The three perhaps most common observational designs are:

• Cross-sectional studies
• Longitudinal studies
• Case-control studies

Cross-sectional studies

Cross-sectional studies are based on data collected at a single point in time. Thus, we measure the exposure and outcome simultaneously.

This kind of study is perfect for estimating prevalence and can be used to explore patterns and associations in the data – but it does not allow us to draw any causal inferences (i.e. we cannot determine that the exposure is actually causing the outcome).

Longitudinal studies

In longitudinal studies, we have data from at least two measurement points: the exposure is measured at what is commonly called baseline, and the outcome at a later time, usually referred to as a follow-up.

There are at least three sub types of longitudinal studies:

Trend study

Following the same population over time.

Example Examining the prevalence of cannabis use among 15-year-olds in Sweden between 2009 and 2019.

Panel study

Following the same cross-section of individuals over time.

Example Exploring the association between exposure to bullying among those who were aged 10-18 in 2009 and the risk of cannabis use ten years later (2019).

Cohort study

Following the same cohort of individuals over time.

Example Studying the association between peer relationships in adolescence and drug dependence in adulthood among everyone born in Sweden in 1970.

A note on the meaning of time

Since time is such a fundamental concept in longitudinal studies, we also want to say something about longitudinal data and how these can be analysed to make full use of the detail. For example, when the outcome can occur at different points across the follow-up, we can use time-to-event analysis such as Cox regression (see Cox regression). Often when our outcome is based on count data, we have collected information for a longer follow-up period; then we can use e.g. Poisson regression (see Poisson regression).

But we might also model the longitudinal data in many other alternative ways. For example, individual developmental outcome patterns over time (in terms of the outcome’s frequency, duration, complexity, and sequencing) can be captured with methods such as latent growth models, group-based trajectory modelling, and sequence analysis. These methods will be elsewhere the guide.

Case-control studies

Then we have the case-control studies. Here, one focuses on a group that has a certain outcome (the “cases”), and matches them to a similar group without the outcome (the “controls”). These two groups are subsequently compared with regard to different exposures. Information about the exposure can reflect the same time point (i.e. cross-sectional data) or at an earlier time point (i.e. longitudinal data).     

Retrospective vs prospective data

The terms retrospective and prospective are sometimes used a bit sloppy when it comes to observational studies. In terms of study design, these terms refer to what one would assess first: the exposure or the outcome.

• In retrospective studies, the outcome is first established, after which one looks backwards in time to examine exposures. Thus, this is what one usually (but not always) do for case-control studies.
• In prospective studies, the exposure is first established, after which one looks forward in time for the outcome to occur. This is typically what we do in longitudinal studies (more specifically panel and cohort studies).

It is nonetheless quite common that cohort studies are retrospective. This means that we define a cohort that has already experienced both the exposure(s) and the outcome(s) of interest, and we collect information on this through e.g. administrative records. 

To make things more confusing, it is quite common that all sorts of observational studies (e.g. cross-sectional studies, cohort studies, and case-control studies) include retrospective questions, such as survey questions about past events and experiences. This is, however, not exactly what is meant by retrospective designs.