When data generate populations

Epidemiology could help redress some ‘intrinsic weaknesses’ of artificial intelligence (AI) by giving more weight to causal inference thinking and by addressing algorithmic fairness, as rightly explained by Sung and Hopper in a recent issue of the International Journal of Epidemiology.1 AI and big data affect considerably the three tasks of epidemiology and data science: description, prediction and causal inference.2 If we want epidemiology to remain consequential, i.e. a science informing how to improve the health of populations,3,4 we argue here that we need to be clear about what is a population and, more specifically, about what are the study and target populations. Overlooking the definition and identification of these populations jeopardizes the external validity of study findings and, consequently, their transportability to target populations.5,6 A population is usually defined as a collection of individuals who share at least one common characteristic, e.g. living in a specific geographical area. Big data provide information from numerous individuals; the problem is that these individuals may not come from a well-defined population and may not be representative of the target population of interest.7 Major challenges for consequential epidemiology to get the full potential of big data are therefore first to identify and characterize the study population from which these data have emerged and, second, to assess how this study population relates to the target population.5,6 The study population is made of the individuals from whom data are collected to conduct the study.5 Nothing is new in the fact that populations are evolving over time due to changes in their composition and moving boundaries. However, for many types of big data, the scale and speed of these changes have increased dramatically. As a result, the populations generating these data are moving targets and difficult to characterize. For instance, although data from social media can offer new insights, e.g. on users' health behaviours, the users of a given social medium are not a fixed population and change rapidly, and one cannot assume the findings from a vaguely defined and ever-changing social media community to be easily transportable to specific target populations. Many analyses of big data are biased because researchers treat them as a census, like a complete and representative collection of the target population, whereas they are a ‘misrepresentative mixture of subpopulations’.7 At the extreme, researchers apprehend these data no longer as the products of identifiable populations, made of individuals with measurable characteristics; the study ‘populations’ are seen as generated by the data, they are eventually the data. The danger is to be blind to the fact that big data are often the product of complex selection processes, and do not emerge from a representative random sample of a well-defined population. These data must be redressed to make them informative about a targeted population. For example, the UK Biobank is a large cohort study with extensive high-quality health information, but not representative of the UK population, raising concerns about its external validity.8,9 The mistake is to think that ‘associations [found in this study] are generalizable to all possible target populations, or relevant to public health and clinical medicine, simply because the sample size is large’.8 Weighting methods are necessary to mitigate the effect of selection bias if one wants these study findings to be transportable to well-defined target populations.9 Epidemiology has tools to tackle the issues of study and target populations which are exacerbated by the rise of big data. First, one should carefully define the research question by specifying explicitly a descriptive or causal estimand and by defining a target population6,10; external validity, not only internal, should be considered a priori. Second, one must understand the selection mechanisms constraining the data available for the analyses, and how the study population relates to the target population. One way to measure the degree of transportability of study findings is by quantifying their target validity, i.e. by explicitly assessing the difference between descriptive or causal effect estimates in the study sample and in the target population.6 This type of analytical framework will help make big data useful to improve population health. A.C. drafted the manuscript which was substantially reviewed by C.C. Both authors agreed on the final version. Swiss National Science Foundation (SNSF) grant 188549. None None declared.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.