Kant argues that we need some kind of criteria to distinguish between pure and empirical knowledge (Kant, 1787, p. 43). To explore the relationship between the two forms of knowledge, one based on pure a priori reasoning and one based on observation, Kant develops an argument about what he calls judgements or rules (Korner, 1955, p. 104), meaning the way we interpret the objects of our perception. Judgements can be analytic or synthetic (Korner, 1955). An analytic judgement is one in which the meaning of the predicate is included in the meaning of the subject (Kant, 1787, pp. 45–49; Ward, 2006, p. 16). A simple example of this is ‘all bachelors are men’. The same meaning is intrinsic in each term – men and bachelor – because the definition of bachelor means an unmarried man. Analytic judgements like this are a priori, and to deny them is self-contradictory (Korner, 1955; Ward, 2006, p. 18).
On the other hand, judgements may be synthetic. In this case, the meaning of the predicate is not contained in the subject (Ward, 2006, p. 16). Two things that are not intrinsic to each other are brought together (Korner, 1955). An example of this would be ‘the book is blue’. We are bringing together two ideas, book and blue. They are not intrinsic to each other because books can be of any colour and the colour blue could apply to many different objects (Guyer, 2006). The blue book on my table is one specific object located in space and time, observed empirically. The book is blue is a synthetic judgement because the idea blue is not intrinsic to the idea book (Guyer, 2006, p. 47). Synthetic judgements are a posteriori because they are established by recourse to experience (Ward, 2006, p. 18), in this instance of books and of colours. Simple a priori knowledge is analytic, and empirical a posteriori knowledge is synthetic.
However, Kant suggested that it is more complex than this because some a priori judgements are synthetic not analytic (Kant, 1787, p. 51). For Kant, these synthetic a priori judgements are the fundamental judgements or precepts of geometry, mathematics, natural science and metaphysics, that is, they are more than just contained in the subject and predicate and cannot be determined on the basis of experience alone (Guyer, 2006, p. 47; Ward, 2006, p. 19). For Kant, the most important ideas of this type relate to space and time (Korner, 1955). A priori rationalism provides the conceptual architecture in the form of synthetic a priori judgements and is the basis for empirical science.
The apparently simple distinction made by Hume between rationalism and empiricism therefore turns out for Kant to be more complex. For Kant, there were three elements involved in knowing and understanding and used in the process of interpreting sense data or scientific evidence – to recap – (i) analytic a priori judgements, (ii) synthetic a posteriori judgements and (iii) synthetic a priori judgements. These judgements or rules are quite helpful for considering the processes and practices of EBM.Footnote 3
Therefore, following Kant's argument, first there are the analytic a priori judgements. These are a very significant part of the EBM armoury. They are the logical and methodological givens. They are self-referential in the sense that by definition they are true and to deny them would be self-contradictory. The most important analytic a priori propositions in EBM are those that give rise to the hierarchy of evidence. The hierarchy is the operationalisation of the a priori analytic principle, and fundamentally rationalist idea, that there is a true and real relationship between phenomena and that extraneous or confounding factors mask that true relationship. By reducing bias, we can get closer to that real relationship. The problems described in the previous sections of this article relating to the biases that arise as a consequence of the process of observation, or the difficulties surrounding uncertainty or the fallibility of the human observer, are thought of, in this view, as masking our ability to see the true nature of things. The methodological task following from this is to try to limit the impact of such things in order to reduce uncertainty and bias and to see things as they really are. The concept of efficacy is premised on this precept.
The idea that there are true and real relationships between things, usually defined as the independent and dependent variable, is the central a priori analytic judgement in EBM. A number of other analytic a priori tools follow from this principle, including the judgements that: RCTs control extraneous or confounding factors in an assessment of an intervention; confidence intervals help distinguish between true effects and chance ones; and that summing results in meta-analyses produces a truer result than a single observation. The hierarchy of evidence is not about the possibility of the elimination of bias, but about the actual or real elimination of bias to reveal a pure relationship uncluttered by other things.
Internal validity is intrinsic to the hierarchy. The precept of internal validity is that it is possible to be certain that the action of the independent variable is the reaction in the dependent variable, that the measure of the reaction is true and that if repeated under the same conditions it will produce the same degree of change in the dependent variable. This is an a priori position because empirically and after the fact, this can never be demonstrated; it is an ideal position. And even empirically, it can only be demonstrated by controlling for all possible confounders, something that is never attained in real life. Moreover, in medical practice, the way disease presents is not as a single entity but frequently as multiple morbidities, that is, the patient has more than one thing wrong with them. Therefore, the idea of simple causes and effects as implied by the focus on internal validity in EBM is seldom the reality of medical presentations. Nevertheless, methods that are strong on control and on internal validity sit at the top of the evidence hierarchy in EBM. The analytic a priori judgement that follows from this is that it is possible to distinguish between types of method on the basis of their ability to eliminate bias. This is an entirely rationalist position that owes little to empirical science because the idea of a real pure relationship between the independent and dependent variable is only possible in the realm of pure reason and is about the relationship between ideas. It is further based ontologically on the idea that single outcomes could have single causes and that other things being equal these could be found and measured precisely. All this reasoning operates in the world of ideas and pure reason, not empirical science. These principles are the basis for deductive reasoning in EBM.
What the hierarchy of evidence attempts to do is describe a science in which the elimination of bias is a real possibility and to privilege those methods that control the extraneous factors out of the equation. They are givens; they are true by reference to their meaning and to deny them would be self-contradictory. None of this is to deny that bias introduced by the act of observation or fallible human observers should not be minimised, or that science should not strive for accuracy and objectivity. In this regard, the hierarchy serves well. However, its fundamental principle is not about eliminating bias deriving from observation, but bias meaning eliminating distortions in the true relations between phenomena.
Kant's second set of judgements or rules are synthetic a posteriori judgements. These are judgements that are made on the basis of observation, can only be made after the fact and are in essence pure evidence, or more precisely the representations of the reality that the evidence describes, or as Kant would put it as they appear to be. Therefore, for example, the statement that comparative effectiveness of compound x over compound y in cases of disease z has n difference in effect size, is judged on the basis of its synthetic a posteriori quality. It involves three after-the-fact observations, which are brought together. The observation of what x does, the observation of what y does and the difference between them. This is the core synthetic a posteriori matter of fact at the heart of EBM, and indeed is often thought of as the very quintessence of EBM and HTA. These matters of fact can then be the basis of induction. The concept of effectiveness as against efficacy is synthetic and a posteriori.
Kant's third set of judgements are the synthetic a priori judgements. These are required because, as both Hume and Kant noted, induction is a process involving judgement and interpretation. Inductive reasoning needs more than matters of fact. For Kant, it needs synthetic a priori judgements. These are plentiful in EBM. Synthetic a priori judgements transcend the empiricist world of synthetic a posteriori evidence of effectiveness and the rationalist analytic a priori world of the hierarchy of evidence. The synthetic a priori judgements bridge the two domains and without them EBM would be impossible.
The first example of synthetic a priori approach to illustrate the point is medical diagnosis. Critical to the process are clinical judgement, disease labels and observation. Diagnostic categories or disease labels are essential to EBM and HTA, as a diagnostic category is normally the focus of the efforts to determine clinical and cost effectiveness. Simplistically, it might be imagined that diagnosis involves fitting the observed symptoms to the agreed disease taxonomy. Taxonomies are familiar enough from any medical textbook, based on a variety of ways of arranging medical knowledge. Attached to disease taxonomies are descriptions of the epidemiology, aetiology (if known), therapeutics and prognostics. Taxonomies change as medical knowledge advances; they are not fixed and immutable (Bell, 2010). The taxonomies are typifications that can only exist in an ideal rationalist sense. This is because the way that pathology manifests itself empirically in the human or in any animal or plant does not follow the strict limits of the disease as described in the taxonomy. The taxonomy is a social product that changes, not a fixed underlying reality.
In any event, there is always variation. In part, this variation is associated with the individual biological and genetic difference referred to above. It also arises because biological and genetic variations interact with different aetiological agents. Empirically, disease is therefore a spectrum of pathology and many patient presentations are a cluster of multiple pathologies and morbidities. These will more or less approximate to the taxonomic ideal (Bell, 2010). Therefore, taxonomies not only change through time, they always have to be used flexibly in diagnosis. Doctoring is not simply about fitting sets of observed symptoms to taxonomies. Clinical judgement involves being flexible with the taxonomies. If all that medicine involved was applying the categories described in the textbook, then anyone who could read and understand the textbook could be a doctor. Clinical training involves learning that the categories are the necessary, but not sufficient condition for diagnosis – clinical judgement is also involved. It is what is not in the textbook that is the basis of the practice of clinical medicine. Therefore, the very concept of the identification via diagnosis of a particular disease is synthetic and a priori as it rests, and is only possible, on the basis of empirical observation and clinical judgement.
The synthetic a priori idea of diagnosis involving clinical judgement is a very different type of judgement to applying the mechanics of the hierarchy of evidence or a statistical test. The two ways of thinking sit decidedly uncomfortably together in EBM. The certainties of the analytic a priori concepts of confidence intervals, hierarchies of evidence and the elimination of bias are a long way removed from the kinds of subjectivities involved in clinical judgement. Therefore, many of the tensions associated with doing EBM can be seen to revolve around the rationalist–empiricist divide and the contrast between the processes of inductive reasoning associated with clinical activity and the deductive reasoning associated with the mechanics and techniques of EBM and HTA (Barnett et al, 2009). Many of the critiques of EBM that arise precisely because of these two ways of thinking – the synthetic a priori clinical judgement versus the rule-driven certainties of EBM grounded in analytic a priori judgements – are different (Egger et al, 2001). Both sides interestingly appeal to the evidence – the synthetic a posteriori concepts – as a rhetorical device to justify their position (see Russell et al, 2008). Kant's eighteenth-century epistemology offers an explanation of this very modern problem.
The second key synthetic a priori idea is that of modelling, which is also at the heart of the EBM/HTA enterprise and an area where the principles are hotly contested. Economic modelling, which is such a central part of HTA, is founded on the juxtaposition of different concepts; say the amount of health improvement or quality of life gained as a consequence of the administration of such and such a degree of medical intervention. This is the synthetic element. The a priori element comes from the association that is known empirically to exist in general terms between these two elements and the ability to predict a priori that these elements will be conjoined in the future. Economic modelling is classically synthetic and a priori.
The third example of the synthetic a priori group of concepts is external validity (Campbell and Stanley, 1966). Unlike internal validity, which is an a priori analytic concept, built entirely out of rationalist principles, external validity is classically a priori and synthetic. External validity is conventionally defined as dealing with the question of whether the results obtained in one setting would apply in another (Campbell and Stanley, 1966). In EBM and HTA classically, it is about determining whether the findings of one trial are transferable or generalisable more broadly. Statistical generalisability is usually corralled in order to help reach the decision. Applied in EBM, it is about dealing with the question of whether the results of study A will help patient B, and can team C in hospital or primary care setting D implement it in such a way that it will work as in study A. In areas where the judgements involve a long causal chain from the intervention to the outcome as in public health or other social care or educational interventions (Kelly et al, 2010; Kelly and Moore, 2010), the problem of external validity is still more vexing (See also Pawson, 2002, 2006). In these cases, we are trying to determine whether the results of study A, which describes a rather loosely defined and often poorly specified intervention B, applied in manner C, in context D, by team E, in organisation F, to sub population G, will produce the same result as it did when it was originally done two decades ago in study A (Moore and Rutherford, 2012). These judgements are quintessentially and inevitably difficult. It is not going to be possible to derive the judgements from better more highly powered studies, covering all sub populations, nor to find all the details of the fidelity of the intervention or the process of implementation (Davidson et al, 2003). Therefore, we have to use synthetic a priori judgements if we are to work scientifically. In short, external validity is empirical evidence conjoined with theory and is about probabilistic statements in the face of real-world uncertainty; internal validity is about rationalist certainty in the world of ideas.