Epidemiology is a fundamental discipline in medical research, as it helps elucidate the distribution and determinants of diseases in populations. In pursuit of this goal, different methodological designs have been developed to facilitate data collection, analysis and interpretation, each with its own advantages, limitations and specific applications.

Epidemiological study designs can be classified into two broad groups: observational studies and experimental studies. The former include descriptive studies, such as case series and cross-sectional studies, which provide information on the frequency and distribution of a disease in a population. They also include analytical studies, such as cohort studies and case-control studies, which allow for the evaluation of possible associations between exposure variables and health effects. Finally, experimental studies, particularly randomized controlled trials, are the most robust tool for establishing causal relationships through direct intervention in the exposure of participants (Table 1).1–3

The selection of the most appropriate design depends on multiple factors, such as the research question, the feasibility of the study and the available resources. This article analyzes the characteristics, applications and limitations of the main epidemiological designs with the aim of providing guidance for their selection and correct implementation in research practice.

Observational studies are those in which researchers act as mere observers, without exerting any influence on exposure. They are divided into descriptive and analytical studies. Descriptive studies provide data collected from a sample of a population, and they include case series and cross-sectional studies. Analytical studies, which have a comparison group, include cohort studies and case-control studies.4

Randomized controlled trials (RCTs) are the prototype of experimental design and their ultimate goal is to evaluate the effects of an intervention in order to establish causality, which cannot be established through observational studies. Fig. 1 summarizes the design of an RCT.

Figure 1.

Flow diagram of the design of a basic clinical trial and potential sources of the most common biases.

They are forward-directed and prospective.

Six key aspects need to be taken into account to guarantee the internal validity of a RCT9:

• -

  Initial clinical question.10 All RCTs must seek to answer a structured clinical question that defines the population, intervention, comparison and outcome.

• -

  Randomization. Participants must be assigned to one of the groups (intervention or control) so that all subjects have the same chance of being assigned to either group.

• -

  Random allocation sequence. It must be concealed so that it cannot interfere with results due to conscious or unconscious behavior of researchers or participants.

• -

  Masking. The compared interventions must be masked so that neither the participants nor the researchers know what each study subject is receiving (blinding). This reduces information bias and interference in the interpretation of results.

• -

  Follow-up. This should be equal and completed by all groups, avoiding excessive losses.

• -

  Analysis of results.11 Whenever possible, subjects should be analyzed according to the group to which they were initially assigned (intention-to-treat analysis) as opposed to the group in which they eventually participated (per-protocol analysis), as the benefits of randomization are lost otherwise.

In RCTs, the strict control exerted in the study design yields data of the highest quality, but it also reduces external validity by limiting the generalizability of the results to real-world settings.

Quasi-experimental studies are controlled trials in which random allocation is not possible. This includes the following circumstances: natural experiments (the intervention occurs naturally or circumstantially), studies with historical controls (the intervention group is compared with another group that was treated with a similar intervention in the past), post-intervention studies (the observation is made after the intervention measure has been carried out), and pre-post studies (certain variables are measured before and after a specific intervention). These studies are simpler and more economical, but they are susceptible to selection bias and confounding, which compromise their external validity and the applicability of their findings, as well as the placebo effect and the Hawthorne effect (improvement occurs simply due to the awareness of being observed).

There are different designs within the RCT category:

• -

  Factorial trial.12 Evaluation of more than one intervention in a single trial (always with independent mechanisms of action and effects).

• -

  Cross-over trial.13 Participants receive the same two or more treatments in a different order (with a “rest” period in between).

• -

  Sequential trial.14 The sample size is not fixed in advance. Subjects are included until a significant difference is detected, establishing the follow-up period and conducting interim analyses.

• -

  Pragmatic trial.15 Carried out under conditions similar to those of clinical practice. Used to assess the overall effect of an intervention in the real world.

• -

  Non-inferiority study.16 The purpose is to demonstrate that a new treatment is not inferior to the standard of care.

In RCTs, potential biases that may confound the interpretation of the results need to be taken into account (Table 2).

Systematic reviews, by combining multiple studies, seek to obtain answers that cannot be provided by individual studies in order to draw firm conclusions about a specific health problem.17

The steps in performing a systematic review are18: framing a research question, searching for and selecting studies, evaluating the quality of the evidence and risk of bias of selected studies, extracting and analyzing the data and presenting the results. The last step can be qualitative or quantitative (meta-analysis).

In framing the research question, it is very important to structure the question adequately following the PICO framework (population or problem, intervention, comparison and outcome). The search strategy should ensure that all relevant and important publications are reviewed and should be described in detail. The selection of studies that will ultimately be included should also be detailed.

To assess the quality of the evidence, a risk of bias evaluation must be performed for each study using validated tools, and the tools used for the purpose reported. This analysis is very important because lower-quality studies yield weaker evidence.

A systematic approach must be applied to the retrieval of data from selected studies and described in the report, as it may have a significant impact on the final conclusions.

When the quality and consistency of the studies allow it, the results of a systematic review can be supplemented with a meta-analysis. The meta-analysis is a set of statistical procedures whose objective is to provide a quantitative summary of the results of a systematic review, obtaining an overall measure of the effect under consideration.

Combining the results of several individual studies yields a larger sample size, which increases the accuracy of the estimate and the power of the review, allowing the detection of differences that went unnoticed in individual studies and drawing of conclusions when the results of individual studies are contradictory, in addition to increasing the generalizability of the findings.

To carry out a meta-analysis, researchers must obtain all the available information. While obtaining exhaustive information is a key objective, it is important to remember that this is not always achieved (Table 2) due to either flaws in the search or selection of sources (information bias) or the selective publication of some findings while leaving other results unpublished (publication bias).19 Researchers also need to assess whether the primary studies are sufficiently homogeneous to be “mixed” in a statistical synthesis (analysis of heterogeneity), select the most appropriate mathematical method to combine the results20 and present and interpret the results of the synthesis correctly, frequently displayed graphically as forest plots (Fig. 2).

Figure 2.

Interpretation of the visual representation of a meta-analysis (forest plot).

The forest plot shows which variables are compared and helps identify the outcome effect measure. A diamond shape in the diagram illustrates, through its position and width, the overall result of the meta-analysis.21 The heterogeneity of the included studies can be assessed at a glance by noting the degree of dispersion or overlap of the boxes representing the primary studies, although ideally a quantitative assessment of heterogeneity with appropriate measures, such as the Cochran Q or the I2, is also performed. Another aspect to consider is where the overall effect diamond and its confidence interval are located in relation to the line of no effect. As long as they do not cross this line, the result is statistically significant, although this does not mean that it is clinically significant, an aspect that is determined based on the relevance of the variable itself.

All of the above refers to meta-analyses of clinical trials, which is the type most frequently encountered in the literature, although there are other types of meta-analysis, such as diagnostic test accuracy or network meta-analyses, which have some differential characteristics.

A diagnostic test (DT) is any procedure that can provide information to guide clinical decision-making. The capacity of a test to provide a correct diagnosis is determined by assessing its validity and reliability.1,22,23

Validity refers to the extent to which a DT measures what it is intended to measure and corresponds to diagnostic accuracy. It is assessed by comparing the results obtained with the test to those obtained with a reference standard (RS), yielding the main indicators of validity (Table 3): sensitivity (Sen), specificity (Spe), likelihood ratios (LRs) or predictive values. In the case of DTs for which different cut-off points can be applied in interpreting the results, these are expressed in the form of a ROC curve (diagnostic performance curve) that reflects the Sen and Spe obtained with different cut-off points as well as the overall performance of the test. The validity of DTs is determined through studies conducted on population samples, so these estimates are subject to random variability (they should be reported with confidence intervals) and susceptible to bias if the methodological design is flawed (Table 4).24

Reliability refers to the capacity of a DT to produce comparable results when used under identical conditions. The variability of the measurements is affected by multiple factors that should be known and controlled. Reliability can be assessed, depending on the nature of the variable measured by the DT, using the Kappa coefficient, the weighted Kappa coefficient or the intraclass correlation coefficient (for tests with qualitative results) and Bland-Altman analysis (for tests with quantitative results).

Diagnostic tests accuracy can be assessed with any type of epidemiological study design. There is considerable heterogeneity in the design of DT accuracy studies, which hinders the interpretation and application of their results. The aspects that need to be assessed when analyzing diagnostic accuracy studies are:

• -

  Timing of data collection: depending on the time that the DT and RS are received, they can be simultaneous (both at the same time), prospective (DT first) and retrospective (in patients already diagnosed by means of the RS).

• -

  Single or multiple sets of candidate eligibility criteria.

• -

  Participant sampling method (consecutive, random, convenience/availability, etc).

• -

  Blinding in performance/interpretation of DT.

• -

  Number of compared reference standards.

• -

  Limited verification design (yes/no).

Based on these aspects, DT accuracy studies could be categorized as follows25:

• -

  Fully paired design: each participant receives all DTs. Offers the greatest statistical power.

• -

  Partially paired design with random or nonrandom subset: some participants receive multiple DTs but others receive only one of the DTs. Higher risk of bias. An attractive option when a DT is expensive, invasive…

• -

  Unpaired randomized design: researchers assign each participant to a DT or RS.

• -

  Unpaired nonrandomized design: extremely high risk of bias.

The Standard for Reporting of Diagnostic Accuracy (STARD) statement26,27 contains a checklist of items that should be reported in any diagnostic accuracy study. There are also tools designed to assess the quality of diagnostic accuracy studies to be included in systematic reviews, such as the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2).28,29

Having discussed the different possible designs for epidemiological studies, we now broach the subject of selecting the most suitable design for a specific research project. In addition to the research question and the objectives of the study, the choice will depend on several other factors, such as the possibility of controlling the intervention of interest, the need to compare different groups and the sampling criteria used to select participants.

When the objective is to assess the effect of an exposure on a specific outcome, the most appropriate approach is to introduce the exposure in a controlled manner. In this case, interventional study designs are most appropriate, among which the randomized controlled trial is the highest standard, as it can be used to establish the causal relationship between exposure and effect. However, in situations where it is not possible to randomize allocation or have a control group, uncontrolled trials or quasi-experimental studies are suitable options.

If researchers cannot control the intervention due to ethical or practical limitations, an observational design is the alternative, in which researchers simply collect data but do not modify the exposure of participants. Within the subset of observational designs, if comparison between groups is not required, the appropriate choice is a descriptive study. A characteristic example is the use of a cross-sectional design, in which a sample is analyzed at a given point in time and who has the advantages of being simple and inexpensive. However, these studies have the drawback that they cannot establish causality. There are other types of descriptive studies, such as case series and case reports, with a lower quality of evidence.

On the other hand, if groups are being compared, the appropriate design is an observational analytical study, which allows the assessment of associations between variables. In this case, the way in which participants are selected determines the type of study to be conducted. If participants are selected based on their exposure to a given factor, a cohort study is conducted, which allows assessment of outcomes over time and yields evidence of greater quality. However, if participants are selected based on the presence or absence of an outcome, such as a disease, a case-control study is performed, a design that is more efficient when the outcome is infrequent, although it is more susceptible to biases than the cohort study.

Another aspect to take into account is the unit of analysis. Research is usually conducted at the level of individual subjects, but, in some instances, there is an interest in analyzing population groups. In the latter case, ecological studies are an option if what is intended is a descriptive analysis, and community trials if the research is interventional in nature.

In conclusion, the selection of the methodological design should be based on the possibility of intervention, the need to compare groups and the sampling criteria (Fig. 3). Randomized controlled trials offer the highest quality of evidence, followed by cohort studies and case-control studies. On the other hand, descriptive studies, while useful for certain purposes, yield less robust evidence.

Figure 3.

Flow diagram of the selection of the most appropriate methodological design.

This research did not receive any external funding.