1. Introduction
On 9 July 2012, Canadian scientists rallied on Ottawa's Parliament Hill, invoking an image of the Grim Reaper and mourning the ‘death of evidence’. What led to this unusual protest? A series of federal government decisions reducing the capacity of federal institutions to collect and bring forward evidence. These decisions included budget cuts to research programmes at Environment Canada, Fisheries and Oceans Canada, Library and Archives Canada, the National Research Council Canada, Statistics Canada and the Natural Sciences and Engineering Research Council of Canada. The federal government was also closing the Experimental Lakes Area research facility and ending funding to the Polar Environment Atmospheric Research Laboratory in Nunavut (CBC Web). The current political situation renders highly relevant this question: what is the relationship between social values and science? Government funding is a powerful expression of a society's values. In this way, at the very least, social values and science interact. But the question then arises as to at what point, and in what way, social values do and should figure in the project of science. And since many, though perhaps not all (see Goldenberg Reference Goldenberg2015), social values are subjective, any account of social values in science must consider the implications for science's objectivity. For the government-of-the-day's political or social preferences to dictate which scientific projects to support clearly threatens science's objectivity. And the scientists who opposed the government's decisions were likewise motivated by values: environmental and other. But what kind of account of the role of social values and the scientific enterprise keeps science's objectivity intact?
Investigating the interweaving of values and science is not new to the philosophy of science. For most of the logical positivists of the early 1900s, the only permissible role for values was through evaluating or estimating against shared standards (McMullin Reference McMullin1982). While evaluating in this manner invoked judgement and could result in different assessments from different individuals, it still excluded any non-rule guided practices. It is now, however, a common view that values – of some description – are vital to the project of science. The post-logical positivist age would come to be defined by these four propositions: (i) that the goal of science is theoretical knowledge; (ii) that theories of science are underdetermined by the empirical evidence; (iii) that assessment of theories involves value judgement in an essential way; and (iv) that observation in science is theory dependent (McMullin Reference McMullin1982). It is in large part to Kuhn that we owe the belief that values – in the form of epistemic values – are essential to science. Kuhn demonstrated that scientists with shared and equal commitment to the core criteria of science can still choose different theories based on both their assessment of the theory vis-à-vis the core criteria and the relative weight afforded to each criterion. In this way, the criteria function not as rules, but as values. For Kuhn, the epistemic values of accuracy, consistency (coherence), scope (explanatory power), simplicity and fruitfulness (novel predictive success), while neither exhaustive nor immutable, provide the agreed upon criteria for theory choice (Kuhn Reference Kuhn and Kuhn1977). These values are called ‘epistemic’ because of their truth-promoting quality.Footnote 1
Having opened the door to epistemic values, the question arises as to whether all sorts of values are let in. In contrast to epistemic values, non-epistemic values include political, moral, social and religious values.
When no sufficient case can be made for saying that the imposition of a particular value on the process of theory choice is likely to improve the epistemic status of the theory, that is, the conformity between theory and world, this value is held to be non-epistemic in the context in question. (McMullin Reference McMullin1982, 19)
Another question concerns the stage at which it is appropriate to consider non-epistemic values – only when knowledge application is at stake, at the earlier point of theory choice or right at the beginning phase of scientific inquiry when research questions are chosen? At one end of the spectrum lies the view that the very distinction between epistemic and non-epistemic values is questionable and that social values permeate everything. For example, Helen Longino's contextual empiricism views science as entirely porous to social context, restrained primarily by critical interactions among scientists (Longino Reference Longino, Nelson and Nelson1996). One popular response to granting that sociopolitical values infuse all of science is to limit them to particular stages of the endeavour. Heather Douglas, for example, allows a legitimate direct role for social values in selecting a research project, making policy, and to some extent, even in adopting a methodological approach. In contrast, if social values were permitted to take a direct role in theory acceptance itself, for Douglas, science's objectivity would be threatened (Douglas Reference Douglas2009). It is against this backdrop that Stathis Psillos suggests in ‘Evidence: Wanted, alive or dead’ that ‘… the key issue is not whether scientific judgement is value-laden but rather what kind of values it is laden with: what kind of social values are the right kind of values’ (Psillos Reference Psillos2015).
Much of the discourse on the role of values and science assumes a clear divide between evidence or raw ‘data’ on the one hand, and social values on the other. And it is only if evidence and values are kept distinct that it can be meaningfully proposed that social values should play no role in the scientific process, or that such values should be limited to a particular phase of scientific inquiry. In my view, the distinction between evidence and values requires scrutiny. Closer examination reveals difficulties with any attempt to clearly separate the two. But this intermingling of values with evidence need not compromise science's unique objectivity. This is because, in my view, it is not the evidence itself that earns science its special place in understanding our world but rather the method by which scientific evidence is acquired. With the focus properly placed on method, the conundrum about the role of values in science dissolves.
In the next section, I will set the stage with a very brief overview of the discourse on science and values that assumes a clear dichotomy between the two. In Section 3, I will discuss the role that initial beliefs play in a model of scientific learning based on updating. Section 4 draws on the medical examples of peptic ulcers and schizophrenia to make the point that initial value-based beliefs shape both the collection and interpretation of evidence. Section 5 situates my thesis as both complementary to, and an extension of, Kuhn's widely accepted claim that values – in the form of epistemic values – are key to science. And in Section 6 I argue that while evidence and values intertwine, the objectivity of science can remain entirely intact.
2. The assumed divide between evidence and values
Most of the literature on the relationship between values and evidence assumes a fairly clear divide between the two. Historically, some of the logical empiricists sought to exclude value judgements from science (McMulin Reference McMullin1982). Arguments from underdetermination then opened the door to values. Today, some philosophers of science see values as all-pervasive. And of course there is also a wide array of variations in between. But what most of these disparate accounts share is the assumption that evidence and values are distinct.
Two common views about the proper place of social values in science, values such as concern for human life, reduction of suffering, political freedoms and social mores are ‘the gap’ and ‘error’ arguments. For example, as Brown (Reference Brown2013) explains, ‘the gap argument’ contends that social values properly fill the space or ‘gap’ created by the underdetermination of theories by the evidence. That is, if it is agreed that evidence alone cannot uniquely determine which of two competing theories is correct (see also Psillos Reference Psillos2015), there must be something else that allows scientists to bridge that gap. That extra ‘something’ may be social values. The argument based on the risk of error, also known as inductive risk and discussed in some detail by Ingo Brigandt, frames the proper role of values somewhat differently. On this account, social values become relevant prior to the point of a theory's acceptance based on an explicit consideration of its social consequences. Since the results of scientific inquiry may always be in error, social values determine the threshold for a theory's acceptance: the more serious the social consequences, the greater should be the proof required for its acceptance (Brown Reference Brown2013).
But these approaches raise the spectre of subjectivity in science. That is, if social values are thought to figure in the reasons to take a theory to be true and in theory acceptance itself, the objectivity of science might be threatened. One response to this is to exclude social values from the scientific process per se and locate them instead at the point of policy-making. That is, since social values aren’t about truth, they should not intrude on the truth-focussed endeavour that is science. It is also possible to reconcile the presence of values within science and its single-minded pursuit of the truth. Heather Douglas's account, which builds on the argument from the inductive risk of error, accomplishes this by distinguishing between direct and indirect roles for values (for a critical discussion of this distinction see Brigandt Reference Brigandt2015). According to Douglas, all science is value laden. However, unacceptable, politicized science occurs when values are allowed to direct the empirical claims made by scientists. By distinguishing between direct and indirect roles for values, Douglas seeks to preserve the fundamental integrity of science. Since our values have no direct bearing on the reality of the world, according to Douglas they should not direct our explanations of it. Furthermore, as evidence mounts, uncertainty declines, as does the need for values to indirectly address that uncertainty. On this account, while science's objectivity does not require it to be value free, it does require these values to be constrained to the indirect role of weighing the importance of uncertainty and evaluating the consequences of error. In this way, values may indirectly guide the decision to do science, to pursue a particular project, to choose particular methods, to characterize and interpret data and even to draw results from the research. But by the time one is at the point of deciding what to believe from this research or what empirical claims to make, values should play no direct role. That is what is required by the pursuit of reliable knowledge about the world (Douglas Reference Douglas2009).
Psillos expresses a closely related view that in the interplay between values and evidence, the world is the final arbiter. Values and evidence get into the scientific inquiry at the same time and presuppose each other. However, since social values might ‘jeopardise rather than promote the objectivity of science … unless, the social values are the right kind of values’, Psillos suggests limiting the role of social values in science to those that are universal: ‘the universalisability of social values, is, for all practical purposes at least, their objectivity’ (Psillos Reference Psillos2015). Another account on the role of social values in science is put forward by Brigandt (Reference Brigandt2015). Brigandt argues for a broader role for social values than merely increasing the quantity of evidence needed to avoid a false theory (as is the case for arguments based on the risk of error). On this account, values directly attach to theories in that they determine a theory's particular conditions of adequacy. Adequacy conditions include scientific considerations about what would make the intended theory explanatory or complete, which is often based on empirical ideas and epistemic values, but can also be influenced by environmental and social values. For example, the man-the-hunter hypothesis prominent in the 1960s viewed the activities and tool use of males as crucial in creating modern humans, while ignoring the evolutionary contribution of females. In a hunter/gatherer account where the men are the hunters and the women the gatherers, it was assumed that hunting furthered the development of bipedalism, an enlarged brain and tool use. While based on true data (about males only), this androcentric theory was incomplete. And so social values matter for delineating what an adequate, complete theory of human evolution should include.
All these disparate views have this in common: they rely on a clear separation between evidence on the one hand, and values, or judgement, on the other. That is, in an attempt to preserve the objectivity of science they hive off values from the otherwise objective and shared proceedings of science. Each account promotes a different role for values as well as different appropriate points of entry. But what they all assume is a fairly clear separation between the two. What I will argue is that the relationship between evidence and values is more complex. At the same time, this need not threaten science's objectivity. That is, it is possible to honour science's objectivity without requiring its evidence to be value free.
3. The role of prior probabilities
Why do I say that evidence and values (both epistemic and social) cannot be neatly kept apart? To answer this, I invoke a model of degrees of belief that builds incrementally on background assumptions. That is, since no hypothesis is ever considered in a vacuum, existing beliefs inform whether a given hypothesis will either be taken seriously or neglected. I use a very minimalistic form of Bayesianism to model these degrees of belief and illustrate the impact of background assumptions in some cases. The ‘Bayesian’ label is used to identify many different kinds of claims and to mark many different distinctions. For example, claims and/or distinctions that are often thought to define ‘Bayesianism’ include that belief is a matter of degree and degrees of belief may be measured numerically; that degrees of belief should obey the axioms of probability; that statistical reasoning requires a prior probability distribution to proceed; that people have prior distributions in their tacit degrees of belief; that probabilities should be updated by conditionalization; or that decisions should be made in accordance with the expected utility maximization rule. To accept the discussion that follows requires a commitment to none of these strong claims. Instead, it invokes only a minimalistic commitment to the idea of updating: as new information (data or evidence) becomes available, we should update our existing beliefs and take into account what we know about potential flaws in the way this information was gathered.Footnote 2 The approach invoked is also ‘informal’ in that updating is taken to occur in situations in which we cannot or do not formally quantify the probability of our prior beliefs or of the measurement errors at stake. In this way, Bayesian updating is what most of us do most of the time. Put simply, the model of learning looks like this: initial beliefs (including beliefs about measurement error) + recent data=a new and improved belief. On this informal model of updating, both the process of interpreting raw data into evidence and the process of updating beliefs based on new evidence invoke prior beliefs.
A clever ‘study’ performed by Leonard Leibocivi and published in 2001 in the British Medical Journal illustrates the importance of prior beliefs (Leibocivi Reference Leibovici2001). In 2000, Leibocivi looked at the medical records of 3393 patients who had developed blood infections at the Rabin Medical Center from 1990 to 1996. He randomly assigned the patients to two groups: the control group and the intervention group. Those assigned to the intervention group were prayed for. Those in the control group were not. He found and reported on a statistically significant difference in the length of stay and duration of fever between the control and intervention groups. It was therefore possible to conclude that ‘Remote, retroactive intercessory prayer was associated with a shorter stay in hospital and a shorter duration of fever in patients with a bloodstream infection’. The study attracted considerable commentary: both positive and critical. Believers in the power of prayer pointed to the study as support for their view. Scepticsfound much to criticize in Leibocivi's methodology. Leibocivi ultimately clarified that his paper was ‘intended lightheartedly to illustrate the importance of asking research questions that fit with scientific models’ (Bishop and Stenger Reference Bishop and Stenger2004).
Many lessons can be drawn from the Leibocivi study. The one I wish to emphasize is that evidence is not collected in a vacuum. Scientists formulate their hypotheses within the context of existing scientific models and prior beliefs. If you assign a prior probability of near zero to either the power of prayer or the ability of present acts to affect the past, then the evidence collected by Leibocivi will not convince you of the truth of either possibility. That is, if you conceive of rational learning as initial beliefs plus recent objective data generating a new and improved belief, your initial beliefs matter. New evidence that is both qualitatively and quantitatively powerful would be required to overturn these initial beliefs. And considering prior probabilities comes naturally to most of us most of the time. If, as the result of experience we come to believe proposition P to a degree of probability x, then when new evidence arises, a good Bayesian must adjust her probability estimate accordingly. The new probability function is obtained from the old by updating the probability in light of the new evidence.
Diagnostic medicine provides a very intuitive and relatively uncontroversial example of the importance of prior probabilities and the intuitive process of updating our knowledge about the world in light of both prior probability and new evidence. Nervous Nelly is over 40 and tests positive for breast cancer. You know this about the test: it correctly identifies the disease 80% of the time. That is in 80% of cases where there is cancer, the test correctly identifies it. Conversely, the test misses 20% of patients who have cancer (false negative). The test also has a 15% likelihood of yielding a false positive. That is, in 15% of cases where there is no cancer, a positive result will arise. What is the likelihood that Nervous Nelly actually has the disease? Should she quit her job, withdraw every last penny from her bank account and ‘live it up’ assuming she has nothing to lose? Or, is she in fact more likely than not to be cancer free?
The Bayesian updating approach requires that you take into account both your existing knowledge of the prevalence of the disease in the population and your knowledge about the error rates of the test in arriving at your answer to these questions. That is, if you already know that 1% of women over age 40 have breast cancer, then your answer to the question, ‘how likely is it that Nervous Nelly in fact has the disease’ will be very different than if you know that 50% of women in that population have the disease. The application of Bayes’ formula yields the result that if 1% of women over 40 have breast cancer, given the parameters of the test accuracy described above, only 5.1% of women who get a positive test result will in fact have cancer. That means that even with a positive test result, Nelly is almost 95% likely to be cancer free!
If, on the other hand, 50% of women in the population have the disease, then given the same test fidelity, there would be an 84.2% chance that Nelly has cancer. That is, the odds of cancer vs. good health are essentially reversed. At first glance, it can appear quite counter-intuitive that a test that correctly identifies the disease 80% of the time and only yields 15% false positives can ever have such a high likelihood of getting it ‘wrong’. But of course, the likelihood of a false positive is only very high where the prior probability of the disease is very low. And where they are known it is simply essential to take both the base rate and error rate into account.
Presented in terms of natural frequencies, the answer becomes more intuitively clear: 10 out of 1000 women over age 40 who participate in routine screening have cancer. Given a test with 80% true positives, you can expect about 8 out of those 10 women to get a positive mammography (an accurate result). But 80% true positives tells you very little unless you also know how many false positives the test yields and how big the relative pools of cancer-ridden and cancer-free patients are to which the true positive and false positive percentages apply. That is, we know too that 15% or approximately 149 out of the remaining 990 women without breast cancer will also get a positive mammography. Since 80% of a small number is a small number and 15% of a big number is a big number, looking only at 80% vs. 15% is highly misleading.
Seen in this way, it becomes quite clear that only approximately 8 out of 157 (8+149) women with a positive test actually have cancer, which immediately appears to be a small percentage (and can quickly be calculated as ~5.1%). This makes a good deal of intuitive sense when you consider that the test has more opportunities to make mistakes in testing the larger pool of cancer-free women. Alternatively, in a population where 50% of the women have cancer, we can think of it in this way: Let us say that 500 out of 1000 women over age 40 who participate in routine screening have cancer. Given a test with 80% accuracy, approximately 400 of those 500 women will get a positive mammography. At the same time, 15% or approximately 75 of the remaining 500 women without breast cancer will also get a positive mammography result. In this way, approximately 400 out of the 475 (400+75) women with a positive mammography result will actually have cancer. This is ~84.2%.
Of course, sometimes you receive information relating to a topic about which you have no prior knowledge. Under such circumstances, you are left to rely only on, for example, the false positive and true negative rate (namely the sensitivity) and the true positive and false negative rate (namely the specificity) of the test. But here too, it is important to recognize how little you know when all you know is test sensitivity. Knowledge of whether the disease is common or rare in the population is critical to making sense of such results. To know test sensitivity alone is, in the case of Nervous Nelly described above, to know that she could have anywhere from a more than 96% chance that she is cancer free to an 80% chance that she is an unfortunate victim of the disease: two radically different scenarios.
The mammography example demonstrates how the process of updating one's beliefs should work. Both data and evidence alone can lead us astray. Tests are prone to error. If the rate of error is known, then we can interpret the evidence appropriately. If you start with the hypothesis that Nelly has cancer, the prior probability, or base rate, is the general incidence of cancer in the population. This is your estimate of the probability that the hypothesis is true, before gaining current evidence – in this case through performing the mammogram test. The posterior probability is what you want to know: the probability of the hypothesis of cancer given the observed evidence of the mammogram result. Bayes’ rule tells us how to derive this chance of actually having cancer, given the evidence of a positive result. It comes down to the chance of a true positive result divided by the chance of any positive result (false positive or true positive). In this way, the probability of any positive result acts as a normalizing constant. What is clear is how evidence alone can lead us terribly astray. The process of Bayesian updating shows the importance of other beliefs to the appropriate appreciation of this evidence. And as we will see in what follows, this opens the door to values.
In the case of the Leibocivi study, the error rate of the ‘test’ is less clear. The ‘test’ in question is a single randomized trial. We may not have a precise false positive or false negative rate available, but what we still should know is that there is room for error. As long as this possibility of error is conceded, prior beliefs come into play. And more than one kind of prior belief may be relevant here. The relevant prior belief may relate to the effectiveness of prayer. If you assign an almost zero probability to the effectiveness of retrospective prayer – then the evidence collected by Leibocivi will do little to change your overall belief. You may also have a healthy scepticism for the reliability of the results of a single randomized trial, believing instead that meta-analyses of large numbers of high-quality randomized controlled trials should inform our beliefs. If so, you would likely informally assign a significant probability to the chance of a false positive to bring down the overall probability of the hypothesis given the evidence. And since there is more than one way to merge results from different studies – approaches for combining individual study results into an overall estimate of effect include inverse variance, Mantel–Haenszel and Peto (Kontopantelis, Springate, and Reeves Reference Kontopantelis, Springate and David2013) – the choice of approach to combining results is yet another way in which both epistemic and social values can enter. In all these ways and more, background beliefs can influence whether certain data is or is not viewed as evidence for a claim.
Both the mammography example and the Leibocivi study demonstrate that data cannot be blindly interpreted. Prior probabilities and normalizing constants – either formally or informally assigned – help us to distinguish causes from correlations and to ignore irrelevant associations that would otherwise be suggested by raw data. But since these priors and constants often have a heterogeneous pedigree – much that went into their formation is often neither known nor knowable – it becomes impossible to cleanly separate evidence from values.
4. The infusion of values through prior beliefs
The section above demonstrated the importance of initial beliefs – both the prior probability assigned to the hypothesis in question and the probability assigned to the risk of error – in the process of learning. In this section, I discuss two examples from the history of science that illustrate that this role is more than theoretical. In the pursuit of science, values and evidence can intertwine in a number of ways. First, values can determine which research questions are considered, and once considered, pursued. If one views confirmation in science as usually contrastive, namely that hypotheses are not tested in isolation, but rather one hypothesis is generally tested against at least one other hypothesis, then values might significantly enter the process at this earliest point of mere consideration.Footnote 3 And while the choice of research question does not directly connect values to evidence, it does so indirectly. This is because the choice of research question affects which data is collected. And evidence is drawn from available data. Second, since raw data is not itself evidence, the interpretative exercise of converting data into evidence also invokes values.
It could be argued that the Bayesian process of updating beliefs keeps value-laden priors and evidence distinct (they are, after all, separate entries in the Bayesian equation). But since in the absence of available data, no updating can occur, and since the transformation of data into evidence requires interpretation against a back drop of existing priors, Bayesian updating itself is no solution to the inescapable infusion of values into evidence. What data you have to work with and the way in which you treat data as evidence allows values to enter the equation.
4.1. The case of peptic ulcers
A relatively recent example of initial assumptions having a remarkably long-lasting effect on the choice of hypotheses pursued comes from general medicine. It concerns the common problem of peptic ulcers: sores (lesions) in the lining of the digestive tract, typically in the stomach or duodenum. Today, peptic ulcers are known to be commonly caused by bacterial infection. Accordingly, the treatment of choice is a course of antibiotics. But for nearly 80 years, bacterial infection was overlooked as a possible cause of the condition. Next to no research was done to investigate the possible connection. How did this happen?
If a hypothesis is, however informally, ‘assigned’ a very low prior probability by the scientific community, certain questions are unlikely to be asked. This affects evidence indirectly by shaping the available pool of data from which evidence may be drawn. The prior probability ‘assigned’ to any proposition lies somewhere on the spectrum between 0 and 1. Here ‘0’ represents certainty in the falsity of a proposition, and ‘1’ represents the equivalent certainty in its truth. To ‘assign’ a prior probability of 0.5 to a proposition is to be perfectly ambivalent about it. Of course, when we’re talking about a scientist's prior beliefs about some specific workings of the world, no actual numerical probability need be assigned. As I write, the window behind me reveals a blue sky with some fluffy light grey clouds in the distance. Turning my mind to the question of whether it will rain this afternoon, I think ‘probably not’. Even though I have not assigned a specific numeric probability to the proposition ‘it will rain this afternoon’, it would still be correct to attribute to my belief a probability that is greater than (and not equal to) ‘0’ and less than (and not equal to) ‘0.5’.
In the late 1800s and early 1900s, so-called spiral micro-organisms were reported as found in animal and human stomachs. Some of the people in whom these micro-organisms were reported also had peptic ulcers or gastric cancer. At that time, the possible connection between the presence of these bacteria and the stomach conditions was considered. Patients are even known to have been treated at that time with an antimicrobial compound. But at some time in the early 1900s, the idea that bacteria were connected to ulcers was somehow summarily discounted. And just like that, a very low probability was ‘assigned’ to the proposition that bacterial infection and ulcers were connected. The exact reason for the summary dismissal of the connection between the two is unknown. One likely explanation however is that spiral bacteria were often found in the guts of healthy individuals (Kusters, van Vliet, and Kuipers Reference Kusters, van Vliet and Kuipers2006).
For nearly 80 years, the question of whether bacteria were connected to peptic ulcers went unexamined. It was assumed by the medical community that peptic ulcers were primarily caused by stress and spicy foods. The treatment of choice was hospitalization, bed rest, and prescribing bland foods. A little later, based on the same causal assumptions, antacids and medications that block acid production became the standard therapy. This treatment was largely ineffective and the rate of recurrence was high (Crane Reference Crane2011). Still, the medical community was not persuaded to probe the cause of this common affliction (Sung, Kuipers, and El-Serag Reference Sung, Kuipers and El-Serag2009).
Despite the combination of a common condition and an ineffective treatment, it was not until the early 1980s that Australian physicians Robin Warren and Barry Marshall performed experiments that would ultimately transform our understanding of ulcers. Warren and Marshall eventually concluded that it was not stress or diet, but the bacterium Helicobacter pylori (H. pylori) that caused ulcers. Slowly but surely this finding was accepted by the medical community. In Reference Yamada, Searle, Ahnen, Aipers, Greenberg, Gray and Joscelyn1994, A National Institutes of Health Consensus Development Conference concluded that there was a strong association between H. pylori and ulcer disease. In 1997, the same institute recommended antibiotics as the treatment of choice for ulcer patients with H. pylori. That year, the Center for Disease Control and Prevention, along with other government agencies, academic institutions and medical industries, started a national education campaign to inform health care providers about the link between H. pylori and ulcers. In 2005, the Nobel Prize in Physiology or Medicine was awarded to Marshall and Warren for their discovery of ‘the bacterium H. pylori and its role in gastritis and peptic ulcer disease’ (Crane Reference Crane2011, 115).
Today, it is widely accepted by the medical community that H. pylori is the primary cause of ulcers. The bacterial infection is considered responsible for 80% of gastric ulcers and 90% of duodenal ulcers. A course of antibiotics is the treatment of choice for the infection, and symptoms usually improve considerably in as little as two weeks (Hopkins web: Sober Reference Sober2008).
The case of peptic ulcers nicely illustrates a common, and I would argue inevitable, phenomenon: the starting point for beliefs greatly influences the course of science. In investigating the world, you must start somewhere. That ‘somewhere’ is what Bayesians call your prior beliefs. And the point from which you begin the journey shapes the course of that process. Most Bayesians would argue that in the long run, the priors ‘wash out’ as the state of knowledge converges on truth. The so-called ‘washing out’ of priors means that despite an unfounded, subjective starting point, objective reality is the ultimate destination. We can certainly hope that this is so. My point is simply this: whether or not you believe we’ll reach the truth in the long run, we can’t expect to have already arrived. In many, if not most, scientific disciplines, we are still in the early days of discovery. This given, the beliefs with which we start still matter. And these starting point beliefs cannot be established as value free. Why? By definition, the starting point lacks the support of objectively gathered evidence. If the ‘starting point’ were already supported by objectively gathered evidence, it would no longer be the ‘starting point’. Furthermore, the values in question are opaque – that is, not overtly held to shared epistemic standards – and cannot be considered free from social, political and moral considerations.
For peptic ulcers, this meant that the unfounded prior belief that ulcers and bacterial infection were unrelated shaped the choice of research questions in the area for 80 years. In fact, as late as in the 1990s, there were fewer than 200 peer-reviewed publications on H. pylori in total. By the early 2000s, once Marshal and Warren's findings had finally been widely disseminated, this number skyrocketed. In 2006, there were approximately 1500 peer-reviewed published articles on the topic in a single year (Kusters, van Vliet, and Kuipers Reference Kusters, van Vliet and Kuipers2006).
If, on the other hand, the potential relationship between bacteria and ulcers had not been summarily dismissed in the 1900s, the connection between the two would likely have been investigated sooner and evidence found to support it. That is, since we now know the connection to exist, mere investigation would likely have yielded results and beliefs would’ve been updated accordingly. Given the inefficacy of conventional treatments, one may also wonder whether this example of peptic ulcers illustrates a failure to adhere to the model of Bayesian updating. That is, in light of significant evidence supporting the inefficacy of the bland foods and reduced stress approach, did not the practice of good updating require diminishing conviction in the initial hypothesis? Certainly this is so. But whether the failure here is characterized in terms of the strength of the initial belief that bacteria weren’t responsible, or in terms of a failure to update in light of available evidence, both possibilities invoke a value-infused account of evidence. A failure to update appropriately is a failed commitment to the epistemic value of rational learning. A strong initial belief that bacteria didn’t cause ulcers simply because they were also found in healthy guts is just an epistemic failure of a different sort: perhaps a mistaken epistemic belief that causes are more often than not accompanied by their effects.
In any case, the choice of research questions in this area was clearly shaped by initial unscrutinized beliefs. And in the absence of questions asked about bacteria related to ulcers, strong evidence to support a connection simply wasn’t available.
4.2. The role of values in initial beliefs
An initial belief that cannot be tested against shared epistemic values is not, however, necessarily value laden. We must consider the possibility that the beliefs in question are based mostly on available data – gathered observationally and informally by the individual, as well as built on the evidence gathered thus far by the medical community. Each scientist after all does not begin her research from a blank slate. As Isaac Newton said, ‘If I have seen further it is by standing on the shoulders of giants’.
Is it possible that the mistaken prior belief that bacterial infections are unrelated to ulcers is attributable to data alone? Was this mistaken belief simply a matter of insufficient, randomly skewed data? I would argue not. First, as I have already discussed, the pool of available data was itself shaped by value-laden initial beliefs. This is because early beliefs informed what research questions were asked and what data was collected: data that could then function as evidence. But in my view, the role of value-laden initial beliefs goes beyond that. That is, values were also clearly inherent in the interpretation of the available data allowing it to function as evidence as well as in the use of this evidence to support the chosen theory. Neither the evidence nor the theory it supported was value free.
Evidence underdetermines not only theories but also initial beliefs. Clearly, there are many beliefs that would be compatible with the early observational data gathered in the case of peptic ulcer. At the very least, we can see that the data available initially was consistent with both the early ‘mistaken’ belief and the contemporary understanding. So if there is something more involved than data, that ‘something’ can best be categorized as a value of some sort.
But are the values in question ‘epistemic’ or ‘social’? It can, and has been, argued that even the core traditional epistemic values have a social valence. For example, if the background theories are sexist, then the epistemic value of coherence with other theories is likely to have undesirable social implications. That is, as Longino (Reference Longino, Nelson and Nelson1996) has argued, if traditional frameworks have played a role in gender oppression by making gender invisible, then the value of coherence will perpetuate this oppression. Instead, the opposite value of novelty would not only encourage new empirical ideas, but (given the prior sexist accounts) also result in politically progressive theories. Similarly, while simplicity can be epistemically as well as socially good in some contexts, it can also have negative social implications. If simplicity is characterized as having the fewest different kinds of entity, this can have the effect of overlooking legitimate differences among people. In the context of gender, this may lead to ignoring gender-related diversity and perpetuating, for example, a family structure in which it is one person – ‘the benevolent patriarch’ – who is decision-maker (Longino Reference Longino, Nelson and Nelson1996, 53). In this way, commitment to the virtue of simplicity can import sociopolitical implications.
In the case of peptic ulcers, it is possible to view the background assumptions as predominantly shaped by social values. The ‘stress-based’ understanding of ulcers blamed anxious people for their poor health. It may have been informed by a prior mistaken belief that ‘poor’ lifestyle choices lead to poor health and that stress and spicy (foreign?) foods should be avoided. The bacteria-based alternative explanation, on the other hand, appears largely devoid of any blame attached to social group or life choices. It is entirely possible then that this difference between the two explanations allowed the stress-based explanation to resonate with prevalent background values in a way that the bacteria-based explanation did not.
It is perhaps more plausible, however, that in the case of peptic ulcers, epistemic rather than social values were paramount. One might suppose, for example, that the epistemic belief that causes are more commonly accompanied by their effects was in play. A belief along these lines would result in the rapid dismissal of a bacteria-based explanation given data that the same bacteria were present both in healthy and diseased guts. That is, if cause and effect almost always go together, then the existence of many healthy individuals infected with H. pylori would properly undermine the explanation that bacteria cause ulcers.Footnote 4
Whether the initial beliefs are seen as infused primarily with epistemic or social values, the peptic ulcer clearly demonstrates two things: (i) that opaque values enter the scientific process at the starting point of initial beliefs; and (ii) that these initial choices and assumptions can have a powerfully long-lasting effect with social implications.
4.3. The case of schizophrenia
The case of schizophrenia demonstrates the role of values even more strongly since it clearly invokes social values at the stage of initial scientific belief. It also illustrates the point made above that the transformation of data into evidence requires interpretation. In the case of schizophrenia, both the choice of research question and the interpretation of existing observational data were affected by prior beliefs. The value-laden belief that the disorder was caused by a troubled family dynamic, delayed investigation into biological causes. It also influenced the interpretation of the available information – namely the disease's tendency to run in families – thereby delaying the discovery of evidence for the biological basis of the disease. This in turn delayed the discovery of effective chemical treatments.
According to the American Psychiatric Association's Diagnostic and Statistical Manual of Mental Disorders (DSM -5), schizophrenia is ‘characterized by delusions, hallucinations, disorganized speech and behavior, and other symptoms that cause social or occupational dysfunction. For a diagnosis, symptoms must have been present for six months and include at least one month of active symptoms’. In the 1940s, it was observed and noted by American psychiatrist Leo Kanner that there was often a ‘lack of warmth’ between parents and their autistic or schizophrenic children. Parents, and in particular mothers, were often blamed for their children's atypical behaviour (Kanner Reference Kanner1943). In the 1950s, the term ‘refrigerator mother’ became a common label for mothers and parents of children diagnosed with autism or schizophrenia. In a 1960 Time Magazine interview, Kanner described the mothers of autistic children as ‘just happening to defrost enough to produce a child’. In 1965, the book Schizophrenia and the Family, reported on a long-term study comparing schizophrenics and their families with non-schizophrenic patients and their families and again emphasized the genesis of the disease in troubled family dynamics (Lidz, Fleck, and Cornelison Reference Lidz, Fleck and Cornelison1965).
Today, however, schizophrenia is recognized as a biologically based disease of the brain. The Centre for Mental Addiction and Health describes the illness as resulting ‘when genetic programming is impaired’, which ‘disrupts the chemistry of the developing brain of the fetus’ (Paterson et al. Reference Paterson, Butterhill, Tindall, Clodman and Collins1999). The illness is treated primarily with neuroleptic or antipsychotic medications as well as psychosocial interventions. It is due to these pharmacological interventions that people with schizophrenia no longer need to be hospitalized for extended periods. While schizophrenics can benefit from supportive psychotherapy, troubled family dynamics are no longer seen as playing a causal role.
And so, as with peptic ulcers, the history of schizophrenia shows prior beliefs – in this case the belief that mothers are to blame for their children's severe malfunctions – shaping the course of science. For decades, it was family dynamics rather than genetics that were the focus of research into schizophrenia. And once data existed showing a correlation between schizophrenic children and troubled family dynamics, this correlation was interpreted as a cause rather than an effect of the disease. In these ways, value-laden beliefs delayed the discovery of evidence supporting the disease's genetic origin.
In the case of schizophrenia, the values associated with the initial beliefs could have been epistemic. That is, they could be more of a kind with values such as internal consistency, empirical adequacy as measured against existing evidence, or the pragmatic values of a theory per se such as scope, simplicity and potential explanatory power. But I would argue that with the belief that ‘refrigerator mothers’ raise schizophrenic children, social (and blatantly misogynistic) values were quite clearly involved. The ‘refrigerator mother’ understanding of schizophrenia blamed women for their children's shortcomings. It is no accident that the theory was not about ‘refrigerator fathers’. This understanding of schizophrenia was most likely informed by, and undoubtedly perpetuated, negative stereotypes about women: that women are both prone to frigidity as well as disproportionately responsible for their children's well-being. The burden was on mothers, never fathers, to ensure that their children turned out okay. With schizophrenia, it is a little harder to speculate as to potentially relevant epistemic values. To give early investigators the benefit of the doubt, it is possible that given the available evidence, family dynamics was the simplest or most coherent explanation available at the time. That is, unlike with, for example, Huntington Chorea which involves mutations in only a single gene (the HTT gene), as we now understand it, there is no single gene associated with an increased risk of schizophrenia. Instead, while people with a close relative suffering from schizophrenia are more likely to develop the disorder than those without, it is believed that no single gene causes the disease. Instead, the current belief is that several genes are associated with an increased risk of schizophrenia and that schizophrenics tend to have higher rates of multiple rare genetic mutations. These genetic differences involve hundreds of different genes and probably disrupt brain development (Paterson et al. Reference Paterson, Butterhill, Tindall, Clodman and Collins1999). So here too, it would not be impossible to explain the earlier understanding of schizophrenia in terms of explanatory coherence. Given the evidence available at the time, perhaps it was just the most coherent view of the disease to attribute it to refrigerator mothers. But this seems unlikely.
Social values may play a large or small role in informing the kinds of prior beliefs that operate in science. But they do play a role. The origin of prior beliefs and the probabilities assigned to them are by definition opaque. For their genesis to be clear would transform them into posterior beliefs. This given, social values cannot be easily removed from the equation.
Both the cases of peptic ulcers and schizophrenia involve value-laden prior beliefs that delayed the collection of new data to serve as evidence and/or informed the interpretation of existing observational data into evidence. I have chosen these examples of now-believed-mistaken prior beliefs because an unfounded belief subsequently exposed as erroneous is more easily seen. Of course, prior beliefs can be equally unfounded but also correct. What is essential to my argument is that in both these cases, the course of science was guided by initial beliefs that neither were, nor could they have been, held to shared epistemic standards. Values of an opaque nature were a key part of the scientific history of these diseases.
Admittedly, it may be possible to attribute these accounts to the context of discovery rather than of justification. That is, while values may have been involved in these particular cases, it need not follow that values – particularly social values – should be involved. Many would say that the practice of science strives to exclude or contain non-epistemic values within its process. My response is that it is one thing to strive and another to achieve. The examples above illustrate how insidiously values enter the practice of science. And even if the process of justification succeeds in excluding them, values will at the very least sneak in in the form of theories available for scrutiny.
5. How this relates to Kuhn's point
My suggestion that evidence and values – both epistemic and social – are inextricably interwoven throughout the process of science is a natural extension of the now widely accepted idea that evidence (or at least data) underdetermines theory.Footnote 5 In his 1977 paper, ‘Objectivity, value-judgement, and theory choice’ Kuhn makes the point that it is possible to draw more than one conclusion from a given body of evidence.
For Kuhn, even if all scientists accept the same epistemic criteria for judging scientific theories, other non-epistemic values will at least enter the picture in the form of the application of these values as well as the relative weights assigned to them (Kuhn Reference Kuhn and Kuhn1977). In this way, epistemic criteria are non-algorithmic (the same input can generate more than one output) and non-epistemic factors inform which output will be generated.
For Kuhn, these non-epistemic features do not render theory choice a matter of non-debatable taste (I like chocolate; you prefer vanilla. End of discussion). Quite to the contrary, Kuhn speaks of judgements or evaluations that are very much debatable against the backdrop of shared epistemic values. As Kuhn says,
The criteria of [theory] choice function not as rules, which determine choice, but as values which may influence it. Two men deeply committed to the same values may nevertheless, in particular situations, make different choices, as in fact they do. (Reference Kuhn and Kuhn1977, 331)
And Kuhn himself accepts some form of basic Bayesian model for theory choice:
… each scientist chooses between competing theories by deploying some Bayesian algorithm which permits him to compute a value for p(T, E), i.e. for the probability of a theory T on the evidence E available both to him and to the other members of his professional group at a particular period of time. (Kuhn Reference Kuhn and Kuhn1977, 109)
At the same time, Kuhn holds that the shared criteria – which include predictive accuracy, internal coherence, external consistency, unifying power, fertility and simplicity – are insufficient to fully determine any algorithm. Each individual's choice of theory is different by virtue of subjective considerations. These subjective considerations are ‘idiosyncratic’ and ‘dependent on individual biography and personality’. Predictive accuracy speaks for itself and is perhaps the primary evaluative criterion. For a theory to be internally coherent requires that it hang together as an explanation, and avoid logical inconsistencies as well as unexplained coincidences. External consistency involves consistency both with other theories and with a general background of expectation. Unifying power speaks to a theory's ability to bring under a single explanatory umbrella a wide set of observations. And fertility refers to the generation of new related explanations about the world. Simplicity is usually viewed as a somewhat lesser, aesthetic, virtue of a theory, though it too can form part of the theory selection process (Kuhn Reference Kuhn and Kuhn1977; McMullin Reference McMullin1982). Since these criteria are shared, for Kuhn there can certainly be debate about how two competing theories stack up on any one criterion as well as about which criterion should matter more.
My account adopts Kuhn's view and then extends it to allow non-epistemic, social values to enter the picture. For Kuhn, subjective considerations may enter into theory selection in a number of ways. For example, Jane may give theory A an ‘8’ on internal coherence and theory B a ‘5’. All other things beings equal, she will commit to theory A. Jill, on the other hand, may think theory B actually excels on the criterion of internal coherence and choose it instead. Since they have shared values, they will have much to discuss. But in the end, they may fail to convince one another and agree to disagree. What will have informed this difference in assessment? It is something other than the shared epistemic values. Alternatively, Jane and Jill may both agree in their assessment of how each theory stacks up on each of the shared criterion, but still come to a different ultimate choice. If theory A is relatively weak on internal coherence, but the stronger of the two when it comes to fertility, the decision will come down to which criterion matters more. And the criteria themselves are unable to help with this decision.
Prior probability assessments are one way in which to understand the heterogeneous body of information that scientists bring to hypothesis assessment (Salmon Reference Salmon1990). And as already mentioned, such prior probability assessments needn’t have precise numerical values. But, in my view, this is not the only way in which values enter the picture: they are also involved in the interpretation of evidence and in other aspects of the model of updating – such as in the informal assessment of the reliability or accuracy of the ‘test’ used to gather the evidence.
My point, which extends beyond Kuhn's admission of subjective factors, is that the infusion of values extends back beyond the stage of theory choice.Footnote 6 It is present when research questions are chosen, when data is interpreted and so on. It is there right from the start with the initial belief(s) that precede all that is to come.
And so, just as Kuhn saw subjective considerations entering into theory choice in the form of assigning different weights to shared epistemic criteria, and in the evaluation of a theory against shared bench marks, I would say that values – both epistemic and non-epistemic – also creep in at the stages of research project choice and the interpretation of data into evidence. The examples of peptic ulcer and schizophrenia discussed in the paragraphs above illustrate this point. And although some would argue that differences based on initial beliefs are eventually ‘washed out’ in the long run, we can’t know that enough evidence has been accumulated for this to occur. In addition, values continue to flow into the process as interpretation and updating continues.
The question remains whether the inevitable infusion of non-epistemic values throughout science undermines its unique value. It does not. Just as Kuhn's view need not undermine the objectivity of science, neither need mine.
6. Focussing on method
My claim is that the while social values infuse its practice, science remains uniquely objective. This is crucial, since there is reason to worry that the influx of values into science might threaten its objectivity. My view is that it is not possible to exclude social values from science – they infuse the process at all stages – but that science's objectivity rests elsewhere.
It is not the objectivity of evidence that lends science its properly privileged place. To rely on evidence to lend science its objectivity is to put that objectivity at risk. As I have argued above, evidence cannot easily be construed as static fact. Instead, interpretation and reinterpretation of data are necessarily involved and non-epistemic factors cannot be reliably kept out of the equation. All is not, however, lost. Science's uniquely objective status can be maintained in the face of the infusion of non-epistemic values throughout the process. This is because, in my view, it is not evidence per se, but rather the method by which science generates it, that justifies science's privileged role.
The scientific method is based on independent (and objective) verification. It is defined by falsifiable hypotheses, repeatable methodology, verifiable results and testable theory. It is this process that earns science its position of privilege. To know more definitely what to make of Leibocivi's study would require repetition of the methodology with another set of patient records. Until the methodology has been repeated and the results replicated, we should withhold judgement. And therein lies the value of meta-analyses that review all relevant studies that meet certain standards of scientific rigour. In the medical field, the Cochrane collaboration features reviews of clearly formulated questions that use ‘systematic and explicit methods to identify, select and critically appraise relevant research, and to collect and analyse data from the studies that are included in the review’ (Cochrane Glossary, web). It is this methodology, unique as it is to the project of science, that leads to objective knowledge.
It is neither necessary, nor in my view possible, to eliminate social values from science. They are present right from the start, often hidden somewhere within inscrutable prior beliefs.
We began with concerns about government funding cuts to science. Some of the arguments raised to defend these cuts cite scientific impediments to economic growth or commercial ends. Psillos (Reference Psillos2015) issues a caution that unless social values in science are universal, they potentially threaten its objectivity. My point is that we must concede that social values – and not just those that are universal – are an inescapable aspect of the scientific process. They sneak in through initial beliefs that simply can’t be held to shared epistemic standards. But all is not lost. The objectivity of science is preserved not by concerning ourselves with the kind of values but by privileging the method of science.
Is science infused with values through the operation of prior beliefs and otherwise? Undoubtedly. But in my view, this is only a matter of what and whose values in so far as the values pertain to the scientific method itself. In some ways, it is that simple: it is the values inherent in the scientific process that should always be promoted. Science pursued according to the scientific method is always good science. The evidence that this process happens to generate is secondary.
Acknowledgements
I would like to thank Brian Baigrie for the invitation to speak at the University of Toronto's Institute of History and Philosophy of Science and Technology's Situating Science Joint Colloquium, April 2014 as well as Stathis Psillos's paper that was the subject of that colloquium and inspiration for my paper. For feedback that led to improvements to this paper, I am grateful to James Brown, the anonymous referees, and most especially to Ingo Brigandt, whose comments greatly improved the paper.
