overmatching: matching on a variable related to exposure but not a confounder can reduce efficiency or obscure a real association.

Matching can control selected confounders at design stage, but matched variables must be handled appropriately in analysis. A case-control design is generally unsuitable for estimating disease prevalence directly.

Cohort studies

A cohort study begins with exposure status and follows exposed and unexposed participants to observe outcome occurrence. Cohorts may be prospective, in which exposure is measured before future outcomes occur, or retrospective, in which existing records reconstruct exposure and follow-up.

Cohort studies establish temporal ordering more clearly than cross-sectional studies and can measure several outcomes arising from one exposure. They are particularly useful for common exposures, multiple outcomes and prognostic questions.

For exposed and unexposed groups, cumulative incidence can be compared using the risk ratio:

The risk difference is:

The risk difference is often more directly useful for planning because it describes the absolute excess or reduction in risk. A negative value indicates lower risk among the exposed or treated group.

When person-time differs, an incidence rate ratio may be used:

For a time-to-event analysis, the hazard ratio compares instantaneous event rates between groups under a specified statistical model. It is not identical to a risk ratio and should not be described as such.

Cohort studies may suffer from loss to follow-up, exposure misclassification, confounding and changes in exposure over time. Loss to follow-up is especially concerning when it differs by exposure and is related to outcome risk. Time-varying exposures and confounders require more advanced methods, such as marginal structural models, when ordinary adjustment is inadequate.

Randomized controlled trials

In a randomized controlled trial (RCT), eligible participants are allocated by a random process to intervention and comparator groups. Randomization aims to balance known and unknown prognostic factors in expectation, thereby strengthening causal inference. It does not guarantee balance in a small trial, and it does not correct poor follow-up, non-adherence or biased outcome measurement.

A well-designed RCT includes a clearly specified intervention, a meaningful comparator, allocation concealment, prespecified outcomes, appropriate follow-up and an analysis plan. Allocation concealment prevents recruiters from predicting the next assignment. Blinding limits differential behavior or assessment after assignment; participants, clinicians, outcome assessors and analysts may be blinded where feasible.

The intention-to-treat principle analyzes participants according to their assigned group, regardless of adherence. It estimates the effect of assignment to the intervention and preserves the main benefit of randomization. Per-protocol analyses may estimate effects among those adhering to treatment but are more vulnerable to post-randomization confounding.

For a binary outcome, the absolute risk reduction is:

When the intervention reduces risk and the value is positive, the number needed to treat is:

The absolute risk reduction must be expressed as a proportion, not as a percentage. For example, if risk falls from 0.20 to 0.15, the absolute risk reduction is 0.05 and the number needed to treat is 20, over the stated follow-up period. The number needed to treat depends on baseline risk and time horizon.

RCTs have high internal validity when implemented well, but explanatory trials may have narrow eligibility criteria, intensive monitoring and unusually high adherence. They may therefore have limited generalizability to routine services. Ethical requirements include informed consent, equipoise, favorable risk–benefit balance, independent review, privacy protection, safety monitoring and fair participant selection.

Pragmatic trials

A pragmatic trial asks whether an intervention works under usual health-care conditions. It may recruit a broad patient population, use routine staff and facilities, compare realistic alternatives and measure outcomes important to patients, providers and payers. An explanatory trial, by contrast, asks whether an intervention can work under ideal conditions.

Pragmatic trials may still be randomized. Their distinctive feature is the context and purpose, not the absence of randomization. The PRECIS-2 framework can be used to consider how pragmatic or explanatory a trial is across domains such as eligibility, recruitment, setting, organization, flexibility in delivery, flexibility in adherence, follow-up, primary outcome and analysis.

For managers, pragmatic trials can provide evidence about implementation burden, equity, cost, workflow disruption and real-world effectiveness. Their limitations include contamination between groups, variation in intervention delivery, missing routine data and difficulty separating intervention effects from contextual influences.

Quasi-experimental designs

Quasi-experimental designs evaluate interventions without random allocation. They are especially important when a law, financing reform, staffing policy or service redesign is introduced for operational or ethical reasons and randomization is not possible.

Before-and-after studies

A simple before-and-after comparison measures an outcome before and after an intervention. It is easy to conduct but weak for causal inference because changes may reflect secular trends, seasonality, regression to the mean, unrelated events or altered measurement.

Controlled before-and-after studies

Adding a comparison site or group improves interpretation. The comparison should be similar and should not be exposed to the intervention. Nevertheless, differences in underlying trends can remain.

Interrupted time series

An interrupted time series uses multiple observations before and after an intervention. It can estimate a change in level immediately after implementation and a change in trend over time. Segmented regression is commonly used:

Here, Y_t is the outcome at time t; β₀ is the baseline intercept; β₁ is the pre-intervention slope; β₂ estimates the immediate level change; β₃ estimates the change in slope after intervention; Time_t is time; Intervention_t is coded 0 before and 1 after intervention; PostTime_t counts time after intervention; and ε_t is the error term. The model assumes that the functional form is appropriate and that serial correlation, seasonality and other concurrent changes are addressed. A sufficiently long and dense series is needed.

Difference-in-differences

Difference-in-differences compares the change over time in an intervention group with the change in a comparison group:

Its key assumption is parallel trends: without the intervention, the groups would have followed similar average trends. Pre-intervention data should be examined to assess the plausibility of this assumption.

Regression discontinuity

A regression discontinuity design assigns an intervention according to whether a continuous score is above or below a threshold, such as age, clinical risk or income. Participants just above and below the cutoff may be comparable. The estimate is local to the threshold, and manipulation of the assignment variable or incorrect modeling can threaten validity.

Instrumental variables

An instrumental variable affects receipt of the intervention but influences the outcome only through that intervention, apart from random variation. The assumptions are demanding and often difficult to verify. The resulting estimate may apply only to compliers—people whose treatment status is changed by the instrument.

Comparing validity, feasibility and usefulness

Internal validity concerns whether the observed association or effect is credible for the study population. External validity concerns whether findings apply to other populations, settings and times. A study can have strong internal validity but limited generalizability, or broad coverage but serious bias.

A manager should consider the value of information, not simply methodological prestige. A large RCT may be unnecessary for estimating current waiting-time patterns, while a cross-sectional survey is inadequate for deciding whether a new clinical pathway caused a reduction in mortality. Mixed-methods evaluation may be appropriate when quantitative effect estimates must be combined with qualitative evidence about acceptability, implementation and equity.

A worked selection example

Suppose a health system wants to know whether introducing pharmacist-led medication review reduces 90-day hospital readmissions among older adults.

The intervention is implementable at clinic level, but individual randomization may be difficult because pharmacists and clinicians could treat patients in the same clinic differently. The manager could use:

1. A pragmatic cluster randomized trial, with clinics randomized to immediate implementation or delayed implementation.
2. A stepped-wedge cluster trial if all clinics must eventually receive the intervention.
3. A controlled interrupted time series if randomization is impossible but reliable monthly readmission data exist.
4. A simple before-and-after study only as a preliminary monitoring exercise.

The primary outcome should be defined before analysis, such as readmission within 90 days. Baseline risk, patient mix, competing interventions, medication adherence and implementation fidelity should be measured. If the intervention is assigned by clinic, analysis must account for clustering. A statistically significant result is not automatically managerial success: absolute risk reduction, staffing cost, workload, equity and sustainability also matter.

Illustrative case: evaluating a hospital discharge reform

Background. An urban hospital introduces a standardized discharge checklist and follow-up telephone call for adults discharged after heart-failure admission. Hospital leaders want to know whether the reform reduces 30-day readmission.

Problem. A comparison of readmission rates in the month before and month after implementation shows a reduction from 18% to 15%. Leaders are tempted to attribute the entire change to the reform.

Analysis. The simple before-and-after result is vulnerable to seasonal variation, changes in bed availability, altered coding, patient case mix and regression to the mean. The evaluation team obtains 24 monthly observations before and 12 after implementation, identifies a comparable hospital without the reform, and examines patient-level risk factors and process fidelity.

Approach. An interrupted time series with a comparison series is used. The team estimates immediate level and slope changes, checks autocorrelation and seasonality, and tests whether the comparison hospital experienced a similar change. It also conducts interviews with staff to determine whether the checklist was actually used and whether follow-up calls reached high-risk patients.

Outcome. The adjusted analysis suggests a modest immediate reduction, but no clear change in long-term trend. The effect is larger among patients reached within seven days, while implementation is inconsistent on weekends. The hospital therefore retains the intervention, redesigns weekend coverage and continues monitoring rather than claiming a universal causal effect.

Learning points. The case illustrates the difference between an observed change and a causal effect, the usefulness of interrupted time series methods, the importance of implementation fidelity and the managerial value of combining outcome, process and qualitative data.

Ethics, governance and data quality

Study design does not remove ethical responsibilities. Researchers and managers should establish scientific validity, proportionality of risk, informed consent or a justified waiver, confidentiality, secure data management and fair inclusion. Vulnerable populations require additional safeguards, but exclusion should not be used automatically because it can worsen inequity in evidence.

Routine administrative data may contain identifiable information, coding errors, missing values and changes in definitions over time. Data linkage requires governance approval, minimum necessary access, encryption and transparent retention policies. Artificial intelligence may assist with case finding, record linkage, risk prediction and automated extraction, but it can reproduce historical inequities, obscure classification errors and create automation bias. Human review, validation against a reference standard, documentation of model versions and monitoring for differential performance are essential.

Researchers should prespecify hypotheses and analysis plans where possible, report denominators and missing data, distinguish exploratory from confirmatory analyses, and avoid selective reporting. Reporting guidelines such as STROBE for observational studies, CONSORT for randomized trials and TREND for non-randomized evaluations can improve transparency. Registration and sharing of protocols, code and suitable metadata strengthen reproducibility while protecting confidentiality.

Best-practice decision process

A defensible design decision can follow these steps:

1. Define the population, exposure or intervention, comparator, outcome and time horizon.
2. Decide whether the purpose is description, association, prediction, causal effect or implementation evaluation.
3. Identify the unit of assignment and unit of analysis.
4. Assess whether randomization is ethical and operationally possible.
5. Select outcomes that are valid, important and measurable with acceptable completeness.
6. Anticipate confounding, selection bias, information bias, contamination and loss to follow-up.
7. Select design-specific analytic methods and prespecify primary analyses.
8. Assess feasibility, cost, workforce, data access and time.
9. Consider distributional effects across sex, age, socioeconomic position, geography and other relevant groups.
10. Interpret statistical findings alongside clinical importance, costs, harms, equity and sustainability.

Chapter Summary

Epidemiological study designs are tools matched to questions. Descriptive and ecological studies map population patterns; cross-sectional surveys estimate prevalence; case-control studies efficiently investigate rare outcomes; cohort studies establish exposure–outcome sequence and measure incidence; randomized trials provide strong causal evidence under allocation by chance; pragmatic trials examine effectiveness in routine practice; and quasi-experimental designs evaluate policies when randomization is unavailable.

Every design has assumptions and vulnerabilities. Internal validity depends on appropriate selection, accurate measurement, control of confounding, adequate follow-up and design-consistent analysis. External validity depends on similarity between the study and the intended decision context. For health-care managers, absolute effects, costs, implementation fidelity, equity and sustainability are as important as statistical significance. Ethical governance, transparent reporting and responsible use of routine data and AI are integral to credible epidemiology.

Key Terms

Analytic epidemiology: Epidemiology that evaluates associations or causal relationships.

Allocation concealment: Protection against predicting the next treatment assignment in a trial.

Case definition: Explicit criteria used to classify an individual as a case.

Case-control study: Study selecting participants by outcome status and assessing prior exposure.

Cohort study: Study classifying participants by exposure and observing subsequent outcomes.

Confounding: Distortion of an exposure–outcome relationship by a third variable.

Cross-sectional study: Study measuring exposure and outcome at approximately one point in time.

Ecological fallacy: Incorrectly applying group-level associations to individuals.

External validity: Generalizability of findings beyond the study population or setting.

Incidence: Occurrence of new cases in a population at risk over time.

Internal validity: Credibility of the estimated relationship within the study population.

Interrupted time series: Repeated outcome measurement before and after an intervention.

Odds ratio: Ratio comparing odds of exposure among cases with odds among controls.

Pragmatic trial: Trial designed to estimate effects under usual practice conditions.

Prevalence: Proportion of a population with an existing condition at a specified time.

Quasi-experimental design: Evaluation design without random allocation but with a structured counterfactual strategy.

Randomized controlled trial: Experimental study assigning participants by a random process to intervention groups.

Selection bias: Systematic error caused by how participants enter or remain in a study.

Temporality: Requirement that a cause precede its effect.

Validity: Degree to which a study supports a correct inference for its intended purpose.

Review Questions

Short Answer Questions

1. Distinguish prevalence from incidence.
2. What is the ecological fallacy?
3. Why are case-control studies efficient for rare diseases?
4. Explain the purpose of allocation concealment.
5. What is meant by a pragmatic trial?

Long Answer Questions

1. Compare cross-sectional, case-control and cohort studies in terms of sampling, temporality, measures of association and major biases.
2. Explain how randomization, blinding, intention-to-treat analysis and allocation concealment contribute to the validity of an RCT.
3. Discuss the strengths and limitations of interrupted time series and difference-in-differences designs.
4. Develop a design-selection framework for evaluating a new hospital quality-improvement program.
5. Explain why statistical significance alone is insufficient for managerial decision-making in health care.

Analytical / Application Questions

1. A district reports lower maternal mortality after increasing facility deliveries. Propose an evaluation design and identify at least four threats to causal inference.
2. A hospital observes that readmissions fell after a discharge intervention, but the intervention began during a period of reduced admissions. Which design and additional data would best assess the effect?
3. A researcher wants to study whether a common occupational exposure is associated with a rare cancer. Recommend a design, specify the principal measure of association and explain how controls should be selected.

Multiple Choice Questions with Answers

Question 1

Which design is most directly suited to estimating the prevalence of hypertension in a population at a specified time?

A. Cross-sectional survey

B. Case-control study

C. Prospective cohort study

D. Randomized controlled trial

Correct Answer: A. Cross-sectional survey

Question 2

In a case-control study, the usual measure of association is the:

A. Risk difference

B. Odds ratio

C. Incidence rate

D. Prevalence

Correct Answer: B. Odds ratio

Question 3

Which feature most directly strengthens causal inference in a randomized controlled trial?

A. Large sample size alone

B. Random allocation to study groups

C. Use of historical controls

D. Measuring exposure after the outcome

Correct Answer: B. Random allocation to study groups

Question 4

A study compares district-level air pollution with district-level mortality. What is the principal methodological concern?

A. Ecological fallacy

B. Allocation concealment

C. Lead-time bias only

D. Blinding failure caused by randomization

Correct Answer: A. Ecological fallacy

Question 5

If risk decreases from 0.20 in a control group to 0.15 in an intervention group, the number needed to treat is:

A. 5

B. 10

C. 20

D. 50

Correct Answer: C. 20

Question 6

The key assumption of a standard difference-in-differences analysis is that, without the intervention:

A. Both groups would have identical outcomes at every time point

B. The groups would have followed parallel average trends

C. All participants would have been randomized

D. The outcome would have been rare

Correct Answer: B. The groups would have followed parallel average trends

Question 7

Which design is especially efficient for investigating a rare disease with a long latency period?

A. Case-control study

B. Cross-sectional survey

C. Ecological census

D. Simple prevalence study

Correct Answer: A. Case-control study

Question 8

In an interrupted time-series model, the coefficient for the immediate intervention indicator generally estimates:

A. The baseline population size

B. The immediate level change after intervention

C. The pre-intervention slope

D. The number of participants lost to follow-up

Correct Answer: B. The immediate level change after intervention

Question 9

A pragmatic trial primarily aims to determine whether an intervention:

A. Works only under ideal laboratory conditions

B. Produces effects in routine practice

C. Has no possible adverse effects

D. Can be studied without defining an outcome

Correct Answer: B. Produces effects in routine practice

Question 10

Which principle analyzes trial participants according to their originally assigned group?

A. Per-protocol analysis

B. Intention-to-treat analysis

C. Ecological analysis

D. Post hoc subgroup analysis

Correct Answer: B. Intention-to-treat analysis