Evidence-based public health demands that study findings provide meaningful insights into improving the health of target populations, making representativeness a hallmark of high-quality epidemiological study [1, 2]. Representativeness is, however, neither necessary, nor sufficient to guarantee the external validity of study findings [1]. Further, big data and the digital health datademic—a concept related to infodemic to describe the overabundance of data [3] and the challenge of using them to produce actionable information—are changing the way target and study populations are defined and how to ensure the external validity of study findings [4]. While minimizing selection bias and ensuring similarity between a study sample and its target population help, they are not sufficient to guarantee external validity. What matters is assessing the degree of transportability of study findings, i.e., how well these findings inform about the target population—the population on which information is searched to assess its health status and disease burden, as well as their causes, and to guide health related decisions [5, 6]. In this education article, I review the gain of shifting the focus away from study representativeness and instead prioritizing the explicit assessment and reporting of transportability.

According to the Cambridge dictionary, representativeness is defined as “the fact of a smaller group of people or things representing a larger group accurately, so that the smaller group is typical of the larger one” [7]. Without a census, epidemiologists and other population health scientists typically consider a study representative if it is conducted on a random sample of the target population [8]. In this context, if the researchers are lucky enough with the sampling, the study’s descriptive or causal estimate—once internal validity biases due, e.g., to measurement errors and confounding are accounted for—provides insight into the true value of the descriptive or causal estimand in the target population, with a known degree of uncertainty. The notion of estimand is meant here to describe the target quantity, i.e., the goal of a descriptive or causal assessment in a target population, while the estimate is what is obtained from a given analysis in a specific study population [9].

However, while random sampling is one approach to achieving representativeness, it is neither necessary nor sufficient to ensure that study findings will apply to a target population [1]. Further, a growing number of epidemiological and population health studies rely on non-random samples that neither represent the reference population—the group from which study participants were drawn—nor the target population relevant for policymaking [4, 10]. This is particularly evident in studies using various types of real-world or big data, which include vast numbers of individuals who do not originate from a well-defined reference population, do not represent the target population of interest, and whose selection process (beyond self-selection) cannot be fully identified.

For instance, many studies using digital-trace data—from smartphones and wearables, search engines, social media, or streaming platforms—cannot identify the populations that generate these data (Box 1) [11]. Beyond basic sociodemographics, high-quality information on users’ characteristics is often lacking, limiting stratified analyses. Access is also constrained by the companies that hold these data. Digital data donation—users sharing the data held by these companies with researchers [12]—can reveal the data-generation process needed to assess transportability, but such donations must be widespread to yield meaningful, population-level insights.

Without a well-defined reference population, the population coverage remains unknown [7], leaving these real-world-, digital-, and big-data as little more than a shifting and misrepresentative mix of multiple, often changing, populations [4, 16]. Unstable or poorly documented study populations contributes to the poor reproducibility of epidemiological research findings [17]. Using this type of data, increasingly accessible in a digital health era, is changing radically the way we define study and target populations, and the way we shape an epidemiological research question (the estimand). Hence, as shown in the Figure 1, in a classical epidemiological study (Panel A), the question (estimand) is used to define the data collected, allowing the production of an estimate. In an era of big data and datademic (Panel B), the question is shaped by the available data, in an iterative process. Data are generated by the population, but they can also define the population [4], and multiple estimates are produced.

Of note, even when a study is conducted in a random sample of a well-defined target population, following a proper sampling frame and plan, declining participation rates increase selection bias and threaten representativeness [18]. For example, the impact of the low participating rate on the representativeness of the famous UK biobank is well known, but its effect on the transportability of findings might have been underappreciated (Box 2) [19, 20]. Ultimately, representativeness is like a benevolent spirit which is haunting studies—often longed for, sometimes desperately [2], yet increasingly elusive in the era of big data [8]. It is timely to reassess representativeness and external validity through the lens of transportability [5, 6, 21].

The major tasks of population health data science are description and causal inference [23], both of which are subject to internal and external validity biases, challenging transportability [6]. The role of representativeness differs, however, between these two tasks. In population health monitoring, where the goal is descriptive—such as estimating disease burden in a target population—representativeness is often necessary or at least expected. In contrast, for causal inference studies, which aim to estimate the effect of an exposure or intervention on an outcome, a representative random sample is generally not considered critical for the transportability to study findings [1].

The differences in representativeness requirements between descriptive population health monitoring and causal inference studies are, however, debatable. What truly matters is determining the extent to which study findings can be transported to the target population, particularly by accounting for the presence and distribution of modifiers of the estimand, being descriptive or causal [5, 6, 9]. Regardless of the data production process, a consequential epidemiology approach calls for explicitly reporting the transportability of study findings [24]. This is essential for health decision-makers [25], who need to assess whether observational or experimental study results are reliably extended to specific target populations, the ones that they are taking care of (Box 3) [26].

Of note, it is frequent to distinguish generalizability from transportability, with generalizability being used to refer to what extent study findings can be applied to the population from which data were collected (the sampled population, see Figure 1) and transportability to what extent findings inform about other potential target populations [8, 25]. Assessing generalizability can be, however, considered as a specific case of transportability assessment, where the target population is the population from which data were collected. In this educational paper, we use transportability to refer to which extend study findings inform about any prespecified target populations, including the population from which data were collected.

On the one hand, ideally, study findings are transportable when the quantitative estimates obtained from the study sample can be generalized to the target population, that is, when the unit-specific quantity of interest can be reliably aggregated over this population [9]. On the other hand, transportability is often based on vague qualitative findings that are only tentatively extended to the target population [8, 29]. While this distinction between these two types of transportability is common, it should not be seen as a fundamental difference. Rather, it reflects varying degrees of confidence in how much—or how far—we can extend descriptive or causal estimates to the target population.

Within that logic, it is useful to apprehend transportability over a continuum [5]. In some studies, transportability is high, meaning the estimate can be confidently applied quantitatively to the target population of interest for descriptive or causal statements. In others, the transportability is low, allowing merely a rough qualitative estimation in the target population. Low transportability reveals a high level of uncertainty and might underscore the need for further research to generate estimates that are more transportable [30].

Strengthening transportability requires accounting for at the study design stage and quantifying it during analysis. Whatever the study design, it is necessary to identify and measure modifiers of the descriptive or causal estimand [5, 9]. Transportability to the population from which data were collected can be strengthened through random sampling of this population. When random sampling is not feasible, alternative approaches such as purposive sampling—where individuals are selected based on expected heterogeneity—or stratified selection based on key modifiers can be used [5]. Randomly sampled study participants from a given population is, however, not enough to ensure transportability of study findings to other target populations [21]. Like with real-world data, information on modifiers of the descriptive or causal estimand will be necessary.

At the analytic stage, transportability is addressed through methods such as matching, outcome regression, standardization, or weighting, with the goal of adjusting the descriptive or causal estimates for different distributions of modifiers in the study sample and the target population [6, 16, 31–33]. Methods based on weighting are popular (see also Box 2) [19, 32, 33]. Calculating these weights requires a clear understanding of the selection and recruitment process, which can vary significantly, for example, between a study based on a random sample of a well-defined population and one relying on online survey participants without probability sampling (see Box 3) [3]. In the latter case, multiple assumptions are needed to approximate the population from which study participants originate, and these assumptions have to be explicitly formulated for estimating the degree of transportability to a target population.

The degree of transportability can be evaluated by assessing the similarity between the study and target populations, e.g., through a generalization score, the differences in standardized mean differences for modifiers, or a propensity score for selection if individual-level data are available for both study and target populations [5]. The key is to assess differences in the distribution of modifiers between the study and target populations and account for them in the analysis [19].