4.5.1 Background
The ICL model, running on CovidSim, was developed by Neil Ferguson and his team and was based on an earlier influenza model (Reference Ferguson, Cummings and CauchemezFerguson et al. 2005, Reference Ferguson, Cummings and Fraser2006, Reference Ferguson, Laydon and Nedjati Gilani2020). The primary purpose of the model was to project the impact of various possible
policy choices on Covid-19 deaths and demand for hospital beds,
including intensive care unit (ICU) beds. Specifically, the report
considers a variety of permutations from a set of possible policy
choices (in addition to a potential ‘do nothing’ policy), consisting of
the following elements, shown in Figure 8.
The subsets of these NPIs considered to be ‘mitigation’ strategies that
the model explored were: PC; CI; CI&HQ; CI&SD; CI&HQ&SDO; and
PC&CI&HQ&SDO. All of these strategies are ‘shown’ to result in
massive overwhelm of the healthcare system, as shown in Figure
9.
‘Suppression’ is defined as being able to ‘reduce R to close to 1 or
below’ (Reference Ferguson, Laydon and Nedjati GilaniFerguson et al. 2020, 10). The report cautions that, at
least, ‘[c]ase isolation, general social distancing, and either school and
university closure or home quarantine’ are required to achieve suppression, but
the only ‘suppression’ strategy whose simulation results are presented in the
report is the combination of all four elements (i.e., ‘[h]ome isolation of
cases, household quarantine, school and university closures, and social
distancing of the entire population’). Reference Ferguson, Laydon and Nedjati GilaniFerguson et al. (2020) show the simulation results of
instituting these policies cycling on and off for the 18 months they expected
it to take to achieve a vaccine (assuming a baseline R0 value of
2.2) (Figure
10).
In building the model, the modellers aimed to steer policy
choices by the UK Government, as well, to a lesser degree, as other governments
around the world, by highlighting the extremely high death toll, and burden on
healthcare systems, that would ensue from pursuing all but the last of those
policy choices (
Reference Broadbent and StreicherBroadbent and Streicher 2022). Reference Ferguson, Laydon and Nedjati GilaniFerguson et al. (2020, 16) said:
We therefore conclude that epidemic suppression is the only viable
strategy at the current time. The social and economic effects of the measures
which are needed to achieve this policy goal will be profound. Many countries
have adopted such measures already, but even those countries at an earlier
stage of their epidemic (such as the United Kingdom) will need to do so
imminently.
4.5.2 What Is Represented? Study Topic and End
Points in Context
Following the basic framework outlined in Section
4.2, we can begin
to outline some big-picture value judgements that went into building the ICL
model. First and foremost, building this model reflects the judgement that it
is an ethically defensible project: in other words, that it is desirable to
seek more information about the likely effects of certain policy choices on
Covid-19 deaths and hospitalizations. It is important to realize that this
needn’t be taken for granted. It is not hard to imagine people having a set of
values according to which many of the policy choices explored in the model,
particularly those that involved restrictions on various liberties, constitute
violations of fundamental rights and should not even be considered. It is also
not hard to imagine having a set of values, and a set of prior beliefs about
the severity of the virus, on which it would have been morally unacceptable to
delay action even long enough to carry out the modelling project – or to accept
any risk at all that the model would erroneously steer policy-makers away from
drastic suppression measures. It is only according to a certain set of values,
and a certain set of prior beliefs given the state of evidence at the time, and
a certain confidence that a minimally informative model could be built, that it
would appear the right thing to build a model like the ICL model and to
consider using its output to evaluate the costs and benefits of the kinds of
policy choices that are explored in the model.
Next comes representational decisions concerning what the primary end
points of the model ought to be. In the ICL model, the variables taken to be
primary end points were Covid-19 infections, hospitalizations, occupancy of
intensive care hospital beds due to Covid-19, and Covid-19 deaths. Including these
as end points in a pandemic mitigation model signifies an ethical judgement
that these are important outcomes to consider when reasoning about what NPIs to
implement to help slow the pandemic. What about the outcomes that were excluded
from the model? Recall that if a model does not include certain
outcomes this could signify one of two things. The fact that the ICL model did
not project the impact of the possible policy choices explored on things like
educational outcomes and economic output could be interpreted in two ways:
either as suggesting that these outcomes were not ethically significant, or as
suggesting that decision-makers would have to look elsewhere to get this
information.
This highlights the degree to which models must be understood in the context
of the intended purposes of their users. It is not built into the ICL model
whether it is intended to be a tactical model that can be used, directly, to
assess different policy choices, or intended to be used alongside other sources
of information about the effects these choices would have on things other than
Covid-19 deaths, hospitalizations, and so on. But insofar as we can read the
famous ‘Report 9’ (Reference Ferguson, Laydon and Nedjati GilaniFerguson et al. 2020), which was the primary document
produced with the help of the ICL model, as a policy guidance document,Footnote 10 we can read it as suggesting that these
outcomes were relatively ethically insignificant in comparison to the outcomes
that were in fact part of the model.
Here is another way to frame the same point. If you happen to believe
that policy choices that were guided with the help of a model were made without
due consideration of a certain dimension of their consequences, it is difficult
to assign moral responsibility for that moral failure. Responsibility could lie
with the model-builders for failing to include those dimensions among their
model end points, or it could lie with policy-makers for relying on only one
kind of expertise (say, infectious disease epidemiology) when in fact two kinds
of expertise (educational psychology as well as infectious disease
epidemiology) were required. Or responsibility could fall on both parties. But
in the case of Report 9, it seems clear that the ICL group judged that their
model results alone could guide policy, and that their recommended strategy
should be implemented despite its social and economic effects: ‘We
therefore conclude that epidemic suppression is the only viable strategy at the
current time. The social and economic effects of the measures which are needed
to achieve this policy goal will be profound’ (Reference Ferguson, Laydon and Nedjati GilaniFerguson et al. 2020, 16).
4.5.3 Finer-Grained Representational Decisions
Other
Variables
While representational decisions concerning what variables to include as
end points in the ICL model (e.g., ‘Should learning loss be an end point in the
model?’) have a particularly obvious ethical importance, decisions about what other
variables to represent are value-laden too. One example is the decision to
represent the ages of inhabitants of the United States and the United Kingdom,
and their status as workers or students, but not their race, income, postal
code, or occupation. As a result of this decision, the ICL model is not
adequate for exploring research questions like: (1) whether a mitigation or
suppression strategy will lead to a racially unjust distribution of infections,
hospitalizations, and deaths; or (2) whether such a strategy will
differentially affect people of different professions. For example, will the
closing of non-essential industries result in the burden of the disease falling
primarily on those who provide essential services? As we noted in Section
1, at the heart of
computational models like climate models and models of disease spread are the
bits of math that jointly create the behaviour of the model, given its starting
state, and the choice of counterfactual scenarios (from among all the infinite
possible counterfactual scenarios) that it is designed to investigate. Usually,
this requires picking other variables besides the ones being studied as end
points, the mathematical relationships between them, and the values of
parameters that feature in the equations that specify those relationships. The
ICL model, for example, does not include among its variables the time of year
and its possible impact on viral transmission. This makes it impossible for the
model to explore the possibility that the virus could come in waves due to an
underlying seasonality of the virus.
Parameters
Overall, the ICL model employed equations with almost 900 different
parameters. Given its purpose, the model needed inputs for the expected
Covid-19 death rate, hospitalization rate, and ICU admittance rate for every
100,000 people infected in each of several different age brackets, given in a
table in Figure
11.
Figure
11 Estimates of
virus morbidity and fatality from Reference Verity, Okell and DorigattiVerity et al. (2020) used in Report 9.
Note: The table is somewhat confusing because
fatalities are given as a percentage of infections, but hospitalizations are
given as a percentage of ‘symptomatic infections’, and ICU beds are given as a
percentage of hospitalizations. If we dig into the code, we can determine that
two-thirds of infections are expected to be symptomatic, and so the percentage
of overall infections that will require hospitalization is 0.66 times the
numbers in this column.
Source: Reference Ferguson, Laydon and Nedjati GilaniFerguson et al. (2020, table 1, 5).
At the time that Report 9 came out, in the middle of March
2020, none of the correct values of these inputs were well known. (Indeed, our
estimates of them at the time of writing now remain imperfect.) One moral
judgement that could have been made at the time was that our confidence
intervals around these input values were too large to make modelling the impact
of different possible policy choices a worthwhile project. If, for example, one
thought that the possible values for the infection fatality rate of SARS-CoV-2
at the time could be anywhere between 0.1% and 3%, and one thought, as the ICL
team clearly did, that you had to pick a single value (at least relative to
each age bracket) then you might think that a model of this kind would be
useless for assessing the costs and benefits of policy choices. That’s because
you might think that it is fairly obvious that, if the infection fatality rate
(IFR) is 0.1%, the most stringent strategies are almost certainly going to look
too costly, and if it’s 3%, they are almost certainly going to look like they
are morally required. If you thought that, you might conclude that getting a
better estimate of the numbers in Table 2 was a much higher priority than
building a model like the ICL model – indeed, it might even be a condition on
the moral permissibility of building it.
Since the ICL team obviously decided to proceed with the modelling
project, how did they choose those values? In early March 2020, researchers had
at least two sources of data available to them for the parameters in Reference Ferguson, Laydon and Nedjati GilaniFerguson et al. (2020, table 1) (Reference Verity, Okell and DorigattiVerity et al, 2020). The first was the Diamond
Princess (DP) cruise ship. This was the first ‘natural experiment’ of a
Covid-19 outbreak, where every single passenger had been tested for infection
and the health outcomes of each passenger were well known. The second was the
data available from the first ‘epicentre’ of the pandemic: Hubei province in
China. The advantages and disadvantages of each data source were clear. The
advantage of the DP data was that they were comprehensive. The exact number of
infected people was known as was the exact number of each health outcome. The
disadvantage was that the number of people was relatively small, and the age
structure of the population was unusual. Most of the passengers were rather old
and most of the crew were very young. There were very few intermediate-aged
people in the data set. The advantage of the Hubei dataset was the mirror image
of the disadvantages of the DP set. The dataset was large and every age
demographic was included. The main disadvantage was that while the numerators
for all of these outcomes were relatively well known,Footnote 11 the denominators were not known at all. That
is, how many people had died of Covid-19 in Hubei and how many had been
hospitalized was known, but it was not known how many infections this
represented. Knowing the ratios of these numbers, however, was crucial
to estimating the parameters needed for the ICL model.
The ICL group responded to this lacuna by looking at repatriation
flights from Hubei into the United States and Europe. That is, at the time of
the outbreak, citizens of the United States and some European countries were
evacuated from Hubei province and returned to their home countries. Each person
on these flights was carefully screened for infection with SARS-CoV-2. The ICL
group used the proportion of expatriates who were infected to estimate the
incidence of infection in Hubei at large. The disadvantages of these methods
were obvious. For one thing, this ended up being a very small sample. In total,
only six people were found to be infected on the flights.Footnote 12 Had, by chance, a seventh person tested
positive, then all the values in Table 2 would have been six-sevenths the size
they are. Another disadvantage was that the method assumed that very wealthy
and culturally outlying expatriates had a degree of infection that was
representative of the underlying population.
So the makers of the ICL model had at least two choices with regard to
this ‘how to represent’ question: they could have generated something like Reference Ferguson, Laydon and Nedjati GilaniFerguson et al.’s (2020) Table 1 from the Hubei data or
from the DP data. In what way did values influence this choice? The
model-builders could clearly see that the IFR and hospitalization rates that
came from the Hubei data were considerably higher than from the data from the
DP.Footnote 13 Thus, choosing the Hubei data made it more
likely the IFR would be overestimated than opting for the DP data, and less
likely that it would be underestimated. Thus, the more serious you consider the
harm of a Covid-19 death and/or hospitalization, and the less serious you
regard the various harms of the mitigation and suppression strategies being
considered, the more inclined you will be to choose the Hubei data, and perhaps
the more inclined you will be to use the repatriation flight method for
estimating the denominator than some other method that produced a lower IFR.
To return to a point we made in Section
4.4 that we called
the inarticulability thesis, it is probably not reasonable for us to
think that Ferguson and his colleagues had in mind exact probabilities that the
Hubei data were closer to reality than the DP data. Recall that what rationally
guides representational decisions like the choice between the Hubei data and
the DP data is utility maximization, and that is a function of the values we
assign to the various states of affairs that might follow from our
representational decisions being good or bad, along with the probabilities we
assign to the various good and bad outcomes that could follow from particular
representational decisions.
Suppose that the members of the ICL group had one, and only one, binary
choice to make: use the DP data or the Hubei data. And suppose that they had in
mind exact probability distributions that they assigned to each dataset
overestimating and underestimating the effectiveness of each intervention to
various degrees, say P(DP) and P(H). And suppose they had a complicated
function of utilities over the space of each of those degrees of overestimation
and underestimation. Even articulating just this amount of information would be
overwhelming. But the situation is far worse than this: they have over 900
representational choices to make (just regarding the items they chose to
represent, excluding those they chose to ignore!), and most of those choices
are not in fact binary. They could, in principle, have had whole probability
distributions over possible values of all the numbers in Reference Ferguson, Laydon and Nedjati GilaniFerguson et al.’s (2020) table 1. This suggests that, in
fact, in a complex modelling situation like the one faced by the ICL team, a
whole host of representational choices have to be made in conjunction with each
other, with only a very coarse-grained and holistic assessment of the expected
utility of a small subset of a nearly infinite set of possible choices they
could have made. This is why the inarticulability thesis cautions that the
decision could only have been made based on a rough expected utility
estimation, without the modellers having precise values of probabilities and
utilities in mind. Their choice of data set is ultimately the result of an
inchoate mix of epistemic assessments and value commitments that cannot be
fully articulated.
4.5.4 Uncertainty
Another ‘how to represent’ question concerns how to represent, in a
model, our uncertainty regarding the best value of the relevant parameters. For
example, the ICL model assumes that when people socially distance, their
probability of getting infected at home increases by 25%. But why 25%? Why not
35%? In fact, there were no data or research to support any particular choice
in the model, since we had few well-established rates for any past virus, let
alone rates for the novel SARS-CoV-2 virus. It is very common in modelling to
attempt to deal with uncertainty about the correct value of parameters by
running a sensitivity analysis. The idea is to run the model on a wide sample
of parameter values, in order to try to figure out how sensitive the model is
to small differences in those values, and to try to figure out which values do
the best job of capturing known data. In climate modelling, this is referred to
as doing a ‘perturbed physics ensemble’. But the ICL model was used to
influence major policy decisions in the absence of any study of parameter
sensitivity.
Fortunately, Reference Wouter, Arabnejad and SinclairEdeling et al. (2021) finally undertook such a study in
November 2020. They wrote:
Here we report on parametric sensitivity analysis and uncertainty
quantification of the code. From the 940 parameters used as input into
CovidSim, we find a subset of 19 to which the code output is most sensitive –
imperfect knowledge of these inputs is magnified in the outputs by up to 300%.
The model displays substantial bias with respect to observed data, failing to
describe validation data well. Quantifying parametric input uncertainty is
therefore not sufficient: the effect of model structure and scenario
uncertainty must also be properly understood.
(Reference Wouter, Arabnejad and SinclairEdeling et al. 2021, 128).
They found, in particular, that almost two-thirds of the differences in
the model’s results could be attributed to changes in just three especially
important variables: the length of the latent period during which an infected
person has no symptoms and can’t pass the virus on; the effectiveness of social
distancing; and how long after getting infected a person goes into isolation.
More importantly, Reference Wouter, Arabnejad and SinclairEdeling et al. (2021) found that for most values of
these parameters, 5–6 times as many people die during ‘suppression’ than the
model predicted using the values that the ICL group used. Thus, Reference Wouter, Arabnejad and SinclairEdeling et al. (2021) show that, had Ferguson’s group
done a sensitivity analysis over the range of parameter values that were
consistent with what was known about the virus, they would have been unable to
show that the suppression strategy they appeared to be recommending would have
had much benefit with respect to Covid-19 outcomes.
It is not hard to see that had Report 9 (Reference Ferguson, Laydon and Nedjati GilaniFerguson et al. 2020) included a sensitivity analysis of
the kind found in Reference Wouter, Arabnejad and SinclairEdeling et al. (2021), its influence on policy-makers
and the public might have been less dramatic. After all, Reference Wouter, Arabnejad and SinclairEdeling et al. (2021) seemed to show that it was
consistent with what we knew in March 2020 that the mitigation and suppression
strategies considered in Report 9 were all going to be equally ineffective. So
we can see two ways in which the choice of how to represent these parameter
values was value-laden. First, the ICL team chose values of parameters that
maximized the projected benefits of the strategy they appeared to be
recommending. Thus, they appear to be judging it to be less serious a mistake
to overestimate the effectiveness of those strategies than to underestimate
them.
Second, the choice not to include a sensitivity analysis in the
characterization of the model output we find in the report was itself highly
value-laden. One possibility is that it reflected the value judgement that
urgent action was required before there was time for the sensitivity analysis
to be carried out. Another possibility is that it reflected the value judgement
that an estimation of the degree of uncertainty regarding the effectiveness of
the measures would be less valuable than the precise, fine-grained projection
they in fact made. This is always a balance of values that model-builders and
interpreters face: how to balance the benefit of the informativeness of a
precise projection against the value of the confidence one can have in a wider,
imprecise estimate of that same benefit (Reference WinsbergWinsberg 2018). Another possible explanation of the ICL
group’s failure to do a sensitivity analysis is that they deemed that doing so
would be too likely to cause governments to wrongly choose to abstain from
maximum suppression (i.e., by emphasizing the degree of existing uncertainty).
This obviously would reflect a value judgement about how bad such an outcome
would be.
4.5.5 Choice of Counterfactuals for Projection
Let’s compare the way that the ICL model and climate models make
projections. The ICL model projections are conditional on policy choices, while
climate models are conditional on representative carbon pathways (RCPs). The
latter are not policy choices: they are outcomes that are conditional on policy
choices and numerous other factors acting in complex interaction with one
another; there are no uncontroversial connections between policy choices and
carbon pathways. In comparison, the ICL model takes policy choices as counterfactuals
for projection. Model developers are thus put in the position of
estimating how, for example, university and school closures will affect social
contact rates – but there is enormous uncertainty around such relationships,
not least because they stand to vary from setting to setting. Imagine putting
climate modellers in a similar position: for example, asking climate modellers to
assume that regulations imposed on nuclear power plant builders are reduced and
subsidies are provided to electric utilities that build out solar power
infrastructure. Our confidence in the model would have to be relatively low:
the results would no longer reflect a causal pathway of which scientists have a
good understanding (Reference Harvard and WinsbergHarvard and Winsberg 2021). So the decision to represent
possible policies in the model, rather than simply putting in contact rates
that would be the outcome of policy choices, forced the modellers to
choose how to represent those choices. This opens up a huge number of choices
to make. And there was enormous uncertainty concerning nearly every one of
these choices, many of which take the form of parameters built into the model’s
coding. Most modelling choices were relatively unconstrained by data or
background knowledge; and when there was data, it was of poor quality.
4.6 Moral Responsibilities in Modelling
In this section, we have explored the consequences of the claim,
developed in Section
3, that
model-building involves a kind of epistemic risk that is fundamentally
different from the kind of epistemic risk involved in endorsing a truth-apt
claim as a fact. In particular, we used the example of the ICL computational
model of Covid-19, called ‘CovidSim’, to highlight the ways in which making
representational decisions – of what to represent and how to represent it – is
highly value-laden.
One thing this discussion highlights is how the value-ladenness of
modelling laces the practice of model-building and model-using with significant
moral responsibilities. Model-builders and model users, especially model users
who are policy-makers, face significant moral responsibilities because the
choice to model at all, and the choices of how to model, can have serious moral
consequences. Thus, model-builders and users are morally responsible for
building the right models for the right purposes and for
making representational choices that embody the right balance of risks
(i.e., that their models will fail to be adequate for purpose in one way rather
than another).
But, of course, every use of the word ‘right’ in the previous sentence will
be highly value-dependent. And when model-builders and model users are working,
as they so often do, on behalf of the public, the question of what constitutes
the right set of values for informing the modelling process can become
overwhelmingly vexed. How can model-builders and model users possibly navigate
these incredibly turbulent waters? How can they ensure they are not imposing
their idiosyncratic values on the public?
One proposal that we find in the general ‘values in science’ literature
is that scientists should strive to make their own reasonable methodological
decisions and then be transparent about what values guided those decisions (Reference DouglasDouglas 2009; Reference ElliottElliott 2017; Reference Elliott and McKaughanElliott and McKaughan 2014; Reference SchroederSchroeder 2017). There are two considerations here that
suggest this is unlikely to do the work it needs to do – to avoid imposing
idiosyncratic values on the public. For example, the ICL group (see Reference Verity, Okell and DorigattiVerity et al. 2020) are relatively ‘transparent’ about
the fact that they chose the Hubei data set over the DP data set. (As we saw, Reference Verity, Okell and DorigattiVerity et al. (2020) showed that the estimate of IFR
from Hubei was at least twice as large as the one you would get from the DP,
and they chose to use the former exclusively.) But for ‘transparency’ to
mitigate the problems discussed here, it should enable members of the public to
figure out whether the choice the ICL group made is or isn’t the one they
would have made, given their values. If the public can see that they would
have made the same choice, then no idiosyncratic values risk being involved. If
they can’t tell that, then the strong possibility exists that the ICL group is
being value-laden in a way that the public would fundamentally object to, and
that this fact remains hidden. Therefore, it is a criterion of success for the
transparency proposal, that transparency leads to members of the public being
able to tell if modellers are making choices that fail to accord with their
values. We can call this the ‘congruence’ criterion.
In the kinds of modelling projects we are looking at here, that is,
complex models that incorporate diverse sources of evidence and aim to directly
inform policy, it seems unclear whether the congruency criterion could be met.
Indeed, the inarticulability thesis seems strongly to suggest otherwise. The
inarticulability thesis, recall, says that it is unrealistic to ask modellers
to articulate all of the values and probabilities that drive their
representational choices. In principle, how could modellers say more than ‘We
chose the Hubei data over the DP data’, in a way that would satisfy the
congruence criterion? What the literature seems to suggest is that modellers
should issue transparent statements in a form such as, ‘We chose the data set
that erred on the side of overestimating the risk of death from Covid-19’
(e.g., Reference DouglasDouglas 2009). In principle, this sort of statement could
be adequate for satisfying the congruence criterion. However, notice that it is
only adequate if what modellers mean by it is that ‘no matter how high a risk
estimate a data set would yield, and regardless of our assessment of the
quality or accuracy of a dataset, we would always pick the data set that erred
on the side of overestimating the risk of death from Covid-19’ and if members
of the public share this extreme view. Otherwise, a statement of this type does
not allow members of the public to determine whether they would have acted as
the modellers did, because it doesn’t tell them the ICL group’s relative
weighting of different harms. It doesn’t, crucially, tell a member of the
public how much more the ICL group values avoiding a Covid-19 death than they
do a job loss, or a child losing a year of education. And unless members of the
public know this, they can’t be sure whether they would have chosen the data
set that produced a higher death rate, irrespective of how high that death rate
is and how likely that data set was to be the better one. Unless the ICL group
can tell members of the public what all their value commitments are,
in a fine-grained form like ‘We think avoiding one Covid-19 death is worth
losing 100 child learning years’ for every single relevant policy consequence,
then members of the public won’t be able to tell whether the ICL group made the
same representational decisions they would have. And the inarticulability
thesis suggests it is not feasible for transparency to take this form.
Instead of merely asking for transparency, we could ask that scientists
simply make the choices that reflect the ‘right’ values (Reference SchroederSchroeder 2017). But what would this mean? Roughly speaking,
there are two things we could mean by the ‘right’ purposes and the ‘right’
balances of risks. ‘Right’ here could mean the ethically correct ones, or it
could mean the ones that are actually held by the public (Reference SchroederSchroeder 2017). Arguably, not being the ethically wrong
purposes and balance of risks is a minimal condition on being the satisfactory
ones. A model that assumes men’s health outcomes are more important than
women’s is not an ethically defensible model. But, also arguably, not being
ethically wrong is not a narrow enough constraint on representational choices
in a model. There might have been, for example, a value commitment about the
relative value of preventing Covid-19 deaths versus preventing job losses, and
all their attendant harms, on which choosing the Hubei data set was the right
choice, and another value commitment in relation to the same consequences on
which the DP data set would have been the right choice. And it might very well
have been the case that reasonable people would disagree about which was the
right set of values to have. If that’s right, then having scientists limit
themselves to ethically permissible representational choices will
underdetermine those choices and leave them open to making choices that do not
reflect the values of the majority of the people on whose behalf decision-makers
will be acting when they make use of the model.
4.7 Public Participation in Modelling
In modelling projects that aim to directly inform public policy, it
seems to us that scientists have an obligation to make the ‘right’ choices in
the sense of ‘right’ that means ‘in accord with publicly held values’.
Something like this line of argument is defended by Reference Alexandrova and FabianAlexandrova and Fabian (2021) with respect to decisions
regarding which ‘thick concepts’ to employ in science. Their idea is that if
scientists are going to theorize about something like ‘well-being’, they need
to use a concept of well-being that accords with the public’s. Here, we are in
broad agreement with them: in fact, we think their basic idea needs to be
extended far more widely to include representational decisions in modelling
generally. To ensure that thick concepts in science reflect public values, Reference Alexandrova and FabianAlexandrova and Fabian (2021) propose a process of
‘co-production’, whereby scientists and members of the public work together to
determine how to construct relevant measures such as ‘well-being’. A similar
process of co-production to build policy-relevant models seems like a fruitful
one to explore, and in fact various research groups across disciplines have
endeavoured to involve members of the public in participatory modelling
projects (Reference Bunka, Ghanbarian and RichesBunka et al. 2022; Reference Gray, Paolisso, Jordan and GrayGray et al. 2016; Reference Staniszewska, Hill and GrantStaniszewska et al. 2021; Reference Voinov and BousquetVoinov and Bousquet 2010; Reference Xie, Malik, Linthicum and BrightXie et al. 2021). Currently, it is unclear whether
participatory modelling projects are meeting the goal of ensuring that models
reflect public values, and various challenges with this type of co-production
will have to be addressed in future research. Among various challenges are the
idea that co-production could become a ‘box-checking’ exercise (Reference Alexandrova and FabianAlexandrova and Fabian 2021), that members of the public won’t
be able to understand what’s going on in the modelling process, or that
modellers won’t be able to articulate the relevant considerations that would
link their values to representational choices.
Prima facie, it seems like asking the public to co-produce a model is a
bigger ask than co-producing a thick concept like well-being. For example,
asking members of the public to weigh in on what aspects of well-being are most
important to them (e.g., having a feeling of purpose in life) requires much
less technical understanding than does expecting them to recognize that
shortening the latency period of a virus will make NPIs look more effective
than lengthening it and therefore implicitly weights as more serious the risk
of allowing too many Covid-19 deaths than the risk of unduly damaging the
economy, and by how much. Or figuring out that one version of a sensitivity
analysis privileges precision over confidence, and by how much. What seems to
be required are normative guidelines for public modelling projects that
articulate how representational decisions should be made collaboratively
between modellers and stakeholders (Reference Harvard and WinsbergHarvard and Winsberg 2023; Reference Husereau, Drummond and AugustovskiHusereau et al. 2022). In addition to this, one might
think that normative guidelines are required not only with respect to how
representational decisions should be made, but with respect to how those
decisions should be implemented in code when the representations in
question are complex computational models. Indeed, this is the focus of Reference Horner and SymonsHorner and Symons (2020), who point to the concrete
challenges involved in software engineering and the various potential errors
that can result from this aspect of modelling practice (cf. Reference PrimieroPrimiero 2014). Reference Horner and SymonsHorner and Symons (2020) argue that software engineering
standards, too, are negotiable matters that should involve public deliberation
concerning trade-offs (e.g., regarding safety, uncertainty, urgency, resources,
risk, etc.). As representational decisions intersect with numerous other
socially significant decisions throughout the modelling process (What software
should modellers use? Should all models for public decision-making be ‘open
source’? How should such models be validated?), an important initial question
concerns the appropriate scope for public participation in modelling
and corresponding normative guidelines. More philosophical and empirical work
is required to conceptualize and address these questions.
4.8 Conclusion
In this section we used ‘CovidSim’ as an example to illustrate the ways
in which representational choices in modelling, and the stage of model
interpretation at which facts are endorsed, both involve values, and we argued
that this places moral responsibilities on model-builders, model interpreters,
and the policy-makers who engage with them. Regarding model-builders, we
canvassed three different ways in which they can discharge their
responsibilities: by being transparent about their values, by using ethically
correct values, or by appealing to publicly held values. We highlighted the
respects in which the third way seems to be the only fully satisfactory one of
the three, but also by far the most difficult to achieve.
·
Jacob Stegenga
·
University of Cambridge
·
Jacob
Stegenga is a Reader in the Department of History and Philosophy of Science at
the University of Cambridge. He has published widely on fundamental topics in
reasoning and rationality and philosophical problems in medicine and biology.
Prior to joining Cambridge he taught in the United States and Canada, and he
received his PhD from the University of California–San Diego.
·
This
series of Elements in Philosophy of Science provides an extensive overview of
the themes, topics and debates which constitute the philosophy of science.
Distinguished specialists provide an up-to-date summary of the results of
current research on their topics, as well as offering their own take on those
topics and drawing original conclusions.
size=3 width="100%" align=center data-v-7036083a=""
aria-hidden=true>
References
Abramowitz, Gab, Herger, Nadja, Gutmann, Ethan, et al. 2019. ‘ESD Reviews: Model Dependence in Multi-Model Climate
Ensembles: Weighting, Sub-Selection and Out-of-Sample Testing’. Earth System Dynamics 10 (1): 91–105.CrossRefGoogle Scholar
Alexandrova, Anna. 2017. A Philosophy for the Science of Well-Being. Oxford: Oxford University
Press.CrossRefGoogle Scholar
Alexandrova, Anna, and Fabian, Mark. 2021. ‘Democratising Measurement:
Or Why Thick Concepts Call for Coproduction’. European
Journal for Philosophy of Science 12 (1): 1–23. https://doi.org/10.1007/s13194-021-00437-7.Google Scholar
Alvich, Jason. n.d.-a. ‘Earth System (ESM4)’.
Accessed 18 November 2022. www.gfdl.noaa.gov/earth-system-esm4/.Google Scholar
Alvich, Jason. n.d.-b. ‘Model Development’.
Accessed 18 November 2022. www.gfdl.noaa.gov/model-development/.Google Scholar
Ankeny, Rachel, and Leonelli, Sabina. 2021. Model Organisms. Cambridge: Cambridge University Press.Google Scholar
Bailer-Jones, Daniela M. 2002. ‘Scientists’ Thoughts on Scientific Models’. Perspectives on Science 10 (3): 275–301.
https://doi.org/10.1162/106361402321899069.CrossRefGoogle Scholar
Bailer-Jones, Daniela M. 2009. Scientific Models in Philosophy of Science. Pittsburgh, PA: University
of Pittsburgh Press.CrossRefGoogle Scholar
Bailer-Jones, Daniela M., and Bailer-Jones, Coryn A. L.. 2002. ‘Modeling Data: Analogies in Neural Networks, Simulated
Annealing and Genetic Algorithms’. Magnani and Nersessian
2002: 147–65. https://doi.org/10.1007/978-1-4615-0605-8_9.CrossRefGoogle Scholar
Basshuysen, Philippe van, White, Lucie, Khosrowi, Donal, and Frisch, Mathias. 2021. ‘Three Ways in Which Pandemic
Models May Perform a Pandemic’. Erasmus Journal for
Philosophy and Economics 14 (1): 110–27. https://doi.org/10.23941/ejpe.v14i1.582.Google Scholar
Bergstrom, Carl T. @CT_Bergstrom. 19 April 2020. ‘I Believe That If #SARS-CoV-2 Is Allowed to Spread …’.
Twitter Page. https://twitter.com/CT_Bergstrom/status/1252075528711860224.Google Scholar
Bergstrom, Carl, and Dean, Natalie. 2020. ‘What the Proponets of “Natural” Herd Immunity Don’t
Say: Try to Reach It without a Vaccine, and Millions will Die’. New York Times, 1 May 2020. www.nytimes.com/2020/05/01/opinion/sunday/coronavirus-herd-immunity.html.Google Scholar
Biggs, Adam T., and Littlejohn, Lanny F.. 2021. ‘Revisiting the Initial
COVID-19 Pandemic Projections’. The Lancet Microbe
2 (3): e91–2. https://doi.org/10.1016/S2666-5247(21)00029-X.CrossRefGoogle ScholarPubMed
Black, Max. 1962. Models and Metaphors: Studies in Language and Philosophy. Ithaca, NY: Cornell
University Press.CrossRefGoogle Scholar
Bokulich, Alisa. 2011. ‘How Scientific Models Can Explain’. Synthese 180 (1):
33–45. https://doi.org/10.1007/s11229-009-9565-1.CrossRefGoogle Scholar
Bokulich, Alisa, and Parker, Wendy. 2021. ‘Data Models, Representation
and Adequacy-for-Purpose’. European Journal
for Philosophy of Science 11 (1): 1–26.CrossRefGoogle
ScholarPubMed
Briggs, Andrew, Sculpher, Mark, and Claxton, Karl. 2006. Decision Modelling for Health
Economic Evaluation. Oxford: Oxford University Press.CrossRefGoogle
Scholar
Britton, Tom. 2010. ‘Stochastic Epidemic Models:
A Survey’. Mathematical Biosciences 225 (1): 24–35. https://doi.org/10.1016/j.mbs.2010.01.006.CrossRefGoogle
ScholarPubMed
Britton, Tom, Ball, Frank, and Trapman, Pieter. 2020. ‘A Mathematical Model Reveals
the Influence of Population Heterogeneity on Herd Immunity to SARS-CoV-2’.
Science 369 (6505): 846–9. https://doi.org/10.1126/science.abc6810.CrossRefGoogle
ScholarPubMed
Broadbent, Alex. 2020. ‘Lockdown Is Wrong for Africa’. Mail & Guardian, 8 April 2020. https://mg.co.za/article/2020-04-08-is-lockdown-wrong-for-africa/.Google
Scholar
Broadbent, Alex, and Streicher, Pieter. 2022. ‘Can You Lock Down in a Slum? And Who Would Benefit If You
Tried? Difficult Questions about Epidemiology’s Commitment to Global Health
Inequalities during Covid-19’. Global Epidemiology
4 (December): 100074. https://doi.org/10.1016/j.gloepi.2022.100074.CrossRefGoogle
Scholar
Bunka, Mary, Ghanbarian, Shahzad, Riches, Linda, et al. 2022. ‘Collaborating with Patient
Partners to Model Clinical Care Pathways in Major Depressive Disorder: The
Benefits of Mixing Evidence and Lived Experience’. Pharmacoeconomics
40 (10): 971–7. https://doi.org/10.1007/s40273-022-01175-1.CrossRefGoogle
ScholarPubMed
Cartwright, Nancy . 1983. How the Laws of Physics Lie. Vol. 34.
Oxford: Clarendon Press.CrossRefGoogle
Scholar
Cartwright, Nancy 1989. Nature’s Capacities and Their Measurement. Oxford: Oxford University
Press.Google
Scholar
Cartwright, Nancy 2019. Nature, the Artful Modeler: Lectures on Laws, Science, How Nature
Arranges the World and How We Can Arrange It Better. Vol. 23. Chicago, IL: Open Court Publishing.Google
Scholar
Chikina, Maria, and Pegden, Wesley. 2020. ‘Modeling Strict Age-Targeted Mitigation Strategies for
COVID-19’. PLoS ONE 15
(7): e0236237. https://doi.org/10.1371/journal.pone.0236237.CrossRefGoogle
ScholarPubMed
Choi, Yeon-Woo, Tuel, Alexandre, and Eltahir, Elfatih A. B.. 2021. ‘On the Environmental
Determinants of COVID-19 Seasonality’. GeoHealth
5 (6): e2021GH000413.
https://doi.org/10.1029/2021GH000413.CrossRefGoogle
Scholar
Covid-19
Forecasting Team. 2022. ‘Variation in the COVID-19 Infection–Fatality Ratio by Age,
Time, and Geography during the Pre-Vaccine Era: A Systematic Analysis’. The Lancet 399 (10334): 1469–88. https://doi.org/10.1016/S0140-6736(21)02867-1.Google
Scholar
Douglas, Heather. 2000. ‘Inductive Risk and Values in
Science’. Philosophy of Science 67 (4): 559–79.CrossRefGoogle
Scholar
Douglas, Heather 2009. Science, Policy, and the Value-Free
Ideal. Pittsburgh, PA: University of
Pittsburgh Press. https://bit.ly/45lqr7P.CrossRefGoogle
Scholar
Downes, Stephen M. 1992. ‘The Importance of Models in
Theorizing: A Deflationary Semantic View’. PSA:
Proceedings of the Biennial Meeting of the Philosophy of Science Association
1992 (1): 142–53. https://doi.org/10.1086/psaprocbienmeetp.1992.1.192750.Google
Scholar
Wouter, Edeling, Arabnejad, Hamid, Sinclair, Robbie, et al. 2021. ‘The Impact of Uncertainty on
Predictions of the CovidSim Epidemiological Code’. Nature
Computational Science 1 (2):
128–35. https://doi.org/10.1038/s43588-021-00028-9.Google
Scholar
Elliott, Kevin Christopher. 2017. A Tapestry of Values: An Introduction
to Values in Science. New York: Oxford University Press.CrossRefGoogle
Scholar
Elliott, Kevin C., and McKaughan, Daniel J.. 2014. ‘Nonepistemic Values and the Multiple Goals of Science’.
Philosophy of Science 81 (1): 1–21.CrossRefGoogle
Scholar
Elliott, Kevin Christopher,
and Richards, Ted. 2017. Exploring
Inductive Risk: Case Studies of Values in Science. New York: Oxford
University Press.Google
Scholar
Ferguson, Neil M., Cummings, Derek A. T., Cauchemez, Simon, et al. 2005. ‘Strategies for Containing an
Emerging Influenza Pandemic in Southeast Asia’. Nature
437 (7056): 209–14.CrossRefGoogle
ScholarPubMed
Ferguson, Neil M., Cummings, Derek A. T., Fraser, Christophe, et al. 2006. ‘Strategies for Mitigating an
Influenza Pandemic’. Nature 442 (7101): 448–52.CrossRefGoogle
ScholarPubMed
Ferguson, Neil, Laydon, Daniel, Nedjati Gilani, Gemma, et al. 2020. ‘Report 9: Impact of Non-Pharmaceutical Interventions
(NPIs) to Reduce COVID19 Mortality and Healthcare Demand’. Imperial College
London. https://doi.org/10.25561/77482.CrossRefGoogle
Scholar
FitzGerald, J. Mark, Arnetorp, Sofie, Smare, Caitlin, et al. 2020. ‘The Cost-Effectiveness of
As-Needed Budesonide/Formoterol versus Low-Dose Inhaled Corticosteroid
Maintenance Therapy in Patients with Mild Asthma in the UK’. Respiratory Medicine 171: 106079. http://doi.org/10.1016/j.rmed.2020.106079.CrossRefGoogle
ScholarPubMed
Frigg, Roman, and Hartmann, Stephan. 2012. ‘Models in
Science’. In Stanford Encyclopedia of Philosophy,
edited by Edward N. Zalta. http://plato.stanford.edu/archives/fall2012/entries/models-science.Google
Scholar
Frigg, Roman, and Nguyen, James. 2021. ‘Seven Myths about the Fiction View of Models’. In Models and Idealizations in Science, edited by Alejandro
Cassini and Juan Redmond, 133–57. Cham: Springer.Google
Scholar
Frisch, Mathias. 2018. ‘Modeling Climate Policies:
The Social Cost of Carbon and Uncertainties in Climate Predictions’. In Climate Modelling: Philosophical and Conceptual Issues,
edited by Lloyd, Elisabeth A. and Winsberg, Eric, 413–48. Cham: Springer. https://doi.org/10.1007/978-3-319-65058-6_14.Google
Scholar
Giere, Ronald N. 1988. Explaining Science: A Cognitive
Approach. Chicago, IL: University of Chicago Press.CrossRefGoogle
Scholar
Giere, Ronald N. 1999. Science without Laws. Chicago, IL: University
of Chicago Press.Google
Scholar
Giere, Ronald N. 2006. Scientific Perspectivism. Chicago, IL: University
of Chicago Press.CrossRefGoogle
Scholar
Giere, Ronald N, Bickle, John, and Mauldin, Robert F. 1979. Understanding Scientific Reasoning.
Belmont, CA : Thomson/Wadsworth.Google
Scholar
Godfrey-Smith, Peter. 2009. ‘Abstractions, Idealizations,
and Evolutionary Biology’. In Mapping the Future of Biology: Evolving Concepts and Theories,
edited by Barberousse, Anouk, Morange, Michel, and Pradeu, Thomas. Boston Studies in the Philosophy of Science
266. Dordrecht: Springer
Netherlands, 47–55. https://doi.org/10.1007/978-1-4020-9636-5_4.Google
Scholar
Gray, Steven, Paolisso, Michael, Jordan, Rebecca, and Gray, Stefan, eds. 2016. Environmental Modeling with Stakeholders: Theory, Methods, and
Applications. Cham: Springer.Google
Scholar
Handel, Andreas, Longini, Ira M, and Antia, Rustom. 2007. ‘What Is the Best Control
Strategy for Multiple Infectious Disease Outbreaks?’. Proceedings of the Royal Society B: Biological Sciences 274 (1611): 833–7.
https://doi.org/10.1098/rspb.2006.0015.Google
ScholarPubMed
Hartmann, Stephan. 1995. ‘Models as a Tool for Theory
Construction: Some Strategies of Preliminary Physics’. In Theories and Models in Scientific Processes, edited by Herfel, William, Krajewski, Władysław, Niiniluoto, Ilkka, and Wójcicki, Ryszard, 49–67. Amsterdam:
Rodopi.CrossRefGoogle
Scholar
Hartmann, Stephan 1998. ‘Idealization in Quantum
Field Theory’. In Idealization IX: Idealization in
Contemporary Physics, edited by Shanks, Niall, 99–122. Amsterdam:
Rodopi.CrossRefGoogle
Scholar
https://www.cambridge.org/core/elements/scientific-models-and-decision-making/B7AC2159C941E7D0A08D9981FC8822F1?utm_source=hootsuite&utm_medium=facebook&utm_campaign=Elements_Philosophy_January_IOC&fbclid=IwAR0Ud2-ZZldn6RKaeSHYehx4l64HWKuUICG-MocZfjdy2XuSn56-n9Qs3fA
