Risk, Uncertainty and Precaution in Science: The Threshold of the Toxicological Concern Approach in Food Toxicology

Abstract

Environmental risk assessment is often affected by severe uncertainty. The frequently invoked precautionary principle helps to guide risk assessment and decision-making in the face of scientific uncertainty. In many contexts, however, uncertainties play a role not only in the application of scientific models but also in their development. Building on recent literature in the philosophy of science, this paper argues that precaution should be exercised at the stage when tools for risk assessment are developed as well as when they are used to inform decision-making. The relevance and consequences of this claim are discussed in the context of the threshold of the toxicological concern approach in food toxicology. I conclude that the approach does not meet the standards of an epistemic version of the precautionary principle.

Introduction

Over the past years, improvements in analytical methods have led to the detection of an increasing number of previously unknown substances in food products and food contact materials. Estimates say that we might be looking at several thousands of chemicals. They might be present due to degradation processes, unintended reaction by-products, migration from packaging material or impurities in the manufacturing process. Usually, these substances occur in low or very low concentrations and their toxicity as well as their potential effects on human health are often unknown. Assessing the risk of non-intentionally added substances in food products has become a major challenge.

Currently, one of the most discussed concepts for the evaluation of potential risks of incidental low-dose substances is the so-called threshold of toxicological concern (TTC). The TTC is a risk assessment tool that rests on the assumption that it is possible to derive risk estimates for substances with unknown toxicity in food products and food contact materials from toxicological data of structurally similar substances. The approach is meant to be a first step in the risk assessment process in order to determine whether a comprehensive risk assessment involving toxicological testing is necessary if a new substance is found, or whether the substance is below its respective threshold of concern.

In what follows, I will explain how the approach works, what it is supposed to accomplish, and what its problems are (Sect. 2). I will show that the TTC is a useful tool for defining threshold values below which toxic effects are highly unlikely. However, the TTC does not inform us about the potential consequences of a case where a substance is wrongly evaluated to be of little toxicological concern. I argue that a feasible science-based risk assessment tool should be able to inform us what can happen if some of its predictions or assessments turn out to be false. By failing to do so, the TTC gives rise to the so-called Tuxedo Fallacy, i.e. the approach focuses merely on quantifiable risk and ignores qualitative uncertainties (Sect. 3). In accordance with Sprenger (2012), I will also argue that the well-known precautionary principle has substantial implications for model building in science. That is to say that the precautionary principle is of epistemic import. Precaution must already be applied when evaluating the epistemic robustness of scientific models, and not only when scientific knowledge is used to inform decision-making. This runs against the widely held view that the precautionary principle is primarily a guiding principle for risk management and not so much for scientific risk assessment. I will conclude that it remains highly questionable whether the TTC provides an adequate tool for the assessment of potential health hazards if one evaluates the approach against the standards of the precautionary principle taken as an epistemic rule (Sect. 4) (readers familiar with the technicalities of the TTC might prefer to proceed directly to Sect. 3).

The Threshold of Toxicological Concern

How the Approach Works

The main theoretical rationale underlying the TTC is that structurally related substances exhibit similar toxicological behaviour. It is assumed that the toxicity of an untested substance can be estimated by comparing it to structurally similar compounds with known toxicity and by defining a threshold exposure value below which the risk of toxic effects is sufficiently low.

In order to determine a threshold of concern for a given substance two things are needed: (1) the chemical structure of the substance, and (2) a valid estimate of exposure in terms of μg/person/day. Given these two pieces of information, we can evaluate the risk of the substance for chronic oral exposure by assessing whether the exposure lies below a threshold value obtained by toxicological tests of structurally similar substances. The application of the TTC leads to a definite answer as to whether or not the estimated exposure to a particular substance is likely to be a safety concern.

Fig. 1

The TTC decision tree according to Kroes et al. (2004) (reproduced with permission from Elsevier)

Figure 1 provides an overview of the TTC decision procedure suggested by Kroes et al. (2004).¹ The first step in the evaluation process is to rule out that the substance is a non-essential metal or a metal-containing compound, a nanomaterial, a radioactive substance, a steroid, or a protein. If the substance belongs to one of these classes the TTC is not applicable, because of their tendency to accumulate in the body or their potent biological activity (some further substances that are known to be highly toxic are also excluded in this step). In the next step, the substance is checked for structural alerts for genotoxicity (i.e. DNA reactivity). If they are present, the estimated intake must not exceed 0.15 μg/person/day. At this dose, the life-time cancer risk for almost all carcinogens is unlikely to be greater than 1 in 10⁶. Certain high-potency carcinogens such as aflatoxin-like-, N-nitroso- or azoxy-structures have to be treated separately. Organophosphates, many of which are potent nerve agents, are also excluded in this step if the estimated exposure exceeds 18 μg/person/day. If the substance does not belong to one of these exclusion classes, it is then allocated to one of three so-called Cramer classes of organic chemicals using the decision tree approach originally developed by Cramer et al. (1978). The Cramer decision tree contains 33 “yes/no” questions, each of which leads either to another question or to a final allocation of the substance into one of the three classes.²

Class I includes chemicals of simple structure with efficient metabolism and low oral toxicity, such as acyclic aliphatic hydrocarbons, common carbohydrates, or the like. Most substances in class II are chemicals with functional groups that are similar to, but somewhat more reactive than functional groups in class I; or chemicals with more complex structures than those in class I, but that are common components of food. Class III includes all chemicals that do not belong to class I or II. Examples of substances in class III are structures that contain elements other than carbon, hydrogen, oxygen, nitrogen or divalent sulphur, as well as heterocyclic or aliphatic substances with more than three different types of functional groups.

Each of the three Cramer classes is associated with a specific threshold value: 1800 μg/person/day for class I, 540 μg/person/day for class II and 90 μg/person/day for class III (or 30, 9, and 1.5 μg/kg body weight/day, respectively). These values represent the 5th percentiles of the cumulative distributions of no-observed-effect-levels (NOELs) derived from data of chronic toxicity studies on 137, 28, and 448 compounds in classes I, II, and III, respectively (Kroes et al. 2005; Munro et al. 1996, 1999). A NOEL is the highest dose in a chronic toxicity test at which no statistically significant effect can be observed. NOELs are always related to specific test procedures. The most common procedure used is oral exposure in rodents. The TTC values for human exposure are then derived by a division of the NOELs from the animal data by a 100-fold safety factor (see Fig. 2). A factor of 10 accounts for uncertainties related to the extrapolation from animal data to humans and a factor of 10 is supposed to account for sensitivity differences between human individuals.³

Fig. 2

Log-normal fitted distribution of the most conservative NOELs for compounds in the reference database grouped into Cramer classes I, II and III (Box class I percentiles; Circle class II percentiles; Triangle class III percentiles.) The TTC values (without safety factor) for each class are the concentrations at the fifth percentile (3 mg/kg bw/day for class I; 0.9 mg/kg bw/day for class II; 0.15 mg/kg bw/day for class III). The values for human exposure are derived by dividing these values by a safety factor of 100. Figure taken from Munro et al. (1996) (reproduced with permission from Elsevier)

Fig. 3

Increase in incidence in testicular (top) and breast cancer (bottom) across Europe. Figures taken from WHO (2012) (original data of top figure from Richiardi et al. 2004; reproduced with permission of the publisher)

Defining the 5th percentile as the threshold value means that an untested substance allocated to one of the three classes has a 5 % chance of having a lower observable effect level than the TTC of the respective class. Or, to put it another way, 5 out of 100 substances submitted to the TTC will be assessed to be of “low concern”, although they might have shown an observable effect below the respective threshold value had they been submitted to a regular toxicity test.⁴ Strictly speaking, the TTC does not provide genuine toxicological threshold values, below which no toxic effects occur, but it defines levels of acceptable risk. Hence the name “threshold of concern” is quite apt. The TTC sets a limit for human concern. It goes without saying that the concept therefore carries a lot of normative weight, because, as we will see later, the determination of a level of acceptable risk is not merely a matter of empirical fact, but it necessarily involves a normative judgement and must ultimately be seen as a convention that is more rooted in scientific habit (a 5 % confidence level is a commonly used measure in statistics) than based on a democratically reached agreement on how much risk society is willing to accept.

A certain degree of normativity is already present at the level of the experimental determination of the NOELs. Determining a NOEL requires a definition of what counts as a “statistically significant observable effect”. An effect is significant if there exists a statistically significant difference between the dosed groups of test animals and the non-dosed control group. This also means that the statistical significance of an effect depends on the sample sizes of the groups, and it can change with varying numbers of test animals.⁵ What counts as an observable effect is also highly dependent on the test protocol, and there is huge variance between different protocols (Crane and Newman 2000).⁶ Defining standards for statistical significance and sample sizes—a standard cancer test is usually done with 100 animals for each dose (Shibamoto and Bjeldanes 2009, p. 51)—always requires an evaluative judgement that cannot be based on purely empirical considerations. The use of uncertainty factors does not change this. The tacit assumption of many is that multiplying a NOEL by an uncertainty factor yields an acceptable level of risk. However, what “acceptable” means in this context is never defined quantitatively.

Advantages of the TTC

An obvious advantage of the TTC concept is that it greatly accelerates evaluation processes for substances that humans are exposed to at low doses by providing a prioritisation scheme for the identification of potential health hazards. Its application makes it possible to focus on potential risks and to avoid unnecessary testing. The approach takes into account the limitations in time, funding, and test animals that would be required in order to properly evaluate every newly identified substance. The TTC therefore provides a cost-effective way for risk assessors to give urgent advice whenever a new substance is detected. Application of the TTC also allows chemicals to be fast-tracked through the risk assessment process leading to market approval within a few months, instead of several years if comprehensive risk assessments were required. This undoubtedly constitutes a huge commercial advantage.

It should therefore come as no surprise that the TTC is supported in particular by the industry and industry-related research institutes such as the Life Science Institute (ILSI). ILSI is an international non-profit organisation that, according to its website, seeks “[...] to provide science that improves human health and well-being and safeguards the environment”. It is funded by its member organisations, which mainly consist of large food and biotech corporations such as Nestlé, Monsanto, Coca-Cola, McDonald’s, and more.⁷ ILSI has also been involved in the development of the TTC (see for instance the acknowledgements in Kroes et al. 2004). Many recent publications arguing in favour of the TTC are related to ILSI, and authors have often received honoraria from ILSI.⁸

Problems and Criticisms

A clear disadvantage of the approach is that it is not substance specific. The application of the TTC concept does not inform us whether the substance under scrutiny is in fact toxic or not, because the assessment is not based on a substance’s intrinsic toxicological properties. The TTC just provides risk thresholds for classes of structurally similar substances.

The TTC also crucially depends on the reliability of exposure estimates. As mentioned, the approach can only be feasibly applied if one can provide robust values of exposure. Estimating the exposure to a given substance, however, is far from trivial. First, the problem of the data source arises. For many food products and packaging materials there are no reliable numbers available. But even if there are sufficient statistical data, there exists huge variability due to local circumstances. Who consumes which product how often is a complex question to answer. And when it comes to individual storage or eating habits, things become even more difficult.

A further problem of the approach is its limited applicability. The TTC is unable to provide risk assessments for toxicological endpoints other than subacute and long-term toxicity (including carcinogenicity and genotoxicity). Endpoints like allergies, hypersensitivity or intolerances are not covered by the approach (SCCS/SCHER/SCENIHR 2012, p. 42). So a substance might be positively evaluated by the TTC as having a low risk of systemic toxicity, but this would not tell us anything about other effects that the substance might exhibit below the threshold.

Another endpoint of major relevance is teratogenicity.⁹ Given the current state of research, little is known about the mechanisms of teratogenic effects, and it is commonly acknowledged that teratogenicity cannot be predicted based on chemical structure (Shibamoto and Bjeldanes 2009, p. 48). This poses a significant theoretical challenge to the structural similarity approach underlying the TTC, because even if there is strong empirical support for the assumption that structurally related substances behave similarly when it comes to acute toxicity or carcinogenicity, the same does not hold for teratogenicity. It is also controversial whether endocrine-disrupting potential can be inferred from a substance’s chemical structure given our limited knowledge of endocrine disruption mechanisms (WHO 2012, p. 35).

A further reason for concern is the growing evidence for the existence of low-dose effects and non-monotonic dose-response relationships. There is epidemiological evidence that shows a drastic increase in incidences of reproductive diseases and endocrine-related disorders over the past decades, such as male infertility, genital malformations, neurobehavioral disorders, and endocrine-related cancers. As an example, Fig. 3 shows the increases of testicular and breast cancers across Europe. The rate at which these increases have occurred cannot be explained by genetic factors alone. There is strong evidence that environmental factors including exposure to endocrine disrupting chemicals (EDCs) play an important causal role and that low-dose effects and non-monotonic dose-response relationships seem to be remarkably common in this context (Vandenberg et al. 2012). A WHO report on the state of the science of EDCs concludes that “disease risk due to EDCs may be significantly underestimated” and that “[n]umerous laboratory studies support the idea that chemical exposures contribute to endocrine disorders in humans and wildlife”, adding that “[t]he most sensitive window of exposure to EDCs is during critical periods of development, such as during fetal development and puberty” (WHO 2012, p. 3). For instance, a recent study has shown that children might be especially susceptible to an exposure to ethylhexyl methoxycinnamate (EHMC), which is used as a UV filter in suncreams and is known to have thyroid-disrupting effects. During summer periods exposures can be as high as 873 μg/kg bw/day (Manová et al. 2015).

All these problems boil down to one crucial question: Can we be confident that the health risk from substances below the TTC is really negligible? We should expect serious doubts to arise in this respect if we were to find even a few examples of substances which can be proven to be toxic below the threshold dose. And these examples do in fact exist. Take for instance Perfluorooctanoic acid (PFOA) and Bisphenol A (BPA). Application of the TTC would result in an allocation of these compounds in Cramer class III and, accordingly, their threshold (in animals, without human safety factor) would be at 150 μg/kg bw/day (see Fig. 2). Studies reveal, however, that PFOA has effects in animals at doses of 10 μg/kg bw/day and lower, and that “[d]ue to the low-dose sensitivity [...] to PFOA in CD-1 mice, a no observable adverse effect level [...] was not identified” (Macon et al. 2011, p. 134). Bisphenol A also has wide-ranging effects in particular on the male reproductive tract. It also has long-lasting developmental effects on the brain of test animals (Richter et al. 2007; see also: Newbold et al. 2009; Welshons et al. 2006; Vandenberg et al. 2012). Estimates say that BPA exposure from food sources is between 0.4 and 5 μg/kg bw/day (Vandenberg et al. 2012; see also Vandenberg et al. 2007; EU Scientific Commitee on Food 2002). In 2012 EFSA launched a re-evaluation of Bisphenol A focussing on exposure and possible low-dose effects. Based on this re-evaluation EFSA’s experts considerably reduced the tolerable daily intake (TDI) of BPA from 50 to 4 μg/kg bw/day.¹⁰ Furthermore, epidemiological studies carried out since 2006 have shown that exposure to low-dose neurotoxicants is the cause for the rise in prevalence of many neurodevelopmental disorders in children, and it is assumed that many more neurotoxicants remain undiscovered (Grandjean and Landrigan 2006, 2014).

In light of these results, it has to be acknowledged that there exist several examples of substances that show large-scale health effects at very low-doses and far below the 5th percentile of the NOEL of their respective Cramer class. It has been questioned whether threshold concepts are feasible at all when it comes to substances that show significant systemic effects at very low doses and that may exhibit significant non-monotonic dose-response relationships. This also poses a challenge to the age old paradigm in toxicology according to which it is the dose that determines the poison. Although the paradigm clearly has its value when applied to acute toxicity, the increasing evidence for low-dose and non-monotonic effects suggests that for many toxic substances, non-threshold concepts would be more appropriate.¹¹

To be sure, a few counterexamples do not undermine an approach that claims no more than to be right in most but not all cases, as the TTC does. On the other hand, it has become clear that the TTC is essentially unable to answer the question as to whether and to what extent an individual substance poses a threat for human health. Only extensive empirical investigations could do so. The TTC merely provides probability based risk estimates for classes of structurally similar substances. It therefore only provides a very rough probability measure of whether a substance might have a toxic effect or not.

Current and Future Uses of the TTC

The TTC approach has been used to evaluate flavouring substances by the Joint FAO/WHO Expert Committee on Food Additives (JECFA) and the European Food Safety Agency (EFSA); for food contact materials by the US FDA; for genotoxic impurities in pharmaceuticals by the European Medicines Agency (EMA); and for the risk assessment of chemicals by the World Health Organization’s International Programme on Chemical Safety (WHO IPCS) (SCCS/SCHER/SCENIHR 2012, p. 10). In 2011, EFSA’s Scientific Committee discussed and recommended using the TTC for areas including: food contact materials; impurities and breakdown/reaction products in food; plant metabolites and degradates of pesticides; metabolites of feed additives formed in target species that are not covered by tests in laboratory species; technological feed additives; flavouring substances in feed; and trace contaminants in food. The Scientific Committee also suggested some modifications of the original decision tree, such as a single TTC value for Cramer classes II and III of 90 μg/person/day.¹² The opinion of the Scientific Committee was based on the suggestions made by a TTC Working Group consisting of 13 members. It should be noted that in March 2014, the European Ombudsman found that some of the members of the working group had a conflict of interest due to links with industry and industry-founded research institutes such as ILSI. Balanced stakeholder representation in external meetings and workshops was not ensured. This was judged as an instance of maladministration on the part of EFSA. The Ombudsman concluded that the suggestions of the Working Group might be seen as biased towards the TTC.¹³

In December 2014, EFSA and WHO held a joint expert meeting, which included a stakeholder hearing on the first day of the meeting. The preliminary report of the meeting concludes that the TTC is a “valid, science-based screening tool useful for the prioritisation of chemicals and for more general applications in chemical risk assessment” (EFSA/WHO 2015). Although the question as to whether the TTC can take into account non-monotonic and low-dose effects is considered relevant, the report concludes that this question belongs to a set of “generic questions in the risk assessment of chemicals that are under discussion in the scientific community”, and that the question is therefore not specific to the TTC (ibid., p. 5). Hence the expert group leaves the question unanswered. As mentioned, NGOs and industry stakeholders were only present during day one of the meeting.¹⁴ The report therefore does not necessarily represent their opinions. Several participants publicly raised serious concerns about the TTC that are not mentioned in the report. The criticisms mainly concern the quality of the data used (parts of which are 50 years old) or the fact that the TTC fails to recognise that current science provides sufficient reasons to believe that biological systems can be responsive to very low doses of chemicals in their environment.¹⁵

Risk, Uncertainty and the TTC

The TTC is defined as a probability-based risk assessment tool. It is generally admitted though that “it is important for both risk assessors and risk managers to keep in mind that it [...] may have additional uncertainty” (SCCS/SCHER/SCENIHR 2012, p. 40). Let us therefore take a look at how the TTC conceptualises risk and what kinds of “additional uncertainties” might be relevant for its application.

There is no single standard definition of risk. Colloquially “risk” refers to situations in which some undesirable event might occur. When we say that the risk of losing a coin toss is 50 %, we identify the risk with the probability of an event (qua its relative frequency). In decision theory and risk analysis “risk” is often taken as the product of the probability of an event and its negative utility, i.e. the degree of harmfulness of its effects. The latter notion of risk also applies, for instance, when the potential negative consequences of two technologies are compared (e.g. nuclear energy vs. coal). Although the probability of an accident at a nuclear plant is very low, nuclear power might turn out to be a riskier technology than coal because the consequences of a nuclear incident are far more drastic. Here “risk” does not merely denote the probability of an event, but rather refers to something that is also called the expectation value of the event. Thus conceived, the term “risk” offers a quantitative representation of the severity of a potential hazard (see Hansson 2005).

A decision made under risk presupposes that the outcomes of an action can be quantified in terms of probabilities and the (negative) utility of each outcome can be evaluated quantitatively. If we are unable to put probabilities on outcomes, or if the set of possible outcomes is unknown, or if we are not able to grade the consequences of these outcomes quantitatively according to their preferability, the decision is said to be made under uncertainty.

The distinction between risk and uncertainty (or ambiguity, as it is often called today) was famously introduced by the economist Frank Knight in 1921 in a book entitled Risk, Uncertainty and Profit. According to Knight, “risk” denotes a kind of uncertainty that is measurable in terms of probabilities, whereas “uncertainty” refers to an unquantifiable type of uncertainty. Knight was among the first to point out that the consequences that arise from the two kinds of uncertainty are fundamentally distinct. Following Knight, John Maynard Keynes also used the term “uncertainty” for matters about which “[...] there is no scientific basis on which to form any calculable probability whatever” (Keynes 1937, p. 214). The distinction between risk and uncertainty has become canonical not only in economics, but also in other fields of science, and in particular in decision theory.

In most real-life decisions we are usually confronted with a combination of risk and uncertainty. This means that even in situations where we are able to generate probabilities and quantify utilities, this is normally not sufficient for eliminating uncertainty altogether, because further uncertainties might be present, for instance uncertainties pertaining to unconceived outcomes or to the scientific models that were used to generate the probabilities. Clear-cut cases of decisions under risk, where all the probabilities and utilities are known, are more likely to be found in textbook examples than in real life. Very often these textbook examples refer to card games or gambling situations. The fact that real-life decisions usually take place in the face of non-negligible uncertainties led Hansson (2009, p. 423) to claim that “[l]ife is usually more like an expedition into an unknown jungle than a visit to the casino”. A visit to the casino is a risky enterprise. But an expedition to the jungle, besides being risky, also confronts us with a lot of uncertainty.

Hansson also believes that there exists a strong tendency in decision-supporting disciplines to tackle all decision problems as if reliable probability estimates were available. He calls this the Tuxedo Fallacy. It consists in treating all decisions “as if they took place under conditions analogous to gambling situations”. The fallacy is dangerous “since it may lead to an illusion of control and to neglect uncertainties that should have a significant impact on decisions” (Hansson 2009, p. 427). I hold that the TTC is guilty of the Tuxedo Fallacy.

In a situation where a potential health hazard is commonly acknowledged, it is a natural reaction of scientists and decision-makers to look for a way to determine the risk quantitatively. But all too often risk is simply reduced to probability. The TTC is a paradigm example in this respect. The TTC allows us to allocate a substance with unknown toxicity to one of three classes and to derive a threshold concentration below which the probability of an adverse effect coming from that substance is reasonably low. But in order to evaluate the actual risk of a substance, we would have to know not only the probability that the substance might have an effect at all, but also the kind and magnitude of the potential effect (e.g. whether the substance has a carcinogenic, teratogenic, or endocrine-disrupting endpoint). What is more, in order to obtain a quantitative measure of the actual risk of a substance, we would have to be able to quantify the consequence of its potential effect by assigning a number to it. But, as was mentioned before, we usually do not know what kind of effect a substance might have unless we submit it to actual toxicological tests. Whereas the TTC might be suitable to assess the probability of there being any effect at all, it is unable to predict, let alone quantify, the potential consequences of the effect.

Here is a hypothetical example: according to the TTC, 5 out of 100 substances allocated to Cramer class III are likely to have an observable effect below an exposure of 90 μg/person/day. Let us assume that the population of the United States is exposed to all five of them on a regular basis at doses below 90 μg/person/day. Let us also assume that each of the five substances has only a small effect, and that the effects differ in severity. One substance will cause a decrease in fertility in 1/100,000 exposed males due to its endocrine-disrupting activity, one will cause 10 cases of colon cancer in 10⁶ exposed individuals, one will have a teratogenic effect on the embryos of 1 in 10⁶ exposed pregnant women, one will cause occasional headaches immediately after exposure, and one will cause extremely rare cases of hiccups. In the population of the United States, this scenario would lead to roughly 1600 men with reduced fertility, 3200 cases of colon cancer, 320 birth defects, many occasional headaches, and very few cases of hiccups. Although these outcomes differ greatly in their severity, each of the substances will be treated equally by the TTC. Each will be evaluated as having a 5 % probability of an observable effect below the TTC value.

Of course, for a regulatory body that is confronted with the task of identifying those substances out of a vast number of chemicals, for which a comprehensive risk assessment is necessary, these missed cases might be within the acceptable range. However, for individuals, who suffer from one of the more severe effects above, the situation is rather different. Their evaluation of the actual risk might have been different from that of the regulator, because they would have evaluated their individual consequences rather differently. In addition, they were exposed to a risk, of which they had not been made aware and because they were not aware of the risk, they were not in a position to avoid it. By defining thresholds of concern, scientists are implicitly making a normative judgment about the risk that other people should be willing to accept.

This shows that it is important to distinguish between a notion of risk that defines “risk” merely in terms of the probability of an event, and a notion of risk which multiplies the probability of an event by a factor representing its consequences. The uncertainties pertaining to the latter are usually not measurable quantitatively, and even if they were, their evaluation would highly depend on the context and on the perspective of the person making it. By just providing rough probability measures based on highly uncertain exposure estimates, and by not taking into account the potential consequences that the exposure to a substance might have, the TTC falls prey to the Tuxedo Fallacy. It ignores uncertainties that should be taken into account when it comes to the question of how to deal with the numerous substances with unknown toxicological properties in our environment. What is more, the TTC does not take into account the severity of missed cases, i.e. substances that are “of low concern” according to the TTC but that do in fact have a toxicological effect. A proper risk assessment tool should be able to do so. It should be able to inform us about the severity of what might happen in cases where its assessments are mistaken.

Precaution and Robustness in Science

Science-based decision-making has to take into account the fact that scientific information can be incorrect or incomplete. This is especially relevant in the context of environmental or health related decisions. The so-called precautionary principle is one of the most important principles in environmental and health-risk-related decision-making. Its application is supposed to guide decision-makers in the face of scientific uncertainty. The Wingspread Conference in 1998 presented the following formulation of the precautionary principle: “When an activity raises threats of harm to human health or the environment, precautionary measures should be taken even if some cause and effect relationships are not fully established scientifically”.¹⁶ According to this formulation, the precautionary principle instructs us to refrain from actions that pose a potential threat to human health or the environment even if their harmfulness is not scientifically established beyond reasonable doubt. In short, the precautionary principle holds the following: in the face of scientific uncertainty and potential hazards, be cautious and choose the precautionary option!

The precautionary principle has been criticised in many ways. A frequent criticism is that it is overly conservative because it focusses predominantly on hazards. According to this line of argument, the principle ignores potential benefits of risky actions. Many novel risky technologies do indeed have beneficial effects, which are sometimes as hard to foresee as are its potential threats. Hence the precautionary principle might have an inhibitory effect on innovation. Furthermore, it has been claimed that the precautionary principle collides with other principles of decision-making. Peterson (2006, 2007) states an impossibility theorem showing that the precautionary principle cannot be reasonably applied to a decision that may lead to harmful outcomes given certain other principles of decision-making.¹⁷

Based on arguments that try to establish its decision-theoretic incoherence, there is a growing tendency in the literature to find other explanations for the meaning and the import of the precautionary principle. Along the lines of Peterson, Sprenger (2012) also argues that it might be misleading to conceive of the precautionary principle first and foremost as a decision-guiding principle, i.e. a tool that helps decision-makers to arrive at reasonable decisions in the face of potential hazards. As a principle of decision-making, it is applied at a stage when science has already provided everything it knows about the causal relationships underlying the potential hazards. If the scientific knowledge is incomplete or uncertain, decision-makers who act upon that knowledge have to apply the precautionary principle. In this reading, the precautionary principle is applied by decision-makers and not by the scientists who inform the decision making process. But, Sprenger asks, what if the precautionary principle were to be applied at an earlier stage? What if it were to be applied to the scientific model-building process itself? What if we think of the precautionary principle not as a decision rule but rather as an epistemic rule?

In this view, the precautionary principle is gaining substantial import for scientific model building, in particular for models pertaining to disciplines dealing with environmental or health issues. When applied to scientific models, precaution can be spelled out in terms of robustness. A model is robust if it bears comparison with alternative models and if an action licensed by the model would still be licensed if certain theoretical assumptions of the model turn out to be wrong. In climate science, for instance, it is standard practice to use multi-model ensembles in order to cancel out individual model biases. So instead of trying to find the best model, climate scientists often focus on averages across structurally different models. A climate model is considered robust if its predictions and conclusions are consistent with those of structurally different models. Potential errors of the individual models become dampened in the ensemble. Ensemble approaches are therefore epistemically more robust than single best model approaches. Hence, using ensembles instead of single best model approaches can be considered the precautionary choice. A decision based on a single, simplistic model would go against a precautionary attitude.¹⁸

Note that it is not at all common to interpret the precautionary principle as an epistemic rule. That is to say that the precautionary principle is usually taken as a guiding principle for risk managers and not so much for scientists who develop the tools for the assessment of risks. The EU Communication from the Commission on the precautionary principle, for instance, highlights the relevance of the precautionary principle for decision-makers and explicitly states that it must not be confused with the epistemic caution that should guide scientists in their work:¹⁹

The precautionary principle should be considered within a structured approach to the analysis of risk which comprises three elements: risk assessment, risk management, risk communication. The precautionary principle is particularly relevant to the management of risk. The precautionary principle, which is essentially used by decision-makers in the management of risk, should not be confused with the element of caution that scientists apply in their assessment of scientific data. (Commission of the European Communities 2000, emphases added).

Taken as an epistemic rule as opposed to a rule for risk management, the precautionary principle instructs scientists to build models that are robust in the above sense. Now what would this mean if applied to the TTC? Can the TTC be considered robust and thereby satisfying the requirements of an epistemic version of the precautionary principle? The answer is a clear no. And the reason for this is very simple. There are no viable alternatives to the TTC. Until now, it is the only tool that science has to offer to decision-makers when they have to deal with the problem of the many thousand low-dose substances in our environment. And since it is the only approach available, it is, by the force of logic, also the best approach available.

It might be objected at this point that the TTC does not qualify as a scientific model in the strict sense insofar as it does not model the causal relationships in a target system in order to predict the value of a given variable, as it is the case in climate models. Rather, the TTC merely provides probabilities based on statistical data. Unlike physical models, which aim at a representation of causalities in real-world systems, statistical risk assessment tools do not have to be considered in ensembles. The approach that is the most sophisticated with respect to the data it integrates, can be considered the best approach, and the TTC satisfies this criterion when it comes to the assessment of intentionally and non-intentionally added substances in food products.

In response to this objection, it is important to keep in mind, however, that the TTC is indeed based on several causal assumptions. The most important of these is the assumption that structurally similar substances have similar toxicity. As a general heuristic this might indeed be true. But as was mentioned above, for certain toxicological endpoints it does not hold. For all we know, teratogenicity cannot be inferred from chemical structure. Furthermore the TTC assumes that non-monotonic and mixture effects can be ignored during the first step of risk assessment. This is justified by the lack of better knowledge: because we do not have enough knowledge about these effects, we have to settle for an approach that ignores them. The application of epistemic precaution would force us to draw the exact opposite conclusion. We do not have enough robust knowledge about the mechanisms of low-dose or mixture effects and non-monotonic dose-response relationships, so we should not ignore them, because we do not know what the consequences might be.

As already mentioned, as of now the TTC is the only viable tool that science provides to policy-makers. The application of the TTC will lead to little action on the side of policy-makers because most non-intentionally added substances in food products and food contact materials will fall below the recommended threshold values. Regulating bodies will legitimately justify their lack of action with the fact that their decision is based on the best available scientific data. Their decisions will be in accordance with the precautionary principle because they are licensed by the best (qua only) available scientific approach, which tells them that their decisions are conservative. Therefore, decision-makers who act on the basis of the tools with which science provides them can legitimately claim to be acting precautiously.

However, if we shift the scope of the precautionary principle down to the level of the scientific model-building process, we realise that the scientific experts, who are involved in the development of the TTC, do not comply with the requirements of the precautionary principle in an epistemic sense. For them to exhibit a precautionary attitude would mean to engage in contrafactual reasoning and to ask whether the actions licensed by the TTC would still be licensed if one or more of its theoretical assumptions turned out to be wrong, for instance if we discover that there exist large-scale low-dose or mixture effects, or that non-monotonic dose-response relationships have a more significant impact than we first thought. It would also mean making explicit the potential harmful consequences of a wide-ranging application of the TTC, even if there are doubts about whether these effects will in fact occur. Epistemic precaution precludes scientists from exploiting the lack of knowledge as a justification for the feasibility of an approach.

Insisting on the application of epistemic precaution is also a way of nudging scientists to place the common good above special economical or political interests. As we have seen above, many of the authors arguing in favour of the TTC are either funded by or at least professionally related to industry, which, if nothing else, increases the plausibility of the assumption that their endorsement of the TTC is not completely free form partisan interests.

Invoking precaution at the level of scientific model building also emphasises the role of non-epistemic values as an integral part of scientific reasoning. As Douglas (2000) convincingly argues, non-epistemic values should be taken into account by scientists in cases where taking inductive risks includes the risk of non-epistemic consequences.²⁰ In other words, when the risk of scientific error is sufficiently large, and when errors can have severe non-epistemic consequences, scientists should be guided by non-epistemic values when they choose methodologies or interpret results. In cases where inductive risk is linked to non-epistemic consequences, the judgement of scientists is incomplete without the consideration of non-epistemic values. Applying precaution at the level of scientific model building is therefore an instance of the application of non-epistemic values in science.

The link between an epistemic precautionary principle and the role of non-epistemic values in science has also been discussed by Steel (2015, ch. 7).²¹ Based on the argument from inductive risk and the rejection of the value-free ideal, Steel claims that “methods for assessing the probability of risk and its severity can be influenced by ethical values, such as those embodied in PP” (Steel 2011, p. 362). He also argues that the uncertainty factors used in toxicological risk assessment are a genuine instantiation of the precautionary principle (Steel 2011; 2015, ch. 8). Seen from this perspective, the TTC would in fact meet the requirements of an epistemic version of precaution, simply because it operates with uncertainty factors. Note, however, that the above argument against the TTC need not necessarily be in opposition to Steel’s claims. Even if uncertainty factors are seen as an illustration of the precautionary principle, the TTC as whole could still fail to comply with an epistemic version of the principle because of the additional uncertainties that go along with the application of the approach. So even if uncertainty factors represent a precautionary aspect in environmental risk assessment, the application of uncertainty factors might not be sufficient for an approach to qualify as precautionary in an epistemic sense. Quite on the contrary, the use of quantitative uncertainty factors can even be seen as another instance of the Tuxedo Fallacy, because, by providing a quantitative measure for the uncertainty pertaining to the extrapolation from animal data to human threshold values, it might actually generate the illusion of precaution rather than provide epistemic robustness. It should also be noted that Steel’s arguments are not so much aimed at supporting an epistemic version of the precautionary principle; rather, they are directed against those who argue that the principle is pertinent only to unquantifiable hazards. The criticism expressed against the TTC in the above considerations does not depend on this claim.

Conclusion

I have discussed the TTC approach in food toxicology in order to demonstrate that scientists should be guided by precautionary principles when they develop models and risk assessment tools that are used in decision-making. That is to say that precaution should also play a role as an epistemic principle and not merely as a decision rule. The TTC was used as a negative example in this respect. Because the approach does not provide even a qualitative assessment of potential negative consequences of its application, it fails the requirements of precaution in an epistemic sense.

Following Sprenger (2012), I have argued that epistemic precaution is related to model robustness. Taking a precautionary attitude as a scientist means to build robust models. Model robustness cannot be universally defined. Epistemic robustness in climate science is not the same as epistemic robustness in high-energy particle physics. In the context of the TTC, robustness would mean taking into account potential negative health effects coming from low-dose chemicals in the environment, even if the existence and the potential mechanisms of these effects are a matter of ongoing research. The TTC fails to do so. Scientists who argue in favour of the TTC are not acting precautiously in an epistemic sense, because they use the insufficiency of our knowledge regarding low-dose, non-monotonic, or mixture effects as a justification for ignoring them in the risk assessment. A large-scale application of the TTC would lead to a deregulation of a vast number of low-dose chemicals in our food products despite the fact that we do not know what their potential health effects are. The TTC therefore serves special groups interests more than it serves the public good. The precautionary alternative would be to regulate the use and emission of chemical substances with unknown toxicity in our food products and food contact materials until science proves their safety beyond reasonable doubt.

Although the discussion in this paper had a strong focus on toxicology, the claim that precaution should be applied at the level of model building holds for other disciplines too, in particular if they have consequences for decision-making. In the case of food toxicology, the decision-relevant implications are obvious. In other fields of science, such as physics, engineering, or economics, these might be less immediate. Nevertheless, the same holds for scientists working in those fields: they too should consider and communicate the potential negative consequences of the application of their models even if the occurrence of those consequences is uncertain. This should already be done while the models are being developed, and not only when they are applied by decision-makers.