In what sense should science be value-free? Can it be?
Introduction
The role of values in science remains a controversial topic. Amongst laypeople and non-scientists in general, there exists a perception that science is, in some meaningful sense, ‘value-free.’ In this essay, I shall explore the role of ‘non-epistemic values’ in scientific reasoning and argue that whilst science must seek to eliminate certain sorts of value arguments, it cannot be thought of as being ‘value-free’ since non-epistemic values form an intrinsic part of scientific reasoning in exemplar cases. This will be illuminated with a thought experiment in epidemiology, and drawing on the work of Douglas (Inductive Risk and Values in Science, 2000) on ‘inductive risk.’ The discussion shall be limited to the natural sciences, since these issues become increasingly complex in the social sciences. I shall also respond to criticisms of this view as brought out by Levi (Must the Scientist Make Value Judgments?, 1960) about the role of value in regard to the acceptance or rejection of a hypothesis.
Epistemic and non-epistemic values
Firstly, the nature of the values to be discussed shall be outlined. McMullin (Values in Science, 1982)defined an epistemic value as ‘Values… presumed to promote the truth-like character of science.’ So a value like ‘you should only believe true statements’ would be an epistemic value. Hence, non-epistemic values are any sorts of values that fall outside of this definition, such as those regarding justice and risk and so on. Though this distinction is not necessarily a straightforward one, (for example, values regarding honesty seem to have both an epistemic and non-epistemic,) this shall be our definition. Epistemic values are largely uncontroversial and shall not be the focus of this essay.
Secondly, I shall briefly outline the philosophical backdrop and the work I intend to draw upon. As Douglas states in Inductive Risk and Values in Science, ‘the common wisdom of philosophers of science has been that only epistemic values have a legitimate role to play in science.’ Hempel in Science and Human Values (1965), states that ‘judgments of value… [should] be dismissed as lacking all relevance to [a] hypothesis since they can contribute neither to its support nor its disconfirmation.’ Hempel then, however, goes on to articulate a sense in which values do seem to be involved in the scientific method. He discusses this using the idea of a ‘rule of acceptance,’ which is a rule for determining ‘how strong the evidential support for a given hypothesis has to be if the hypothesis is to be accepted into the system of scientific knowledge.’ He goes on
‘when a scientific rule of acceptance is applied to a specified hypothesis on the basis of a given body of evidence, the possible "outcomes" of the resulting decision may be divided into four major types: (1) the hypothesis is accepted… in accordance with the rule and is in fact true; (2) the hypothesis is rejected… and is in fact false; (3) the hypothesis is…, but is in fact false; (4) the hypothesis is rejected…, but is in fact true.’ (p 92).
Hempel supposes that since science has to find some way to produce outcomes (1) and (2) but not (3) and (4), science requires some way of assigning ‘values or disvalues’ to these possible outcomes, and hence in this sense ‘presupposes valuation.’ Hempel took this to be an issue of scientific methodology rather than one of scientific reasoning (Douglas, p 561), however Douglas argued, and I further argue, that such valuations are a part of the process of scientific reasoning and cannot be extricated from it, either by the epistemic-non-epistemic distinction, nor by Hempel’s ‘method-reasoning’ distinction.
Inductive risk
Hempel’s ‘acceptance rule’ must account for what he called ‘inductive risk,’ which is simply the possibility that the rule produces outcomes (3) or (4) as above. Douglas attempts to show that this ‘inductive risk’ is in fact pervasive throughout scientific reasoning by drawing on examples from studies of the carcinogenicity of dioxin in rats. A thoroughgoing exploration of Douglas’ paper is beyond the scope of this essay, but I shall endeavour to briefly outline Douglas’ position as to how inductive risk makes non-epistemic value considerations intrinsic to the process of scientific reasoning.
Her paper presents three radically different interpretations of the same data with regards to the carcinogenicity of dioxin in rat models. It is Douglas’ assertion that these radically different interpretations arise due to the influence of non-epistemic values on the reasoning of the scientists involved in the production of each different interpretation. Douglas states that the experts concerned must contend with the fact that
‘[F]or [the evidence] there is significant uncertainty in whether a judgment is correct. With this uncertainty comes significant inductive risk and the need to evaluate the consequences of potential errors.’ (p 571).
These judgments are affected by knowledge of both the public health impact and economic impact of the regulatory decisions that may flow from the research as it is conducted. This is most striking in comparing the 1980 and 1990 groups (p 570) in Douglas’ table. The 1990 results were produced by a private contracting company concerned with the economic impact of dioxin regulation on various industries, whereas the 1980 results were produced by the Environmental Protection Agency (EPA) in the U.S. with an eye to regulating dioxin for the benefit of public health. The evidence examined and the statistical standards of significance remained the same, the only difference between these studies was the interpretative function that the scientists involved undertook. It is Douglas’ argument, one that I accept, that the differences produced can and should be explained by the scientists involved hedging against inductive risk. In the case of the EPA study, the scientists wish to hedge against false negatives, that is, that dioxin causes increased cancer risk but that this is not confirmed by the data. So, when the EPA scientists are confronted with a borderline case, it makes good scientific sense for them to hedge towards increased cancer risk, since they are concerned with protecting public health. In the case of the consortium case, the opposite is true. They will hedge against false negatives, since they are concerned with the economic impact of over-regulation of dioxin. Hence, in borderline cases, it makes good scientific sense to hedge against false positives, since they are concerned with the impact of dioxin regulation on the private sector.
Here it might be claimed that this is simply bad science, and that there is a matter of the fact about dioxin’s carcinogenicity and these studies are not adequately controlling biases. A thorough refutation of this point is beyond the scope of this essay, but I wish state that I find Douglas’ argument compelling, and accept her assertion that ‘inductive risk’ partakes in science that has non-epistemic consequences. I shall attempt to illuminate my position further with a though experiment in epidemiology that hopefully will prove a convincing case for non-epistemic values in scientific reasoning.
A thought experiment in epidemiology
Imagine that there is a pathogen that has recently jumped from primates to humans called philosophitis. It has a high mortality and morbidity rate and there is currently no cure, and no treatment beyond pain alleviation. All cases have been limited to zoo-goers in a small town, which has been quarantined. The pathogen is spreading amongst those under quarantine, but it is noticed that it is spreading more quickly amongst certain groups of relatives than it is other groups. A battery of genetic tests is ordered and it is found that those with a genotype, A1, appear to be contracting the illness at a greatly increased rate and account for most of those currently infected and the deceased. Two studies are funded to explore the link between philosophitis and the A1 genotype:
Study Y will use primates and intentionally infect both those with A1 and non-A1 genotypes and observe the spread of the pathogen, as well as a control group. Since a live pathogen is being used, the researchers are not able to be fully blinded: they can likely deduce what they are working on and are aware of the ongoing quarantine.
Study Z will use computer modelling of the pathogen and the protein A1 codes for, producing many variations in different scenarios to examine whether A1 is a risk factor or if other factors (such as proximity, physical contact, other genetic traits) might be a better explanation. This research will be blind, in that the computer scientists who run the actual models will not know what exactly the data refers to.
If A1 is a significant risk factor in contracting the disease, the Government decides that the quarantine can be scaled back to those families with the genotype and those who already have the illness. Otherwise, the quarantine must remain in place. These studies will be the only available data to determine the Government’s actions, and they will listen to the advice given by a body of Centres for Disease Control (CDC) scientists drawn from the international community. When the results come in, they are contradictory. The primate study shows a strong, statistically significant link between the A1 genotype and the risk of contracting philosophitis. The computer model, however, posits that in fact close personal contact and shared living space better accounts for the spread amongst those who are closely related. Now, I shall posit two scenarios in which the above occurs, in which the consultation of non-epistemic values leads to good science producing differing conclusions.
Scenario 1: Genotype A1 occurs randomly throughout the general population and those who have it and are in the quarantine zone are no more or less likely to be in a minority than the general population. The CDC advisors recommend easing the quarantine, since the evidence from study Y seems to be stronger than the computer model.
Scenario 2: Genotype A1 is possessed almost exclusively by a minority, group B, that has historically been oppressed, and are frequently compared with ‘monkeys’. While group B recently gained equal legal standing, the quarantine effects almost exclusively those of group B. The CDC panel concludes given the very possible bias present in the researchers in study Y, the blind nature of study Z, and that stricter quarantine is the safer option until more data is available, that the computer model should be accepted going ahead until more studies can be published and recommends the quarantine remains in place.
Inductive risk and the epistemic responsibility of those in Scenario 1 versus Scenario 2
The above thought experiment is perhaps a little far-fetched, nonetheless, I think it serves to demonstrate a real point. Only matters of fact regards politics and genetics vary. The studies, those who produce said studies, and their results remain the same. It might now be complained that this example turns on the blindness of the researchers, and that if study Y was truly blind then this would not be a problem. True enough, but it is often not possible to completely blind researchers, and more pertinently, the researchers in Scenario 1 are not blinded either. The work in Scenario 1 carries greater epistemic weight because there is no obvious reason why they ought to be biased one way or the other. In Scenario 2, however, we have credible reasons to expect bias, including the history of comparing group B with primates, which may skew them toward positive results. They may have also been skewed unconsciously by a desire to see group B more heavily quarantined than others, given the history of oppression of group B. Nevertheless, I must emphasise that it is not at this point that it is proper for non-epistemic values to enter. In fact, this is an example of the sorts of non-epistemic values that ought tobe kept out of science (and why I chose this thought experiment). (I would like to expand considerably on this point, but the limitations placed upon me prevent me from doing so.) The proper place for non-epistemic values to enter is on the part of the CDC panel reviewing the evidence and deciding which conclusions to accept or reject. They must consider the social environment in Scenario 2 in a way that they do not in Scenario 1 if they are to produce responsible analysis of the studies conducted. Since in Scenario 2, the appropriate scientific position to take is to permit that Study Y could be coloured by bias, that is, the inductive risk is greatly increased, the quarantine should remain in place until further data is available. Furthermore, it is only by consulting a non-epistemic value (that of the risk to the non-A1 genotype population) that we can reach such a conclusion, and it would be epistemically irresponsible to fail to account for the flaws in Study Y. In Scenario 1, no such epistemic responsibility is produced by the risk to the non-A1 genotype population, since we have no grounds for thinking it flawed.
More detailed examination of Scenarios 1 & 2
It is worth noting that in both Scenarios, the CDC is not, strictly speaking, choosing which to accept hypothesis Y or hypothesis Z. In both Scenarios, the CDC is choosing between accepting the conclusions of study Y, (which is always the stronger of the two studies,) in the light of study Z, and awaiting further confirmation of either hypothesis. The non-epistemic value at play in the decision is the level of risk acceptable to those who are non-A1 genotype. The theoretical level of risk is unchanged across both Scenarios, however, the practical risk in Scenario 2 is increased, because of the bias that could potentially be influencing scientists, in study Y. Briefly, the risk of a false positive in Scenario 2 is increased. This must influence the CDC panel to be more cautious in their assessment of the situation. McMullin argues that such decisions are in fact related to the ‘utilities typically associated with the application of science,’ (p 8) and not in fact intrinsic to the theoretical framework of science itself. However, it seems that the CDC panel are making a wholly scientific decision: is study Y, in the light of study Z, sufficient to conclude that the A1 genotype is a significant risk-factor for contracting philosophitis? This decision does not seem to me to at all concern the application of science, it concerns the very foundations of a hypothesis in both Scenarios. When presented with contradictory evidence, can we favour this study over that one, in the light of everything we know about them? It is my assertion that, with Douglas, we must consult a non-epistemic value when there are non-epistemic consequences, the level of risk to non-A1 genotype, in deciding which evidence ought to be given credence in the two different Scenarios, producing different results depending on the various relevant data.
Levi’s argument against values
Levi (Must Scientist Use Value Judgments?, 1960) argues that reasoning much like the above is unacceptable, summarizing the argument, in which he is paraphrasing Rudner, as follows:
‘(1) The scientist qua scientist accepts or rejects hypotheses.
(2) No amount of evidence ever completely confirms or disconfirms any (empirical) hypothesis but only renders it more or less probable.
(3) [Because] of (1) and (2), the scientist must decide how high the probability of a hypothesis relative to the evidence must be before he is warranted in accepting it.
(4) The decision required in (3) is a function of how important it will be if a mistake is made in accepting or rejecting a hypothesis.’ (p347).
He refutes this line of reasoning as implausible, stating it must be argued for further, adding the following assumptions:
‘(5) To choose to accept a hypothesis H as true (or to believe that H is true) is equivalent to choosing to act on … H relative to some specific objective P.
(6) The degree of confirmation that a hypothesis H must have before one is warranted in choosing to act … H relative to an objective P is a function of the seriousness of the error relative to P resulting from basing the action on the wrong hypothesis.’ (p348).
Levi rejects (5) because there are cases where accepting a hypothesis need not have an objective, and even when there is, ‘it often seems appropriate to distinguish between acting [based on] a hypothesis relative to [some] objective and accepting the hypothesis as true.’ (pp 350-1). Therefore, it is not clear that scientists must consult (non-epistemic) values when accepting or rejecting a hypothesis, especially in highly theoretical or low-stakes situations. Levi’s argument may hold some weight against my argument, since it certainly seems that the CDC in my example is acting with an objective in mind; that of protecting public health. And it certainly seems like the CDC panel are concerned with the probabilities of being wrong, that is, the inductive risk involved. They also seem to be deciding whether to accept or reject the hypothesis that the A1-genotype significantly increases the risk of contracting philosophitis, in the light of both studies available to them. So, does Levi’s objection hold against me?
To some extent I think it must, since there is obviously a theoretical sense in which either A1 is or is not a risk factor for contracting philosophitis, and the CDC wants to have a firm, ‘factual’ answer in this regard. Also, this is a high-stakes case in the presence of a great deal of under-determination of the hypotheses by the evidence, so in some respects Levi’s argument is not designed to apply here. This said, we might find that Levi’s argument becomes less biting when we imagine a scenario that is intellectually ‘minimal risk.’ Let’s imagine philosophitis is instead a new strain of a flu-like pathogen, being tested in a theoretical setting whereby the risks are simulated and the studies produced are done so based on a known simian pathogen. Let Scenario 3 be otherwise identical to Scenario 1, but translated into this new flu-like hypothetical, and Scenario 4 be otherwise identical to Scenario 2, likewise translated. Would this change the epistemic balance? It isn’t immediately obvious that the CDC panel would change its position from the Scenarios outlined above if pushed to provide one. It seems that the CDC in Scenario 4 should still reject hypothesis Y considering study Z until there is more data available, based on the hypothetical risk to non-A1 genotype population. There is no action to be taken, per (5), yet in Scenario 4, it seems that their submission ought to reflect their position in Scenario 2, at least to some extent. Thus, I take it that although Levi’s argument has some traction against mine, even in a purely hypothetical scenario, it appears non-epistemic values ought to enter consideration somewhere in scientific reasoning, even when no action is called for.
Concluding remarks
The place of non-epistemic values in science will remain a controversial one, however, I hope to have demonstrated, using the work of Douglas and Hempel, that it the matter should not be decided with undue haste. There is a case to be made for the necessity of some non-epistemic values within the scientific method, and indeed touched on the objectionable nature of others. Even considering Levi’s objection, it still appears that are at least two sorts of value-laden-ness in science; one that must be eliminated (as exemplified in Scenario 2’s study Y,) and one that must be preserved (as exemplified by the CDC panel in Scenario 2.) This essay leaves unclear exactly how to tease apart such these sorts of value-laden-ness, and it may be that future work could unpick this distinction.
Bibliography
Douglas, H. (2000). Inductive Risk and Values in Science. Philosophy of Science, 67(4), pp.559-579.
Hempel, C. G. (1965). ‘Science and Human Values’. In Aspects of Scientific Explanation and other Essays in the Philosophy of Science. New York: The Free Press pp 81-96.
Levi, I. (1960). Must the Scientist Make Value Judgments?. The Journal of Philosophy, 57(11), p 345.
McMullin, E. (1982). Values in Science. PSA: Proceedings of the Biennial Meeting of the Philosophy of Science Association, 1982(2), pp 3-28.