Uncertainty Complexity and Concepts of G PDF

UNCERTAINTY, COMPLEXITY AND CONCEPTS OF GOOD SCIENCE
IN CLIMATE CHANGE MODELLING: ARE GCMs THE BEST TOOLS?
SIMON SHACKLEY1 , PETER YOUNG 2 , STUART PARKINSON 2 and

BRIAN WYNNE1
1
Centre for the Study of Environmental Change (CSEC) Bowland Tower East, Lancaster University,
Lancaster, LA1 4YN, U.K.
2
Centre for Research into Environmental System (CRES), Institute of Environmental and Biological
Sciences, Lancaster University, LAI 4YQ, U.K.
Abstract. In this paper we explore the dominant position of a particular style of scientific modelling
in the provision of policy-relevant scientific knowledge on future climate change. We describe how
the apical position of General Circulation Models (GCMs) appears to follow ‘logically’ both from
conventional understandings of scientific representation and the use of knowledge, so acquired, in
decision-making. We argue, however, that both of these particular understandings are contestable.
In addition to questioning their current policy-usefulness, we draw upon existing analyses of GCMs
which discuss model trade-offs, errors, and the effects of parameterisations, to raise questions about
the validity of the conception of complexity in conventional accounts. An alternative approach to
modelling, incorporating concepts of uncertainty, is discussed, and an illustrative example given for
the case of the global carbon cycle.
In then addressing the question of how GCMs have come to occupy their dominant position, we
argue that the development of global climate change science and global environmental ‘management’
frameworks occurs concurrently and in a mutually supportive fashion, so uniting GCMs and envi-
ronmental policy developments in certain industrialised nations and international organisations. The
more basic questions about what kinds of commitments to theories of knowledge underpin different
models of ‘complexity’ as a normative principle of ‘good science’ are concealed in this mutual rein-
forcement. Additionally, a rather technocratic policy orientation to climate change may be supported
by such science, even though it involves political choices which deserve to be more widely debated.
‘The multiplicity of models is imposed by the contradictory

demands of a complex, heterogeneous nature and a mind that
can only cope with few variables at a time; by the contradic-
tory desiderata of generality, realism, and precision; by the
need to understand and also to control; even by the opposing
esthetic standards which emphasize the stark simplicity and
power of a general theorem as against the richness and the
diversity of living nature. These conflicts are irreconcilable’
(Richard Levins) [1].
1. Introduction
The information available on climate change for policy making purposes is plagued
by large inherent uncertainties. This includes uncertainties in the climate change
models, as well as in the models of climate change impacts, economic costs and
policy responses. It also includes those uncertainties which affect the social and
Climatic Change 38: 159–205, 1998.

c 1998 Kluwer Academic Publishers. Printed in the Netherlands.
160 SIMON SHACKLEY ET AL.
policy contexts into which such knowledge is intended to be used. In such situa-
tions of endemic social and political uncertainty, policy analysts have frequently
recommended a strategy of pluralism. They have reasoned that a diversity of policy
approaches should be adopted to increase the likelihood that at least some of these
will prove useful and successful despite the uncertainty that characterises both the
present and future in this most challenging area of study [2]. This argument is, how-
ever, rarely extended to the provision of scientific knowledge itself. That is because
objective criteria for defining ‘good science’ are widely held to be available which
can be used to ‘sort out’ competing approaches to the production of new scientific
knowledge, such that only the most successful and promising are supported. Thus
consolidation around General Circulation Models (GCMs) in much contemporary
climate change research is explained by the higher rating that the climate science
community gives to GCMs in comparison with alternative approaches.
In this paper, we present a different, less widely known, viewpoint according
to which the criteria defining good science for policy are not solely derived from
science, with its currently cherished paradigms, but also incorporate social and
policy judgements. Little has been written about the role of such non-scientific
judgements in the provision of climate science for policy [3]. Past studies of envi-
ronmental policy have suggested that the successful reception and implementation
of environmental policy is dependent upon a coalition of assorted actors (such as
scientists, policy makers, environmentalists and industrialists), often with diver-
gent expectations and agendas, yet who come together for the specific support
of certain policy measures (albeit it sometimes for different reasons) [4]. Whilst
shared scientific knowledge and understanding is frequently important, such stud-
ies suggest also the major role of shared social commitments about the issue and its
solution. This insight is important in two respects: it disputes the impression, still
found amongst some, that science by itself is sufficient to determine environmental
policy; and it suggests that, for policy determination purposes, a comparison of
competing interpretations of knowledge might come to include more than purely
scientific considerations.
We suggest that although GCMs are widely considered to be the ‘best science’
for the study of future climate change, this judgement is significantly influenced
by factors which are not purely scientific. For the following reasons, reflection on
these non-scientific factors which, we argue, also support their dominant position,
is important to the overall resilience of the scientific research base for global
environmental change policy making.
Firstly, it is easier to argue a case, say, for research funding, if all the reasons
in support of that case are fully appreciated.
Secondly, the validity of non-scientific elements, especially when used in
contexts of application different from those in which the criteria were originally
developed, can be better appreciated and evaluated.
UNCERTAINTY, COMPLEXITY AND CONCEPTS OF GOOD SCIENCE 161
Thirdly, research policy can be made more resilient in the face of changes in
the social and policy worlds if it takes into account the role of social and policy
factors in contributing to notions of ‘good science’.
Fourthly, without the critical scrutiny of non-scientific elements, there is a
danger that climate change science will unconsciously come to support a
technocratic approach to policy which always puts its main emphasis on an
instrumental management and control of the environment and society. While
this eventuality may be sensible, depending upon the circumstances, it always
needs to be considered carefully, and more widely debated, in terms of its
possibility, negative consequences and desirability, as a way of relating to the
environment and to each other.
But the importance of non-scientific elements is not the only factor to be considered
when questioning the role of GCMs in the study of future climate change. Another
important dimension of our argument is that there are other scientific approaches to
the study of climate change which are more robust and promising than their current
(low) standing at the science-policy interface suggests. We argue that the relative
lack of attention to alternatives is caused, in part, through their apparent weaker
connection to policy and social goals and agendas (than GCMs). But these social
and policy goals are not necessarily sovereign or unchanging. Hence, analysis of
the non-scientific elements involved in defining ‘good science’ enables a better
appreciation of alternative policy and scientific approaches and argues for more
plurality in science. Our position is that, while GCMs clearly have an important
role to play, other scientific alternatives should be further developed for possible
use at the science-policy interface so as to support and complement the information
derived from GCMs.
In Section 2, we support our claim that GCMs are presently dominant in produc-
ing knowledge of possible future climate change. Sections 3 and 4 both critically
examine the conventional set of reasons in support of GCMs as being dominant,
and find them questionable. In Section 5 we provide a brief illustrative example of
an alternative modelling strategy based on different epistemic and methodological
commitments to representing complexity and uncertainty. This serves to reinforce
the point that more discerning scientific and other criteria should be involved in
defining ‘good science’ in the climate change domain, and these are analysed more
systematically in Section 6.
2. The Climate Modelling Pyramid
This section briefly considers the current and emerging future role of GCMs in
climate change research, using the common metaphor of the climate modelling
pyramid, and clarifies the criteria by which models are ranked by climate scientists.
2.1. THE ROLE OF GCMs IN CLIMATE RESEARCH
One of the most prominent climate change research activities has been the develop-
ment and application of General Circulation Models (GCMs) of the global climate.
GCMs have been used to simulate the consequences of an increase in atmospheric
CO2 on the mean global climate, both in instantaneous doubling of CO2 equilibri-
um simulations, and time-dependent simulations in which the CO2 concentration
is increased incrementally over a number of model years. Coupling of atmospheric
GCMs (AGCMs) with 3-D dynamical ocean models (OGCMs), together with more
or less complex models of the land surface vegetation and sea-ice, represents the
present ‘state of the art’ in such scientific research. Because of their low resolution
and large errors in the simulation of regional feedbacks, however, GCMs can-
not presently provide detailed regional climate simulations or estimates of future
regional climate change. Nevertheless, it is assumed by many GCM modellers
(‘GCMers’) that the regional errors do not compromise the validity of the global
response to CO2 forcing. This is because, provided the major, large-scale features
are simulated and energy at the top of the atmosphere (TOA) is balanced, then the
model will produce a new global equilibrium position which is largely independent
of what occurs regionally.
Support for this latter view arises from the reasonable level of agreement
between the majority of GCMs in terms of climate sensitivity, despite the wide
divergences in their different simulations of regional climate change [5]. Many
climatic feedbacks, especially ones related to atmospheric chemistry and biologi-
cal processes, are not currently included in GCMs and much current work at the
interface of climatology, geophysics, environmental physics and ecology is aimed
at understanding such feedbacks and addressing how to include them in GCMs [6].
The development and application of GCMs is not only a major component of
the climate change research field in financial terms; there is also a powerful logic
in their role as the ‘lynch-pin’ of such a growing area of research. This arises from
the scientific and policy consensus that GCMs provide more reliable scientific
inference of future climate change than any other model or scientific methodology
currently available. And, given that computer models have been highly significant
in providing ‘evidence’ for the enhanced greenhouse effect, this puts GCMs at the
crux of the current scientific mission to provide reliable scenarios of future climate
change. ‘Reliable’ in this context means two things: firstly, that the proponents of
GCMs consider that they represent the current climate system more realistically
than other models; and secondly, because of this, that GCMs are more likely
than other models to incorporate accurately those climatic processes which will
be important in simulating the climatic changes ensuing from the atmospheric
accumulation of greenhouse gases (as well as from other changes in radiative
forcing).
2.2. THE RELATIVE RANKING OF CLIMATE MODELS
GCMs are widely considered by climate scientists to represent the climate system
more realistically than other models, such as the 0, 1 or 2 dimensional energy bal-
ance models, radiative-convective models and statistical-dynamical models. They
do so, according to Shine and Henderson-Sellers, because of ‘the relative merit and
complexity with which these four [basic climate] elements [of dynamics, radiation,
surface processes and resolution] are included’ [7]. These authors consider that
different climate models can be positioned in a pyramidal space, as shown in Fig-
ure 1, according to the completeness with which the basic elements are included.
GCMs emerge as the most integrated of models, even if: ‘they do not : : : incorpo-
rate all aspects of the climate system and are, therefore, not at the absolute apex
of the pyramid’ [8]. Greater realism is defined by Shine and Henderson-Sellers,
therefore, in terms of the complexity of the interaction of the primary processes at
the highest possible resolution, rather than the complexity per se with which the
primary processes are represented.
Thus one-dimensional radiative-convective models (RCMs), which treat the
atmosphere as a one-dimensional column, sometimes have a more sophisticated
handling of radiation and other physical parameterisations than do GCMs. And
Wigley and Raper’s one-dimensional STUGE energy balance model included the
effects of CO2 fertilisation, feedback from stratospheric ozone depletion and the
radiative effects of sulphate aerosols; whereas contemporaneous GCMs did not
include these key climatic processes [9].
Shine and Henderson-Sellers remark that: ‘The apex [of the modelling pyramid]
suggests that when all facets are correctly and adequately incorporated at a high
enough resolution a model, presumably identical to the real climate, results’ [10].
Hence, the complexity of interaction of the dynamics and physical parameterisa-
tions, properly formulated, is held to generate realism (and these criteria can be
used to sort out the adequacy of different models).
The hierarchy of the climate modelling pyramid is legion. In the volume Climate
System Modeling, for example, Schneider notes that: ‘Many scientists believe
that the ultimate goal of climate modeling should be fully comprehensive, three-
dimensional models of all elements of the climate system including very high
resolution and as much detail as possible’ [11]. In the same volume, Kiehl organises
a discussion of AGCMs around the model hierarchy, defining model superiority in
terms of the extent to which models avoid spatial averaging of the basic hydro- and
thermodynamic equations for the atmosphere: more averaging is clearly equivalent
to greater approximation and less realism [12].
GCMers readily admit to the omissions of some climate-related feedbacks in
their current models and, for this and other reasons, opinion is divided on how dis-
tant GCMs are from the pyramid’s apex. What is widely agreed upon, however, is
that the key, first-order climate feedbacks are incorporated, and that feedbacks can:
‘be predicted credibly only by physically-based models that include the essential
Figure 1. The climate-modelling pyramid. The position of a model on the pyramid indicates the
complexity with which the three primary processes interact. The base of the pyramid can be considered
hollow since there is essentially no interaction between the primary processes. Progression up the
pyramid leads to greater interaction between each primary process. The vertical axis is not intended
to be quantitative. (a) The positions of the four basic model types. (b) Particular climate models
are seen to be based upon different methods of incorporating the primary processes and the level of
complexity of the interactions (after Shine and Henderson-Sellers, 1983). (From Henderson-Sellers
and McGuffie, 1987 [8]; Figure 2.1, page 37.)
dynamics and thermodynamics of the feedback processes. Such physically-based

models are the general circulation models’ [13]. In semi-public contexts, the physi-
cal basis of GCMs, compared to other methods, is a frequently heard claim for their
superiority (though a number of physical laws, such as conservation of energy, are
also incorporated into most simple climate models). When modellers are commu-
nicating to non-scientists – for example to journalists, during television broadcasts,
and in front of Congressional Committees – it is not uncommon for them to respond
to scepticism by pointing to the fundamental physical laws underpinning GCMs
[14].
Our interviews with climate modellers provide further support for the perceived
superiority of GCMs compared to other climate models. One modeller discussed
the findings of the first transient, coupled AOGCM runs in the early 1990s compared
to earlier, simpler models: ‘the box-diffusion model [used in the Intergovernmental
Panel on Climate Change 1990 assessment] we couldn’t take too seriously. There
are things you could anticipate but couldn’t prove without the coupled AOGCM
model’ [15]. Hence AOGCMs provided proof of climatic change, whereas the one-
dimensional model only indicated a possible effect. For another GCM modeller:
‘Simpler models are useful for isolating and illustrating points in an easily
understandable manner. : : : They are essential for understanding but useless
for prediction’ [16].
Yet another modeller, who has worked extensively with a range of climate models
including GCMs, described how objections to the use of simple models emerged
in the peer review process, both at the proposal and paper submission stages [17].
2.3. THE PERCEIVED ROLE OF SIMPLE MODELS
The existence of a model hierarchy does not imply that simple models are not
valued within the GCM community. On the contrary, they are typically regarded as
extremely useful tools in model development and testing [18]. Simpler models are
portrayed as useful heuristics to generate insights about the processes involved in
climate change, and to aid further understanding for teaching or policy purposes.
But their scientific sophistication is inevitably compromised by the requirements
imposed by such teaching, policy or open-ended exploration requirements, so that
they are not seen as credible simulations of reality as GCMs are claimed, or aim, to
be. In the dominant view, completeness and accuracy can only be traded-off against
tractability and economy in the learning stages of the model development process,
not when realism is sought. Often, the same judgement is made by many users and
advocates of the simpler models, who routinely calibrate the sensitivity of simpler
models to CO2 -doubling by using the simulations of GCMs. As one modeller, who
normally uses simpler climate models, put it:
‘: : : ultimately one would have to say that [GCM development] is perhaps the
only way, in the end, that we’re going to generate better answers ultimately
because, in principle, the way they [GCMs] are constructed and conceived
enables them to take on board all components of the climate system in a
dynamic way; in a physically realistic way and with a capability for looking
at feedbacks. The fact that they haven’t done so to date doesn’t mean that they
are incapable of doing so. In terms of intellectual progress, then ultimately
that is going to be probably the route that is most successful’ [19].
The use of the words ‘ultimately’ and ‘in principle’, and the general tone of the
above passage, suggests that GCMs are assessed, to an important extent, by their
future potential to realistically represent the interaction of primary processes.
Whilst the perspective of Meehl, Henderson-Sellers, and Kiehl is reasonably
typical, there are other climatologists and modellers who take a more pluralist
stance, advocating the validity of different models from across the hierarchy, or a
more pragmatic, instrumental position towards climate models, in which, as one
climatologist put it: ‘Different models are valuable for different purposes. At the
end of the day you can’t say that one is better than the other’ [20]. Certainly, it is this
view that models should be ‘objective-oriented’, that is they should be developed
and utilised in different ways for different purposes, that has most appeal to the
present authors.
2.4. REASONS IN SUPPORT OF THE CLIMATE MODELLING PYRAMID
The climate system is widely perceived as being complex, not only in terms of the
characteristics of the processes involved (such as their non-linearity and the level
of interconnection between sub-systems) but also with respect to their sheer range:
i.e., the broad scope of interactive physical, chemical and biological processes in
earth systems which influence climate. It is also widely accepted that this level of
complexity is far beyond the power of any present climate models.
One dominant response to these arguments is the inclusion of ever more detail
so as to increase the perceived accuracy of the model; i.e., to meet complexity
in nature with model complexity, as a way of improving understanding of the
interactive dynamical physical processes, as well as for thorough testing of the
model’s sensitivity to certain processes and possible policy scenarios. This strategy
assumes that a more complex model may lay claim to be a better representation of
a complex system and, hence, that it has a greater truth-content than other models.
From this standpoint, redundancy in model detail is assumed to be a very long way
off.
Thus, the complexity of model representation becomes a central normative
principle in evaluating ‘good science’ in the climate research domain. It is argued
that the predictive potential of the model is enhanced; the uncertainty reduced or
better defined; and, as a further bonus, the policy-usefulness of the model output
is strengthened relative to simpler models. It is this ‘central dogma’ – that greater
complexity equals greater realism, equals greater policy-utility – which we set out
to explore critically in this paper.
3. Challenges to the Climate Modelling Pyramid I: Applying the Scientific

Criteria of Evaluation
A common criticism of pluralism is to argue that not all options are as good as one
another and that there are objective scientific criteria of evaluation which sort out
the adequacy of the various model options. Two such criteria advocated by many
climate modellers, which emerge from the discussion in Section 2, are:
to what extent is the model physically-based?; and,
what is the complexity with which the physical and dynamical processes, and
especially their interactions, have been modelled?
We share the general perspective that different options should be critically eval-
uated, but question whether the above criteria alone can be used to differentiate
between climate models. For this set of evaluative scientific criteria to be credible
in explaining the dominance of GCMs, they would clearly need to be demonstrably
applicable and widely accepted amongst the peer-group. Otherwise who decides
what the criteria should be, how these might change as model applications change,
and how they are applied and interpreted? In this section, some of the practical
difficulties of applying criteria for climate model evaluation are discussed, using
examples from the published literature.
3.1. PARAMETERISATIONS IN GCMs
Many climate phenomena occur on smaller scales than explicitly resolved in the
three-dimensional grid of a GCM (which is typically several hundred kilometres
in the horizontal directions). Parameter representations, or parameterisations, are
used in GCMs to represent physical processes, such as clouds, collectively at the
sub-grid scale. Parameterisations often consist of a statistical relationship (based
on climatological data) between variables that are resolved at the grid scale and
ones that are not, and which can then be used to simulate the unresolved physical
process. The modeller uses his or her judgement in developing a parameterisation,
based on physical theory and understanding, empirical data, intuition, other facets
of the model, and so on. Parameterisations have frequently been presented more
publicly as a problem involving a lack of sufficient detail in the model; a problem
which can be partially solved by increasing model resolution [21]. However, this is
not the view of other modellers, especially when writing in more specialist contexts.
For example, Stephen Schneider comments that:
‘Modelers, of course, strive to keep their parameterizations as physical and
non-empirical and scale-independent as practical. Thus, the validity of para-
meterization and overall model performance as well, depends ultimately on
empirical tests, not only on the inclusiveness of the first principles. In other
words, even our most sophisticated “first principles” models contain ‘empirical
statistical’ elements within the model structure’ [22].
In reality, the key parameterisations, whilst they are usually based partially on the
physical understanding of the scientist, are rarely fully theoretically-based. As a
result, scale-dependence and some degree of arbitrariness is the rule rather than
the exception [23]. Moreover, whilst the common portrayals of GCMs in public
contexts present parameterisations as introducing an uncertainty at the sub-grid
scale, hence making simulation at that level alone suspect, their effect also filters
through at the resolved-scale level. As one modeller put it in a letter to us: ‘Since the
simulation of the whole model is determined by the fidelity of the parameterizations,
a climate model as a whole is not deterministic’ [24].
There is an equivocal use of the word ‘deterministic’ here: each run of the GCM
is clearly deterministic in the usual scientific meaning of the word, since the effects
of uncertainty are not accounted for. If uncertainty prevails in the thinking and
practice of the GCM modellers and different parameterisations, or scenarios, are
utilised to reflect this uncertainty, then all well and good. But such procedures are
certainly not rigorous exercises in uncertainty analysis and it is determinism that
permeates the overall process of the mathematical analysis.
It would seem, therefore, that even if one accepts that the physical basis of
a model is the key criterion by which to assess its adequacy, GCMs do not rate
quite so highly as their frequent, semi-public representation indicates. Although
they might rate more highly than other climate models in terms of the attempted
physical representation of interactive processes, this does not correspond with
physical realism in any straightforward way for two reasons. Firstly, as already
noted, particular physical processes may be more completely physically-based in
simpler models. Secondly, a reductionist approach to the representation of physical
processes, as in GCMs, does not automatically guarantee physical realism [25].
3.2. PARAMETERISATIONS AND MODEL INTERACTIONS
Stone and Risbey analysed the meridional flux of heat in the atmosphere simulated
by three AGCMs (versions of the GFDL, GISS and NCAR AGCMs, descriptions of
which were published in the early to mid-1980s) [26]. It turned out that such fluxes
were poorly related to observations (with errors of typically 50%), notwithstanding
the models’ successful simulation of observed surface air temperatures (SATs). If a
more ‘realistic’ meridional heat transport is used to constrain the models, however,
the SATs then become very inaccurate.
This raises the question as to whether it is best to maximise the simulation
of either the SATs or the meridional heat flux; or whether to find a compromise
between such objectives. Some modellers have argued that, since the observational
record of meridional heat fluxes is inferior to that of SATs, discrepancy of the
model with the former is a less reliable measure of its accuracy. From a statistical
standpoint, there is no doubt that these differences in measurement uncertainty
should be taken into account when modelling on the basis of such observations. Yet
Stone and Risbey rebuke modellers for always assuming that it is the observations
which are at fault rather than the structure and parameterisations within the models,
without providing detailed reasons for this decision.
For example, many GCMers have stated a preference for the calculated merid-
ional heat transports of the European Centre for Medium Range Weather Forecast-
ing’s (ECMWF’s) model over the observations of the rawinsonde network analysed
by Oort; where the latter shows greater discrepancy from the GCMs than the former
[27]. The ECMWF’s approach is based on operational analysis, i.e., one in which
the model’s simulation and available observational data are merged. The implicit
argument of the GCMers has been that the ECMWF’s data are superior because
the model generates reliable data where there are no observations. But this view
can be questioned in two ways: firstly, it is effectively using a model to validate a
model; secondly, in a paper published in 1978, Oort actually tested the sensitivity
of the heat transport to the observational rawinsonde network by supposing that
the model was, in fact, a representation of reality, interpolating what would be the
energy fluxes at the specific positions of the rawinsonde network, and comparing
the heat transport implied by these few points to that of the fully resolved model.
He found that the current observational network was, in fact, adequate (this being
because the processes by which heat is transported are large-scale).
Stone and Risbey also analysed the sub-grid scale vertical flux (associated with
moist convection and boundary-layer processes) which largely determines the ver-
tical profile of temperature. To test the interactions between the sub-grid scale
parameterisations and the vertical temperature profile in the absence of observa-
tional data on vertical heat fluxes, Stone and Risbey performed a sensitivity analysis
by comparing the vertical dynamical heat fluxes in the GISS Model 1 using two
different convection schemes: a moist adiabatic adjustment scheme and a penetra-
tive convection scheme. As expected, the results were completely different for the
small-scale convective flux, since the latter is a function of the parameterisation. Yet
the large-scale flux was also different for the two parameterisations (50% smaller
when the penetrative convection scheme was used). According to Stone and Risbey
this finding is significant since:
‘If, then, current GCMs do have an advantage (over other models) in simulating
vertical dynamical heat fluxes, it must be because they can explicitly and
accurately calculate the vertical fluxes associated with large scale motions.
: : : GCMs’ large scale heat fluxes interact so strongly with the parameterized
subgrid scale physics that their calculations of the large scale fluxes are no
more accurate than their (uncertain) calculation of the subgrid scale physics’.
And they conclude:
‘: : : we doubt that the GCMs currently in use for climate change experi-
ments are better at simulating global scale climate feedbacks and temperature
changes than simpler models that do not simulate large-scale processes from
first principles’ [28].
This is a very important caveat, for it recognises that models should be related
to the scale of the processes involved: the reductionist argument that large scale
behaviour can be represented by the aggregative effects of smaller scale process has
never been validated in the context of natural environmental systems and is even
difficult to justify when modelling complex man-made processes in engineering
[29].
3.2.1. Response from GCMers

One GCM modeller reacted to Stone and Risbey’s result in the following way:
‘That is something you can argue either way : : : say “well if the convective
scheme is parameterising what it says it is simulating, the precipitation for
different schemes should be the same” : : : but I don’t think we would nec-
essarily expect two convection schemes to produce the same amount because
they inevitably interact with large-scale processes : : : What you’re trying to
do is to get the net effect to be right’ [30].
In other words, for this modeller, the convection schemes should be evaluated
by assessing how well they simulate global patterns of precipitation: this being
the net result of many processes (both sub-grid and large scale processes) acting
in different directions. The above modeller went on (though perhaps somewhat
tongue-in-cheek) to explain how he thought another, world-famous, GCM modeller
might react to Stone and Risbey’s finding:
‘He would take the view that the model knows what it wants to do and whether
it does it [distribution of water vapour] through the convection scheme or
through the large-scale rain, for example, is not so critical. The main thing is
that you’re getting the patterns right and the right overall sensitivity’ (ibid.).
This is effectively stating that it is satisfactory to achieve the right result for the
wrong reason, a position which would be challenged by many modellers. Never-
theless, these views create a dilemma for assessing the complexity of interaction
upon which the modelling pyramid is commonly based. If the analytical reduction
of the model to its constituent parts compromises the special interactive properties
of the model, how can the complexity and accuracy of the interaction between the
primary processes be assessed properly? Clearly, in delegating the issue of what is
the appropriate interaction of dynamics and parameterisations to the model ‘itself’,
an a priori commitment to the validity of GCMs (and their dominant position on
the modelling pyramid) is entertained.
According to other modellers, including Stone and Risbey, this response is
unsatisfactory because, if the division of flux between the sub-grid and resolved
large-scale processes is changed, different dynamical processes become involved,
which are also likely to have different feedbacks. Thus, the response of the model
to a change in forcing, such as that from greenhouse gases, is likely to be affected
by the presence of significant model errors [31]. In addition, some commentators
feel that there is an unscientific circularity in some of the arguments provided by
GCMers; for example, the claim that GCMs may produce a good simulation sits
uneasily with the fact that important aspects of that simulation rely upon somewhat
arbitrary parameterisations, tuning and prescribed boundary conditions.
3.3. FORMAL AND INFORMAL STYLES OF MODEL EVALUATION
In summary, there appear to be two different strategies in operation within the sci-
entific community for evaluating GCMs. One involves a relatively explicit analysis
of model errors, including attempts to understand the detailed nature of the mod-
el interactions. This we term a ‘formal evaluation strategy’. The other approach,
more common within the GCM community, is a more intuitive, model-trusting,
and subjective approach, which we term an ‘informal evaluation strategy’ [32]. In
the formal strategy, the desire is to explicitly calibrate the detail and adequacy of
interaction within the model – to get ‘inside’ the model, so to speak. In the informal
method, more trust is placed in the GCM modeller’s personal judgement as to the
complexity and adequacy of model interactions.
Our point is not to choose one strategy of evaluation over the other, since both
would seem to have some value. It is rather that their existence poses a problem
for the evaluation of GCMs, and their positioning vis-a-vis other climate models,
according to any one set of criteria. Also the choice of the particular scientific style
adopted by the modeller comes to influence the sorts of testing that can readily
be performed on the model. The complexity of GCMs tends to limit the extent to
which the model processes, interactions and uncertainties can be understood and
analysed. This is because of the range and interdependence of such processes and
the practical and analytical limitations thereby imposed upon our understanding of
the model.
We have focused here upon several aspects of GCMs which are puzzling and/or
troubling to the claims frequently made for such models, especially as far as their
position in a modelling pyramid is concerned. There are additional features of
GCMs which raise very similar questions and problems [33]. A more in-depth
analysis of such additional facets of GCMs would be important in ascertaining
the generality and significance of the results we have drawn upon above. We are
not suggesting that other models are somehow devoid of these and other prob-
lems. We are simply pointing out an inconsistency in the arguments of many
climate modellers. They frequently consider the output from simple models to be
unreliable because they are ‘over-parameterised’, in the sense of over-relying on
parameterisations to simulate climate processes [34]. GCMs by contrast are held
to yield quantitatively meaningful projections of future climate change. Yet, this
reasoning sits uncomfortably with the crucial dependency of GCMs on their own
parameterisations [35].
4. Challenges to the Climate Modelling Pyramid II: Complexity and Model

Construction
This section advances the challenge to the climate modelling pyramid by question-
ing whether the evaluative criteria which inform the climate modelling pyramid
are themselves the only appropriate set of criteria.
4.1. ARGUMENTS FOR SIMPLICITY
The arguments about the strengths and limitations of GCMs tend to centre on
whether or not one believes that the apparently complex global climate system
can only be represented adequately by a complex simulation model. And yet one
dominant theme in scientific thought has been to espouse a preference for simplicity
in scientific explanation (sometimes called the principle of ‘Occam’s Razor’) [36].
In theoretical physics, for example, there has been a strong feeling that simplicity
underlies seemingly complex behaviour. As Einstein once stated it: ‘In a reasonable
theory, there are no (dimensionless) numbers whose values are only empirically
determinable : : : this world is not such as to make an ‘ugly’ construction necessary
for its theoretical comprehension’ [37]. The underlying principle of parsimony
maintains that, all else being equal (e.g., assuming it could explain or correctly
predict as much behaviour), the simpler representation of a system’s behaviour is
preferable to unnecessary complexity. Popper, in like-mind, wrote that:
‘The method of science depends on our attempts to describe the world with
simple theories. Theories that are complex may become untestable, even if
they happen to be true. Science may be described as the art of systematic
over-simplification: the art of discerning what we may with advantage omit’
[38].
Popper’s emphasis on the testability of models reflects his central idea that falsifica-
tion of knowledge-claims is the only secure route to improving the knowledge-base.
Because simpler models are easier to falsify they are preferable from a Popperian
standpoint.
Apparently simple models can, however, display highly complex behaviour
(as in chaotic models, such as the classic Lorenzian attractor model). Indeed,
falsification of an apparently simple model may be an artefact of the domain in
which the model resides under strict test conditions. It has been suggested that
‘falsification’ may, in any case, be a problematic concept for two reasons: first,
scientists normally adjust their theories in small ways to avoid discrepancy with
observations; and second, falsification relies upon faith in the reproducibility of
the falsification of a low-level empirical hypothesis. Some detailed studies of
scientific practice have shown how theory and model adjustments are routine and
reproducibility is highly indeterminate [39]. There is always, therefore, more than
logic and methodology involved in deciding whether the falsification of a hypothesis
or model has been achieved or reproduced.
Hence, whilst intuitively plausible, it remains an open question whether model
complexity makes falsification any more problematic. But there is, possibly, a
sociological and even political sense in which Popper might often be right. In the
case of a simple model, there is often greater transparency to the outside observer in
the assumptions imported into the model, including greater openness in the relations
between different scientific disciplines involved in its production, than is the case
for a complex model [40]. Arguably, such transparency improves understanding and
the possibility of model analysis and testing, as well as increasing the modellers’
self-awareness of the commitments entered into and their conditionalities, all of
which would improve the opportunities for falsification (or confirmation).
4.2. COMPLEXITY AS A MORE HOLISTIC SCIENCE?
Those who advocate complex models respond to the above point by arguing that the
traditional physical approach, and its demonstrated use of parsimony in scientific
explanation, is not applicable to open systems. The real world is being observed
passively and planned experiments, of the kind conducted in the laboratory, are
the exception rather than the rule. Indeed, the apparent beauty and simplicity of
physical theory is sometimes even regarded as an artefact of the controlled space
and conditions of a laboratory, or of the epistemic commitments of scientists
[41]. Against this, scientific representations of open-ended environmental systems
may well appear ‘ugly’; but surely only if they are judged against received and
inappropriate normative criteria.
Some advocates of GCMs point to such models as representing a more scientific
approach to a new kind of scientific problem, namely the modelling of multivariable,
open environmental systems; in contrast to modelling the kind of closed systems
that are associated with planned experimentation in the laboratory sciences [42].
They argue that it is only through complex models, which are able to simulate the
subtle, often non-linear, interactions between dynamics and thermodynamics, that
counter-intuitive, unexpected, and abrupt processes of change (such as those which
litter the palaeoclimatological literature) will be successfully simulated [43].
An illustrative argument of this type in favour of GCMs can be put in the follow-
ing way (after Palmer [44]). The non-linearity of the climate system means that the
sensitivity pattern of the global atmosphere to perturbations is not coincident with
its response. The consequences of increasing greenhouse gases will be transmitted
through changes in the frequency and stability of different regime structures in the
atmosphere. It is, therefore, imperative to understand and document the various
possible spatio-temporal regime structures, so demanding a 3-D spatio-temporal
representation which is capable of adequately simulating natural modes of vari-
ability, stability or instability, and multimodality. As Palmer puts it: ‘the nonlinear
perspective suggests that there is no simple short cut to the comprehensive mod-
elling studies currently under way, if one requires a quantitative knowledge of
the effects of enhanced CO2 ’ [45]. Taken to its extreme, the implication of these
arguments is that even the global response of simple models to a change in forcing
is unreliable [46].
Even if spatio-temporally distributed systems, such as the global climate, do
require a modelling approach which is substantially GCMs-based, it is unclear
whether current GCMs and coupled AOGCMs are adequate to the task anticipated
by Palmer (a point he also makes). For them to be so, we must assume that
climate GCMs can and do reliably reproduce the complex, non-linear dynamics
and thermodynamics of the real climate system. Yet this is, to some extent, unknown
and has to be taken on trust, or in the faith that future model development will deliver
such an improved model [47].
Model approximations and tuning are routinely performed in GCM work which
invest heavily in pre-suppositions about the character of the response. An example
is the adoption of the assumption that the difference between the doubling of CO2
perturbation and the control runs is the same as that between the ‘real’ climatology
(the mean value of climate variables over a specified time-period, usually 30 years)
and the ‘real’ perturbed climate. Therefore, the errors which result in the difference
between the control run and climatology are not believed to move the climate
feedbacks out of a linear regime. Hence, it is questionable whether the current
generation of GCMs, as used in climate change experiments, can provide what
Palmer outlines as necessary in order to properly simulate the spatio-temporal
response to a change in forcing. This limitation of GCMs is not meant as a criticism,
however; it is more an acknowledgement of the difficulty of the scientific task and
is one, therefore, also shared across similar scientific fields and modelling.
The more general issue here is that the propensity of complex models to generate
surprises is queried when the necessary judgement of the modeller in producing and
running a model which is plausible (according to the prevailing understanding of
the GCM community) is acknowledged [48]. Hence, some of the force of arguments
in favour of complexity, such as those of Palmer, refer not to the current generation
of GCMs, but rather to the anticipated, potential achievements of future, more
complex, GCMs. To observe that this is a further judgement by GCMers is not
to claim that it is right or wrong, but rather to open the way for other scientific
judgements on such issues to be heard.
5. Challenges to the Climate Modelling Pyramid III: The Need for

Objective-Oriented, Stochastic Models
In this paper, the model building process used to formulate and construct the GCM
is considered as a prime example of ‘deterministic reductionism’. By reduction-
ism, we mean here the process of ‘reducing’ a complex system to the sum of its
perceived component parts (or sub-systems) and then constructing a model based
on the assumed interconnection of the sub-models for these many parts. This is
not, of course, a process which necessarily reduces the model in size at all: on the
contrary, it normally leads to more complex models, like the GCM, because most
scientists feel that the apparent complexity that they see in the natural world should
be reflected in a complex model: namely a myriad of ‘physically meaningful’ and
interconnected sub-systems, each governed by the ‘laws of nature’, applied at the
micro-scale but allowed to define the dynamic behaviour at the macro-scale, in a
manner almost totally specified by the scientist’s (and usually his/her peer group’s)
perception of the system. This reductionist philosophy of the GCM model is ‘deter-
ministic’ because the models are constructed on purely deterministic principles.
The scientist may accept that the model is a representation of an uncertain reality
but this is not reflected at all in the model equations: the GCM is the numerical
(always approximate) solution of a complex but purely deterministic set of non-
linear partial differential equations over a defined spatio-temporal grid, and no
attempt is made to introduce any quantification of uncertainty into its construction.
The admission of uncertainty often only enters the scene when the model is later
utilised in a predictive sense, normally by considering a range of different, but again
deterministic, simulation ‘scenarios’ that reflect the model-user’s lack of precise
knowledge about the model parameters and its inputs. However, the propagation
of uncertainty within the non-linear GCM, and its effect on the model predictions,
is not quantified by any systematic, statistical analysis of the kind considered later
in this section.
Attempts to quantify model uncertainty have been made in relation to other
global models; e.g., those which attempt to represent global carbon cycle dynamics
[49]. But the consequences of such uncertainty are then only evaluated using
fairly rudimentary methods of deterministic sensitivity analysis. In particular, there
appears to have been little research on the effect that stochastic influences, such
as those arising from uncertain inputs and model parameters, might have on the
model outputs and predictions. One notable exception is the work of Gardner and
Trabalka (1985) who used Monte Carlo Simulation (MCS) analysis to investigate
the implications of such stochastic effects on the Oak Ridge National Laboratory’s
World Carbon Cycle Model [50].
Deterministic reductionism, as discussed above, is by no means the norm in
mathematical modelling. In many areas of environmental science, engineering,
economics and statistics, a mathematical model is considered only as an approxi-
mate representation of reality, so that the quantification of uncertainty and its effects
on model predictions is considered sine qua non for successful modelling. Also,
whilst there is usually little argument over the fundamental physical laws which
underlie the formulation of the model, the way in which these laws are utilised in
model construction can differ widely, depending upon the perception of the model
builder and the use for which the model is intended. Under this more pluralistic
and objective-oriented modelling philosophy, the concept of a completely gener-
al, all-encompassing model, which is accepted by all scientists and satisfies all
requirements, is considered an idealistic fantasy. Rather, in any particular practical
application, a variety of models are likely to be required, each of which is tailored to
a different, fairly specific objective and defined at an appropriate scale of behaviour
and measurement.
This pragmatic, objective-oriented view of model building sits uncomfortably
within the concept of a modelling pyramid: the idea that one model in the ‘family’
of plausible models has some special characteristics which make it universally
‘better’ than others is difficult to sustain in rational terms (though at any one
time a specific model might be the best for a particular purpose). At this point
in time, the GCM might well be considered as the ‘best’, albeit far from ideal,
currently available model for certain objectives, such as the exploration of different
hypotheses about the detailed, regional nature of the global climate dynamics, and
the causes of observed patterns of climate change, or the generation of speculative
regional forecasts that cannot be provided by other currently available and more
aggregated models. But, as a vehicle for making probabilistic forecasts about the
global mean temperature variations into the next millennium, it clearly has certain
severe disadvantages, not the least being its totally deterministic nature and the
complexity of its solution for such long period forecasts. It should be recalled that,
even if the theoretical equations of the GCM were able to provide exact descriptions
of the global climate, which surely most scientists would consider impossible, the
numerical solution of these equations in the computer would still be approximate
since an accurate, analytic solution of the equations is not possible. Moreover, the
nature of the error propagation in this approximate solution is not clear, so that
systematic errors become increasingly likely as the predictive interval is increased.
And such systematic errors could well be similar to, and therefore not identifiable
from, other errors that are endemic in the coupled AOGCM and are currently
handled by devices of questionable validity such as ‘flux adjustment’ [51].
Perhaps the most publicised use of the GCM is for predicting the future levels of
climate variables, such as global and regional mean temperatures with an assumed
increase in atmospheric CO2 concentration. There is a large technical literature
on probabilistic forecasting and attention to this must raise some doubts about the
manner in which the GCM is currently used for forecasting purposes. It would
certainly suggest that the predictions obtained from the GCM should be treated
with more circumspection than seems to be the case at the moment. It is simply
a matter of ‘horses for courses’ and the highly bred, expensive GCM filly is not
necessarily the best horse for the uncertain and hard ‘going’ that characterises the
global climate racecourse into the next century. This is not, we hasten to add, meant
as a criticism of the GCM or GCMers in total terms; merely we wish to emphasise
that the supreme position of the GCM in global climate research, particularly in
relation to the forecasting of global mean temperature variations into the far distant
future, is not fully justified. What is justified, we believe, is a more balanced use
of models in global climate research, with the GCM playing a very important but
not so dominant role in the future.
Bearing the above comments in mind, this section explores the need for plurality
in modelling the global climate by considering briefly a specific global climate
modelling example: namely, an investigation of the Enting-Lassey (EL) (1993)
model of the global carbon cycle [52]. We would have liked to utilise a GCM in these
studies but, as pointed out previously, the facilities required for conducting such
research are expensive and were not available to us. Moreover, the complexity of the
GCM model would have militated against the completion of the research within
a reasonable Ph.D. time-frame. Nevertheless, while the (EL) model is nowhere
near the size and complexity of the GCM and takes the form of twenty three,
interconnected, non-linear ordinary differential equations, rather than the partial
differential equations of the GCM, it is still quite complex by more ordinary
modelling standards. And it has played an important part in the studies carried out
for the IPCC, so that its critical evaluation does raise important issues that relate
also to the GCM. For detailed description of the EL model see the references in the
endnote [53].
The EL model is formulated in deterministic terms and so, to start with, we
consider how it may be converted into a probabilistic form which recognises
explicitly the uncertainty that must be inherent in any scientific evaluation of the
global carbon cycle. The implication is, of course, that this modified form of
the model is more appropriate for predictive purposes, since it can consider the
propagation of predictive uncertainty arising from the lack of accurate knowledge
about the parameters and inputs of the model.
After this stochastic study, another mathematical technique, Dominant Mode
Analysis (DMA), is utilised to explore whether the relative complexity of the EL
model is really required for predicting global mean atmospheric CO2 variations;
or whether the apparent complexity of the model is belied by the simplicity of
its dynamic behaviour in this regard. It should be noted that, within the context
of our discussion on GCMs, the main purpose of introducing this example is to
emphasise that, if models (criticised as being overly-simplistic by some GCMers)
present certain problems when considered in terms of their treatment of uncertainty
and their complexity, then there can be no doubt that the GCMs themselves present
even more difficult problems in this regard; problems that we believe demand
urgent attention.
5.1. QUANTIFYING THE CONSEQUENCES OF UNCERTAINTY: MONTE CARLO

SIMULATION (MCS) ANALYSIS
While it is a technique which is rarely used in climate research, possibly due

to the reductionist view that it is ‘second best’ [54], stochastic modelling, in its
various forms, is a powerful tool for investigating poorly defined systems with
limited observational data, such as the global climate. In this era of the powerful
digital computer, probably the most useful, general approach to the stochastic
modelling of non-linear systems is Monte Carlo Simulation (MCS) analysis [55]
. It is an approach that, rather conveniently in the present context, can be applied
to physically-based models that have, in the first instance, been formulated in
deterministic reductionist terms [56].
In this kind of MCS analysis, the physically-based, deterministic model is
converted into stochastic form by assuming that its parameters and inputs are
inherently uncertain: the parameters being characterised by probability density
functions (pdfs) and each input being expressed in terms of a stochastic time
series model [57]. The subsequent analysis then exploits MCS; i.e., the repeated
simulation of the model, with the parameter and input values for each simulation
run (or ‘stochastic realisation’) randomly drawn from the defined pdfs. Typically,
MCS involves several hundred, and sometimes several thousand, simulation runs:
the exact number being dependent on the required accuracy of output statistical
measures, such as cumulative distribution functions for each output variable [58].
Adopting a Bayesian approach, the pdfs are defined by reference to the current
literature on the parameters values and their associated uncertainty. Since knowl-
edge is often lacking in this area, however, a certain amount of subjectivity will
naturally enter into the analysis and this must be taken into account when evaluating
the results. On the other hand, if this aspect of the research is carried out carefully,
the measures of uncertainty will reflect the prevailing opinions of the scientific
community and it is important, from a Bayesian standpoint, that they should be
considered as prior information to be introduced systematically into the modelling
process. In the absence of any evidence to the contrary, the pdfs for the parameters
are assumed to be Gaussian in nature and independent of each other. Of course,
if the parameters are, in fact, not statistically independent, then this assumption
may lead to higher estimates of uncertainty in the model responses [59]. Again,
however, this reflects the uncertainty that characterises the scientific community’s
knowledge of the climate system and it is as well that the consequences of such
limitations in knowledge are fully realised.
In the case of the EL model, the uncertainties in the model parameters and inputs
required for the MCS analysis were obtained by reference to the latest literature on
the subject and from information supplied by scientists working on global climate
change. Here we present only the most pertinent results of the detailed research
programme that followed [60].
The uncertainty associated with the ensemble of atmospheric CO2 realisations
generated by the MCS analysis was always much greater (by up to twenty times)
than that in the observational record. Even taking into account the caveats men-
tioned above regarding the specification of the pdfs, this is a large discrepancy,
which suggests that more research is required both to reduce the levels of uncer-
tainty associated with the model parameters, and to consider whether the assumed
values of the parameters are entirely appropriate to the scale of the model: it is not
obvious that parameters inferred at the micro-scale are applicable at the macro-
scale, particularly when the macro-scale, as in this case, is the entire globe.
Continuing with the MCS analysis, a special form of generalised sensitivity
analysis reveals that the EL model is very dependent on the values of a small number
of important parameters, especially the pre-industrial level of atmospheric CO2. The
high level of variance of the MCS ensemble of atmospheric CO2 realisations is due,
in part, to such high sensitivities. However, it could also arise because the model
is more complex than is justified by the nature and paucity of the observations.
This possibility is given further credence by the Dominant Mode Analysis (DMA)
discussed in the next sub-section [61].
Table I compares the uncertainty measures associated with the EL model pre-
dictions of atmospheric CO2 in the years 2050 and 2100, as generated by the MCS
Table I
Comparison of the uncertainty in future CO2 levels due to IPCC sce-
nario IS92a between three deterministic modelling exercises (Wigley
and Raper, 1992; Enting and Lassey, 1993; and Schimel et al., 1995) and
the stochastic methods used in this paper
Year Uncertainty in Source

atmospheric CO2
(ppmv)
Range
2050 494 to 510 16 Schimel et al. (1995)a

520 to 550 30 Wigley and Raper (1992)

491 86 172 This paper, original simulation

494 22 44 This paper, modified simulation
2100 667 to 719 52 Schimel et al. (1995)a
740 to 800 60 Wigley and Raper (1992)
615 to 683 68 Enting and Lassey (1993)

664 144 288 This paper, original simulation

669 48 96 This paper, modified simulation
a
The results presented in this paper are derived from Enting et al. (1994).
analysis in the case of the IPCC scenario IS92a; and compares them with results
obtained in deterministic studies, including both Enting and Lassey, and Wigley
and Raper, the latter using their STUGE model [62].
These results show that the stochastic uncertainty is larger than the range of
variability found in the other studies. However, two criticisms could be levelled at
the way the MCS was carried out: (i) no account was taken of recent atmospheric
CO2 observations, which are more accurate than pre-industrial estimates used to
constrain the model; and (ii) as discussed before, it is possible that our estimates
of the uncertainties in the parameters and inputs have been exaggerated by the
assumption of independence in the pdfs. To take some account of these shortcom-
ings, then, the MCS analysis was repeated with each realisation constrained to the
uncertainty in the observed CO2 level in 1990 (353 2 ppmv), and with parametric
and input uncertainties reduced by an arbitrary factor of two [63]. The results of
this revised simulation are also shown in Table I. As we can see, the results of
the stochastic simulation are much reduced, but they remain 50% wider than the
widest of the deterministic results. This is a significant amount: moreover, it would
probably be larger still if the stochastic methodology could be applied to a complete
set of global carbon cycle models, as in the IPCC’s comparison of deterministic
simulation studies [64].
By modifying the deterministic EL model into a stochastic form and applying
MCS analysis based on parametric uncertainty estimates supplied by the research
community, we have produced estimates of uncertainty which are significantly
greater than those gained by deterministic model studies alone. It is recognised

that this kind of stochastic analysis is still in its early stages and more research,
especially on parametric uncertainty, is required. Nevertheless, the present results
suggest quite strongly that significant sources of uncertainty are not being con-
sidered at present by many climate researchers. This would not be surprising in
the case of GCMs since their huge computational demands make such uncertainty
analysis difficult, limiting any objective assessment of GCM results in relation
to uncertainty. But it suggests that uncertainty analysis of this, or other, types is
overdue in relation to GCMs and that, in predictive applications, GCMers need to
address more rigorously and quantitatively than heretofore, the consequences of
predictive uncertainty on the GCM forecasts.
5.2. SIMPLICITY OUT OF COMPLEXITY: DOMINANT MODE ANALYSIS (DMA)
Dominant Mode Analysis (DMA) seeks to analyse a given, physically-based, deter-

ministic model by identifying objectively the small number of dynamic modes
which appear to dominate the model’s response to perturbations in the input vari-
ables. In contrast to the traditional reductionist modelling, this normally results in
a considerable simplification of the model, which is simultaneously both reduced
in order and linearised by the analysis.
The DMA methodology involves perturbing the complex and usually non-linear,
physically-based model about some defined operating point, using a sufficiently
exciting signal, i.e., one that will unambiguously reveal all the dominant modes of
behaviour [65]. A low order, linear model, in the form of a transfer function, is
then fitted to the resulting set of simulated input-output data, using special methods
of statistical estimation that are particularly effective in this role [66]. As might
be expected from dynamic systems theory, a low order linear model obtained in
this manner reproduces the quasi-linear behaviour of the original non-linear model
about the operating point almost exactly for small perturbations. Perhaps more
surprisingly, as revealed in the present example of the EL model, the reduced order
model can sometimes also mimic the large perturbational response of its much
higher order progenitor.
To carry out the DMA, the full, non-linear EL model is initially set to an
equilibrium condition with the (pre-industrial) atmospheric CO2 concentration set
at 275 ppmv. Then a small perturbation in the fossil fuel input is applied to the model
and the resulting response is monitored over a period of 3000 years. The statistical
identification and estimation analysis yields the 4th order, linear transfer function
model, one interpretation of which is as a parallel connection of three, first order
systems and an integrator (the latter is required because of the mass conservation
assumption in the EL model) [67]. This simple dynamic model explains almost all
(99.98%) of the simulated model response.
More significantly, Figure 2 compares the response of this 4th order model with
that of the full, non-linear model over the entire industrial period [68]. As can be
Figure 2. Comparison between the responses of the reduced fourth order, linear model and the 23rd
order non-linear simulation model (Enting-Lassey Model) to fossil fuel input (error + 325 ppmv
shown above).
seen, the error is very small: never greater than 0.5 ppmv, even though the model
has moved 45 ppmv above the operating point at which it was estimated. These
rather surprising results demonstrate the robustness of the low order, dominant
mode model and suggest that the non-linearities in the original model are hardly
being excited by these fairly substantial perturbations over the industrial period.
Moreover, further DMA results at a range of different operating points reveal that
the objectively identified, 4th order model structure does not change at all, and
the behaviour of the non-linear model can be reproduced very closely with only
small variations in the parameter values. This is quite a dramatic result which
demonstrates the dominance of a small number of modes of behaviour in the
system, as defined by the eigenvalues of the reduced order, linear model at any
defined operating point.
An interesting aspect of these results is the size of the three time constants
(residence times), Ti , (where i = 1, 2, 3) in relation to the original physically based
model: T1 (= 4:3y ) is not dissimilar to the stratospheric turnover time of 6 2y ;
whilst T2 (= 19:1y ) is fairly close to the air-sea exchange time of 11 4y ; and,
finally, T3 (= 467y ) is of similar size to the deep ocean turnover time, of 900 250y
[69]. Naturally, the complex interconnected nature of the simulation model, with
its inherent feedback processes, modifies the effective ‘closed loop’ values of these
time constants when the model is integrated and so the identified dominant mode
time constants seem very reasonable [70].
What can we learn from the above DMA results? It is clear that the variations in
atmospheric CO2 produced by the high order, non-linear EL model over the whole
historic industrial period can be reproduced very adequately by a 4th order linear
model whose modal structure has reasonable physical significance. As regards the
prediction of atmospheric CO2 , therefore, this must call into question the need for
such a complex representation of carbon balance in the EL model and suggests that
a much simpler representation with fewer, or at least lower order, sub-systems (e.g.,
the EL model has 18 compartmental levels in the ocean sub-system alone) could
have produced very similar results. Moreover, such a reduced order model would
be more appropriate to the amount of observational data available in this example
which, in itself, makes the assumption of a high order model rather questionable
on statistical grounds. Indeed, there is evidence in the literature that the EL model
can only be fitted to the available data with constraints applied to many of the
parameters, a common indicator of severe over-parameterisation [71].
It should be noted, however, that we have only carried out DMA for a single
input and a single output of the EL model. Using more inputs and/or outputs may
well produce a more complex reduced order model. The number of inputs and
outputs in any DMA analysis should relate to the goals of the analysis. In this
particular case, if it were felt that the use of CO2 input from fossil fuel emissions
and atmospheric CO2 concentration output were adequate, then our example would
suffice; if not, then our study would need to be extended. This is straightforward in
methodological terms.
5.3. IMPLICATIONS FOR GCM DEVELOPMENT AND USE
Within the more general context of the present paper, the results outlined briefly
in the previous two sub-sections suggest the need for more checks and balances
in the deterministic reductionist approach to modelling which gives rise to such
complex representations of decidedly uncertain physical systems. Clearly, how-
ever, the alternative modelling methodologies used in the above analysis are still
being developed and themselves have shortcomings: e.g., the stochastic simula-
tion modelling is limited by the difficulty of quantifying fully the parametric and
input uncertainties; and by the fact that it has not been used to estimate output
uncertainty due to uncertainty in the model structure (though it could perhaps be
developed to do so). In addition, such alternative methods have as yet to be applied
to spatio-temporal models described by partial differential equations, as in the case
of GCMs.
Nevertheless, all of these methodologies (amongst others) offer useful additions
to conventional deterministic reductionist approaches, with considerable potential
for revealing new information and insight within the area of climate change, as
they have in other fields. Such methods could be exploited, for example, to develop
an interactive, multi-disciplinary modelling strategy in which they help to raise
meaningful questions about the quantification of uncertainty and the relevance of
processes or feedbacks in more traditional reductionist models; questions which
could then be addressed and answered, on a continuing basis, by the reductionist
modeller. This would help to rationalise the large models by avoiding excessive
complexity and over-parameterisation (in the sense given in note 34) and it would
ensure that sufficient attention is given to the consequences of uncertainty.
These alternative modelling methods also seem particularly attractive when
one considers the comparatively low demand on funding and computational facil-
ities made by them in comparison to those required for GCMs. In addition, the
new analyses of uncertainties are likely to be important inputs to policy debate
which, given the potential importance of global climate change, must surely take
cognisance of any available quantifiable measure of uncertainty when evolving risk-
averse strategies. For these reasons, the development of such techniques should,
we believe, be placed on a more equal footing with the more commonly accepted
reductionist modelling philosophy that underlies GCMs.
Of course, in highlighting alternative techniques we are not in any sense denying
the relative importance of GCMs. It is simply our objective to point out that they
score well with respect to certain criteria, but not to others. For example, if the goal
is defined as providing long-term climate predictions which are based on current
scientific perception of the physical mechanisms involved, then GCMs are certain-
ly a strong candidate. But while GCMs are, therefore, capable of exploring some
key features of climate change, including variable time-horizons and, potentially,
regional scales and variability, they are much less suitable for integrating with other
physical and socio-economic models, or for performing uncertainty analysis and
associated stochastic prediction. Moreover, they are resource-intensive and rather
intractable. There is a wide-spread experience with them although, as we have
discussed, confirmation of their reliability, especially for the purpose of making
projections, remains a difficult issue. Finally, they are not very accessible or trans-
parent and feedback from other scientific communities as to their validity is not
readily achieved.
Our main argument is that the role of GCMs can only emerge through consid-
ering fully the other possible methods that are available. An impediment to this is
that, at any one time, research tends to coalesce around certain methods as if they
are the only ones applicable for the given purpose, even though judgements might
change radically with the benefit of hindsight. Several alternative approaches to
studying future climate change that have been proposed and, in part, developed over
the past several decades, now seem to have ‘gone out of fashion’ or disappeared; or
else they are only recognised in small research enclaves (and seem to be especially
marginalised when research results are presented to policy makers). Alternatively,
they might be ascribed a role as ‘second-rate’ to GCMs or as ‘merely’ a heuristic
or educational tool.
In the following section we go on to discuss the wider policy and social dime-
nions of GCMs which we earlier defined as also critical to an understanding of the
relative role of different methods.
6. Why are GCMs Perceived as the Best Climate Models?
To account fully for the dominance of deterministic reductionism in climate mod-

elling would require a detailed study of such a scientific style in its rich historical,
institutional and intellectual context. Our purpose here is less ambitious, namely to
consider the role of social, institutional and policy relationships in helping secure
the position of GCMs. We explore whether GCMs provide unique and critical infor-
mation for policy making (and hence are seen as ‘better’ models for this reason),
as well as a more sociological interpretation which emphasises the role of GCMs
in linking-up different scientists and policy makers.
6.1. ARE GCMs ESPECIALLY POLICY USEFUL?
At first sight, GCMs appear to be at a distinct advantage for policy purposes:

relative to simpler climate models, they provide regional climate change scenarios
which can be used in assessing climate impacts and, hence, in devising suitable
policy responses. But, as already noted, since GCMs do not currently produce
robust regional simulations, this apparent advantage must be scheduled into, and
stored-up for, the future – at which point such simulations will become robust. Or
so it is anticipated. The promise of reasonably robust regional climate simulations
can, nonetheless, strengthen the case for current continued development of GCMs
as the only models that are potentially capable of producing regional simulations.
Yet, GCMs would also seem to be at a disadvantage in policy-making because of
their complexity which: (a) demands massive computing and personnel resources,
so restricting the number of GCMs and modelling centres; (b) prevents the rapid
inclusion of new processes into the GCMs as new scientific insights, understand-
ing or hypotheses emerge; (c) precludes the opportunity of performing extensive
sensitivity and uncertainty analysis; and (d) prevents the use of GCMs in a ‘respon-
sive mode’ to rapidly explore the implications of different model assumptions and
policy scenarios [72]. Evidence that these disadvantages have been experienced
emerges from the demonstrated policy-usefulness of simple models in the IPCC
process: because of their greater flexibility and ability to address specific policy
questions, simple models have a clear advantage in this role when compared to
GCMs [73]. As one government policy maker put it: ‘He [a scientist using simple
models] could come up with the goods very quickly with his simple box model’
[74].
If models are sufficiently user-friendly and simple, advisory scientists and even
policy makers can run their own simulations, sensitivity and scenario analyses.
By contrast, GCMs create a dependency on just a few centres and their resident
experts. It is possible that this dependency is not welcomed by many policy makers
(though it may be accepted as inevitable), since they (or their scientific advisors)
are less able to gauge the merits of the model outputs through, for example, not
having direct access to the modellers’ advice and interpretation [75].
On the other hand, such dependency might be welcomed by scientists, given
that non-specialist users of models are more liable to run the model uncritically
or misinterpret model output through not appreciating its framing assumptions,
uncertainties and limitations. And so the policy maker can be misled into assuming
that the apparent simplicity of the model implies a simple interpretation of the
implications of the model’s output for the purposes of ’reality’ and hence policy
decisions [76]. Use of GCMs, by contrast, are more likely to involve the judgement
of the modeller as an integral component of the scientific advice [77]. But let there
be no doubt that we feel such involvement is essential if models are to be used
for policy purposes: regardless of their complexity, all scientific models should
be used, and their results interpreted, carefully by scientists who are able to fully
understand their nuances and, if possible, have been involved in their synthesis.
6.1.1. Sununu and Simple Models

A pertinent example of how simple models may be employed in policy terms is
the use made by John Sununu, President Bush’s Chief of Staff, of a simple climate
model provided to him in the late 1980s by Warren Washington of the National
Center for Atmospheric Research (NCAR) [78]. Sununu explored the effects upon
surface temperature of changing the depth of the model ocean’s mixed layer; and, in
tandem with the results of Washington’s early AOGCM transient runs, concluded
that any warming would be taken-up by the ocean. His exposure to climate models
apparently led him to characterise them as ‘bad models’ – a charge which would
presumably be refuted by climate scientists on the basis that Sununu was guilty
of having over-interpreted or misinterpreted the output of a simple climate model
[79].
Yet, the problem of interpretation faced by Sununu was not exclusively related
to simple climate models, since Washington also exposed Sununu to GCMs. As
Washington now tells the story, the early transient run of the coupled AOGCM
which he showed to Sununu in 1989 did not show a slowing down of the thermo-
haline circulation (THC) with an increased atmospheric concentration of CO2 with
consequent effects on the take-up of heat into the deep ocean. The slowing-down
of the THC only occurred subsequently in the apparently improved transient runs
[80].
We cannot conclude much about the use of simple or complex models in policy
from just one example. The limited disputes over the use of knowledge from climate
models in policy have not, to date, revolved around the simple/complex distinction,
but more about conflicting interpretations of the same simple and complex mod-
els, and their relationship with other sources of knowledge, such as instrumental
temperature records [81]. Simple and complex models may, nevertheless, serve
different purposes in the climate change policy domain. GCMs, perceived as the
most authoritative, ‘realistic’ models sustain the legitimacy of climate change as a
central issue for international environmental policy. Simple models, on the other
hand, serve a more instrumental role in addressing policy-driven questions; for
example, concerning the effects of specific scenarios and projections. In conclu-
sion, the climate modelling pyramid cannot be explained by the special policy-role
of GCMs (even if policy usefulness were to be admitted as a valid criterion in this
context).
6.2. TOWARDS A SOCIOLOGICAL UNDERSTANDING OF THE POSITION OF GCMs
A more sociological understanding of the role of GCMs, is to point out the way in
which they link-up a diverse range of sciences and policies, both now and potentially
in the future. The suggestion here is that the more extensive are the links between
GCMs and other sciences and policy issues, then the more support is implicitly
being provided to GCMs. Their position as the ‘best science’ is buttressed by the
implicit or explicit recognition and uptake of GCMs by others, both scientists and
policy actors.
We argue that this occurs through a number of processes of mutual exchange
and reinforcement of the rationale and validity of a scientific or policy enterprise.
The most important of these are summarised in Figure 3, to which the reader should
refer throughout this section. Three sets of relations are stressed:
those between GCMers and the policy community, both of which share a
commitment to the task of long-term climate projections at global and regional
scales;
those between GCMers and the climate impacts community, both of which
share a commitment to producing regional climate change simulations;
those between GCMers and surrounding sciences, both of which share a com-
mitment to including more processes and detail into GCMs.
The mutual reinforcement which emerges out of these, partially over-lapping,
shared commitments is an important element in understanding the role and position
of GCMs in climate change research.
The end-result may be that GCMs become enmeshed in a network which con-
stitutes a powerful and resilient set of relationships: it is perhaps this resilience
which gives the impression that GCMs are ‘naturally’ at the top of a climate mod-
elling pyramid. Another way of putting this is that GCMs (contra other models or
methods) come to act as a sort of common currency between groups of scientists
and policy makers – each considers they have something to gain in intellectual,
scientific, funding and social terms – from being involved in their development and
use – and that this commonality serves as a way of linking-up such groups into
UNCERTAINTY, COMPLEXITY AND CONCEPTS OF GOOD SCIENCE
Figure 3. The interdependencies between climate modellers, surrounding disciplines and the policy community. The direction of the arrow on the lines
indicates the message received by one group from the other, e.g., the polilcy community receive from the GCM modellers the key messages that long-term
climate projection is ‘do-able’ and that anthropogenic climate change is a ‘real’ issue.
187
loose coalitions. Hence, GCMs come to have a wider symbolic significance than
implied by their scientific credentials alone.
A necessary part of our argument is that these mutually reinforcing commitments
emerge in part from the lack of knowledge and understanding of each group
when dealing with the other. For example, GCMers typically (and understandably)
operate with a somewhat narrow conception of what policy ‘needs’ from science and
this comes to limit the sorts of research they promote in their dealings with policy
makers. The subsequent response of the policy maker in ‘requesting’ particular
sorts of science is already partially conditioned by the assumptions of the scientist.
As far as the scientist is concerned, however, this merely reinforces his or her beliefs
about what scientific advice and research is needed by policy! However, as apparent
from Section 5, it is also part of our argument that most scientists themselves
are frequently (and again understandably) limited in their acknowledgement of
the possible range of scientific approaches. The existence of particular scientific
commitments (say to deterministic reductionism) also comes to limit what the
scientists communicate to each other and to policy makers.
Exactly the same argument applies to policy makers; i.e., they too presuppose
the role and capabilities of science as well as routinely operating with a limited
appreciation of the possible range of policy approaches. The existence of conver-
gences between these limited accounts of science and policy, which tend to be
assumed and not subjected to critical evaluation, helps explain their reinforcement
of each other, as if they are the only possible versions of science and policy. It
also has a major influence upon how GCMs are evaluated relative to other climate
models [82].
6.2.1. Involvement of Other Scientists

GCMs appear to have great, but only partly realised, potential to interconnect
diverse areas of environmental science. For example, Figure 4 is an ironic version
of these interrelationships from the perspective of the research co-ordinator at one
GCM centre. No other model promises this degree of integration or collaboration
between scientists. A modeller using a simple model who wants to include the
carbon fertilisation feedback or the effect of aerosols in the model, for example,
does not require an extended collaboration with biologists or atmospheric chemists.
What is needed are a few ‘best estimates’ from the world’s experts on these subjects,
probably accessible from the published literature.
In GCM ‘extension work’, by contrast, ecologists and hydrologists are needed to
model, for example, the intricacies of the movement of water from the soil through
vegetation to the boundary layer, to extend micro-level models of catchments so
they can be applied at the resolution of GCM grid-squares; and so on through a
myriad of other possible examples. And again, without the help of atmospheric
chemists, GCMers would be unable to attempt the integration of the dynamics
and chemistry of different greenhouse gases into the three-dimensional structure
of GCMs.
Figure 4. The Hadley Centre perspective of the science base for climate prediction research (after
Copernicus). The outer ring contains the international framework programmes with some of their
constituent projects listed in the next ring. Inside these there are main U.K. programmes followed by
the U.K. institutions with which the Hadley Centre is most closely associated. The inner ring gives a
brief description of the products of highest priority to the Hadley Centre programme. (From A Review
of the Science Base Underpinning Climate Prediction Research, Hadley Centre, U.K. Meteorological
Office 1992, page 1.
There is a clear role, then, for many scientists, other than GCMers, in the con-
tinuing development of GCMs such that they are seen to be ever more inclusive,
realistic and comprehensive. By perceiving the advantages and opportunities of
collaboration with GCMers in their own research work, these other enrolled scien-
tists are effectively endorsing and advocating GCMs. Their own work practices and
disciplinary pre-occupations then change in the service of the greater model (and,
for this reason, they can be considered as mavericks or treated with suspicion from
the perspectives of their native disciplines) [83]. This, then, is a vision of GCMs
as a sort of ‘collective good’ amongst environmental and earth system scientists
of many disciplinary persuasions [84]. They act as integrators of diverse empirical
and theoretical pursuits [85]. Hence the set of links between GCMs and other fields
should be especially dense.
The extended GCM community includes also those meteorologists and climatol-
ogists who specialise in handling observational data-bases, and who increasingly
argue that their data-sets are essential to the proper validation of GCMs. This
includes the satellite agencies, who have made persuasive arguments about the
value of global remote sensing and have secured a large proportion of the total
funds for global environmental change research, as indicated in Table II. Some
observational scientists have gone so far as to pin the full rationale for data collec-
tion upon the need for improved GCMs, sometimes to the discomfort of GCMers
[86] Other observational scientists have become linked with GCM groups, identi-
fying the validation of GCMs as a new role for themselves, the variable spatial and
temporal scales of GCMs allowing flexibility in what analysis is actually done; a
role which supports the relevance of data collection and collation and provides a
respectable home for their data-bases. It is true that simpler models also need data
for validation; but they need much more highly aggregated data organised into a
few global variables. By contrast, because GCMs are simulating climate at many
thousands of grid points and over many time-scales, there is a much greater demand
for disaggregated, comprehensive regional data, whose interpretation and extrapo-
lation to grid points (to provide the model’s boundary-conditions or for validation
purposes) requires detailed interaction with observational scientists. [87].
In addition to satellites, remote sensing and other forms of instrumentation,
GCMs and GCMers are connected to the technological enterprises of super-
computing, networking and software development, including massively parallel
processing. By contrast, most other climate models can now be run on a desk-
top computer; and even computationally intensive procedures such as Monte Carlo
Simulations, which are virtually impossible for GCMs, can be applied to quite com-
plicated (non-GCM) models by linking several such simple processors in parallel
[88].
6.2.2. Scientific ‘Users’ of GCMs

Further participants in the construction of the GCM-network are the various scien-
tific communities using the output from GCMs. For example, the ‘climate impacts
community’ explores the effects of climate change on agriculture by deriving data
from individual grid-points of GCMs and feeding it into crop productivity models
[89]. Such impacts models are reliant upon the regional output of GCMs, which
varies a great deal between different models, and is dependent upon the specific for-
mulation, such as the parameterisation of land-surface processes. Impact modellers
become, to an extent, dependent upon the judgements of the climate modellers
Table II
Approximate expenditure on global change (U.S.) and global climate change
(U.K.) research by subject area
Approximate expenditure in millions

of U.S. dollars for 1993–1994
Broad area U.K. (1993/94) U.S. (1993)
Observations
Satellites 100 560
Ground-based observations 50 25
Informational management 25 275
Sub-total 175 860
Detection 1 Not known
Understanding global change 65 455
Predicting global change 20 50
Impacts/evaluating the consequences 20 40
Response strategies 40 Not known
Assessment tools Not known 15
TOTAL 320 1420
These figures are highly approximate and rounded off to the nearest 5 million
dollars. They should be regarded as no more than a rough guide: for example
other estimates for the U.K. differ by approximately 15% for reasons which are
not known. The source for the U.S. figures is Global Change Research Program:
1995, Our Changing Planet, U.S. Government Printing Office, Washington, DC.
For the U.K., estimates have been made from information assembled by the U.K.
Department of the Environment; a conversion ratio of $2 to £1 has been used in
the table.
as to which GCM model runs to use and how to construct suitable future climate
scenarios.
Climate impacts work is not the only sub-discipline to take the output of GCMs
‘on trust’. Economists who study the costs of climate change, for example, also
require information on factors such as: radiative forcing factors, greenhouse gas
life-times, damage functions, the rate of climate change, and the likely time-scale
for the reduction of critical scientific uncertainties, some of which information is
derived from GCMs [90]. As already noted, those who use simple climate models
also rely upon GCMs for their calibration; and this includes some economists
and engineers developing integrated assessment models. By using GCM data, both
impacts scientists and economists are also engaged in constructing and maintaining
the network, and the pivotal position of GCMs. At the same time, GCMers are
lending their implicit support to the impacts and economics work by providing
seemingly reliable scenarios of ‘realistic’ climate change.
6.2.3. Policy Makers

The rationale for climate impacts and economic studies is, ostensibly, to contribute
to better climate change policy-making. Hence, the GCM network becomes central
to knowledge-based policy-making aspirations. Yet, can it be claimed that govern-
ment or other policy makers attach themselves particularly to GCMs in preference
to other modelling or climate research activities? Funding reveals some preferences
(see Table II) although a direct numerical comparison is not especially meaning-
ful given the different costs of different sorts of research (impacts and integrated
assessment research are much cheaper than GCM modelling, which again is cheap-
er than satellite development and application). Nevertheless, it is noticeable that,
for example, all of the climate modelling funded by the U.K.’s Department of the
Environment is that involving GCMs, even though it could support much simpler
model development for relatively little money.
Why, then, should the U.K. DoE attach itself to GCMs rather than simpler mod-
elling? Clearly part of the reason derives from the scientific community’s evaluation
of GCM work vis-à-vis simpler modelling. The argument that scientific credibility
(and preferably certainty) is necessary for the accomplishment of political consen-
sus appears to hold powerful sway in many political and policy cultures and, if
GCMs are held out as the most reliable and robust of climate models, it follows
that they will enjoy an elevated status in policy circles. The wish to maintain a
specific national influence in international policy negotiations is also seen by some
policy makers to depend upon the national government being amongst the global
scientific leaders which, assuming the existence of the modelling-pyramid, means
supporting GCMs.
Given the above situation, part of the scientific evaluation of the GCM science
as ‘the best’ climate science derives from this policy status and influence [91].
Furthermore, of all climate models, only GCMs seem to offer the promise of the
detailed, regional, ‘hard’ scientific knowledge necessary to implement proposed
local policies for sustainable development (as specified by Agenda 21 of the Rio
Earth Summit), despite the far-from-achieved efficacy of such regional modelling.
As a result, the understandable criteria amongst scientists for the evaluation of
different model types, in terms of future potential simulations, may have been
extended to an evaluation of the current policy usefulness of different models.
A further possible and implicit advantage for policy makers of using GCMs
may be their sheer complexity and impenetrability. As such, they cannot easily be
challenged by critics, whether environmentalists or industrialists, most of whom
cannot possibly comprehend the complexities of such a large computer model,
let alone articulate its deficiencies. There is no conscious deception here since, as
far as the policy makers are concerned, the models are impenetrable because they
represent the most sophisticated science.
Certain models or areas of science, other than GCMs, could even be a threat
to the vision of using predictive knowledge of natural (and social) systems for
the purposes of global environmental risk management. For example, stochastic
models might be seen to fit into this category because they take the uncertainty of
our knowledge as their starting point, so problematically questioning the attempted
deterministic closure of networks around the current science-policy nodes. Chaos
theory and some palaeoclimatology research also seem to qualify as problematic
scientific partners to a global environmental management policy because they
question the very predictability of climate. Not only do these other areas of scientific
research not ‘gel’ with the global environmental change project in its policy-
oriented formulation of the last decade or so, to an extent they have to be actively
resisted or ignored by it [92].
6.3. IMPLICATIONS
We have tried to show that the emergence of consensus around particular concepts
of ‘good science’ are neither predetermined by nature, nor are they without larger
political significance. This is because they inadvertently assist the development of
political consensus as to the appropriate policy response. We have argued that it
is partly these political implications of the global scientific culture which place
GCMs at the top of a ‘natural’ hierarchy of ‘good science’ and valid knowledge
for international policy actors [93].
Sciences which promise integration may be particularly appealing in the context
of increasing senses of policy and political flux and insecurity globally. From this
perspective, GCMs can be seen as critical instruments by which policy agendas
aimed at the control and management of global climate change are shaped and sus-
tained. They are perceived to lend much credibility to the prospect and legitimacy
of uniting and globalising diverse activities, including national and regional poli-
cy making. They achieve this through the application of standardised techniques,
inputs and assumptions, as well as through promoting and legitimising the devel-
opment, in other related domains, of such standard methods and techniques. So,
for example, GCMs provide justification for the creation of greenhouse gas emis-
sion inventories, as well as for changing social practices in a way which reduces
such emissions, whether through energy efficiency programmes, or a change in
agricultural practices. If they are believed, deterministic models such as GCMs and
their applications could become enormously powerful, both in technical and social
terms, because they stand to acquire the power of all those diverse constitutive
activities carried out over decades in different scientific sub-fields.
But is such power deserved? And could it not be dangerous? After all the efficacy
or otherwise of the GCM in predictive terms cannot be tested more rigorously
in scientific terms until sometime into the next millennium. A large part of the
scientific community believes that climate change is unlikely to be predictable
in a precise fashion. Hence, climate change management strategies based on the
availability of precise-knowledge of short- and medium-term events are unlikely
to succeed [94].
The impression that climate change can be so predicted and managed is not
only misleading, but it could also have negative repercussions should policy mak-
ers act on this assumption. Such negative consequences might occur through the
subsequent failure to develop policies which encourage societal resilience to unpre-
dictable change. Additionally, there is a danger that policy authorities will appear
to others to have ‘solved the problem’, whereas they will have done so only in the
unlikely event that the predictions or projections made are correct. The societal per-
ception that the ‘climate change problem’ is being adequately handled could inhibit
the emergence of, and support for, creative social, policy and economic responses
to the challenge of coping with a possibly inherently unpredictable system such as
climate.
7. Conclusions
We have argued that the common relative ordering of different climate models –
which champions GCMs as the ‘ultimate’ climate model – is not fully justified by
scientific arguments alone, at least not for the simulation and projection of future
climate change. The criteria by which GCMs are conventionally assessed as being
superior to other climate models are at least questionable from without (and to
some extent from within) the GCM community. In assessing GCMs one is drawn
into an open-ended discussion concerning:
how the inclusion of perceived complexity should be traded-off against a
better understanding and analysis of dominant model processes, interactions,
feedbacks and uncertainties;
what likely scientific and technological advances can be anticipated;
what complexity is and how it can or should be modelled and measured; and,
the impossibility of validating such large, complex models.
The existing dominant notion of how to model complexity in the climate change
field is accepted by climate modellers as the most adequate. However, complexity
and uncertainty are explored using different modelling philosophies and styles in
other areas of environmental science. These alternative approaches are taken as
defining good science in those fields. Yet such methods are not being sufficiently
assimilated into ‘state of the art’ global climate modelling.
In a situation where the system is nominally very complex and the data are
scarce, it often makes good sense to construct high order simulation models which
represent the ‘state-of-the-art’ scientific understanding of the system. Indeed, large,
complex simulation models play an important part in many fields of science and
engineering and the second author has been involved in many studies which have
exploited such models to good effect [95]. Consequently, let there be no doubt
that we believe models such as the GCMs and their less complex relatives, the
global carbon cycle models, are valuable elements in climate research. But, as we
have tried to demonstrate, such large deterministic simulation models have their
limitations and are not the only type of model that should be considered when the
uncertainties, or even ambiguities, inherent in the climate system are taken into
consideration. There are, we believe, clear advantages in a ‘spectrum’ of models
that are designed to satisfy different objectives and which together provides the
scientific input to policy making exercises.
Some may argue that the diversity we call for already exists in practice. That is
clearly true to an extent. But only to an extent, and one which is not reflected in
policy and user circles. What we are arguing, therefore, is that the choice of method
should depend more explicitly on the sort of multi-dimensional criteria suggested
here and that there should be a more explicit debate on goals, epistemology, expe-
riences, and the policy context. Additionally, the reasons for adopting particular
methods should be better appreciated and acknowledged. This, we believe, will
assist in a better role of science in policy making and in public debate, and allow
for science to be used more intelligently (and realistically) by the user community.
For example, if we accept that GCMs are supported in part because they help to
create an interactive and international community of scientists and policy makers
– and not just because they are ‘the best’ model – then we can at least question
and discuss whether that is the sort of community we want to emerge to inform our
political processes.
We have suggested some possible sociological reasons why GCMs have come
to occupy a predominant position, especially their unique ability to integrate and
implicitly support diverse areas of scientific research and climate policy. We have
hinted that the lack of recognition of such factors is especially acute for complex
and comprehensive models, since the judgements and choices taken – scientific,
social and policy – become more easily hidden in the complexity of the model.
The range of model components and interlinkages, the craft-ladenness of their
construction and use, and their history (having being produced by teams, often
over decades) means that they are enormously difficult to comprehend at an over-
arching level, or in terms of the dominant model interactions, even for many of the
modellers themselves.
Integration of sub-models within GCMs provides a pragmatic solution to the
horrendous problems of delimiting and characterising the system to be modelled,
which is itself ever more comprehensive and globally inclusive. It is also a way of
coping with uncertainties, ambivalence, scientific judgements and disagreements,
though it frequently does so largely through hiding such issues in the complexity of
the model. The effect of the networking around the ‘hub’ of GCMs is that implicit
political and value judgements, assumptions and commitments are being incor-
porated, though not through deliberate debate or the conscious intent of any one
actor: rather, these emerge from the network. The incorporation of such implicit
commitments and beliefs into models also contributes to their removal from wider
discussion and debate in more political contexts (though we are not suggesting this
is the intended outcome). The achievement of an apparently resilient universality
for the scientific culture across all its original disparity, and despite its remaining
incoherence, is nonetheless an enormous social, political and intellectual accom-
plishment.
At one level, the success of such model-based networks in policy may depend
on their denial of the more open-ended scientific and social choices, in favour of
their presentation as objectively given by natural science, in which politics and
values play little or no part. This is largely because of the consequent exclusion
of all social and meta-scientific issues from the realm of legitimate debate, as
well as of the actors promoting them; a delimitation which facilitates political
consensus-formation.
Yet this apparent success may be earned only at the cost of a lack of effectiveness
at different, and arguably more important, levels of the policy process; in particular,
at the stages of influencing change in individual and collective perceptions and
behaviour to – inter alia – reduce greenhouse gas emissions, reduce energy use,
change consumption patterns and so forth. For example, the more the scientific and
social choices are effectively ‘delegated’ to the model [96], the more ambivalence
is likely amongst those policy actors implicated. In consequence, subtle political
and institutional changes could result in radical shifts in the perception of GCMs. If
our analysis is at all valid, however, the credibility of GCMs would suffer, perhaps
greatly, were political and policy actors to display relatively little interest in them.
The cost of impenetrability is fragility of the entire edifice.
Even the conscious deliberation of all scientific social and policy judgements
and choices by the present community of modellers, other research scientists,
research managers and policy-makers, could not possibly illuminate adequately
the validity and wisdom of many of the assumptions and judgements involved in
the construction and scientific policy use of GCMs. That task requires at least an
understanding of how climate and climate change is perceived by a range of policy
actors and the public. Such an understanding could probably only come from a
more open and reflective discussion in a newly constituted public domain [97].
Nor does robustness in the social understanding incorporated into the poli-
cy use of climate science imply correctness, since many of the pre-suppositions
will be legitimate interpretations based on judgement and, frequently, more than
one alternative will present itself. In the absence of a single ‘correct’ analysis,
responsiveness to the outcome of policy based on a particular set of assumptions
or judgements is important. In other words, iterative learning from trial and error
should be a necessary part of the policy and research process; and such openness
to learn and change is what constitutes a resilient policy and scientific approach to
the climate change issue. This implies no less than the development of new public
institutions and processes for the interaction of science, policy and the public. We
envisage that many of these institutions will be largely exploratory and developed
incrementally to better define who might be involved in exploring scientific and
social assumptions and commitments, as well as critically reviewing the effects of
policy based upon them.
If our analysis of the mutual reinforcement of particular scientific and policy
approaches has any purchase, the above processes will also raise questions about
the relative merits of alternative scientific methods and styles. Indeed, we anticipate
that such a discussion will become vitally necessary to the provision of insights and
understanding into the range and potentialities of intellectual tools for the support
of policy options and public debate.
In this spirit, we finish by inviting climate scientists to take part in such discus-
sions: in a sense, to (re)consider the scientific and implicit social and policy com-
mitments they have entered into through pursuit of a particular scientific research
programme. Only through such debate will we be able to more fully address the
question of whether the lack of recognition of the alternative scientific, social and
policy elements in the definition of ‘good science’ has held back some of the
alternative approaches which could be further developed and used profitably in the
widening climate science and policy domain.
Acknowledgements
This paper has benefited greatly from the help and advice of James Risbey (Carnegie
Mellon University) and Peter Stone (Massachusetts Institute of Technology). Ian
Enting (CSIRO, Division of Atmospheric Research) provided us with a copy of
the Enting-Lassey global carbon cycle model for which we are most grateful.
We are also grateful to the following for their generous assistance and readi-
ness to respond to our questions: David Bennnetts, David Carson, Chris Folland,
Klaus Hasselmann, Paul Jarvis, Geoff Jenkins, John Houghton, Mick Kelly, Adri-
an Kelsey, Mike Hulme, Richard Lindzen, Jerry Mahlman, Suki Manabe, Paul
Mason, John Mitchell, Tim Palmer, Roger Pielke, Colin Prentice, Stefan Rahm-
storf, Michael Schlesinger, Edward Schneider, Ron Stouffer, David Viner, Hans
von Storch, David Warrilow, Warren Washington, Robert Watson, Ian Woodward,
Tom Wigley, and Carl Wunsch. We are especially grateful to Stephen Schneider,
Paul Edwards, Naomi Oreskes and two anonymous reviewers, for their further
advice and guidance. It goes without saying that the opinions expressed in this
paper are not attributed to any of the above. The research was funded by the Eco-
nomic and Social Research Council (grant number Y302/28/3001) and the Natural
Environment Research Council (grant number GT4/90/AAPS/27) both of the U.K.
Notes and References
1. Levins, R.: 1966, ‘The Strategy of Model Building in Population Biology’, Amer. Sci. 54 (4),
421–431.
2. Lindblom, C. and Cohen, D.: 1979, Usable Knowledge, Yale University Press, London, p. 129;
Hogwood, B. and Gunn, L.: 1984, Policy Analysis For The Real World, Oxford University Press,
Oxford, p. 289; Brunner, R.: 1996, ‘Policy and Global Change Research: A Modest Proposal’,
Clim. Change 32 (2) 121–147.
3. See, however: Glantz, M.: 1979, ‘A Political View of CO2 ’, Nature 280, 189–190; Stewart, T.
and Glantz, M.: 1985, ‘Expert Judgement and Climate Forecasting: A Methodological Critique
of “Climate Change to the Year 2000” ’, Clim. Change 7, 159–183; Schneider, S.: 1985, ‘Science
by Consensus: The Case of The National Defense University Study “Climate Change to the Year
2000” – An Editorial’, ibid., pp. 153–157; Edwards, P.: 1996, ‘Global Comprehensive Models in
Politics and Policymaking’, Clim. Change 32 (2), 149–161; Pielke, R. Jr.: ‘Usable Information
for Policy: an Appraisal of the U.S. Global Change Research Program’, Policy Sci. 28, 39–77;
Jamieson, D.: 1991, ‘The Epistemology of Climate Change: Some Morals for Managers’, Soc.
Natural Res. 4, 319–329.
4. For example, Hajer, M.: 1995, The Politics of Environmental Discourse, Oxford University Press,
Oxford, p. 332; Haas, P.: 1992, ‘Introduction: Epistemic Communities and International Policy
Coordination’, Int. Organ. 46 (1), 1–35.
5. Grotch, S. and MacCracken, M.: 1991, ‘The Use of General Circulation Models to Predict
Regional Climate Change’, J. Clim. 4, 286–303; von Storch, H. and Hasselmann, K.: 1995, Cli-
mate Variability and Change, Report no. 152, Max-Planck-Institut für Meteorologie, Hamburg,
p. 26.
6. Natural Science and Technology Council: 1996, Our Changing Planet: The FY 1996 U.S.
Global Change Research Program, A report by the Subcommittee on Global Change Research,
Committee of Environment and Natural Resources Research, A Supplement to the President’s
Fiscal Year 1996 Budget, Washington, D.C., p. 152.
7. Page 82, Shine, K. and Henderson-Sellers, A.: 1983, ‘Modelling Climate and the Nature of
Climate Models: A Review’, J. Climatol. 3, 81–94.
8. Page 55, Henderson-Sellers, A. and McGuffie, K.: 1987, A Climate Modelling Primer, John
Wiley, Chichester, p. 217.
9. Wigley, T. and Raper, S.: 1992, ‘Implications for Climate and Sea Level of Revised IPCC
Emissions Scenarios’, Nature 357, 293–300.
10. Page 84 in [7], emphasis added.
11. Page 17, Schneider, S.: 1992, ‘Introduction to Climate Modeling’, in Trenberth, K. (ed.), Climate
System Modeling, Cambridge University Press, Cambridge, pp. 3–26.
12. In a further contribution to the book, Meehl extends the concept of a model hierarchy to coupled
models, relating comprehensiveness to realism and reliability of model output. (Meehl, G.: 1992,
‘Global Coupled Models: Atmosphere, Ocean, Sea Ice’, in Trenberth, K. (ed.), Climate System
Modeling, Cambridge University Press, Cambridge, pp. 555–581; Kiehl, J.: 1992, ‘Atmospheric
General Circulation Modeling’, in Trenberth, K. (ed.), Climate System Modeling, Cambridge
University Press, Cambridge, pp. 319–369.)
13. Page 760, Schlesinger, M. and Mitchell, J.: 1987, ‘Climate Model Simulations of the Equilibrium
Climatic Response to Increased Carbon Dioxide’, Rev. Geophys. 25, 760–798. Similar sentiments
are expressed in the IPCC 1995 report, e.g., see Gates, L. et al.: 1996, ‘Climate Models –
Evaluation’, in Houghton, J. et al. (eds), Climate Change 1995: The Science of Climate Change,
Cambridge University Press: Cambridge, pp. 572.
14. In evidence before the Subcommittee on Environment of the Committee on Science, Space and
Technology of the U.S. House of Representatives (in 1992) the following statements have been
made by research managers: ‘I think we are confident in their [GCM’s] fundamental physics.
They are based on fundamental physical principles that we believe in. Of course there are many
uncertainties associated with the parameterisations : : : . The clouds, the oceans, the ecosystems,
and simply just a matter of resolution’ (oral evidence of Dr. A. Patrinos, U.S. Department of
Energy, U.S. Congress, House, 1992. Committee on Science, Space and Technology. In U.S.
Global Change Research Program, Hearing before the Subcommittee on Environment, 102nd
Congress, 2nd Session, 5 May, 1992, p. 91. Washington, D.C., U.S. Government Printing Office,
p. 200). And: ‘Right now, those GCMs incorporate the best scientific knowledge, the best
understanding of the physics and transport of the atmosphere, and some coupled models include
representations of the ocean as well that the science community can give you’ (oral evidence
of Dr. C. Riordan, U.S. Environmental Protection Agency, ibid, p. 92). Meanwhile, in the U.K.,
Dr. John Houghton in a speech delivered at the opening of the Hadley Centre (at which Prime
Minister Thatcher was in attendance) stated that: ‘The numerical model is an effective way –
in fact the only way we know – of dealing in a meaningul way with all these non-linearities
involved. : : : these is a good expectation of being able, through the use of models, to predict at
least in general terms the change in climate which is likely to occur because of man’s increased
burning of fossil fuels’ (text of speech delivered 25th May 1990).
15. Interview with a GCMer by first author, 11 October 1992. In this paper we use evidence obtained
from semi-structured interviews with climate scientists and policy makers. Where quoting inter-
viewees in this paper, we have sometimes slightly edited the verbatim statements to clarify what
we perceived were the intended meanings. Where the citation is to a ‘personal communication’
this indicates a written communication.
16. Personal communication to first author from a GCMer, 22 May 1995.
17. He told us: ‘I frequently find that when I submit a proposal or a paper which involves some of
the simpler models I run into the criticism – what’s the point of doing this? Why didn’t you
use a GCM? A GCM is obviously better!’. The same scientist told us of an application he had
submitted to a funding agency which was turned down on the basis of peer-reviews, one of which
assessed the proposal as using an over-simple model, whilst another thought it intended to use an
over-complex model. He described a subsequent conversation with an agency official as follows:
‘We had a long conversation and it was a little frustrating because he was not willing to be too
explicit. : : : He talked at great length about all the GCM efforts that he is supporting. : : : He
also mentioned that he was aware of my criticisms of GCMs and I couldn’t help wondering if
that had something to do with this decision’. (Interview with first author, 19 April 1994.)
18. E.g., p. 18, Schneider [11]. Many well known climate scientists have developed and used simple
models including Stephen Schneider, Martin Hoffert, Danny Harvey, Tom Wigley, and Peter
Stone.
19. Interview with two climatologists by first author, 5 and 6 August 1992.
20. Ibid. Also, Land, K. C. and Schneider, S. H.: 1987, ‘Forecasting in the Social and Natural
Sciences: An Overview and Analysis of Isomorphisms’, Clim. Change 11, 7–31. Our argument
is an inevitable simplification of how GCMs are perceived and in reality perceptions of GCMs
are often more finely-textured and undergoing change. For example, new relationships have
been emerging between GCMers and other scientists, such as ecologists and hydrologists, in
which GCMs are less predominant (e.g., Root, T. L. and Schneider, S. H.: 1995, ‘Ecology
and Climate: Research Strategies and Implications’, Science 269, 334–341); and in the U.S.A.
there are not only more research centres than in Europe, but greater diversity in the key issues
pursued. However, GCMs still dominate in the U.S.A. when it comes to providing policy-relevant
knowledge.
21. For example, the Policymakers Summary of the IPCC’s 1990 report commented that: ‘These
models (GCMs) are based on the laws of physics and use descriptions in simplified physical
terms (called parameterisations) of the smaller-scale processes : : : ’ (IPCC, 1990: xxv). In the
IPCC’s 1995 report the Technical Summary states that: ‘Many physical processes, such as those
related to clouds, take place on much smaller spatial scales and therefore cannot be properly
resolved and modelled explicitly, but their average effects must be included in a simple way by
taking advantage of physically based relationships with the larger scale variables (a technique
known as parametrization)’ (p. 31, Technical Summary, IPCC 1996, Climate Change 1995: The
Science of Climate Change, Cambridge University Press, Cambridge).
22. P. 19 in Schneider [11]. The IPCC’s presentation is not inconsistent with Schneider, but rather
incomplete and with a different emphasis, e.g., upon the physical basis of parameterisations
rather than their empirical statistical dimension.
23. Kiehl, in [12]. See also Kiehl, J. and Williamson, D.: 1991, ‘Dependence of Cloud Amount
on Horizontal Resolution in the National Center for Atmospheric Research Community Climate
Model’, J. Geophys. Res. 96, 10955–10980; Slingo, T.: 1990, ‘Sensitivity of the Earth’s Radiation
Budget to Changes in Low Clouds’, Nature 243, 49–51; Randall, D., Harshvardhan, Dazlich, D.,
and Corsetti, T.: 1989, ‘Interactions among Radiation, Convection, and Large-Scale Dynamics
in a General Circulation Model’, J. Atmos. Sci. 46 (13), 1943–1970.
24. Personal communication to first author from a GCMer, 7 February 1995.
25. Risbey has elaborated on this point: ‘GCMs aim to account for physical processes from first
principles, i.e., from the known laws governing the behaviour of the relevant phenomenon.
Simpler models generally do not make this attempt, accepting more aggregate and empirically
based representations. In practice however this distinction is not as strong as it sounds in practice,
since even GCMs are highly dependent on empirical data in their parameterizations, albeit at
smaller scales. This is (one reason) why the so-called “reductionist approach” of GCMs “does
not automatically guarantee physical realism” ’ (personal communication, September 1996).
26. Stone, P. and Risbey, J.: 1990, ‘On the Limitations of General Circulation Climate Models’,
Geophys. Res. Lett. 17 (12), 2173–2176; Stone, P. and Yao, M-S.: 1991, ‘Vertical Eddy Heat
Fluxes from Model Simulations’, J. Clim. 4 (3), 304–317.
27. Interview with a climate modeller by first author, 10 April 1994. (Michaud, R. and Derome, J.:
1991, ‘On the Meridional Transport of Energy in the Atmosphere and Oceans as Derived from
Six Years of ECMWF Analyses’, Tellus 43A, 1–14; Oort, A. H.: 1978, ‘On the Adequacy of the
Rawinsonde Network for Global Circulation Studies Tested through Numerical Model Output’,
Mon. Wea. Rev. 106, 174–195; Oort, A. H.: 1983, Global Atmospheric Circulation Statistics,
1958–1973, NOAA Professional Paper 14, U.S. Government Printing Office, Washington, D.C.;
Keith, D. W.: 1995, ‘Meridional Energy Transport: Uncertainty in Zonal Means’, Tellus 47A (1),
30–44; the on-going debate is reflected in Trenberth, K. and Solomon, A.: 1994, ‘The Global
Heat Balance: Heat Transports in the Atmosphere and Ocean’, Clim. Dyn. 10, 107–134). A major
comparison of models with respect to their heat transport has now been conducted (Glecker, P.
et al.: 1995, ‘Cloud Radiative Effects on Implied Oceanic Energy Transports as Simulated by
Atmospheric General Circulation Models’, Geophys. Res. Lett. 22 (7), 791–794) which reveals
the extent of the probable errors).
28. Pp. 2175 and 2176 (our italics) respectively, in [26].
29. Young, P., Parkinson, S., and Lees, M.: 1996, ‘Simplicity out of Complexity: Occam’s Razor
Revisited’, J. Appl. Statis. 23, 165–210.
30. Interview by first author with a GCM modeller, 11 March 1993.
31. Some evidence for this view comes from research with relatively simple models (e.g., Marotkze,
J. and Stone, P.: 1995, ‘Atmospheric Transports, the Thermohaline Circulation, and Flux Adjust-
ments in a Simple Coupled Model’, J. Phys. Ocean. 25, 1350–1364; Nakamura, M., Stone,
P., and Marotzke, J.: 1994, ‘Destabilisation of the Thermohaline Circulation by Atmospheric
Eddy Transport’, J. Clim. 7, 1870–1882). Note, however, that not all GCMers would accept this
evidence since simple models were used in these experiments.
32. It has been suggested to us that the evaluation of climate models has become increasingly
formalised since the early 1990s, with projects such as the Atmospheric Model Intercomparison
Project (AMIP) and its off-shoots (Mike Hulme, personal communication, February 1997). An
interesting question is the extent to which such formalisation is related to the perceived policy
requirement for validation in a context of uncertain science for policy, rather than such a process
being driven solely by the climate modelling community. There are generic similarities to the
strategies of molecular biologists and particle physicists (Knorr-Cetina, K.: 1991, ‘Epistemic
Cultures: Forms of Reason in Science’, Hist. Political Econ. 23, 105–122).
33. A discussion and classification of the limitations and uncertainties of climate and integrated
assessment models can be found in van der Sluijs, J.: 1997, Anchoring Amid Uncertainty: On
the Management of Uncertainties in Risk Assessment of Anthropogenic Climate Change, Ludy
Feyen, Leiden, p. 260.
34. Note that in mathematical statistics and engineering, ‘over-parameterisation’ is usually taken to
mean that the model has more parameters (coefficients) than can be justified by the available
data on which the model is based: this is, of course, quite different from the usage here.
35. Personal communication by first author with Dr. James Risbey, 6 July 1994.
36. Hempel, C. G.: 1966, Philosophy of Natural Science, Prentice-Hall, Englewood Cliffs, NJ, p. 116;
Hesse, M.: 1974, The Structure of Scientific Inference, Macmillan, London, p. 309; Popper, K.:
1959, The Logic of Scientific Discovery, Hutchinson, London, p. 480.
37. P. 234 in Hesse [36].
38. P. 44 in Popper, K.: 1982, The Open Universe, Hutchinson, London, p. 185; Tennekes, H.: 1992,
‘Karl Popper and the Accountability of Numerical Weather Forecasting’, Weather 47, 343–346.
39. Gieryn, T.: 1995, ‘Boundaries of Science’, in Jasanoff, S. et al. (eds.), Handbook of Science and
Technology Studies, Sage, London, pp. 393–443; Collins, H.: 1992 (1985), Changing Order:
Replication and Induction in Scientific Practice, Chicago University Press, London, p. 187.
40. Shackley, S. and Wynne, B.: 1995, ‘Integrating Knowledges for Climate Change: Pyramids, Nets
and Uncertainties’, Global Environ. Change 5 (2), 113–126.
41. Cartwright, N.: 1994, ‘Fundamentalism vs. the Patchwork of Laws’, Proc. Aristotelian Soc. 94,
279–292; Hacking, I.: 1992, ‘The Self-Vindication of the Laboratory Sciences’, in Pickering, A.
(ed.), Science as Practice and Culture, Chicago University Press, Chicago, p. 474.
42. Interview by first author with a GCMer, 10 September 1992.
43. We avoid use of the term ‘reductionism’ here because of its commonly-held definition by climate
(and other environmental) modellers. At least two definitions seem to be used. The more common
definition of reductionism – namely that the environmental system is analytically decomposable
to the governing physical laws – is widely accepted by modellers (a viewpoint we will critically
explore in the next section). At the same time, however, many climate modellers are keen to
distance themselves from the ‘high-physics’ approach to representing systems by employing
a few rather abstract physical principles, a stance which they also appear to have negatively
characterised as reductionism. Historically, part of the reason for the development of this second
perspective may have been to counter the improper extension of a physics-derived commitment
to simplicity into the subject matter of the environmental and earth sciences. Physical complexity
was at least a step away from such overt reductionism. (Personal communication to first author
from Dr. Naomi Oreskes, 5 February 1996.) Discussing the implications for model validation,
Oreskes et al. have noted that the more sophisticated the model becomes, the more difficult it is
to provide confirmation or falsification of the model (see also, Oreskes, N., Shrader-Frechette,
K., and Belitz, K.: 1994, ‘Verification, Validation, and Confirmation of Numerical Models in the
Earth Sciences’, Science 263, 641–646).
44. Palmer, T. N.: 1993, ‘A Nonlinear Dynamical Perspective on Climate Change’, Weather 48 (10),
314–326.
45. Ibid.: 324.
46. As one modeller puts it: ‘there is a view that if you are only interested in global average
temperature then you don’t really need GCMs – that one dimensional models will be more or
less correct. And that you only need the GCM models if you are interested in homing-in on
a specific region. But I think that is a very simplistic view of the world and you just cannot
make that assumption in a non-linear system. That is really important’ (interview by first author
with a GCMer, 21 June 1993). A problem with Palmer’s analysis, however, is that it is based
on transient results only, whereas thermodynamic argument would suggest it applies much less
forcefully for the case of an equilibrium response (personal communication by first author with
a climate modeller, 6 June 1994).
47. A further example which possibly illustrates the need for GCMs is the model representation of the
thermohaline circulation. Simple models tend to produce a much more unstable thermohaline than
models with higher degrees of freedom, though the modelling evidence is somewhat mixed. The
reliability of the representation of the thermohaline circulation in current coupled AOGCMs, and
its sensitivity to changes in freshwater flux and temperature, remains questionable, however (see,
e.g., Rahmstorf, S.: 1995, ‘Bifurcations of the Atlantic Thermohaline Circulation in Response to
Changes in the Hydrological Cycle’, Nature 378, 145–149).
48. Cf. Ascher, W.: 1981, ‘The Forecasting Potential of Complex Models’, Policy Sci. 13, 247–267.
49. Enting, I. G. and Pearman, G. I.: 1987, ‘Description of a One-Dimensional Carbon Cycle Model
Calibrated Using Techniques of Constrained Inversion’, Tellus 39B, 459–476; Wigley and Raper
[9]; Enting, I. and Lassey, K.: 1993, Projections of Future CO2 , Division Of Atmospheric
Research, Technical Paper 27, CSIRO, Australia, p. 42; Schimel, D., Enting, I. G., Heimann,
M., Wigley, T. M. L., Raynaud, D., Alves, D., and Siegenthaler, U.: 1995, ‘CO2 and the Carbon
Cycle’, in Houghton, J., Meira Filho, L., Bruce, J., Lee, H., Callander, B., Haites, E., Harris,
N., and Maskell, K.: 1995, Climate Change 1994: Radiative Forcing of Climate Change and
an Evaluation of the IPCC IS92 Emission Scenarios, Cambridge University Press, Cambridge,
p. 339.
50. Gardner, R. H. and Trabalka, J. R.: 1985, Methods of Uncertainty Analysis for a Global Carbon
Dioxide Model. Publication DOE/OR/21400-4. U.S. Dept. of Energy, Washington D.C., p. 41.
51. Schneider, E.: 1996, ‘Flux Correction and the Simulation of Changing Climate’, Annales Geo-
physicae 14, 336–341.
52. Enting and Lassey [49].
53. Further details can be found in Parkinson, S.: 1995, The Application of Stochastic Modelling
Techniques to Global Climate Change, Ph.D. Thesis, Lancaster University, U.K., p. 263; Young
et al., 1996 [29]; Parkinson, S. and Young, P.: 1997, ‘Uncertainty and Sensitivity in Global
Carbon Cycle Modelling’, submitted manuscript.
54. Huggett, R.: 1993, Modelling the Human Impact on Nature: Systems Analysis of Environmental
Problems, Oxford University Press, Oxford, p. 192.
55. See e.g., Scavia, D.: 1993, ‘Lake Ecosystem Modelling’, in Young, P. (ed.), Concise Encyclopedia
of Environmental Systems, Pergamon Press, Oxford, pp. 318–320.
56. One GCM group recently wrote a paper entitled: ‘Monte Carlo climate change forecasts with a
global coupled ocean-atmosphere model’ (Cubasch, U., Sausen, R., Maier-Reimer, E., and Voss,
R.: 1994, ‘Monte Carlo Climate Change Forecasts with a Global Coupled Ocean-Atmosphere
Model’, Clim. Dyn. 10, 1–19). In this study they ran the AOGCM model only four times, which
according to the standards of many modellers would not count as MCS at all! Nevertheless
the significance of the model’s initial conditions was clearly indicated, so it had some positive
contribution.
57. We feel it is necessary to clarify our definitions of ‘deterministic’ and ‘stochastic’, since there
appears to be some variation in their meanings. A deterministic model for us is one whose
temporal evolution can be calculated exactly, given a set of initial conditions. A stochastic
model, on the other hand, is one whose parameters and/or inputs are defined probabilistically,
so that the model response can only be defined in probabilistic terms. Thus, a model of the
global climate system can be deterministic or stochastic, depending upon the views of the model
builder. But uncertainty is surely unavoidable in any investigation of the natural environment and
it seems more appropriate that it should be considered in stochastic terms.
58. See [53].
59. It is also possible, however, that this assumption may also lead to lower estimates of uncertainty. It
depends very much upon the nature of the model being considered – see discussion in Parkinson
and Young (1997) [53].
60. See [53].
61. This empiricist approach to model evaluation would not be accepted by those using a more
formal structural epistemology.
62. Schimel et al., Enting and Lassey [49]), Wigley and Raper [9], and Parkinson and Young [53].
The uncertainty in each of the deterministic exercises is due to the following: variation between
different models; variation in one model due to different assumptions about feedbacks; and
limited parametric uncertainty.
63. The uncertainty (66% confidence limits) of the pre-industrial level of CO2 , (pa )pi is thus reduced

from 275 15 ppmv to 275 7:5 ppmv which is closer to the uncertainty discussed by Schimel
et al. [49] – a value published too late to be considered explicitly by this analysis.
64. Schimel et al. [49].
65. See [53].
66. Young, P.: 1984, Recursive Estimation and Time-Series Analysis, Springer-Verlag, Berlin, p. 300;
Young, P.: 1985, ‘The Instrumental Variable Method: A Practical Approach to Identification and
System Parameter Estimation’, in Barker, H. and Young, P. (eds.), Identification and System
Parameter Estimation: Vol. 1 and 2, Pergamon Press, Oxford, pp. 1–16.
67. See [53]. Note that the structural decomposition of the transfer function is ambiguous: other
decompositions are possible, including feedback structures that may provide a better physical
interpretation than the parallel decomposition. However, such discrimination is not important to
the discussion here.
68. The input due to land-use changes is set to zero in the original model in order to carry out this
comparison. Hence, a lower atmospheric carbon dioxide level than expected is seen in Figure 2
during the simulation.
69. These values come from the continuous-time reduced order model reported in Young and Parkin-
son (1996), which explains the EL model data a little better than the discrete-time model described
previously in Young et al. (1996) [29]. However, the time constants of both models are quite
similar and the mechanistic interpretations are the same. (Young, P. and Parkinson, S.: 1996,
Simplicity out of Complexity in Forecasting Climate Change, Technical Note Centre for Research
on Environmental Systems and Statistics, Lancaster University, p. 37.)
70. This final stage of interpreting the reduced order model in physically meaningful terms is part of
a technique which we term Data-Based Mechanistic (DBM) modelling; a technique which can
be applied within DMA analysis to simulated data and, more generally, to real data (see Young, P.
and Beven, K.: 1994, ‘Data-Based Mechanistic Modelling and the Rainfall-Flow Nonlinearity’,
Environmetrics 5, 335–363 (Special Issue on ‘Environmental Time Series Analysis’)).
71. Enting and Lassey, Enting and Pearman [49].
72. The IPCC 1990 report itself stated that: ‘Because running coupled ocean-atmosphere GCMs is
expensive and time-consuming, many of our conclusions about global trends in future climates
are based upon simplified models’ (Houghton, J., Jenkins, G., and Ephraums, J.: 1990, Climate
Change: The IPCC Scientific Assessment, Cambridge University Press, Cambridge, p. 365).
Henderson-Sellers has presented similar arguments to ours. (Henderson-Sellers, A.: 1996, ‘Cli-
mate Modelling, Uncertainty and Responses to Predictions of Change’, Mitigat. Adapt. Strat.
Global Change 1, 1–21; and ‘Bridging the Climate Gap’, in Giambelluca, T. and Henderson-
Sellers, A. (eds.) 1996, Climate Change: Developing Southern Hemisphere Perspectives, John
Wiley, Chichester, pp. 35–60.)
73. We have also observed the enthusiastic response at a meeting of IPCC advisory scientists and
policy makers to Wigley’s 1-D-model, because of its perceived policy-usefulness.
74. Interview by first author with a government official, 23 February 1993.
75. Such dependency could have serious implications vis-à-vis the commitment of developing coun-
tries to climate change as a policy issue. Given a context of distrust between policy makers in
the North and South, the question must be asked of whether GCM climate change scenarios
developed in the North will be trusted, especially given the difficulty of independently evaluating
the judgement which goes into construction of those scenarios.
76. Polanyi, K.: 1958, Personal Knowledge, Routledge and Kegan Paul, London, p. 428.
77. Leading scientists themselves may prefer the level of control over scientific assessment offered
by fewer centres. For example, a previous director of the U.K. Meteorological Office, Sir John
Mason, stated before the House of Lords Select Committee on Sustainable Development in 1994
that: ‘: : : the number of people or the number of groups around the world involved in climate
modelling has increased rather dramatically in the last two or three years and not wholly for the
better in my view. It is easier when you have only three or four major groups, particularly if one
of them is in the United Kingdom, to make judgments than when there are a lot of people in
the field and perhaps introducing more turbulence into the subject’ (p. 15, Minutes of Evidence
taken before the Select Committee on Sustainable Development, Tuesday 3 May 1994, London,
HMSO, HL Paper 66-ii).
78. Interview by first author with a GCMer, 11 April 1994. The episode is mentioned in Gore, A.:
1992, Earth in the Balance, Houghton Miflin Company, New York, p. 407. A similar point is
made by Parsons, E.: 1995, ‘Integrated Assessment and Environmental Policy Making: In Pursuit
of Usefulness’, Energy Policy 23, 463–476.
79. Technology Review: 1992, ‘The Political Pleasures of Engineering: An Interview with John
Sununu’, Technol. Rev. 95, August/September, 22–28.
80. Research results do not clearly indicate the effects of changes in the THC on the take-up of
heat, and subsequent feedbacks upon surface temperature and precipitation. See, e.g., Harvey,
L. D. Danny: 1994, ‘Transient Temperature and Sea Level Response of a Two-Dimensional
Ocean-Climate Model to Greenhouse Gas Increases’, J. Geophys. Res. 99 (C9), 18,447–18,466.
81. A further example is the argument by Schlesinger and Jiang that a ten year delay in greenhouse gas
emission reductions would have a minimal effect on potential warming. Schlesinger and Jiang
were criticised by Risbey, Handel and Stone for basing strong policy conclusions on simple
climate models which failed to take account of non-linear, possibly abrupt climate change, as
well as probably under-recognising the extent of regional climate change. The debate between
Schlesinger and Jiang and Risbey, Handel and Stone subsequently revolved, however, around
different and conflicting interpretations of the same GCM model runs, indicating that it was, in
practice, difficult to divorce the evaluation of the simple models from more complex ones, and
vice versa. A further difference between Schlesinger and Jiang and Risbey, Handel and Stone was
the degree to which models, whether simple or GCMs, were held to be sufficiently robust to act
as the basis of policy-decisions to delay action. Schlesinger and Jiang defended the use of both
types of models for this purpose, whilst Risbey, Handel and Stone implied that neither simple
models nor GCMs were currently adequate to act as the basis for such decisions. (Schlesinger,
M. and Jiang, X.: 1991, ‘Revised Projection of Future Greenhouse Warming’, Nature 350,
219; Schlesinger, M. and Jiang, X.: 1991, ‘A Phased-In Approach to Greenhouse-Gas-Induced
Climatic Change and Climatic Response to Increasing Greenhouse Gases’, Eos Trans. A.G.U.
72 (53), 596–597; Risbey, J., Handel, M., and Stone, P.: 1991, ‘Should We Delay Responses to
the Greenhouse Issue?’ and ‘Do We Know What Difference a Delay Makes?’, Eos Trans. A.G.U.
72 (53), 593.)
82. In this section we draw upon theories developed in the field of social studies of science. See,
for example, Callon, M.: 1986, ‘Some Elements of a Sociology of Translation: Domestication
of the Scallops and of the Fisherman of St. Brieuc Bay’, in Law, J. (ed.), Power, Action and
Belief, Methuen, London, pp. 196–233; Latour, B.: 1988, The Pasteurisation of France, Harvard
University Press, London, p. 273; Latour, B.: 1992 (1987), Science in Action, Harvard University
Press, London, p. 274.
83. Interviews by first author with global vegetation dynamics modellers, 15 December 1993 and 16
September 1993.
84. Compare, Perry, J.: 1992, The United States Global Change Research Program: Early Achieve-
ments and Future Directions, National Academy Press, Washington D.C., p. 20.
85. As a senior scientist at a GCM modelling centre put it: ‘This (modelling centre) of course, is in
a sense the tip of an iceberg though. It depends on a community of international and national
research to actually achieve its task : : : if you don’t have something like (the modelling centre),
you actually run the risk that you make a lot of investment but as a nation, actually, you lack
perhaps the primary benefit of that integrated output. : : : And I think that’s one of the frustrations,
perhaps, of some of the smaller European nations, that they actually contribute quite a bit to the
science base of it, but they don’t get the high profile of the actual prediction activity’. (Interview
by first author with a research manager, 6 May 1993.)
86. Interview by first author with a GCM research manager 5 April 1993.
87. Interview by first author with an observational climatologist, 10 February 1993.
88. This was done by Parkinson and Young [53] in generating the results quoted in Section 5.
89. Parry, M.: 1990, Climate Change and World Agriculture, Earthscan, London, p. 157; Carter, T.,
Holopainen, E., and Kanninen, M. (eds): 1993, ‘Techniques for Developing Regional Climatic
Scenarios for Finland’, Publications of the Academy of Finland 2/93, Painatuskeskus, Helsinki,
p. 63; Viner, D. and Hulme, M.: Climate Change Scenarios for Impact Studies in the U.K.,
Climatic Research Unit, University of East Anglia, p. 70.
90. Admittedly, economists also make extensive use of simple models. Nordhaus, W. D.: 1994,
Managing the Global Commons: The Economics of Climate Change, MIT Press, London, p. 213.
91. It might be objected that this is a circular argument, since the status of GCMs in policy derives
itself in part from the prominence and credibility of GCMs within science. However, what we are
proposing is a process of mutual reinforcement of status in which both processes occur concur-
rently. (Shackley, S. and Wynne, B.: 1995, ‘Global Climate Change: The Mutual Construction
of an Emergent Science-Policy Domain’, Sci. Publ. Pol. 22 (4), 218–230.)
92. It could be countered that versions of the science emphasising the climate system’s chaotic
nature might act to strengthen the rationale for global environmental policy action, especially if
the precautionary principle is accepted. Note, however, that a dominant response to the suggestion
that the climate system might face abrupt, chaotic and unexpected changes, has been model-based
analyses of whether this feature affects our ability to find an optimal solution to the problem of
managing climate change. Hence it attempts to re-impose a control and management ethos at a
subsequent level. (e.g., see: Lempert, R., Schlesinger, M., and Hammitt, J.: 1994, ‘The Impact of
Potential Abrupt Climate Changes on Near-Term Policy Choices’, Clim. Change 26, 351–376).
Nevertheless, as James Risbey points out, Palmer’s analysis may point to a happy marriage
between chaos and GCMs, since the latter are needed to represent regime structure, identify
climate attractors and perform ensemble climate forecasts. A similar point has been made by
a GCM modeller, who noted that: ‘GCMs provide the most practical means of investigating
instability, given that the details of the “mean” climate or attractor determine the nature of the
instability. : : : to get the right answer the detailed shape of the attractor may be important’
(personal communication, November 1995).
93. An alternative argument, suggested by one of our referees, is that policymakers are not so much
interested in ‘good science’ as in ‘good scientists’. Although we can think of a few scientists
who do not personally use GCMs and are also considered within the IPCC to be amongst the
leading scientists, they still draw extensively upon GCMs in their provision of scientific advice.
Moreoever, none that we can think of departs from the consensus view that GCMs are the ‘best’
amongst climate models (for the reasons discussed in Section 2.4). This cursory observation does
not confirm the argument either way, but raises the significant question of whether the evaluation
of scientists can be divorced from the methods and tools used by those scientists.
94. Many scientists would agree with this of course and it appears to be a major rationale for the
development of new integrated assessment models (IAMs) as heuristic tools. However, IAMs
are still considered by many scientists to be primitive when compared to GCMs and the extent
to which they should be trusted as a aid to decision-making in climate negotiations surely needs
to be critically discussed. See also van der Sluijs [33].
95. Young et al. [29].
96. Van Asselt, M.: 1994, Global Integrated Assessment Models as Policy Support Tools, Masters
Thesis, University of Twente, Netherlands, p. 94.
97. For example, there is good reason to believe that the overwhelmingly technical framing of
the climate change issue structures the policy agenda in a way which fails to engage with the
diverse constituencies (such as local government, industry and lay members of the public) whose
commitment would be necessary for any serious policy on global climate change. (Macnaghten,
P., Grove-White, R., Wynne, B., and Jacobs, M.: 1995, Public Perceptions and Sustainability
in Lancashire: Indicators, Institutions and Participation, Lancashire County Council, Preston,
p. 96; Macnaghten, P. and Jacobs, M.: 1997, ‘Public Identification with Sustainable Development:
Investigating Cultural Barriers to Participation’, Global Environ. Change 7 (1), 5–24.)
(Received 17 July 1995; in revised form 30 June 1997)

Uncertainty Complexity and Concepts of G PDF

Uploaded by

Copyright:

Available Formats

Uncertainty Complexity and Concepts of G PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Uncertainty Complexity and Concepts of G PDF

Uploaded by

Copyright:

Available Formats

UNCERTAINTY, COMPLEXITY AND CONCEPTS OF GOOD SCIENCE

IN CLIMATE CHANGE MODELLING: ARE GCMs THE BEST TOOLS?

SIMON SHACKLEY1 , PETER YOUNG 2 , STUART PARKINSON 2 and

‘The multiplicity of models is imposed by the contradictory

Climatic Change 38: 159–205, 1998.

2. The Climate Modelling Pyramid

2.1. THE ROLE OF GCMs IN CLIMATE RESEARCH

dynamics and thermodynamics of the feedback processes. Such physically-based

2.3. THE PERCEIVED ROLE OF SIMPLE MODELS

2.4. REASONS IN SUPPORT OF THE CLIMATE MODELLING PYRAMID

3. Challenges to the Climate Modelling Pyramid I: Applying the Scientific

3.1. PARAMETERISATIONS IN GCMs

3.2. PARAMETERISATIONS AND MODEL INTERACTIONS

3.2.1. Response from GCMers

3.3. FORMAL AND INFORMAL STYLES OF MODEL EVALUATION

4. Challenges to the Climate Modelling Pyramid II: Complexity and Model

4.1. ARGUMENTS FOR SIMPLICITY

4.2. COMPLEXITY AS A MORE HOLISTIC SCIENCE?

5. Challenges to the Climate Modelling Pyramid III: The Need for

5.1. QUANTIFYING THE CONSEQUENCES OF UNCERTAINTY: MONTE CARLO

While it is a technique which is rarely used in climate research, possibly due

Year Uncertainty in Source

2050 494 to 510 16 Schimel et al. (1995)a

greater than those gained by deterministic model studies alone. It is recognised

5.2. SIMPLICITY OUT OF COMPLEXITY: DOMINANT MODE ANALYSIS (DMA)

Dominant Mode Analysis (DMA) seeks to analyse a given, physically-based, deter-

5.3. IMPLICATIONS FOR GCM DEVELOPMENT AND USE

6. Why are GCMs Perceived as the Best Climate Models?

To account fully for the dominance of deterministic reductionism in climate mod-

6.1. ARE GCMs ESPECIALLY POLICY USEFUL?

At first sight, GCMs appear to be at a distinct advantage for policy purposes:

6.1.1. Sununu and Simple Models

6.2. TOWARDS A SOCIOLOGICAL UNDERSTANDING OF THE POSITION OF GCMs

6.2.1. Involvement of Other Scientists

6.2.2. Scientific ‘Users’ of GCMs

Approximate expenditure in millions

6.2.3. Policy Makers

Notes and References

(Received 17 July 1995; in revised form 30 June 1997)

You might also like