Abstract
Using an analysis of the British Medical Journal over the past 170 years, this article describes how changes in the idea of a population have informed new technologies of medical prediction. These approaches have largely replaced older ideas of clinical prognosis based on understanding the natural histories of the underlying pathologies. The 19th-century idea of a population, which provided a denominator for medical events such as births and deaths, was constrained in its predictive power by its method of enumerating individual bodies. During the 20th century, populations were increasingly constructed through inferential techniques based on patient groups and samples seen to possess variable characteristics. The emergence of these new virtual populations created the conditions for the emergence of predictive algorithms that are used to foretell our medical futures.
Today, anyone between the ages of 40 and 70 who wants to know their chance of dying during the next five years can answer a few simple questions and receive an estimate based on a predictive model derived from the mortality experience of half a million people recruited to the UK Biobank cohort (Ganna and Ingelsson, 2015). Two centuries ago, those curious about their lifespan might have turned to ‘Astrology the Art that teaches us the influences and effects of the Celestial Planets upon the bodies of men, women and children’ (Pater, 1795). So what changed between a time when one looked to the study of the movement of the planets and now, when one looks to predictive models of the 21st century such as the one derived from UK Biobank data? The answer lies in the idea of a population, or rather the changing idea of a population.
Method and context
The idea of a population seems to have held increasing significance over the last half-century. On the one hand, demographers have explored the origins of their discipline from the point when populousness – the extent to which a place was populated – gave way to the idea of population as an aggregation of individual bodies (Glass and Eversley, 1965). On the other hand, the ‘discovery’ of population in the 18th century has been used to explain modernity through mechanisms such as biopolitics and liberal modes of government (Foucault, 2001a, 2001b). Studies of the history of statistics and the development of number also, inevitably, engage with the concept of population (e.g. Daston, 1994; Desrosières, 1998; Hacking, 1990; Porter, 1995). In the early 21st century, however, the thesis that the population ‘existed’ and was waiting to be ‘discovered’ has been challenged (Curtis, 2002). Indeed, the very nature of a population, it seemed, could be seen as problematic (Krieger, 2012); Hinterberger (2012) concludes that ‘there is no single answer to the question what is a population? Indeed, there are many.’ (p. 75)
Here I attempt to explore this problem of ‘population’ genealogically. By investigating a single source, the British Medical Journal (BMJ), first published in 1844 (as the Provincial Medical Journal), I cut one slice through the changing meaning of the term. The BMJ provides a contemporary chronicle of medical thought and practice, a record of clinical perception: In articles, editorials, correspondence and notices, individual doctors documented what they could see – and by inference, what they could not see – at any particular time. To further ground the analysis, I explore the concept of population in the BMJ in the context of the perennial medical challenge of predicting the likely course of a patient’s illness. In effect, I try to provide an account of how the population concept was operationalized, by tracking its changing contours through the pages of a single journal.
A study of reports directed at medical practitioners over almost two centuries will necessarily be schematic and largely descriptive. I do not attempt to ‘explain’ changes in the idea of a population in terms of external events and context but instead describe the ‘internal logic’ of how medicine grappled with the concept through the problem of prediction. Besides, many of the forces – individual, political, economic and cultural – that might be used in any explanatory model might equally be construed as consequences rather than causes of the changes described below. For this reason, as well as for reasons of space, I ignore contemporary events and studies outside the main chronicle of the BMJ.
The BMJ archive is word- and phrase-searchable. The initial analysis consisted of counting the use of terms such as ‘population’, ‘prediction’ and ‘prognosis’ by decade and deriving graphs of changes over time. The actual texts using these terms were then examined to clarify patterns around important changes in use. This process enabled the identification of other related terms such as ‘diathesis’, ‘p-value’ and ‘probabilistic’, that were treated in the same way. An account was then constructed that explored the changing meaning of population in the context of clinical prediction. The BMJ database is large (over 170 years of a weekly journal) so illustrative quotes are used to support the narrative; only the date of each quotation is provided for this analysis and further details can be found by searching the BMJ using the given date and text (http://www.bmj.com/search/advanced). As this analysis focuses on commentaries intended for a general clinical readership, anonymized quotations have the advantage of being unencumbered by the known attributes and contexts of well-known authors. The public accessibility of the database provides an important degree of transparency for the analysis I present here.
Pathology and prognosis
During the 19th century, the new framework of pathological medicine transformed the character of prognosis by enabling prediction of the courses of the illnesses from known natural histories of the underlying disease pathologies. Once a clinician had identified a pathology, a prognosis depended not only on its known natural history but also on the extent to which it followed an unvarying course. Some diseases, such as cerebral apoplexy, showed an unpredictable course [1895], while regularity was a key characteristic of other diseases: ‘as the period characteristic of the particular case recurs, the attack is predicted with the greatest certainty, and never fails to appear at the right time – never misses, never anticipates, never postpones’ [1859]. At best, however, prognoses were expressed in very general terms; the outcome of the illness might be ‘favourable’ or ‘unfavourable’, ‘grave’, ‘guarded’ or ‘gloomy’, ‘cautious’ or ‘hopeful’ due to the apparent uniqueness of the individual patient.
Prognosis depended on the pathological life of the disease but also the age, constitution and habits of the patient. The art of medicine therefore relied less on identifying regularity and more on recognizing singular events, particularly in relation to the uniqueness of each individual patient. ‘The same disease in different persons presented often quite distinct features, and it was for the physician to recognise this, and not to rely upon stock formulas, but to treat the individual as a whole’ [1871]. Central to this notion of individuality was the patient’s constitution, which in its turn was related to temperament, idiosyncrasy, susceptibility and diatheses: ‘The realities, which these portentous terms … represent, are nothing less than a man’s constitution, or the individuality of his physical and vital nature’ [1862].
Clinicians recognized four broad temperaments – sanguine, phlegmatic, choleric and melancholic – said to derive from Hippocratic and humoral thought. These could be defined as ‘the sum of the physical peculiarities of an individual, exclusive of all definite tendencies to disease’. The latter seemed better described by the concept of a diathesis: ‘any condition of prolonged peculiarity of health giving proclivity to definite forms of disease’ [1907]. Clinicians also frequently used the term ‘idiosyncrasy’ to capture individual response to disease: ‘each case has its own idiosyncrasy, which must be taken into account’ [1849]. As every patient responded differently, the clinician, particularly when prescribing medicines, had to pay attention to ‘the state of the nervous system, and of the circulation, the constitution, general health, habits, residence, and possible idiosyncrasy of the patient’ [1868].
Whereas prognostic claims based on pathology and its natural history had some chance of identifying regularities, the individualistic characteristics of every patient – constitution, diatheses and idiosyncrasies – complicated the predictive process. ‘We cannot even approach to certainty in attempting to determine what place a given individual shall hold in his class – whether he in particular is to live or to die’ [1862]. Yet at the same time clinical medicine was realizing its predictive limitations, the potential for using knowledge about populations to foretell death was beginning to emerge. The new figures on population numbers arising from the British decennial Census (first conducted in 1801) provided a denominator through which to frame, standardize and stabilize the apparent randomness of life and death.
Numbers of birth and death events taken from parish records, or, after mid-century, national registers, allowed the calculation of birth and death rates. In turn, these rates enabled the derivation of life or mortality tables that showed the life expectancy at any age: ‘It would formerly have been considered a rash prediction in a matter so uncertain as human life to pretend to assert that 9,000 of the children born in 1841 would be alive in 1921’ [1844]. The ability to ascribe population numbers to specific geographical localities also enabled the estimation of the population density for any particular area and this figure, when related to mortality rates, could be expressed as a mathematical equation. ‘It is proved beyond doubt that, if the population be the same in other respects, an increase of density implies an increase of mortality’ [1844]; or, stated as a simple aphorism: ‘Give me the density of your population, and I will tell you your death rate’ [1888].
The 19th-century focus on population numerators and denominators, on life tables and population density as the primary predictors of mortality, reflected the aggregation of individual bodies into a population. Yet the population as a simple accumulation of bodies began to disappear in the following century, along with interest in using its density as a predictive tool (though with some short-lived continuing interest in over-crowding). In its place, a new form of population emerged that would inform every aspect of clinical practice and provide the basis for predictions formerly sought through oracles, horoscopes and fortune-tellers.
Difference and variability
The population that emerged during the 19th century might be described as atomistic, because individual bodies, newly collated in the decennial Census, were localized and counted. On the one hand, bodies hardly differed from each other – they were all anatomical bodies as illustrated in the contemporary anatomical atlases – and the perception of their accumulation in the form of a population simply involved enumeration. At the same time, each body was distinctive to the extent that no two individuals – with their constitutions, diatheses and idiosyncrasies – could ever be the same. This tension between sameness and uniqueness informed the perception of individual patients and their aggregation in the population. In effect, medicine could only see difference based on distinctiveness: from the foundation of the world until now, have there ever existed two individuals who, in outward presentiment and an intellectual and moral attributes, were the exact counterparts of each other? … There are certainly as many varieties of man as there are individuals [1862].
The idea of difference that underpinned the 19th-century BMJ view of the population precluded certain forms of perception and analysis. For example, idiosyncrasy implied distinctiveness in each patient to whom it was applied: At no point could those idiosyncrasies be summed, ranked or compared. Recognition of a patient’s individuality through their idiosyncrasies reflected a view that patients were so unique that comparison was meaningless. In the same way, for medicine, the idea of a population as the sum of all these unique individuals could say little about the mean or the distribution of any characteristic, as these calculations were no more possible than deriving the average of an apple and an orange or of a tumour and an infection.
In the closing decades of the 19th century, the concepts of temperament, diathesis and idiosyncrasy began to decline, as a new discourse on variability emerged.
The subject of temperaments and their relation to diatheses is one of the most warmly debated … It is, nevertheless, true that many of our most distinguished physicians deny entirely the truth of the doctrine of temperaments, and stigmatise it as a gross superstition transmitted from the darkest ages [1880].
Later, older clinicians could look back and note that: ‘It is true that the word “diathesis”, like the word “temperament”, is not so frequently on medical lips as it was in the youth of some of us’ [1930]. By the end of the 19th century, idiosyncrasy was being described as serving ‘merely to cloak our ignorance’ [1899]; an explanation ‘in our ignorance we term idiosyncrasy’ [1900].
The terms ‘variable’ and ‘variability’ had been used during the 19th century to refer to clinical observations, but never to patients. Indeed, at the beginning of the 20th century, it could be noted that ‘the systematic study of variation was of very recent date’ [1901]. It was no longer the differences among patients but their variability that caught medical attention: ‘variability of the individual’ [1899]; ‘individual variability’ [1903]; ‘considerable variability in the individual cases’ [1908]. This was more than simply a change in nomenclature: It was a new way of viewing patients and their interconnectedness. Between the language that described ‘patients of different idiosyncrasies’ [1879] and ‘a common illustration of this variability is the effect of alcohol in different subjects’ [1910] lay a major shift in perception.
The disappearance of the old individualizing strategies of temperament, diatheses and idiosyncrasies did not mean that the individual also faded from the medical landscape. Individuality continued to exist for medicine, but in a new form: These new individuals could be compared, ordered and viewed relative to one another. The shift from the individuality of difference to the variations among individuals opened up a new possibility for mapping and re-conceptualizing the nature of a population, along with technologies of prediction. The old search for regularity, as indicated by its mention in the BMJ, reached its peak in the first decade of the 20th century and then declined as clinical attention turned to different methods for describing patterns of disease.
Comparisons and patient groupings
Within the new framework of patient variability, differences between individuals indicated the existence of common characteristics that could be identified and measured. The term ‘individual difference’ was rarely used (and then only in relation to practitioners) until the very end of the 19th century, when it was applied to patients. Ironically, when every individual was so different as to be unique there was no need to refer to ‘individual differences’, as the concept was embedded in individuality. However, once individuals stood in comparison, within a space of variability, then ‘individual differences’ gained traction as part of the new perceptual framework.
Conceptualizing patients as related to one another through their differences, rather than separated by them, meant that patients needed to be compared with each other if their positions in clinical space were to be known. The healthy or normal patient provided a key comparator in this process. Of course, the healthy patient had provided the anchoring referent or contrast for the patient with pathology during the 19th century, but this was an implicit comparator: healthy/ill were two sides of the same coin, even though medicine dealt primarily with the one side. Reference to ‘healthy patient’ was rare in the 19th century (four mentions in the BMJ before 1880) and mainly took the form of describing the patient before illness intervened, as in: ‘many cases had no history but that of a blow or fall having been sustained in a perfectly healthy patient’ [1878].
In the early 20th century, a new form of comparison emerged between two or more different patients to complement comparison of the same patient over time. In part this could be an idealized patient, as in the representation of a ‘healthy normal individual’ in an anatomical atlas [1902], but an increasing number of contrasts between the patient with pathology and the healthy or normal patient appeared in the early decades of the century. For example, the blood of a patient suffering from chlorosis contained less haemoglobin than did that of a healthy patient [1907], or to give another example, ‘a slight failure of asepsis which would have no ill effect in a normal patient’ could be compared against those cases and populations in which a subsequent infection ensued [1924]. In fact, during the inter-war years it was less the healthy patient who served as the point of comparison than the ‘normal patient’: the diabetic had ‘as good a chance as a normal patient’ [1927], for instance, or the ‘normal patient (had) a sharper distinction between the actual pains and the period of quiescence’ [1929].
The principle of variability combined with a comparative method allowed the identification of clusters or sub-groups in which similar individuals could be collected together. These new assemblages of patients marked a new way of constructing a population. Instead of the geographical sub-division of the 19th-century population into towns, cities and counties, and later by hospitals and wards, the new populations were fashioned by identifying and aggregating sub-groups and clusters. One corollary of this perspective was that the terms ‘general population’ and ‘whole population’ came into wider use to capture what had simply been ‘the’ population.
Clinical diagnoses implied a form of grouping, and in the 19th century these were commonly reported as ‘case series’. Individual clinicians accumulated ‘series of cases’, one by one, to illustrate clinical presentations and the natural history of pathological forms. The latter might lead to better prognosis but this was rarely a specific goal of recording case series. The new principle of patient variability, however, supplanted the case series with a new practice and language of accumulating patients by diagnosis. The causes of mental deficiency in children, for example, could best be understood by dividing cases into two groups, deviations from the normal and the degenerative [1905]; patients could ‘for convenience sake be classified into three sub-groups’ [1906]; ‘the cases may be sub-divided into two main sub-groups’ [1911]; or a ‘group of cases may be considered in three sub-groups’ [1914].
While comparisons of individual patients had underpinned the proliferation of sub-groups, attempts emerged to compare the sub-groups themselves. In part, this process simply recognized sub-groups for what they were: groupings derived from some patient characteristic that was used to both aggregate and separate them. This process began to replace the accretion of case series with an explicit comparison of one group of patients with another: a comparison of cases of a similar character treated with and without alcohol [1902], for example, or a comparison of cases operated with and without rubber gloves [1905].
The sub-group displayed some characteristics of a population, as it could be used as a denominator to establish rates (and therefore a potential space of prediction). Early in the 20th century, for example, knowledge of numerator and denominator enabled estimation of the recovery and death rates in an asylum [1905] or the association of length of heart disease with prognosis [1916] or the role of age in recovery from meniscotomy [1923]. In effect, a new proto-population could be generated using any characteristic or variable. Although case series were still described in the early 20th century, they referred to larger and larger groups of patients, drawn not from one clinician’s practice but from wider insurance or hospital data. Reporting of case series in the BMJ peaked in the 1930s, then rapidly declined.
Distributions and the hand of chance
Unlike the absolute differences implied by diatheses and idiosyncrasies, the measurement of the gaps between individuals allowed patients to be ordered and ranked – some patients had more of an attribute, others had less. This ordering could be expressed as a distribution (such as in height or blood pressure). Identification of distributions of individual differences became more common in the first decade of the 20th century and, indeed, formed the basis for the new science of differential psychology. Patients could be distributed by the nature of their diseases, across a histogram, by age, or according to erythrocyte sedimentation rate. When the underlying characteristic was a continuous variable, these distributions could be plotted on paper as a frequency curve. In mathematics, the frequency curve of a chance event (such as tossing a coin) had been described in terms of a characteristic bell-shaped or Gaussian distribution, and early in the 20th century many biological characteristics (and psychological attributes such as IQ) were also seen as being governed by chance and so had similar distributions. The ‘natural’ distribution of a varying patient characteristic could be captured with a symmetric frequency curve or (from the mid-1920s) ‘normal curve’, the product of a non-human world of chance.
For about 60 years, between the mid-1920s and the mid-1980s, in the BMJ the measurement of variability was embedded in a world of chance. The ‘ideal’ distribution of any characteristic was seen to be determined by chance, and comparison of two distributions was assessed against chance. At the very close of the 19th century, for example, it was observed that ‘multiple deaths from cancer would have been below what we should expect to occur by chance’ [1899]. Equally, the effect of vitamin therapy could be inferred because the distribution of disease was ‘not likely to occur purely by chance’ [1932] or a ‘difference as large as that recorded between the average placenta weights in the pre-eclamptic and normal groups would occur by chance on rather less than 1 in 40 occasions’ [1942]. From the late 1930s, the probability of a chance event was captured by use of the statistical ‘p value’ [1939]. Between the 1950s and the late 1980s it was unusual to find research papers published in the BMJ without a p-value that signified whether the difference between groups or distributions had occurred by chance.
In summary, by the second half of the 20th century it had become increasingly common to find patients distributed in a conceptual space with axes of age, sex, occupation, disease type, occupation, social class, illness duration, diagnostic sub-groups, smoking habit, civil status, etc. Moreover, these differences both defined sub-groups (such as smoker or non-smoker) and enabled their comparison. By the mid-1960s patients could clearly be seen as collections of characteristics, each of which could be analysed, distributed and compared. Whereas the revolution in perception at the beginning of the 20th century was marked by the shift from difference to variability, that of the mid-20th century was to see different proto-populations whenever variability occurred. Nevertheless, as long as chance held sway over medical events, the place of prediction remained constrained.
Samples and imaginary populations
The 19th century idea of a population had served as a denominator for its individual members in terms of births and deaths. This function multiplied considerably during the early 20th century. Reporting ‘rate per thousand’ of the population became increasingly popular as new numerators were identified in the context of more selective denominators derived from population sub-groups. The emergence of an inferential connection between group and population opened the possibility of a new predictive technology based on population samples. An indicative step was the first appearance of the term ‘sub-population’ in the BMJ in 1966.
In the late 19th century, the term ‘representative sample’ had been used only twice in the BMJ, once concerning a case of different drugs being sent to India [1880], the other referring to collections of water from a spring in Bedfordshire [1893]. Both examples demonstrated the existence of variability; had every instance been the same there would have been no need to select a representative sample. In other words, for most of the 19th century, when a population consisted of similar bodies/identities there were only individuals and their aggregation in the population; a ‘sample’ of 200 people could not be used to make inferences about people uncounted. When those bodies and identities were perceived to vary, however, as at the beginning of the 20th century, the idea of taking a representative or random sample to capture that variability became possible.
Although for much of the early 20th century medicine focused on distributions and sub-groups, together with their comparisons, there was often a hint of a whole to which these patients belonged. The idea of a ‘sample of the general population’ [1905] emerged early in the 20th century and it was often implied that any sub-group could also be construed as a sample (of a whole). ‘It is, of course, impracticable that every member of one’s class should be investigated; one can only deal with a sample, and care must be taken that one’s sample is truly representative of the class’ [1905]. Whereas the 19th century case series was an accumulation of individual patients, in the 20th century there was increasing awareness of a thread that linked sample and population: ‘it is needful to start with a “general population” or a random sample of a general population … or in some manner reconstruct this sample of the general population’ [1905].
The random sample marked out the ghost of the population from which it was drawn and inferences from other than such a sample were suspect: ‘the sample is not a random sample … [and] consequently no inference at all should be drawn’ [1920]. The main virtue of the random sample was that it gave access through inference to the whole without the latter being measured.
To enforce the point, we may note that, within the limits of sampling, the death-rate of a district will be precisely the same whether we deal with the whole population or with a random sample of it of any given size [1910].
The ‘whole population’ of the 19th century still provided the point of emergence for these first samples, but in the early 20th century their relationship began to reverse: the sample was not so much a sub-division of the population as the population was an artefact of the sample. Indeed, the term ‘population sample’ occurred twice in the BMJ in the first half of the 20th century but 300 times in the second half. Just as in the early 20th century, when the normal or healthy patient had provided the comparator underpinning a machinery for ranking patients, so the comparison of samples in the second half of the century informed a technology (particularly the clinical trial) for evaluating treatments.
The idea of a ‘control group’ (from 1904), together with notions about sample representativeness and chance, began to inform a new discourse on group comparisons. The concept of a ‘control population’ first appeared in 1945. Increasingly large samples began to enable greater precision: ‘follow-up study of over 1,000 patients with hypertension have shown clearly that life expectancy diminishes with even small increments of blood pressure’ [1959]. These greater numbers – the group, the sample, the series – emerged in the 1960s as ‘a population’, not ‘the’ population of the Census enumeration but a multiplicity of populations that reflected on the different dimensions and characteristics of the individuals that in aggregate made up these new phenomena.
In 1984, it was reported that ‘the presentation of variability in medical journals is a shambles’ with the continuing use of what should be described as ‘chance frequency tests’ rather than their usual ‘significance tests’ [1984]. The problem was that a ‘significant’ result was only concerned with group comparisons and related to a population only indirectly, if at all. Moreover, a significant finding, as measured by the conventional cut-off of p < 0.05, did not necessarily indicate a ‘real effect’; indeed, with large samples it was possible to find a significant result that had no clinical importance. Instead of using significance tests, it was proposed that researchers should use ‘confidence intervals’ [1980] that established ‘estimates of results that would be obtained if the total population were studied’ [1986]. The confidence interval was therefore a way of inferring the characteristics of an otherwise unknown population; it was a way of creating a population. ‘We have always needed to go on to generalise to the population at large and so to decide what the true difference and its clinical importance might be. Confidence intervals help us to do so’ [1986]. This population was not, as in the 19th century, a collection of bodies, of individuals methodically counted, but was rather the inferential construction of a virtual group of all patients who had the characteristics in question. By the middle of the 1980s, mention of the term ‘by chance’ began to decline as populations were framed, constructed and analysed through other means.
In the 19th century, patients were unique individuals who, if enumerated, constituted a population. The population in the late 20th century was something quite different. First, it was virtual. There was no method, nor inclination, actually to measure its extent, especially as it could be inferred more efficiently from small numbers. Second, it was multiple. There were patient populations with different diseases, with different psychological and physical characteristics, with different identities. The task of clinical medicine was not to count patients but to focus on the gaps between them, and from knowledge of these differences construct imaginary populations, not of individual physical bodies anchored to a geographical place, but of identities distributed in a virtual space. This population was dynamic, multiple, virtual, constantly changing. Then medicine fixed on the mathematical properties of this population to develop new strategies of prediction.
Framingham and predictive models
During the 1920s, there were five articles in the BMJ describing the Framingham Demonstration, an experiment conducted in a small Massachusetts’ town to test whether periodic examination and close surveillance of a population could arrest the spread of tuberculosis. In the late 1940s, the Framingham Demonstration gave way to the Framingham Study as the surveillance technology was extended to investigate hypertension. Investigators hoped ‘to clarify the beginnings of hypertensive, coronary, and other cardiovascular diseases of middle age in a “normal” population of sufficient size’ [1952]. The methods in part resembled those of enumeration in the 19th century, but Framingham concerned both a ‘real’ geographically located population and, through inference and generalization to other similar communities, a virtual one.
The new Framingham Study began to report the associations between baseline measures and subsequent cardiac disease in the post-war decades (four mentions in the BMJ in the 1950s, 26 in the 1960s, 53 in the 1970s and 294 in the 1980s). The measurement of ‘risk factors’ (first mentioned in the BMJ in 1967) could help establish prevalence figures for both the causes (e.g. hypertension) and the outcomes (e.g. death from heart disease) in the town’s population. The prevalence data for diseases that were formerly only known through hospital records could then be generalized to infer estimates for other populations, in other settings. Yet while the observation that, say, ‘the higher the level of serum cholesterol the greater the risk of developing coronary artery disease’ [1964] might suggest a predictive potential, these early findings informed the old framework of ‘natural history’ rather than a future foretold. A decade later, with recognition that ‘inadequate knowledge of the prognosis and outcome of many diseases is a continual handicap in day-to-day practice’ [1978], it became clear that the solution lay not in clinical experience but ‘in follow-up studies and record linkage (of which) the Framingham study is one of the best known’ [1978].
Once the role of risk factors as predictive tools had been recognized, medicine began to address the new challenge of achieving greater accuracy by combining different risk factors. Data manipulation techniques such as multiple logistic equations and discriminant function analysis were used to derive ‘risk scores’ or ‘predictive indices’. In the 1980s, the titles of papers published in the BMJ no longer addressed the prognostic significance of individual clinical signs but of the predictive power of populations. ‘Does change in blood pressure predict heart disease? (investigated using data collected in the Framingham Heart Study on 5209 subjects) [1983]; ‘Are low cholesterol values associated with excess mortality?’ [1985]; ‘Systolic and diastolic blood pressures as predictors of coronary heart disease mortality in the Whitehall study’ [1985].
By the early 1990s, the equations and algorithms derived from the original Framingham data together with later population cohorts had been distilled into new technologies of prediction. These were described as ‘prognostic models’ (first used in the BMJ in 1993) and ‘predictive tables’ as patients were scored on their risk of disease. Patients could now be informed of their risk profile or of the new ‘number needed to treat’ statistic (how many needed to be treated before a single patient might benefit).
This new attention to index creation implied a different view of the population. Framingham was no longer a town but a population (and was increasingly referred to as ‘the Framingham population study’); more important, the newly synthesized data applied to other towns and other populations. These new population targets were, in a sense, virtual, artefacts of the predictive algorithm. The Framingham equations spread throughout the Western world, sometimes completely, sometimes with corrections for local population characteristics. The future possibilities of life, death and disease for individual patients were now contained in the population-based probability coefficients.
New tools emerged, such as multi-causal analysis, multi-level modelling and Markov models. The term ‘probabilistic’, rarely used before 1950, rapidly increased in popularity. By the turn of the century, risk factors had become essential mechanisms for understanding the nature of a population. Indeed, new descriptors of a population emerged: instead of the ‘general population’, there was now the ‘normal population’ and, reflecting the new uncertainties, the idea of a ‘population at risk’.
Here, then, was a predictive technology, a machine for knowing the future. It did not necessarily have any utility – there was considerable evidence that providing a risk score did not change behaviour – but patients had a glimpse of their destinies. Populations and sub-populations proliferated as the predictive engine generated and identified ‘local’, albeit virtual, accumulations of patients and their characteristics, from which the future could be made known. The sample, such as UK Biobank, could be generalized to an invented population and clinical prediction for an individual patient derived from those population parameters. This was not, however, the individualized prediction of the palm reader or horoscope, but a future expressed and constrained by population probabilities.
Populations and predictive science
On the one hand, the population is a sum of individual identities, but as those identities change the population provides a fluid denominator, comparator, context and analytic space. On the other hand, over the past two decades, the population has come to define those very individuals. The mechanism that binds and stabilizes these two constructs is prediction. Clinical medicine in the 19th century had moved away from a future determined by external events, such as used by astrologers and almanacs, to foresight based on the known natural history of a pathological form. This reading of inner body entrails predicted life and death, or good and bad prognosis, though the individuality of the patient constantly undermined the accuracy of this foretelling. Two centuries ago, the emergence of the population as a denominator established a new potential for predicting events that eventually grew into the application of population-based probabilities to individual patients, a technology that constituted anew both populations and their constituents. In effect, over the past three decades, the future has been exteriorized once again, but this time our future lies not in our stars but in those populations, real and virtual, that map the details of our lives and deaths.
Footnotes
Acknowledgements
The first draft of this paper was given at a symposium on Statistical Prediction and Decision-Making organised by Silja Samerski at the Hanse-Wissenschaftskolleg in Delmenhorst, Germany. I am grateful to participants for feedback. I would also like to acknowledge the very helpful comments and advice of Kirsten Bell.