The need to extrapolate stems from a tension between the epistemic and pragmatic aims of science and the nature of the objects of investigation of the life sciences. Generalization is at the same time a central aim of science and of paramount importance for applying scientific knowledge in practical contexts. However, the objects of inquiry of biomedical research are heterogeneous in nature. There are theoretical reasons for assuming biological diversity, for instance because the theory of evolution by natural selection requires it or because of the stochastic nature of the molecular interactions underpinning biological activity. However, before being a theoretical requirement in need of further empirical and conceptual elucidation, biological diversity is a crude experimental and medical reality. Scientists are constantly reminded that, ultimately, there is only a multitude of individual biological systems. To give just a few examples: the overall pattern of gene expression measured by analyzing a large population of genetically identical cells may fail to describe the pattern of gene expression of any of the individual cells within that population; what happens in vitro does not always reflect what happens in vivo; a treatment developed in an animal model may not translate into a successful treatment of human disease; and two patients suffering from the same condition may respond very differently to the same treatment. As a result of this tension between the aims of science and the nature of biological systems, generalizing and applying knowledge takes the form of extrapolations extending a claim from the system in which it was actually documented to systems known or expected to be different in potentially relevant ways.
- Types of extrapolations
The first part of the project seeks to provide a systematic overview of the problem of extrapolation. This includes a classification of the distinct types of extrapolative inferences in scientific practice and an analysis of the strategies for evaluating the validity of each type of extrapolative inferences. Drawing on classifications proposed in our previous work we will investigate in more detail:
1a) Statistical extrapolations. These are sample-to-population extrapolations typical of epidemiological studies validated by statistical methods. In addition to reviewing previously raised issues concerning the interpretations of statistical data, randomization, blinding and uniform distribution of confounding variables across test and control groups, we are also interested in the extent to which statistical methods can be used to validate other types of extrapolations.
1b) Similarity-based extrapolations. These include individual-to-individual, species-to-species or cross-model extrapolations (e.g., from in vitro to in vivo, from cell models to animal models, from animal models to humans, etc.). Extrapolative inferences are common in all life sciences, basic and applied. A better and more systematic understanding of how researchers in various domains of investigation tackle the problem of extrapolation is expected to generate new strategies transferable across disciplines.
1c) Applicability to individuals. These include mixed extrapolations from populations to individuals, which are critical for clinical practice, yet remain highly problematic. We are particularly interesting in exploring the possibility of validating correlations identified via subgroup analysis by means of parallel interventions on animal models in order to demonstrate causation. This shift from correlation to causation is raising interesting questions about its assumptions and philosophical foundations, while remaining of outmost interest in bridging the gap between public healthcare and clinical practice by providing potential solutions to the problem of population-to-individual extrapolations.
- Complex extrapolations
Scientific reasoning may involve simultaneous extrapolations along several dimensions. We will investigate the extent to which complex extrapolations can be decomposed into chains of simple extrapolations, as well as the limitations of this approach before considering rival approaches involving holistic validation. To our knowledge, this question has never been addressed in the literature, yet it is of decisive methodological significance for scientific practice.
2a) The general form of extrapolation. Current philosophical and scientific debates tend to construe extrapolations in a simplified form, typically as a type of inference where the premise ‘in an animal model, X led to Y’ grants the prediction that ‘X will lead to Y in humans as well’, where X and Y are assumed identical in both statements. In practice, however, the extrapolated claim seldom remains unmodified when transferred from source to target. For instance, a mouse is never given the same doses of a drug that a human would be given, a surgeon operating in a clinical trial is not the average surgeon, the endpoints of a clinical trial are often not the same as the clinical outcomes sought after in routine healthcare. Our goal for this part of the project is to bring conceptual clarity on issues related to complex extrapolations by correctly identifying their relevant dimensions and by characterizing the most general form extrapolation statements can take.
2b) Methodological issues. The fact that some extrapolative inferences may in fact amount to simultaneous extrapolations across several dimensions raises an interesting methodological question. Is it possible–and if so, is it recommendable from a methodological standpoint–to decompose extrapolations into sets of distinct extrapolative claims that can be evaluated on an independent basis? In some cases, research methodology adopts a step-wise compounding of distinct extrapolative inferences. For instance, results obtained in laboratory models are first extrapolated to humans, then tested in clinical trials before being generalized to the general population and finally applied to specific patients. On the other hand, proponents of Evidence Based Medicine are particularly distrustful of indirect evidence, and openly advocate methods that allow for the direct testing of a claim. Arguably, if there is a small risk of error associated with each extrapolative inference, then compounding more and more inferences can only increase the overall risk of error, and a single weak link in the chain of indifference can lead to catastrophic results. In line with this argument, it has been proposed that clinical practice should be informed by pragmatic trials evaluating the effectiveness of interventions in routine practice conditions. However, a major drawback of pragmatic studies is that they promote external validity at the expense of internal validity by increasing the intra-group variability to a degree that can easily hamper the detection of significant effects, and thus requires larger sample sizes. Further investigation is required in order to correctly assess the strengths and weaknesses of these competing methodological approaches to complex extrapolations. We will want to confirm the general intuition that the internal and external validity of a study tend to be inversely related, and explore the potential for performing an optimization analysis that could address such conflicts in a more informed and beneficial manner. This is of great interest both philosophically and scientifically, and would provide a completely new angle on how extrapolations are validated.
Discussions surrounding the problem of extrapolation are spread out across many fields of investigation, yet lack of communication between disciplines led to a divide between basic and clinical research, as well as between research and medical practice. In particular, the conflation of distinct types of extrapolations generated considerable friction and confusion, as some authors came to recommend generic guidelines for validation without paying sufficient attention to differences in methodology and epistemic goals which require the use of specific types of extrapolations.
3a) The basic/clinical research interface. In the recent years, two distinct kinds of extrapolative inferences emerged. One is hinging on the statistical methods developed in epidemiology in the 1960s and subsequently popularized by the Evidence Based Medicine Movement in the 1990s. A parallel strategy, developed in basic experimental science, relies on phenotypic and mechanistic similarities between various models (e.g., animal, cell, in vitro) used to study a phenomenon of interest. Unfortunately, lack of understanding of the distinct challenges related to basic and clinical research led to a conflation of distinct types of extrapolations. Of particular interest are current debates surrounding the poor translatability of studies on animal models to humans. The hierarchy of evidence proposed by Evidence Based Medicine places a marked emphasis on the results of Randomized Controlled Trials, which specifically aim to validate sample-to-general-population extrapolations by means of statistical analysis, while demoting evidence from mechanistic reasoning. It has been proposed that basic science should adopt the methodological standards of epidemiology, for instance, by conducting randomized preclinical trials on larger and more heterogeneous samples of animal models. Yet the assumption that the extrapolative inferences underlying translation from animal models to humans follows a sample-to-population pattern of generalization amenable to statistical analysis is often times disputed in basic science, which views the differences between the animal model and humans more relevant than the internal heterogeneity of a population of tested subjects. Furthermore, the proposal represents a major encroachment on the research methodology in basic science, which relies heavily on standardization (e.g., use of clone cell lines, inbred mice, etc.) rather than randomization in order to reduce the potential effect of confounding variables, and makes use of mechanistic similarities in order to validate extrapolations. Our objective is to analyze the methodological foundations underlying the clash between the two styles of research and provide more rigorous guidelines needed for the particular use of any given validation method in a specific context.
3b) The research/medical practice interface. A second type of conflation occurs in medical practice, when it is assumed that aggregative statistical data describing a population can be converted into predictions about the probabilities of a given outcome in a particular patient. This kind of reasoning relies on the assumptions that the patient belongs to the relevant population, and that all members of the population respond equally well to the treatment, which is precisely the kind of homogeneity that cannot be automatically assumed when dealing with biological systems. Sample-to-population generalizations and population-to-individual applicability are two distinct types of extrapolations, that need to be evaluated independently. Further work needs to be done in order to better understand the relationship between the two types of extrapolation.