The
Journal of Philosophy, Science & Law
Volume 11, April 4, 2011
www.miami.edu/ethics/jpsl
Scientific
Evidence and the Law: An Objective Bayesian Formalization
of the Precautionary Principle in
Pharmaceutical Regulation
Barbara Osimani,* Federica Russo,** Jon Williamson***
* Università Cattolica del Sacro
Cuore, Milano
** Philosophy, University of Kent
*** Philosophy, University of Kent
Abstract The paper considers the legal tools that
have been developed in German pharmaceutical regulation as a result of the
precautionary attitude inaugurated by the Contergan decision (1970). These
tools are (i) the notion of “well-founded suspicion”, which attenuates the
requirements for safety intervention by relaxing the requirement of a proved
causal connection between danger and source, and the introduction of (ii) the
reversal of proof burden in liability norms. The paper focuses on the first and
proposes seeing the precautionary principle as an instance of the requirement
that one should maximise expected utility. In order to maximise expected
utility certain probabilities are required and it is argued that objective
Bayesianism offers the most plausible means to determine the optimal decision
in cases where evidence supports diverging choices.
1. Precautionary attitudes in
response to uncertain knowledge
Historically, the precautionary principle
arose in response to the lessons learnt from environmental disasters and
injuries to human and animal health caused by chemical compounds (x-ray
radioactivity, benzene, asbestos, PCB, halocarbons, DES sulphur dioxide, etc.).
These tragedies could have been avoided if signals of alarm had been taken more
seriously. This hindsight, together with the increasing awareness of the
unpredictability of environmental and health effects created by the chemical
industry, stimulated the development of juridical instruments for the
management of lack of knowledge—see, for instance, the work done by the
European Environmental Agency (2001). These instruments are meant to be able to
enlarge the powers of intervention for authorities against sources of possible
harm, even in the absence of scientific proof of a causal link. Causal link is the central notion in
this context because it is the basis both for action on the one hand, and for
responsibility attribution on the other hand.
The first programmatic documents advocating
administrative intervention before a causal connection be scientifically
established were developed in relation to North Sea pollution. The first
Conference for the Protection of the North Seas (Bremen, 1st November, 1984) asserts that the States “must not wait for proof of harmful
effects before taking actions”; the second Conference (London, 24-25 November,
1987) makes reference to the precautionary attitude and insists that “[...] a
precautionary approach is necessary which requires to control input of such
substances even before a causal link has been established by absolutely clear
scientific evidence” (art. VII). The Earth Summit in Rio de Janeiro, 1992, ratified
the precautionary principle and extended it to the global environment. The key
feature of the precautionary principle thereby ratified was the reversal of the
burden of proof between intervening authority and potentially polluting agent.
It is not the authority that must demonstrate that some human activities cause
serious harm to the environment in order to be allowed to adopt adequate
preventive measure; on the contrary, in order to postpone these measures, it
must be proved that these activities do
not cause any serious harm to the environment (principle 15).
In general, the precautionary principle
responds to the needs of a complex society where uncertainty is endemic. Dupuy (2004: 80) even uses the term
“radical uncertainty” (see also Tallacchini, 2005 and 2008 on this point).
Science is called on to accomplish two interconnected tasks. The first is the traditional endeavour to
increase knowledge about reality and its physical, chemical, biological
principles; the second one is the
prediction of the effects that the applications derived from this knowledge
produce on the environment, as well as human and animal health (see for
instance toxicological sciences). This second task is enormously more complex
than the first; in fact, by studying the mechanisms of a phenomenon, the
scientist focuses on a particular aspect of reality and on the potential
consequences that technological interventions can produce (jointly or
individually) on the environment. Such effects are by and large unpredictable for
two reasons. First, nature is an integrated system where any “external” action
can produce domino effects at any level; second,
it is practically unfeasible to exhaustively detect all dependencies,
independencies and interference relationships working in nature. Therefore, the
greater the progress in science, the exponentially greater the uncertainty
generated by lack of knowledge about the potential effects of its technological
applications. Not only sociologists (Beck, 1986), but also ethicists (Jonas, 1979),
and philosophers of science (Hacking, 1986) have emphasised a new
epistemological era where the unknown and lack of information should turn
into topics of research on their own.
In the pharmaceutical context, the
multiplicity and sophistication of medical technologies has reached such a
level that the term “medicalization of society” has been coined in order to
describe the pervasiveness of health care at all societal levels (Zola, 1972;
Domenighetti, 2005). After a first phase of unlimited confidence in medical
progress, a cautious attitude followed the numerous pharmaceutical scandals
that have marked the history of pharmacology. For instance, the tranquillizer
Contergan© (thalidomide), marketed in Germany between 1957 and 1961, caused
severe birth defects to more than 6000 children—mainly produced by drug
inducted phocomelia—and fatally injured 2500 people. This and other tragedies
(see the Cronassial© case in Germany, but also worldwide marketed products such
as Lipobay©, Vioxx©, and Bextra©) have contributed to enhance efforts towards
the development of strict pharmaceutical safety regulation grounded in the
criterion of the precautionary principle, which has led to an extension of care
duties for dangerous entities, to an
enlargement of intervention powers for the authority in charge, and to an
amplification of the responsibility spheres for all concerned parties.
Notwithstanding these efforts directed at
making cautious decisions to license drugs, cases of product retirement are
more common than it may be thought, and proposals for improved monitoring
systems are regularly advanced in the literature (see for instance Olivier and
Montastruc 2006; Gassner and Reich-Malter 2006; Laupacis et al. 2003; Waller
and Evans 2006; Talbot and Nilsson, 1998). As a matter of fact, it is not rare
that pharmaceuticals are withdrawn from the market only too late, i.e. when
extensive damage on the population of users has already been produced, which
happened, for instance, with the above-mentioned products.
Various factors are at the origin of the
discontent about how pharmaceutical decisions are taken both by responsible
authorities and by the pharmaceutical industry (see for instance Reiss and
Kitcher, 2008; Abraham and Davis, 2005; Abraham and Reed, 2001; Demortain,
2008): the complexity and inconsistency of data documenting drug efficacy and
risks, the conflict of interest affecting the principal investigators of
chemical entities and information deliverers (pharmaceutical sponsors), as well
as time pressure in the approval procedure. Besides incentives and deterring
instruments aimed at more transparent and safer pharmaceutical marketing (for
instance through fiscal and financial regulation), formal instruments are
needed in order to provide clear guidelines for the application of the legal
norms developed in the area of safety protection and liability attribution.
[3]
2. The Contergan decision:
origin of the precautionary attitude in pharmaceutical regulation
Drugs have a Janus character as healing
promise and poison at the same time. Debates about pharmaceutical products
focus on one and the same principle: health as an individual and societal good,
which drugs contribute both to promote and to endanger. Pharmaceutical
regulation is the consequence of a growing awareness about the uncertainties
surrounding the short, medium and long-term effects of chemical entities on the
human organism.
In the Sixties and Seventies, the trial and
decision about Contergan set the basis for thorough reflection on the specific
epistemic status of pharmaceutical knowledge in relation to the health risks
posed by pharmaceutical products. The Contergan trial has been one of the
longest in the history of German jurisprudence. The Contergan sentence was
rather inconclusive with regard to imputing responsibilities and torts to the
defendants: it ended up with a suspension of the trial by invoking § 153, abs 3
of the Strafprozessordnung – code of criminal procedure.
[4]
Yet, it settled the future standard of
conduct for pharmaceutical firms and their employees, thereby deeply changing
the framework of responsibilities concerning the management, disclosure and
possible prevention of pharmaceutical risks (see below). More recently, the IInd amendment law for compensation and the 2002 amendment to the German Medicines
Act have further developed these sorts of considerations — more on this later
on.
The Contergan sentence (18 December 1970)
contributed to the development of a precautionary attitude in pharmaceutical
regulation by (i) increasing the responsibility scope for pharmaceutical
sponsors through the introduction of strict liability for pharmaceutical
products (until then, only tort liability required the assessment of negligence
for responsibility attribution), and (ii) the principle of well-founded
suspicion. The Contergan sentence establishes that, because positive proof of
damage causality requires time, a large epidemiological basis, and can never be
definitively assessed, a scientific proof of causality cannot be a valid criterion for determining the threshold of safety
countermeasures:
[5]
Before a
risk suspicion can be founded scientifically, enough time may pass as to
produce damage in some consumer. During this vacillation time, the risk has to
be undertaken by the pharmaceutical firm. Moreover, for the principle of
inverse proportionality inherited from danger prevention regulation, even very
low probable suspicions ask for timely countermeasures.
In fact, the principle of inverse
proportionality says that the higher the value assigned to the endangered good,
the lower the probability of damage for approving intervention in its defence.
General criteria for determining the opportunity and entity of safety actions
are therefore:
1.
Severity of the suspected health damages: the more
dangerous the drug is supposed to be, the earlier and prompter must the firm
react to risk news concerning the product;
2.
The nature of the damage: irreversible damage requires quicker
reactions than transitory disturbances;
3.
Side effects’ frequency: the higher the observed frequency, the lower the
suspicion needs to be in order to require safety countermeasures from the firm
and/or the responsible authority;
4.
Therapeutic
importance:
the higher the interests of patient in the availability of the drug, the higher
the permissible risk in allowing the drug to stay on the market. Therapeutic
importance is determined by the lack of alternatives and the severity of the
illness.
General safety measures include the adoption
of warning actions towards the health professional and the end-user, the
introduction of stricter prescription requirement, up to product retirement
from the market.
[6]
3. Legal instruments of safety
protection in the German Medicines Act
The German Medicines Act (AMG) was the first
law in western countries to translate the precautionary principle into concrete
and legally binding norms. It was enacted in 1976 and is the result of the
political debate that followed the Contergan tragedy and of the implementation
of the European directives 75/318/EEC and 75/319/EEC that also followed this
pharmaceutical catastrophe.
[7]
Since 1976, AMG has undergone 14 amendments
(the first one being ratified in 2005, while a fifteenth amendment has been
drafted in 2008-2009 and is still under approval). The explicit purpose of AMG is the safety of
drugs administered to the public through the establishment of criteria for the
efficacy, quality and safety (“Unbedenklichkeit”) evaluation of candidate
drugs.
[8]
The term “Unbedenklichkeit” refers to a
safety judgment based on a risk/benefit assessment and is explicitly translated
with the English term “safety” in the European regulation (Scheu, 2003). It
warrants for safety through drug approval, surveillance and liability norms.
Risk prevention is managed through two control systems: drug approval and
post-marketing control. It is worth noting that drug approval status is, by
default, prohibition, with reserve of permission (“Verbot mit
Erlaubnisvorbehalt”). Liability norms also constitute an indirect incentive to
safety beyond their principal compensatory aim.
The precautionary principle is embodied in
the AMG through two articles that received much attention in the legal
literature: article 5 (prohibition of unsafe medicines: “Verbot bedenklicher
Arzneimittel”) and article 84 (strict liability: “Gefährdungshaftung”). In the
following, we shall focus on article 5.
Article 5 of AMG establishes safety criteria
of drug circulation and withdrawal.
[9]
This norm stipulates prohibition of
circulation for “unsafe drugs” and provides a definition thereof: unsafe drugs
are those for which there is well-founded
suspicion that, by adequate use, will have damaging effects exceeding a
tolerable threshold, according to the knowledge of the
medical science (§ 5 II) on the basis of available scientific data.
[10]
This norm has three main components:
1.
The degree of causal association between risk
and danger source required for intervention need not be certain (“well-founded
suspicion”);
2.
The level of causal association required for
intervention is linked to the tolerance threshold, i.e. to the (un)balance
between drug risk and benefit (and the related notion of “residual risk”);
3.
The tolerance threshold is established
through reference to the state of the art of relevant medical knowledge.
In sections 3.1 and 3.2 we will analyse the
first two criteria. For the sake of argument, we will take the third component
as unproblematic in this context. In sections 4 and 5 we will outline the
general features of Bayesian epistemology and propose an objective Bayesian
formalisation of the principle of well-founded suspicion. It is worth noting
that although the legal setting considered here is the German one, risk
management strategies are indeed very similar across different geographic areas
such as Europe, U.S. and Japan. Indeed, responsible authorities and
pharmacovigilance agencies are engaged in a continuous effort towards the
harmonization of procedures and policy and of related legal tools (see the
International Harmonization Conference
[11]
). Therefore, our arguments, although based
on the German setting, are not strictly limited to it and can indeed contribute
to these harmonization efforts.
3.1 The principle of well founded
suspicion
Because drug reactions are idiosyncratic and
depend on several environmental, biological, and genetic factors, knowledge
about the effects of any drug grows with the number of its users. This means
that even many years after approval, any pharmaceutical is still an
“experimental product”, the information about which is neither exhaustive nor
conclusive. As the Health Minister Dr. Focke declared in the ministerial
statement for the provision of the German Medicines Act, drugs are products
under constant testing (“Arzneimittel sind Produkte in Dauererprobung”)
[12]
. The general recognition of the limited and
fragmentary knowledge related to chemical and pharmaceutical technologies has
contributed to the awareness that criteria for the management of partial and uncertain knowledge are
needed.
The concept of “well-founded suspicion” has
been introduced into risk-management regulation, in order to decrease the
threshold level for signal detection and alerting measures. This instrument represents an answer to the
opacity associated with the pharmaceutical product: given that knowledge of
possible unintended effects is limited and that the causal nexus can seldom be
proven, waiting for the causal connection to be established before intervening
would most times lead to late intervention and irreparable damage. The
historical importance of the principle of well-founded suspicion is related to
its fundamental role in moving pharmaceutical regulation from a “danger
avoidance” system into a “risk prevention” (precautionary) system.
Both systems function according to the
principle of inverse proportionality: the more severe the expected damage, the
lower the probability of its occurrence need be in order to intervene. The
difference between the danger avoidance and the risk system lies in the kind of
causal link between danger source and damage required for action. Whilst in the
danger avoidance system a causal connection between danger source and damage
needs to be established with certainty before the authority can intervene, in a
risk prevention system it suffices to have a suspected causal connection between danger source and damage.
Notice, moreover, that the principle of well-founded suspicion integrates the
principle of inverse proportionality too: the higher the expected damage, the
lower the probability of causal
connection in order to require for intervention measures. In fact “well-founded suspicion” is defined
as “hypothesis of causal connection”
(Scheu, 2003: 113) and, as such, it is quantified in probabilistic terms, i.e.
the probability P that the drug (D) causes harm (H) can be less than 1 (
), that is less than certain. In other words,
the more severe the harm, the lower
need
be.
This means
that the authority and industry are not justified in not intervening because a
causal connection between damage and source has not been conclusively
established, i.e.
. Instead, they are supposed to act as soon as
the probability of a causal connection is sufficiently high with respect to the
potential harm in relation to the potential benefit.
A major problem with the principle of well
founded suspicion, however, is that neither a practical nor a formal rule has
been defined in order to provide standards of conduct and accountability
criteria for pharmaceutical marketing and policy. Differently put, no formal
guidance is given as to how and where to set the relevant P-threshold.
Indicators of suspicion are rather vague and
prone to a biased interpretation, as the facts supporting the suspicion need
not necessarily be concrete cases of damage. Also, the acquisition of new
substantial theoretical knowledge can be a ground for risk suspicion,
especially when there is little or no experience with the drug that could
refute the theory (Di Fabio, 1993: 126-127). New theoretical insights might
contribute to deeper pharmacological understanding and favour or contradict
established knowledge about the effects of a specific substance (Di Fabio,
1993: 126-127; Räpple, 1991: 90-91); also, suspicion about potential damage
begins as soon as a doctor assesses an association between a side effect and a
drug (Di Fabio 1993: 125).
Indeed, the above-mentioned cases of
unjustifiably late product withdrawals testify that more detailed yardsticks
should be provided to the responsible authority and industry so as to decrease arbitrariness
in decision making and establish stronger accountability constraints.
[15]
Furthermore, it is advocated that in such a
complex field as pharmacology, decisions be taken on the basis of all available evidence. A useful
statistical paradigm for the integration of data coming from heterogeneous
sources is constituted by the objective Bayesian methodology, to be discussed
later in the paper (Williamson 2010b). This paradigm can also be used for
integrating knowledge of different experts, i.e. as a knowledge-integration
platform (see Hughes et al., 2007, Cowell, 2007).
3.2 Risk-benefit assessment
Given the ambiguous character of
pharmaceuticals and the consequent impossibility of absolute safety, the
evaluation of drugs cannot result in a distinction between riskless and harmful
products, but rather between an acceptable (“zumutbar”) and an unacceptable
(“unzumutbar”) risk. A risk tolerance
threshold is established such that, below the threshold, the drug is
considered “safe”.
This threshold is relative to the benefit
expected from the drug through a risk-benefit evaluation: this decides how much
risk is to be accepted in the face of how much benefit. The risk-benefit
assessment is made on the basis of known risks and benefits, therefore the proportion of
ignorance surrounding the drug (epistemic uncertainty) is only indirectly
relevant here. The risk-benefit evaluation is affected by “ecological”
uncertainty
[17]
in the sense that pros and cons should be
weighed against each other, and sometimes they are both equally strong. The
main implication of this procedure is that, if the drug is approved, then the
related risk is considered part of the bargain: this is called “acceptable”,
“tolerable”, “unavoidable” or “residual”.
The necessary condition for market approval
is a positive result of risk-benefit assessment in “absolute” terms—that is
when no other drugs in the market compete with the candidate drug—as well as
“relative to” the pharmaceutical environment—that is in relation to the
treatments already present in the market for the same indication.
[18]
This evaluation is based on a comparative
weighting of therapeutic importance and efficacy on one side, and of risk
severity and frequency on the other. The definition of risk traditionally
adopted by safety regulations has been inherited from natural sciences and
engineering and consists in the product of the two dimensions of damage—severity and probability—where the damage is any injury caused to goods
protected by the law (Räpple, 1991: 49).
In decision-theoretic terms, drug approval
can be formalized as follows:
.
The expected utility (EU) of drug approval (D)
should be higher than that of drug refusal (¬D), where both utilities are computed out of the formulae below.
The
expected utility of drug approval is
the sum of the utility times probability products for all relevant attributes {i1, … in} associated with the drug
(benefits, ADRs). Attribute utilities have a positive sign for the drug
benefits and a negative sign for the adverse drug reactions.
The
expected utility of drug rejection is
the sum of the utility times probability products for all relevant attributes {j1…jn} associated
with this option: the negative consequences of not treating the illness with
the drug on one side, and the avoidance of drug side effects on the other.
For any drug to be approved, the expected
utility associated with it must be superior to that of not approving it. The
ecological uncertainty affecting the risk-benefit assessment increases to the
extent that the difference among the inequality factors approaches zero; i.e.,
when EU(D) ≈ EU(¬D). This might be due to
compensatory attributes present in both the risk and the benefit side.
Whenever a new risk possibly associated to
the drug is detected (development risk), a risk-benefit assessment needs to be
made in order to determine whether this risk asks for intervention (for
instance access restriction, approval suspension or product withdrawal) or not.
If the risk-benefit balance remains favourable for the product, then the
detected risk can be considered irrelevant. Instead, if the newly detected risk
changes the risk-benefit balance so as to make it unfavourable, then
appropriate measures need to be considered. Provided that, especially in the
phase of signal generation, the causal connection between drug and ADR is
uncertain, risk prevention/minimization measures should be determined by taking
into account both the importance of the unbalance and by the evidence of
causality.
AMG § 5 prescribes that safety measures
should be undertaken whenever there is well-founded suspicion that by adequate
use, the drug will produce damaging effects exceeding a tolerable threshold: the
greater the unbalance, the lower the probability of association between
expected damage and candidate cause. However, this threshold cannot be just the
unfavourable balance in the risk-benefit, as this way the probability of the
hypothesis of causal link between danger source and damage—which is the
cornerstone of the principle of well-founded suspicion—does not enter the
decision to intervene or not.
In the following two sections we propose that
the precautionary criterion for safety intervention and pharmaceutical policy
stated in § 5 of AMG be translated in the terms of the maximum entropy
principle within an objective Bayesian approach. To state it informally, the
principle of well-founded suspicion prescribes that the decision-maker believes
in the existence of a causal link between drug and damage proportionally to the
positive unbalance of the risk-benefit assessment. In other words, the higher
the unbalance of therapeutic benefit against suspected damage, the stronger the
belief that there is a causal link between drug and damage. Thus, it is not the
result of the risk-benefit analysis alone that triggers action, but the belief in a (suspected) causal link between drug and damage, which is based on a
risk-benefit analysis.
4. Bayesian epistemology to
rescue?
In this section we suggest that Bayesian
epistemology offers promising conceptual and formal tools for the principle of
well-founded suspicion. As will become clearer in the discussion, the advantage
of Bayesian epistemology, and in particular objective Bayesian epistemology, is
twofold. First, objective Bayesianism allows us to deal with probabilistic
inferences, particularly with those that are supposed to trigger action, such
as pharmaceutical decisions. Second, objective Bayesianism reflects the
precautionary stance that has to accompany causal attribution (for instance
about the danger of a drug) and the decisions to be taken as a consequence of
such attribution.
4.1 A crash course in Bayesian
epistemology
Bayesianism is an epistemological position
concerning scientific reasoning, but also reasoning more broadly construed.
[19]
The core of Bayesianism has been formulated
in the framework of the formal theory of probability. Thus, the two main
assumptions behind Bayesianism are that (i) aspects of scientific reasoning,
for instance, confirmation of a hypothesis or of a theory, can be quantified
and constrained by the formal principles of probability theory; (ii)
Bayesianism provides an account of how we should learn from experience. The
formal apparatus of probability theory
[20]
serves to impose coherence constraints on rational
degrees of belief and, typically, uses conditionalisation as a fundamental
probabilistic inference rule for updating probability values according to
Bayes’ theorem. Bayesianism is also taken to be a
methodology that allows inductive reasoning from data, that is, probabilities
of hypotheses in the light of data. Of course, not all reasoning in science is
formalised in terms of probability theory nor is all reasoning inductive in
character. The extent to which the core assumptions of Bayesianism really grasp
the essential features of scientific reasoning goes beyond the scope of the
present paper. What will be key for the following discussion is that
Bayesianism explicitly deals with uncertain reasoning (which is exactly what
happens in pharmaceutical contexts) and that, since it deals with how to learn
from experience, a particular version of it (objective Bayesianism) will be
useful when it comes to make decisions (such as in drug regulation).
Bayesianism, as an epistemological position
about scientific reasoning, is accompanied by an interpretation of probability in
terms of rational degrees of belief. Here, probabilities are quantitative
expressions of the strengths of an agent’s beliefs. Such an interpretation was
famously championed by de Finetti (1937) and Ramsey (1926), who analysed
probabilities in terms of betting behaviour: probabilities are identified in
terms of the betting odds that a rational agent is willing to accept. Without
going into unnecessary technicalities, it can be formally argued (by means of
the so-called ‘Dutch book theorems’
[22]
) that conforming to the probability calculus
is a necessary condition for rationality.
According to strictly subjective Bayesian epistemology, it is sufficient that an
agent’s degrees of belief satisfy the axioms of probability. Other than that,
the agent is free to adopt whichever degrees of belief she wished. The typical
objection is that this account leads to arbitrariness: two agents may assign
different probability values to the same event (given the same background
information) and be equally rational, provided that they do not violate the
axioms of probability. A solution to the objection of arbitrariness is
attempted by empirically-based and objective Bayesian epistemology.
In a nutshell, these two positions impose
further constraints on an agent’s degrees of belief before they can be deemed
rational. Early proponents were Salmon (1967) and Jaynes (1957). There are two
types of constraints: empirical and logical. Those
constraints amount to taking into account any information and lack of
information, when shaping degrees of belief.
Salmon emphasises the role of empirical
constraints and requires knowledge of relative frequencies to assign prior
probability values. This characterises empirically-based subjective Bayesian epistemology. The frequency interpretation is yet
another interpretation of probability, traditionally classed as an empirical or
physical interpretation of probability. Physical probabilities, unlike Bayesian
probabilities, take probabilities to be quantitative expressions of some
features of the world, not of our knowledge or belief about them. A simple form
of the frequency interpretation states that the probability of an attribute A in a finite reference class B is the relative frequency of the
actual occurrence of A within B. Further developments of the frequency
interpretation are due to von Mises (1928) and Reichenbach (1935), who
considered infinite reference classes and identified probabilities with the
limiting relative frequency of events or attributes therein.
Jaynes (1957) goes beyond this
empirically-based approach and puts forward a maximum entropy principle, which
might be thought of as an extension of the principle of indifference. This is known as objective Bayesian epistemology. Thus, whilst empirically-based
Bayesian epistemology contents itself with the adoption of empirical
constraints, i.e. knowledge of observed frequencies is sufficient to shape
degrees of belief, the objective Bayesian approach requires that both empirical and logical constraints
be satisfied. Also, although both the empirically-based and objective Bayesian
interpretations shape degrees of belief using knowledge of observed
frequencies, the two interpretations significantly differ in that the objective
Bayesian approach requires choosing the middling or most equivocal probability
value in case of lack of evidence (e.g., concerning observed frequencies; see
Williamson 2006). We shall explain objective Bayesianism in more detail in the
next section.
It is worth noting that Bayesian
interpretations, whether subjective, empirically-based or objective, interpret
single-case rather than generic probabilities. (On the other hand the frequency
interpretation of probability only makes sense of probabilities of generic or
repeatably-instantiatable outcomes.) In fact, degrees of belief are associated
with bets and a bet in a generic outcome does not make sense. This turns out to
be a useful feature of Bayesianism in general, because decisions in pharmaceutical
contexts are single-case and therefore a Bayesian interpretation of probability
ipso facto proves to be better suited than other approaches (notably,
frequentism).
4.2 Objective Bayesianism
Objective Bayesianism has been subject to
several criticisms since its inception by Edwin Jaynes (1957). The approach has
been criticised for a variety of reasons, ranging from the foundational and
motivational—e.g., it is hard to articulate how evidence constrains degrees of
beliefs—to more technical ones—e.g., a worry that using the maximum entropy
principle engenders serious computational problems. Recently, the approach has
been defended and developed by Williamson (2005, 2006 and 2010). The peculiar
features of objective Bayesianism, in Williamson’s approach, are the following. First, it doesn’t require a separate
updating rule, as probabilities can be determined afresh on each change of
evidence (though updates are often consistent with the results of
conditionalisation). Second,
probabilities are not fully determined by evidence—language and context also
play important role in shaping degrees of belief, and even when these are taken
into account there may remain some room for subjective choice. Objective
Bayesianism is characterised by three norms: the Probability Norm, the
Calibration Norm, and the Equivocation Norm. Of the three, the Equivocation
Norm is key to the suitability of the objective
Bayesian framework formalising the principle of well-founded suspicion.
Simply put, the Probability Norm says that an
agent’s degrees of belief should be representable by a probability function
defined over the sentences of her language. The Calibration Norm states that
those degrees of belief should fit with her evidence – in particular, should
match empirical probabilities where known. Finally, the Equivocation Norm says
that in case more than one probability function is compatible with evidence,
the agent should choose one that is not too extreme – i.e., should choose one
that equivocates sufficiently between the basic possibilities expressible in
her language.
The first norm would be unproblematically
endorsed by all Bayesians; the second norm would certainly be endorsed by
empirically-based Bayesians, since they require degrees of beliefs to be shaped
upon available evidence. The third norm is what sets objective Bayesianism
apart from other flavours of Bayesianism. It requires that the agent’s
probability function be sufficiently close to the equivocator
,
which is the probability function that gives the same probability to each of
the basic possibilities that the agent can express. Suppose the agent can
express elementary propositions
. Then the basic possibilities (‘possible worlds’)
that she can express take the form
, where each instance of
is either just
or its negation,
. Distance between probability functions is measured
by Kullback-Leibler divergence,
Probability functions that are sufficiently
close to the equivocator are those that have sufficiently high entropy:
As to what counts as sufficiently high
entropy is a pragmatic question, guided by considerations to do with the
required accuracy of predictions and so on. In the extreme case we have Jaynes’ Maximum Entropy Principle
,
which says that the agent’s degrees of belief should be representable by a
probability function, from all those that are calibrated with evidence, that
has maximum entropy. Since this is the standard formulation of objective
Bayesianism, we shall presume this formulation in what follows.
5. An objective Bayesian
formalisation of the principle of well founded suspicion
As we mentioned earlier, the principle of
well-founded suspicion in pharmaceutical regulation can be considered as an
instantiation of the precautionary principle. The precautionary principle can be stated in very simple terms as
follows: the decision to withdraw a drug should not wait until strong causal
links between drug and harm are established with certainty, but action may
follow already from the well-founded suspicion of causal link between the
two. Before the sentence of the Contergan case (section 2), the German legal
system required a proven causality
nexus between danger source and possible damage in order to allow for
administrative or punitive actions. As mentioned earlier, with the Contergan
sentence, the legal system shifted from a danger avoidance system
(Gefahrabwehr), where causal connection needs to be certain, to a risk
prevention system (Risikovorsorge), where a hypothesis of causal connection suffices for intervention. According to the precautionary
principle, then, withdrawal should be considered as soon as harm is suspected.
Stated in these terms, the precautionary principle introduced into legal and
administrative theory the concept of probabilistic causal links.
Since according to the precautionary
principle, the causal connection between danger source and effect need not be
certain and scientifically proven, administrative actions, such as suspension
or prohibition of the drug under consideration, can be enforced before scientific proof is eventually
provided. The threshold point for action is established by reference to the
risk/benefit assessment on the one hand and to the probability of causal
connection on the other: the higher the risk in comparison to the benefit, the
lower the probability of causal connection between potential damage and
suspected source can be in order to allow for risk prevention/minimization
strategies.
Now, there are several reasons why objective
Bayesianism is a good candidate for formalising the principle. First, (health) technologies are always
surrounded by a considerable amount of uncertainty in relation to their medium-long
term effects. By providing a machinery of probabilistic inference, objective
Bayesianism can put these considerations into practice. Second, the precautionary principle is very generally formulated,
and even its pharmaceutical concretization, “the principle of well-founded
suspicion” does not provide any concrete reference point that can help to
establish the tolerance threshold in a standard fashion. Withdrawal decisions
are made on the basis of a consensus procedure grounded on empirical data,
expert opinions and contextual factors (availability of alternative treatments,
pressure of interest groups, influence of patient groups and public opinion):
this may lead to biased procedures thwarting any precautionary effort. Third, objective Bayesianism formally
takes into account not only all available evidence, but also lack of evidence,
precisely in the spirit of the precautionary principle. This has to do with the
Equivocation Norm. In the following we present how this works concretely.
5.1 Precautionary principle, expected
utility, and risk-benefit assessment
Decisions
concerning the approval, marketing, suspension, and withdrawal of
pharmaceutical products are generally justified on the basis of a favourable
(or unfavourable) risk-benefit balance. The expected benefits of a drug are
weighed against its potential drawbacks and this comparison determines the
decision outcome. In this respect, these decisions follow the general rule that
an agent should act so as to maximise her expected utility. The example we propose below is
by no means intended as the description of a particular case, but aims to
provide a formalization which clearly distinguishes the different roles played
by the risk-benefit evaluation (utilities) and by the causal assessment
(probabilities associated with the various utilities) in risk management
decisions, and thereby illustrates where the precautionary principle really
intervenes along this process.
Suppose a
drug has been licensed for market. Consider variable D which takes value d if
the drug is taken as prescribed and value ¬d otherwise. R signifies recovery (with
values r and ¬r) and H signifies harm
(taking value h if there is harm
sufficient enough to warrant withdrawing a drug that caused that harm, and
taking value ¬h otherwise). W signifies withdrawal of the drug from
market (with values w and ¬w).
Now since the
drug has been licensed for market, there must be good evidence that
(i) the drug
positively causes recovery, written,
and
(ii) the drug does
not positively cause excessive harm,
.
(Here
‘positively causes’ is taken as the opposite of ‘prevents’, so that D causes R if and only if D positively causes R or D prevents R or is a mixed cause of R. A mixed cause sometimes positively causes
and sometimes prevents.) Let us suppose that the evidence is such that the
probability that the drug positively
causes recovery reaches some threshold,
, and that
the probability that the drug does not positively cause harm reaches another
threshold,
, where
and
are small.
Consider a
utility matrix for withdrawing the drug, given the case in which the drug
positively causes harm and the case in which the drug does not positively cause
harm:
[24]
If
then the expected utility of withdrawing the
drug is
While the expected utility of not withdrawing
is
According to
the principle of maximising expected utility, one should withdraw the drug if
, i.e., if
i.e., if
i.e., if
This is
indeed a precautionary approach: if the probability that the drug positively
causes harm is more than ¼, the drug should be withdrawn. Only a little
evidence of positive causality is required for withdrawal; the causal claim
need neither be established beyond reasonable doubt nor even on balance of
probabilities. Note however that, so far, nothing has had to be said about what
the probabilities mean. The precautionary principle itself is thus independent
of the interpretation of probability. As we shall now see though, the
implementation of the precautionary principle does depend on the interpretation
of probability.
5.2 Objective
Bayesianism and the precautionary principle
In order to
determine the conditions under which
, an
interpretation of probability and an interpretation of causality must be
provided. And different interpretations will warrant different decisions. Since
we are primarily interested in comparing interpretations of probability, let us take a simple
probabilistic account of causality as our reference point: here
if and only if
for some state c of the other possible causes of H, and
for every other such
state c’. (Accordingly a preventative lowers the probability of harm
for some state c and raises it for
none, while a mixed cause raises it
in some contexts c and lowers it in
others.)
[25]
Let us turn, then, to interpretations of probability.
[26]
So far, we
know that
, i.e.,
. Consider
first the case in which
.
A physical
interpretation of probability, such as the frequency theory, would deem the
physical probability that the drug raises the physical probability of harm to
be some undetermined point within the interval
. Since some
points within the interval would trigger withdrawal and others would not, no
decision can be made as to whether to withdraw the drug.
Consider next
an empirically-based Bayesian interpretation of probability. Here an agent's
rational degree of belief that the drug raises her rational degree of belief in
harm is also some fixed point within the interval
, but in this
case the point in question is not out of reach of the agent—rather, the agent
must simply choose some point within the interval. As long as the agent remains
within the interval, all points are deemed equally rational. In this case a
decision will be made as to whether to withdraw the drug, but the decision is
entirely up to the subjective whim of the agent—it is not objectively
determined as to which course should be taken.
Finally,
consider an objective Bayesian interpretation. According to this
interpretation, an agent should believe that the drug positively causes harm to
the degree within the interval
that is as equivocal as possible. Here, ‘as
equivocal as possible’ means as close as
possible to the value given by the maximally equivocal probability function
(the equivocator). As explained above, the equivocator gives each basic
possibility that the agent can express the same probability. As the basic
possibilities take the form
, it
equivocates as to whether
is greater or less than
for any state c of the other possible causes of H – i.e., the equivocator gives probability ½ that
. If the
agent supposes that there are k other
causes of H, all binary variables,
and therefore 2k states of such causes, the maximally equivocal
function yields the probability that the drug positively causes harm to be
. (That the
drug is a preventative has probability
and that it is a mixed cause has probability
.) Now there
is a unique point in the interval
that is closest to
. The
decision for withdrawal should be taken according to whether this point is
greater than or less than ¼. The objective Bayesian interpretation has the
advantage, then, that a decision will be taken, and that decision is
objectively determined by her evidence.
Next consider
the case in which
. In this
case all the interpretations considered above will license the decision not to
withdraw the drug, since no probability in the interval
exceeds the level ¼ that triggers withdrawal.
What sets the interpretations apart here is the case in which new evidence
becomes available. Suppose that new evidence e increases the interval for
from
to
, where
. As before,
a physical interpretation will not license any decision. The empirically-based
Bayesian interpretation will again license some decision, and this decision is
objectively determined by the evidence and the agent's prior degrees of belief:
if, when conditionalising on e, the
agent's degree of belief in positive causation increases above ¼ then the
decision to withdraw the drug will be warranted, otherwise the drug should be
retained. Note, however, that the prior probabilities—including the
probabilities conditional on the evidence—are entirely subjective, so the
decision as to whether to withdraw remains entirely subjective. Finally the
objective Bayesian will again withdraw according to whether
is greater than or less than ¼. In sum, then,
the objective Bayesian interpretation has the same advantage in this case too:
a decision will be taken, and that decision is objectively determined by the
agent's evidence.
In
conclusion, the precautionary principle arguably follows from the principle
that one should act as to maximise expected utility. But the implementation of
this principle depends on how probabilities and causal relationships are
interpreted. Under a standard probabilistic account of causality, a physical
interpretation of probability suffers in that there are situations in which no
decision can be taken. This is a problem because in a court of law a decision must be taken on the basis of available
evidence. On the other hand, a subjective Bayesian interpretation of
probability always licences a decision, but in certain circumstances the decision
taken is entirely a matter of subjective choice. Thus this interpretation can
also fail to provide normative guidance. But an objective Bayesian
interpretation licenses a decision that depends on the evidence. Only on an
objective Bayesian account, then, can the precautionary principle be
implemented in an objective way.
While we have
presented this argument in the context of a specific utility table, it should
be clear that the procedure is fully general. All that is required is that
utilities can be set out in order to determine a threshold for withdrawal.
6. Concluding remarks
Drug decisions are especially
difficult to make because of the high goods at stake, because of the
uncertainty surrounding both the pharmaceutical products as well as the health
damage eventually produced by the disease, and because of the related
difficulty to establish a risk tolerance threshold. Whenever new risks possibly associated with a drug are
detected, then the question is raised as to whether
these can be considered to be part of the bargain, and if not, whether they are
indeed conditioned by drug intake. According to the principle of well-founded
suspicion, the causal nexus between drug and harm need not be certain in order
for safety measures (such as new labeling or product retirement) to be
enforced. Indeed, following the precautionary principle, the probability of a
causal connection can be as low as the possibly associated harm is supposed or
known to be severe. Thus the probability of the causal link constitutes the
critical measure for intervention, together with the harm suspected to be
associated with the drug. Provided that the former falls within a given
interval, objective Bayesianism provides the formal tools for determining,
within this interval, the probability value that is maximally equivocal,
thereby grounding the decision both on the available evidence and on the lack of information, as
required by a precautionary attitude.
References
Abraham, J., T. Reed (2001). Trading
risks for markets: The international harmonization of pharmaceutical
regulation. Health, Risk & Society, 3 (1): 113-128.
Abraham, J., C. Davis (2005). Risking
public safety: Experts, the medical profession and “acceptable” drug injury. Health, Risk & Society, 7 (4): 379-395.
Buckingham Stephens, M. D. (1997). From
causality assessment to product labeling. Drug Information Journal, 31:
849-857.
CIOMS Working Group IV (1998). Benefit Risk balance
for marketed drugs: evaluating safety signals. Cioms, Geneva 1998.
Cowell, R.G., A.P.
Dawid, S.L. Lauritzen, D.J. Spiegelhalter (2007). Probabilistic Networks and Expert
Systems. Exact Computational Methods for Bayesian
Networks. New York: Springer.
Davidson, Barbara Lorraine (1981). The philosophy of probability : an examination of the fundamental nature of the concept of epistemic
probability. La Trobe University.
De Finetti, Bruno (1937). La
Prévision: ses lois logiques, ses sources subjectives. Annales de l'Institut
Henri Poincaré. Translation: Foresight: its Logical Laws, Its
Subjective Sources. in H. E. Kyburg and H. E. Smokler (eds), Studies in Subjective Probability, New
York: Wiley, 1964.
Demortain, D. (2008). From drug crises to regulatory
change: The mediation of expertise. Health, Risk & Society, 10 (1): 37-51.
Dettling, H.-U. (2005).
Arzneimittelrecht als Sicherheitsrecht – Zugleich ein Beitrag zur
Rechtfertigung von Freiheitsbeschränkungen. PharmR, 4: 162-173.
Di Fabio, U. (1993). Gefahrbegriff
und Nachmarktkontrolle. In: Damm/Hart, 1993: 109-131.
Di Fabio, U. (1994). Risikoentscheidungen im Rechtsstaat
: zum Wandel der Dogmatik im öffentlichen Recht, insbesondere am Beispiel der
Arzneimittelüberwachung. Tübingen : Mohr.
Domenighetti, G. (2005). Grandeur et misère des systèmes universels de santé. Bulletin des Medecins Suisse/ Schweizerische Ärztezeitung/ Bollettino dei
medici svizzeri. 86, n. 4: 221-226.
Dubucs, Jacques-Paul (eds) (1993). Philosophy of Probability. Springer.
Dupuy, J.P. (2004). Complexity
and Uncertainty a Prudential Approach to Nanotechnology.European Commission, A Preliminary Risk Analysis on the Basis of a Workshop Organized by the Health
and Consumer. Protection Directorate General of the European Commission, in
Brussels 1 – 2 March 2004 (http://europa.eu.int/comm/health/ph_risk/documents/ev_20040301_en.pdf ).
Eagle, Anthony (eds) (2009). Philosophy of Probability: Contemporary
Readings. Routledge.
European Environment
Agency (2001). Late lessons from early warnings: the precautionary principle
1896-2000. Environmental issue report n. 22.
Garattini, S. (2010). Evaluation of Benefit-Risk. Pharmacoeconomics, 28 (11): 981-986.
Garrison, L.P., A. Towse, B.W. Bresnahan (2007).
Assessing a Structured, Quantitative Health Outcomes Approach To Drug Risk-Benefit Analysis. Health Affairs, 26 (3): 684-695.
Gassner, M., M. Reich-Malter (2006).
Die Haftung bei fehlerhaften Medizinprodukten und Arzneimitteln – Recht und
Rechtsprechung. Medizinrecht, n. 3: 147-152.
Gillies, Donald
(2002). Philosophical
Theories of Probability. Routledge.
Hacking, I. (1986). Culpable
ignorance of interfernce effects. In D. MacLean (ed.) Value at risk
Rowman & Allanheld, Totowa, NJ: 136-154.
Hart, D. (2005). Die
Nutzen/Risiko-Abwägung im Arzneimittelrecht. Ein Element des Health Technology
Assessment [1]. Bundesgesundheitsblatt-Gesundheitsforschung- Gesundheitsschutz,
48: 204-214.
Hart, D. (1998). Zum Management von
Arzneimittelrisiken durch Ärzte und Unternehmen – Rechtsverfassung,
Sozio-Psychologie, Empirie. In V. Preuss (ed.) (1998) Risikoanalysen. Über den
Umgang mit Gesundheits- und Umweltgefahren. Vol II. Heidelberg: Roland Asanger.
Hauber, A.B., F. Reed, E.B. Andrews (2006).
Risk-Benefit Analysis Methods for Pharmaceutical Decision-Making – Where are we
now? ISPOR Connections, 12 (6): 3-5.
Holden, W. L., J.
Juhaeri, W.Dai (2003a). Benefit-risk analysis: a proposal using quantitative methods.
Pharmacoepidemiology and Drug Safety(12): 611-616.
Holden, W. L., J.
Juhaeri, W.Dai (2003b). Benefit-risk analysis: examples using quantitative methods.
Pharmacoepidemiology and Drug Safety(12): 693-697.
Howson, C., P. Urbach (2006). Scientific Reasoning: The Bayesian Approach.
Open Court.
Hughes, D.A., A.M. Bayoumi, M. Pirmohamed
(2007). Current Assessment of Risk-Benefit by Regulators: Is It Time To Introduce decision Analyses? Clinical Pharmacology &
Therapeutics, vol. 82 (2): 123-127.
Jaynes, E. T.
(1957). Information theory and statistical mechanics.The
Physical Review, 106(4):620–630.
Jonas, H. (1979). Das Prinzip
Verantwortung: Versuch einer Ethik für die technologische Zivilisation.
Frankfurt am Main: Suhrkamp.
Kolmogorov, A. N. (1933). Grundbegriffe der
Wahrscheinlichkeitsrechnung. Ergebnisse der Mathematik. Berlin: Springer.
Laupacis, A., J.M. Paterson, M. Mamdani, A.
Rostom, G.M. Anderson (2003). Gaps in the evaluation and monitoring of new
pharmaceuticals: proposals for a different approach. Canadian Medical Association Journal, 169 (11):
1167-1170.
Lynd, L. (2006). Quantitative
Methods for Therapeutic Risk-Benefit Analysis. Issues Panel: Health
outcomes approaches to risk-benefit analysis: how ready are they? 11th International Meetings of the International Society for Pharmacoeconomics and
Outcomes research (ISPOR) May 2006.
Meyboom, R.H.B., A.C.G. Egberts, I.R. Edwards, Y.A.
Hekster, F.H.P. Koning, F.W.J. Gribnau (1997). Principles of
Signal Detection in Pharmacovigilance.Drug Safety, 16: 355-365.
Mises von R. (1928). Wahrscheinlichkeit, Statistik und Wahrheit.
Wien, New York: Springer Verlag. English
translation: 1957, Probability,
Statistics and Truth. Dover (19812, NY).
Olivier, P., J-L Montastruc (2006). The nature of
scientific evidence leading to drug withdrawals for pharmacovigilance reasons
in France. Pharmacoepidemiology and drug safety, 15: 808-812.
Osimani, B. (2007). Probabilistic information and
decision-making in the health context: Package leaflets as basis for informed
consent. PhD Thesis, University of Lugano.
Osimani, B. (2010). Pharmaceutical risk communication:
sources of uncertainty and legal tools of uncertainty management. Health Risk
and Society, October, Volume 12 Issue 5: 453-469.
Osimani, B. (forthcoming). Rationally
ignoring in the health context. Journal of
Socioeconomics.
Ramsey, F. P. (1926). Truth and
Probability. In B. Braithwaite (ed.) (1931) Foundations
of Mathematics and other Essays. London: R. Routledge & P. Kegan:
156-198; reprinted in H. E. Kyburg, Jr., H. E. Smokler (eds) (1980) Studies in
Subjective Probability, Melbourne, Florida: R. E. Krieger Publishing Company:
23-52; reprinted in D. H. Mellor (ed.) (1990) Philosophical Papers, Cambridge: Cambridge
University Press.
Räpple, T. (1998). Das Verbot
bedenklicher Arzneimittel. Eine Kommentierung zu § 5 AMG. Nomos Verlagsgesellschaft. Baden-Baden.
Reed, F.J., A.B. Hauber, C. M. Poulos (2009). A Brief
Introduction To The Use of Stated-Choice methods to
Measure Preferences for Treatment benefits and Risks. RTI Press, Research Report,
September 2009.
Reichenbach, H. (1935). Wahrscheinlichkeitslehre: eine Untersuchung
über die logischen und mathematischen Grundlagen der
Wahrscheinlichkeitsrechnung. English translation: 1948, The theory of
probability, an inquiry into the logical and mathematical foundations of the
calculus of probability. University of California Press.
Reiss, J; P. Kitcher
(2008).
