Lecture 5: Models and Modelling

1. Introduction
What scientists call “models” come in many different forms. To give a few examples, architects,
engineers and archaeologists often use scale models. Biologists and medical researchers use
model organisms – non-human animals that are extensively studied to understand particular
biological phenomena, with the expectation that discoveries made in the organism model will
provide insight into the workings of other organisms. Ecologists, geologists and meteorologists
use computational models that require extensive computational resources to study the behaviour
of a complex system by computer simulation. Many empirical scientists use statistical models,
describing stochastic relations between random variables. These are often used to describe data.
Physicists and economists use mathematical models to describe the behaviour of certain factors
in isolation.
Many sciences, it seems, make use of models, although not to the same extent, and not
necessarily for the same purposes. When talking to scientists, in fact, it seems that while few are
involved in building theory, many see model building, model application and model testing as an
inherent part of their work. So the first question I want to address in this lecture is what are
models? Given the astounding varieties of models and modelling practices, any attempt to
characterize modelling will necessarily remain incomplete. What I aim at here, instead, is to
discuss some important characteristics of models of their use, which help distinguishing models
from theory as well as from phenomena, and which help distinguishing modelling practices from
other theoretical practices. So the second question I address is what scientific uses do models
have?
Once we have taken stock of the main characteristics of models and their uses, it will become
clear that one of the main features of models is that they idealise – they make assumptions about
the world that are obviously false. This is usually not an accident: rather, modellers intentionally
seek to make these idealizing, false assumptions.
This idealizing characteristic of models poses a deep philosophical problem. You have heard that
science pursues knowledge, knowledge is commonly understood as true, justified belief, and that
science pursues knowledge by observing reality and by testing its theories against these
observations. Now it turns out that models always include patently false claims about the world.
So how can hope to obtain knowledge from models, given their idealizing assumptions? This is
certainly something that bothered me a lot when I began as an economics undergraduate: Here I
was, eager to learn about the working of the economy, but what I spent most of my time on
doing the first year was to learn how prices adjust in a two-person economy with only one good!
Those models seemed not to resemble the real economy particularly closely, and I found little
evidence that these “toy models” later on would become more “realistic”.
One way to answer this concern is to point out that models can be true, despite their many
idealizing assumptions. I will discuss whether models can be true and when? later on in this
lecture. But truth may not be the only virtue that models may possess. As we will see when
discussing the different uses of models, ease of understanding, mathematical tractability, and
generality are also important properties, and these properties may conflict with the truth or
accuracy of a model.

1
The final question in this lecture is therefore How should one build models? If the different
scientific virtues of models may conflict, there will be important trade-offs that one has to
consider when choosing a model for one’s research project. Furthermore, even if truth is
considered the dominant goal, it is often not easy to determine the truth or falsity of a model (at
least not in the ways relevant for models, as you will see). So what kind of model properties
could help us determine whether a model is more useful, or adequate, or at least more promising
than other possible models?

2. What are models?

A common observation is that models are models of something: the architect’s model is a model
of a building; the meteorologist’s model a model of weather phenomena; the Crick-Watson
model a model of DNA; the econometric model a model of stochastic relations between variables
in a data set. Models thus have targets.
Note that commonly, we would not say that the model is the same as its target: the computational
model is not a weather phenomenon, nor is the Crick-Watson model a piece of DNA. Rather,
although the model has an identity different from its target, it stands for the target and takes its
place in various scientific practices. For example, scientists often manipulate a model instead of
its target, and observe and analyse the model’s behaviour, in order to learn something about the
target. Philosophers call such a relation of standing-in-for a representation. A model can thus be
said to represent its target, depicted in this very simple graph:

Model represents Target

Figure 1
Once we have established that models are fundamentally representations, a lot depends on the
nature of models and targets, and the meaning of representing.
Let’s start with the nature of targets. In the simplest case, such targets are objects or events that
exist or occur in the real world at present. The Crick-Watson model, for example, represents the
double-helix structure of DNA found in almost all living organisms. Yet other models do not
have such straightforward targets. Climate models, for example, often represent alternative future
scenarios, which may occur if certain states pertain. For example, if global average temperature
rose by about 2oC, all else equal, then the map of Florida may look as follows:

Figure 2
Whether these conditions will ever pertain is not certain, so the model represents only a future
possibility. Even more extreme, models can represent objects or situations that are unlikely or

2
even impossible to ever exist or occur. A two person economy with only one good and money,
for example, is unlikely ever to occur (historically, money only emerged in an economy with
many goods). Yet you will find models of such economies in every microeconomic textbook.
Other examples are models of three-sex-biologies or models of perpetual motion machines.
Depending on one’s metaphysical leanings, one may deny that possible, unlike or even
impossible states have any form of existence, and therefore speak of “models without targets”.
My main point here, however, is that although models have targets, often these targets are not to
be found in the real world.
If we focus on those targets that can be found in the real world, an important distinction between
kinds of real targets offers itself. When people describe real objects and events, they describe
their characteristic features. Think for example, how you would describe a gas explosion, or an
economic recession, or a disease: you would probably mention the explosion noise, the decrease
of unemployment in the recession, and a key symptom of the disease. These descriptions are of
course based on observations of such events, but they differ from the observations themselves:
they are not mere reports of data points and distributions of these, but rather identify relatively
stable features that characterize such events.
Philosophers have therefore suggested distinguishing descriptions of phenomena from
descriptions of data (Bogen and Woodward 1988). Data is of course used in order to establish
phenomena. As you have learned in lecture 4, observation and measurement is not a trivial
process, and requires considerable attention. But the aim of science – what scientific theories and
models are about, what they seek to explain, predict or control – is not data, but the phenomena.
Scientists want to explain explosions, predict recessions and control diseases. For that they often
need to measure and analyse data, but they do not seek to explain, predict or control that data by
itself.
Models are used both for representing phenomena and for representing data. As we will see,
these models are quite different, and therefore we need to distinguish between models of
phenomena – for example scale models, model organisms, or analogical models – from models
of data – for example econometric models or diagrammatic scatter models.
Let’s now move on to the nature of models. Models can be made of many different things. Scale
models and model organisms are made of material objects. The representational function of such
material models relies on the similarity between some of the model’s and the target’s material
properties. A successful scale model, for example, materially exhibits the proportions of an
object. A medical model organism relies on a similarity of its physiological reaction to a drug to
that of its target. Thus material model make use of their material and causal properties in
representing a target.
Mathematical and statistical models, on the other hand, do not rely on their material properties
for representing a target. Instead, they specify formal properties, for example the mathematical
properties of a function, like continuity, slope or curvature. Mathematics is not the only way to
describe such conceptual models: diagrams, pictures, analogies, informal descriptions or
axiomatisations are alternative ways to specify the properties of a formal model.

3
 As an example of axiomatisation, consider the economic notion of a rational consumer. A consumer in
economics is identified with a preference relation > over a set of alternatives X. Rational choice theory makes
two assumptions about individuals' preferences for actions:
• Completeness – all actions can be ranked in an order of preference
• Transitivity – if action x1 is preferred to x2, and action x2 is preferred to x3, then x1 is preferred to x3.
Together these assumptions form the result that given a set of exhaustive and exclusive actions to choose from,
an individual can rank them in terms of his preferences, and that his preferences are consistent.
Note that “consumer”, “preference” and “rational” have meanings different from these in everyday language.
The axiomatic specification of a model prevents these everyday terms to have an impact on the model.
Accordingly, it does not make sense to say something like “but rational consumers often have non-transitive
preferences!”. The axiomatic specification thus removes ambiguity about the model properties. By the same
token, this advantage in clarity often complicates the relation of the model to its intended targets.

Note that these specifications do not describe the target directly. Rather, philosophers of science
now widely agree that these specifications describe the properties of a model, and that this model
in turn represents the target, as shown in figure 3 (Giere 1988, 83).

Axiomatisation

Diagram describes represents

Target
Model
Description

Function
Figure 3
Why should one think of conceptual models in this way? Partly, because it offers a unified
account of both material and conceptual models – it is always the model that represents, and the
model is either specified by its material properties, or by some form of description. But more
importantly, this account allows us to deal with the fact that all models idealise. There is hardly
ever a function, diagram of axiomatisation that is entirely true of its intended target. If we
considered these specifications descriptions of intended targets, we would have to admit that they
are false descriptions. But falsity does not come in degrees: we cannot distinguish between a
“more false” and a “less false” description. This would make it hard to separate those models
that usefully represent important properties of their targets from those that do not. Instead, if we
consider these functions, diagrams, etc. specifications of models, we can then compare the thus
specified models with the intended targets, and assess whether they are similar in the desired
ways and degrees. I will come back to this in a minute.
In practice, the division between material and conceptual models is often not clear.
Computational models, for example, are somewhat in between these two categories.

4
 Clearly, computer simulations start out with an conceptual model: something in the programmer’s mind that she
wants to implement on a computer. But this implementation turns the model into something material, involving
the material properties of the machine on which the conceptual model is implemented. In many fields, there is a
special name - verification - given to the process of collecting evidence that a computer simulation can indicate
solutions that are close enough to those of the programmed equations and/or the preferred model equations.
Verification activities are concerned not only with such things as the adequacy of the numerical methods that
are employed to estimate solutions to the programmed equations, but also with the design and actual
implementation of the computer program that is run to generate the simulation. Possible errors in the coding of
the program are considered important and are investigated as part of the verification process. And if the
computer hardware itself is not working properly or as expected, that too can sometimes be revealed through
verification activities (Parker 2009).

My main point here is that the properties of models relevant for representation can be specified
in different ways, and that diagrams, functions, etc. specify model properties, but do not describe
targets directly.
Now that we have discussed and categories different kinds of targets and models, how are we to
understand the representational relations between them? The simplest case is that of a material
model representing a real-world target. By stressing that some feature of the model resembles a
feature of the target, a particular representational relation is established. The scale model, for
example, resembles the real-world building in that it exhibits the same proportional relations
between doors, windows, roof, etc. It does not resemble the house in other ways, for example in
its size, or the material it is constructed with. The representational relation between model and
house is thus one of similarity in a particular property, to a particular degree. Note that both
model and target have infinitely many properties, and similarity is a matter of degree. Thus, on
the one hand, we can always find a property by which any two objects resemble each other to
some degree; and on the other hand, there will always be properties in which two distinct objects
differ, to some degree. Hence we can never legitimately say that “this model cannot represent
that target”. All we can say is that “it is not useful for our purpose to represent that object with
this model”.
A slightly more complicated case is a material model representing a non-existing target. An
architect, for example, may design a model of a building that she envisions, but which has not
been built. Then the model represents the architect’s imagination, and the respective similarity
relation is between properties of the scale model and properties of the building that she imagines.
Given our discussion around figure 3, we can give a similar treatment to conceptual models.
Diagrams, functions or axioms specify certain properties of the model, and the similarity
between models and their targets are compared according to these properties. Note that one can
use different means to specify the same properties: one can state a function mathematically, or
diagrammatically, for example. An equation and a diagram therefore may describe the same
model. Note also that many of the properties of a model may not be known, because the modeller
has not described them.

3. What are models used for in science?

Models are representations – they are used in place of their target, to learn about or control the
target, to predict its behaviour or explain some of its features. In the simplest case, models are
used as convenient partial representations of targets too complex to handle by themselves. Maps

5
are a prime example of such models. They considerably simplify the geographical properties of
an area in order to be useful for navigating this area. Note that which simplifications are made
depends on what kind of navigation is sought.
Take the current map of the London underground (figure 1c). Clearly, it makes a lot of “unrealistic”, “absurd”
or “false” assumptions about the properties of the London underground system. This includes neglecting some
properties, for example where the trains run underground and where above ground. Also, it apparently states
properties that are incorrect, like the directions of the lines, and the distance between stations. Consequently,
many conclusions that one could derive from this map about the geography of London would be false. Yet this
is widely regarded as an extremely useful map, and it is regarded so presumably because one learns from it
about the real London and its underground system.

Figure 4 Map of the London Underground (a) 1932, (b) 1956, (c) 2002
The evolution of this map offers further interesting observations. Maps until 1932 (figure 1a) largely conformed
to geographical realities. Yet this was abandoned in 1933 because the more schematic approach was considered
more comprehensible. Consecutive map designed tended to be rather “pure” rectangular schemes. In
comparison to them (figure 1b), the current design is more “realistic”, offering some information about the
relative distance between different lines.

What one wants to use a map for thus is an important motivation for which aspects of reality to
simplify. For example, distance between stations can be neglected in the underground map,
because the map is not meant to be used as information how to walk between stations.
Another motivation for making simplifying assumptions, particular in mathematical models, is
tractability. When representing a target with a mathematical function, for example, it is much
easier to analyse this function if it is continuous and differentiable. If the purpose of modelling
lies in such an analysis – e.g. optimality analysis – then it may be useful to represent the target in
this way, even if it is not known whether the target really has these properties, or even if it is
known that the target does not have these properties. Of course, if one makes such idealizing
assumptions, one must be careful to not derive conclusions that substantially depend on them –
just in the same way as one should not use the distances shown in the subway map as a guide
how to walk between stations.
Similar considerations apply when constructing models of data. When scientists speak of data,
they often do not mean “raw data” – the immediate product of observation. Rather, they have in
mind corrected, rectified, regimented, and in many instances idealized version of the data - so-
called ‘models of data’ (Suppes 1962). Models of data are typically derived from data in two
steps. First, presumed errors (from presumed faulty observation, or intentional misleading
(“strategic”) behaviours or answers) are eliminated. Secondly, the data is presented in a ‘neat’
way, for instance by drawing a smooth curve through a set of points.
Take the example of figure 4. Here we have observational data that has only two dimensions: an
observation of variable X associated with variable Y. All but one data point in (a) are closely
fitted by a linear function. The ‘outlier’ is eliminated. In (b), the data points are fitted by a more
complex polynomial function, and an even more complex one in (c).

6
 Figure 5
By “fitting”, is meant an optimization process, for example the minimisation of the sum of
distances between curve and data point at each argument of the x-axis. But how does one choose
between eliminating some data points, and fitting a simpler curve, or keeping these points in and
fitting a more complex curve? I will return to this question later.
At this point, note that the fitted function is the model, which is made as similar as possible to
the underlying data-generating process. In actual data modelling, the underlying process is of
course not known. All one has are scatters as in the top row of figure 4. It is the task of the data
modeller to select the most appropriate structure to fit to the data. Note also that the similarity
comparison is somewhat simpler than in models of phenomena: the data model is obviously only
compared to the data with respect to those properties that are contained in the data. Under ideal
conditions, we may therefore be able to speak of the unique best representation of a given data
set – independent of purpose. For models of phenomena, no such possibility exists.
So models are used to represent phenomena and data, to make them accessible to us, to analyse
them and understand them. But isn’t that what theories do? How are theories different from
models? Or are they the same, and it’s just a recent fashion to speak of models instead of
theories?
Although the last suspicion is not correct, it is a true observation that certain “theories” at some
point were called “models”: for example the “Bohr theory”, that is now commonly called the
“Bohr model” and the “free electron theory” which is now commonly called the “free electron
model”. To call something a theory implies that one indeed has reasons to believe that the claims
by this theory are not false, and that the object of the theory is indeed governed by the principles
stated in the theory. Both the Bohr model and the free electron model were first interpreted this
way. But later, it was realized that they only provided approximations of the objects in question,
useful for certain purposes, and that it was possible to find alternative representations of the same
objects useful for different purposes (Archinstein 1968, 215). And thus they were demoted from
theory to model.
Because they are possible approximating representations useful for some purposes, models can
be more concrete than theories can. They can fill in where a theory remains abstract, making
claims about inner structure, composition or mechanism of objects or events. These aspects can
often not be directly observed, so the model makes proposals about possible states or processes.
Furthermore, theories are usually principled statements expressed in a language – e.g. “all Xs are
Ys”. They can only be true or false of an object or an event. As we learned, models, in contrast,
are objects, either material or abstract, formally specified objects. They are more or less similar
to real-world objects or events, with respect to certain features.

7
This distinction between theory and model becomes particularly clear when one realises that
models are often used to interpret theory. This is particularly pertinent in physics, where general
laws—such as Newton's equation of motion—lie at the heart of a theory. These laws are applied
to a particular system—e.g. a linear oscillator—by choosing a special force function, making
assumptions about the mass distribution of the pendulum etc. The resulting model then is an
interpretation (or realization) of the general law.
Take the example of a harmonic oscillator (Giere 1988, 68-70). Here the force on a particle is proportional to
the negative displacement of the particle from its rest position x. The law for this case is:
(a) F = -kx
where k is the constant of proportionality.
The vast majority of texts introduce harmonic motion through the example of Hooke's law, which states that the
force exerted by a spring is proportional to the amount it is stretched. The constant, k, is then interpreted as a
measure of the stiffness of the spring. Figure 6 shows a typical representation of the problem.

Figure 6
Most authors are careful to indicate that applying equation (a) to the setup pictured in figure 6 requires some
"simplifying assumptions." One text (Wallace and Fenster 1969, 175) goes so far as to list the following
"idealizations": (1) The spring is subject to neither internal nor external frictional forces. (2) The spring is
without mass. (3) The force displacement characteristic of the spring is linear. (4) The mass is subject to no
frictional forces. (5) The wall is rigid, the wall recoil due to motion of the mass may be neglected.
This list of idealizations is especially interesting because it makes clear that the authors are not talking about
any particular real mass-spring system. There are no springs without any mass whatsoever or without internal
frictional forces. Instead, the authors are dealing with an ideal mass-spring system that perfectly satisfies
equation (a). The idealizations are required to ensure that the conditions of the equation are satisfied.

Some philosophers indeed have suggested that theoretical laws only apply to the highly idealized
situations presented in models, not to the world directly at all (Cartwright 1983). Models then
acquire the important function of mediators between theories and the world: theories state
principles that are true in models, and models are similar to objects or events in the world. This
of course complicates the simple idea of theory testing considerably!
Because models are objects (or processes) one can do many things with them. Scientists not only
passively look at models, to see what they present. Rather, they actively manipulate models,
wringing as it were any possible insight about them and thus about the objects or processes they
represent. You do things with models, you don’t just contemplate them or put them in
correspondence to reality.
This is obvious in the case of material models.

8
 When the US army corps of engineers built a scale model of the San Francisco Bay, they did not plan to merely
look at it.

Figure 7
Rather, they increased and decreased the water level, introduced water supply from different angles and
sectioned off parts of the model with a dam. They then observed the resulting changes in the model –
specifically, change in water flow, water exchange and sedimentation – and concluded from this that damming
the real Bay would likely lead to a radical deterioration of its water quality and ecosystem.

But one can also do things with conceptual models, albeit in a roundabout way. Manipulation
here is usually performed by making changes in the specifications that describe some of the
models features and by demonstrating (perhaps mathematically) that certain interesting
consequences follow from these changes. With both material and conceptual models, we
demonstrate: we trigger a mechanism and observe what it brings about. The mechanism may
very well be purely logical (a set of rules of inference) and the consequence a theorem. In this
sense, theorists do on paper something analogous to what experimenters do in the lab (Hughes
1997).
Schelling’s checkerboard model is a model of an urban neighbourhood, whose whole point is that it can be
experimented upon.
Schelling describes the conception of this model (Schelling 2006, 249-53). He wanted to teach his class how
people’s interactions could lead to unintended and surprising results. He had in mind spatial patterns that
resulted from preferences about whom to associate with in neighbourhoods, clubs or classes. But no illustrative
material was to be found in the sociological literature. “I found nothing I could use, and decided I’d have to
work something out for myself”.
Schelling worked something out for himself with a checkerboard and two sets of different coins. By random
placements, he distributed the coins on the board, leaving about a fifth of spaces blank. Then he postulated that
each coin wanted at least half its neighbours to be of the same kind. Coins who were discontent would move to
a place that satisfied their want.

Figure 8
Experimenting with this set-up, Schelling found that it didn’t matter in what order the discontent coins were
selected. The same results obtained no matter what: coins of similar kind would cluster in certain areas of the
checkerboard. He continued experimenting, and found additional interesting results: if coins of one kind were
in majority or minority, or if they had greater or lesser demand for the same kind, the sizes and densities of the
respective clusters would differ accordingly.
Having obtained these results on the checkerboard, Schelling suggested his little game as a model of
segregation in US cities. One kind of coins are “blacks”, the other kind “whites”, the moving rule a preference
not to be in the minority, the checkerboard with its 64 cells an urban environment, and the emerging clusters
segregated neighbourhoods. The result of Schelling’s modelling efforts then include conclusions about the
possibility of urban segregation without the involvement of racist preferences.

9
By demonstrating with models, we hope to learn about the model’s target. But often, it is not
easy to identify a real target for a particular model. Schelling’s checkerboard model, for
example, is not particularly similar to any city. Cities, after all, are not constituted by 64 cells
arranged in an 8*8 grid. So instead of interpreting the model as a representation of any concrete
city, or a representation as a particular type of city, one might conclude that it presents a mere
possible scenario, a city in a “parallel world” (Sugden 2000). Many have concluded that such
models are merely toy models, good to demonstrate a theoretical principle (as in the linear
oscillator, above), from whom nothing, however, can be learned about the real world.
That conclusion seems a little hasty. We can learn from models that represent mere possible
worlds. In particular, we can learn about the deeper structure of phenomena, and about possible
causes for relevant effects. Schelling’s checkerboard model is a good example for this. While
many might have thought that segregation necessarily involves some intention of city dwellers to
live in segregated areas, the model shows that segregation can emerge from the much more
innocent preference to not be in the minority. Segregation can possibly arise, the model shows,
from such preferences alone, and need not involve overtly racist or discriminatory inclinations.
Showing such possibilities can constitute important information: policy-makers, for example,
learn from this model that anti-discriminatory educational programs may not contribute reduce
urban segregation (Grüne-Yanoff 2009).
Even if one cannot identify such learning effects, models may be useful tools to help along the
broader scientific enterprise. Although epistemologically inert (i.e. without contributing directly
to the increase of our knowledge), they function as important heuristic devices (Wimsatt 2007,
94-132). Of the many possible heuristic usages, let me just give you two illustrations.
Two or more models may be used to define the extremes of a continuum of cases in which the
real case is presumed to lie, but for which the more realistic intermediate models are too complex
to analyse. Game theorists have sometimes argued in this way. They have defended the very
strong epistemic assumption of their classical models, and have pointed out that their results
often match with evolutionary game models, which are based on implausibly weak epistemic
assumptions. They then proceed to proclaim something like “the vast majority of human beings
would fall somewhere in between these two extremes” (Binmore). Because of the difficulty of
specifying the right level of epistemic requirements, classical game models like the PD are
justified as specifications of the extremes of a continuum.
An incorrect simpler model may be used as a reference standard to evaluate causal claims about
the effects of variables left out of it. Take the following example:

10
 For example, the Artificial Anasazi model simulates the history of a pre-historical settlement of Ancestral
Puebloans (often called Anasazi) in Long House Valley, northern Arizona, from 800 to about 1300 AD. The
computation takes as input paleo-environmental data, including meteorological, groundwater and sediment
deposition and fertility estimates for the reconstructed kinds of farmland. On the basis of this input, it
reproduces the main features of the settlement’s actual history, as witnessed by archaeological evidence –
including population ebb and flow, changing spatial settlement patterns, and eventual rapid decline.

Figure 9. Best single run of the Anasazi model according to the L1 norm. (c) Nature 2002
As shown, this ‘best fit’ still does not accurately replicate the historical findings. In particular, it simulates a
higher population early on, and does not replicate the complete eclipse of the settlement in around 1300. The
authors therefore concluded that a further functional component had to be introduced into the model:
The fact that environmental conditions may not have been sufficient to drive out the entire
population suggests that additional push and pull factors impelled the complete abandonment
of the valley after 1300. (Axtell et al. 2002, 7278, my emphasis)
Thus, the models help identify additional causal factors that are necessary to fully account for the Anasazi
population dynamics. Although the model thus makes false assumptions, it potentially contributes to the
development of a better theory.

4. Can models make true claims?

Because models are objects, not descriptions in one language or another, they cannot strictly be
true or false – just like paintings or photographs cannot strictly be true or false. Rather, models,
paintings or photographs can be more or less similar to the targets they are supposed to represent.
But one way to use models is to formulate hypotheses with them: to make claims about the world
with the help of models (Giere 1988, 80). For example, one may say that “this spring is a
harmonic oscillator, hence its force is proportional to the amount it is stretched”. This hypothesis
is a statement – namely that a target X is identical to the model M of the harmonic oscillator –
and therefore it is either true or false. To say that a model M is true thus means that the claim
“Target X is similar to M with respect to properties P to the degree dP” is true.
Note that this definition is empty as long as the set of relevant properties P and the respective
degrees dP have not been specified. A simple claim is that all properties of M and X must be
identical for M to be true. Yet such a reading is not fruitful – the only object that is identical in
all properties to an object X is X itself! Because models are distinct from their targets, this would
leave all models to be false. An alternative claim (specific to conceptual models) is that all those
properties specified in the description of M must be identical to those of X for M to be true. But
even this is an extremely stringent claim: neither the rational agent model, nor the underground
map, nor the linear oscillator, nor the checkerboard model would be true of anything in the world
under this reading.
Some scientists at this point conclude that the truth doesn’t really matter. What matters is what
works – leave the vexing question to the philosophers and get on with the work! While such an
attitude may do for some practitioners, it completely leaves in the dark how some models “work”
and others don’t. If science pursues knowledge, and knowledge is commonly understood as true,
justified belief, then the only way that we can account for models “working” is to find a way

11
how they can contribute to true claims about the world. So we should try a little harder in
seeking out how and when models can be true – because truth matters for the scientific project,
and it even holds the key for some important methodological choices.
An often-heard response is that while a currently used model of target X may be false, once more
is known about the target this model will be made more realistic. Here is a famous example from
an economist:
“… we look upon economic theory as a sequence of conceptional models that seek to express in simplified
form different aspects of an always more complicated reality. At first these aspects are formalized as much as
feasible in isolation, then in combinations of increasing realism.” (Koopmans, 1957, p 142)
In such a sequence, a highly idealized model would be subsequently de-idealised. A model
making false assumptions would be justified in such a sequence if the sequence led to a notable
progress in approximating the truth. In that case the advancement of science would correct the
distortions effected by idealizations and add back the discarded elements, thus making the
theoretical representations become more usefully concrete or particular. The idealized, false
assumption was only a temporary deviation from truth, necessitated by limited computational or
comprehension abilities.
De-idealisation is obviously a very important process in science. Many models that start out as
heuristic devices are improved and sophisticated upon to an extent that they can be fruitfully
used as representations of real targets. Yet it is difficult to explain how de-idealisation can lead
to true models. Obviously, it is absurd to claim that de-idealisation continues until all model and
target properties are identical. That is definitely not happening in scientific practice. And it is not
even clear that it could happen, even in principle – nor that it is desirable (I will come back to
this in the next section).
Instead, de-idealisation, as Koopmans said, is about “increasing realism”, about approximating
reality. So the model is to be made somewhat more similar to the target, with respect to certain
properties. But which properties, and to what degree? Obviously, making the model similar to
the target with respect to any property does not do the trick. To add a correct description of the
distribution of hair colour of a particular population to the rational agent model would make it
more similar to that population, but that is not meant by de-idealisation. Instead, we need an
account of which properties de-idealisation should focus on.
One approach is to identify those properties of a model that need not be de-idealised for the
model to be true. It starts from the claim that some of the assumptions of conceptual models are
misinterpreted (Musgrave 1981). Assumption (5) of the linear oscillator, for example, does not
really specify that the modelled wall is rigid. Rather, it states that the recoil of the modelled wall
is negligible for the operation of the spring. Although there are no fully rigid walls in the real
world (at least not above 0 K), a thus interpreted model may still be similar to many real-world
targets: namely all those in which the recoil is indeed negligible for the working of the spring.
Another example concerns the domain of a model. While the preference transitivity axiom of the
rational agent model may not be true of any individual human’s total preferences, there may be
domains – for example those relating to money and financial transactions – in which it is true. So
the transitivity axiom can be re-interpreted as a stating that the model is only applicable where
this property is met.
Reinterpreting the specifications of conceptual models in this fashion relieves the modeller from

12
de-idealising all assumptions in her quest for a true model. But it is obvious that many obviously
false model assumptions cannot be re-interpreted either as negligibility or domain assumptions.
Instead, we need an account that addresses head-on why models can make many false
assumptions and still be true of certain real targets. The idea of isolation offers such an account.
Isolation is based on the idea that reality is constituted by many interacting causes. Such a messy
mêlée makes it hard to identify any regularities, as the effect of any causal force is inhibited or
biased by numerous other forces, depending on specific contexts. Theoretical isolation is the
practice of constructing and studying theoretical models that single out only the core causal
factors that give rise to a phenomenon, eliminating all others.
Theoretical or ideal isolation … is manifest when a system, relation, process, or feature, based on an intellectual
operation in constructing a concept, model or theory, is closed from the involvement or impact of some other
features of the situation. (Mäki 1992, 325)
Thus, isolation is a process, an ‘intellectual operation’ that consists in ‘constructing’ and in
‘closing off from’. This process, Mäki suggests, is analogous to scientific experimentation. An
experimenter causally intervenes in a process occurring in the world, and thus closes off the
target entity from causal interferences of other entities. Yet while the experimenter causally
manipulates real entities, the theoretical isolator manipulates representations. Mäki therefore
conceives of theoretical isolations as thought experiments, as opposed to laboratory experiments:
‘isolation takes place in one’s ideas, not in the real world’ (Mäki 1992, 325).
Isolation proceeds via omitting or idealising (distorting) the impact of inhibiting factors.
Consequently, an isolating model makes false claims about the world. Yet one must be careful to
distinguish the process of idealisation from the product of isolation. When isolating a factor F
from intervening factors G1,…,Gn, one may either omit or idealise the operations of the Gi, but
not the operation of the factor F itself. This way, one makes false claims about the Gi; but the
purpose of the theoretical process – to isolate the operation of F – remains intact. This is at best
illustrated with an example:

13
The simple model of agricultural land use distribution given in Johann Heinrich von Thünen’s famous classic
Der isolierte Staat in Beziehung auf Landwirtschaft und Nationalökonomie (1826/1842) is sometimes called the
world’s first economic model. von Thünen’s model employs numerous unrealistic assumptions in envisaging an
extremely simple situation that appears to have next to nothing to do with real world situations. These
assumptions include:
1. The area is a perfect plain: there are no mountains and valleys.
2. The plain is crossed by no navigable river or canal.
3. The soil in the area is homogenous in fertility.
4. The climate is uniform across the state.
5. All communication between the area and the outside world is cut off by an uncultivated wilderness.
6. Transportation costs are directly proportional to distance and to the weight and perishability of the
good.
etc.
Idealizing assumptions 1-6 serve the function of neutralizing a number of causally relevant factors by
eliminating them or their efficacy. Assumption 1 eliminates the impact of mountains and valleys on land use.
Assumption 2 eliminates the impact of rivers and canals on land use. Assumption 3 eliminates the impact of
variation in soil fertility while assumption 4 eliminates the impact of variation in climate. Assumption 5 isolates
the area from the rest of the world, eliminating the impact of trade (hence “the Isolated State”). Assumption 6
eliminates the impact of roads and railways and any sort of preservation technology (von Thünen envisaged that
delivery to the town takes place by oxcart). And so on.
We were invited by von Thünen to imagine a highly idealized system characterized by a set of idealizing
assumptions. Right after having outlined a few of those assumptions, he asks what happens in the imagined
situation, that is, in the model.
"What pattern of cultivation will take shape in these conditions?; and how will the farming system of
different districts be affected by their distance from the Town?"
He then immediately answers the question, that is, he describes the outcome that emerges within the model. It is
the land use pattern under the peculiar circumstances characterized by the idealizations of the model.
“... near the town will be grown those products which are heavy or bulky in relation to their value and
hence so expensive to transport that the remoter districts are unable to supply them. Here too we shall
find the highly perishable products, which must be used very quickly. With increasing distance from
the Town, the land will progressively be given up to products cheap to transport in relation to their
value. For this reason alone, fairly sharply differentiated concentric rings or belts will form around the
Town, each with its own particular staple product.”
Thus the model yields the famous von Thuenen rings:

Figure 10
The model isolates the major cause of land use, the functioning of the wirkende Kraft of distance, from all its
other causes. And what the model isolates theoretically is also present in the real system even though here it is
not isolated from other influences. In short, the wirkende Kraft in the imagined model world is the truth bearer
and the respective wirkende Kraft in the real world is its truth maker. Whatever else the model contains, and
whatever else modifies the manifestation of the Kraft in the real world does not participate in the truth making
of the model. This is von Thünen’s solution to the issue of how models can be true despite the false
assumptions they make (Mäki 2010).

14
Thus while such a model makes many false claims about the G, it strives to make a true claim
about the F. Such a model partially resembles the world, and hence is true in that partial sense.
Consequently, one must distinguish two truth concepts, where a model resembling its target in
every way expresses ‘the whole truth’, while a model resembling a target only in its intended
property is ‘nothing-but-true’.
Models that are nothing-but-true may contribute to our knowledge about the world, because they
teach us about the operation of a real factor in isolation. Models that are nothing-but-true
therefore may be legitimate endpoints of a scientific project, and do not necessarily require
further de-idealisation. In such a case, we have to distinguish between a quest for realism and a
quest for realisticness: a nothing-but-true model satisfies realism, as it says something true about
reality, but it may be highly unrealistic, because it idealises many other aspects of any concrete
situation.
Of course, such a nothing-but-true model need not be the endpoint of a modelling project. There
may be good reasons for de-idealising some of the assumptions about the intervening factors, for
example in order to see how the isolated factor operates in a more realistic environment. But this
does not affect the truth of a model – that can be had as soon as a model successfully isolates an
aspect of the real world. How one can determine whether a model is successful in this way is
obviously a difficult question that goes beyond the scope of this lecture. My point here was
instead to unfold the conceptual possibility how a model can be true even though it makes many
false assumptions about the world.

5. How should one build models?

Model building is a creative process. Its ultimate goal is to contribute to and advance scientific
knowledge. Thus the truth of models matters. Yet, as we saw, true models can be obtained in
different ways: by de-idealisation, by re-interpretation of assumptions, or by isolation. Which of
these strategies are chosen, how they are combined, and to what part of the model they are
applied, will depend both on the concrete problem to be solved and the ingenuity of the modeller.
For this reason alone, there can be no fixed scheme according to which model should be built –
too much depends on context, and too much depends on ingenuity. Model building, as one
philosopher quipped, is like “baking a cake without a recipe” (Boumans 2005).
While I will thus avoid presenting you with a recipe how to bake your model, a philosophical
discussion can at least point out important epistemic virtues that models should possess to some
extent. Truth, as I discussed, is one, and possibly the prime, of these virtues. But my preceding
discussion of truth and models already made clear that one might have to compromise on the
truth of a model, and make due with nothing-but-truth instead of the whole truth. The latter,
understood as the identity of all model and target properties, is not only never reached by
science, but would not even be desirable. Imagine one would seek such a massive de-idealisation
for the underground map if figure 4. It would render the resulting map utterly useless. Such a
map would not allow us to comfortably infer what train to take, where to change and where to
alight – and it would certainly not fit into our pocket!
Which is to say that although truth is an important virtue for models, other virtues like simplicity,
precision or generality also matter. The simpler a model, the more easily it can be understood,
manipulated, and its results analysed. The more precise a model, the finer the specifications of

15
the model description’s parameters. The more general a model, the more phenomena a model
relates to. This is by far not a comprehensive list of epistemic virtues of models, but you get the
point.
A particularly interesting thing about these different epistemic virtues is that some trade off with
others. That is, if one wants to increase one epistemic virtue of a particular model, one may only
be able to do that by inadvertently reducing another virtue of the same model. This has been
formally shown to hold between a particular kind of generality and precision (Matthewson &
Weisberg 2009): the number of all possible targets to which a set of models applies exhibits a
strict trade-off with the precision of models in that set. That is, the increase in the magnitude of
one virtue necessarily results in the decrease in the magnitude of the other.
These trade-offs characterise optimal model design. Once all the context-dependent constraints
are in, a well-informed and creative modeller must determine which combination of magnitudes
of different virtues is optimal for her purpose. Of course, this would first require specifying
which virtues should at all be considered relevant, and second determining potential trade-offs
between these virtues. Research on these two questions is still largely outstanding. But once that
is completed, we would be a lot closer to an actual recipe how to design models.
In the meantime, we will have to make due with a more piecemeal approach, specifying some
virtues and relying on the trade-offs we know about. For example, data modelling exhibits an
interesting trade-off between goodness-of-fit and simplicity. As you recall from figure 5, data
points can be fitted with different kinds of functions. The more complex the function, the higher
the goodness-of-fit (typically measured as the least-sums of squares, the differences between the
Y-value of a data point and the Y-value of the curve directly above or below it). If goodness-of-
fit were the sole criterion for truth of model, we would almost always prefer bumpier curves over
smoother ones.
Yet we know that measurements also contain error. Maximising goodness-of-fit would
incorporate all these measurement errors into the data model. Such perfectly-fitting curves would
overfit the data: because of the incorporated errors, they will perform poorly when making
predictions about new data sets. What therefore is needed is a method for separating trends in the
data from the random deviations from those trends generated by error. Philosophers and
statisticians have suggested that a measure of simplicity (measured as the number of the model’s
parameters) is the appropriate corrective to the tendency of goodness-of-fit to overfit the data.
The simpler the model, the less likely is a model to overfit the data. Of course, a model can also
be too simple: therefore, a numerical trade-off between goodness-of-fit and simplicity of the
model must be used (Forster and Sober 1994). Note that in this proposal, simplicity is not an
apriori extra-empirical standard, but an empirically-based consideration of which curves is
closest to the truth.
Another model virtue derived from our epistemic limitations is that of model robustness.
Specifically, a model is said to be derivational robustness if the same relevant model result
ensues, despite variation in some of the auxiliary assumptions. Take the following example:

16
 An influential two-region, two sector georgraphic economic model shows that distribution of the manufacturing
sector across the two locations is determined by the balance between centripetal and centrifugal forces. The
centripetal forces arise from a positive feedback process: the more firms and workers there are in the region, the
more attractive the region becomes for further firms and workers (market-size effect). The centrifugal forces
arise from the need to serve the immobile factor, which remains equally distributed across the two regions, and
by competition effects: in the larger region, firms face tougher competition and higher input prices.
These results are very interesting, but they are shown only under the very specific assumptions of the CP
model. Following the initial proposal of the CP model in the early 1990s, one can therefore observe a whole
series of papers that investigate the model and its results under slightly varied assumptions. In particular,
following papers investigated model results under quadratic utilities, different transportation costs and with
both skilled and unskilled labour present. They found that the same model results ensue. They then argued that
this robustness result increased our confidence in the model being true. (Kuorkiski, Lehtinen & Marchionni
2010)

Scientists have a set of models, all of which are false, because all of them make idealizing
assumptions. Yet they also believe that some of the elements of some of these models are true –
in the sense in which a successfully isolating models is true of certain aspects of a phenomenon.
The modellers, however, do not know which of these models is true in this sense, and can’t find
out by testing against observation. If these conditions are satisfied, then robustness analysis can
strengthen the scientists varying degrees of confidence in truth of different assumptions.
Specifically, the scientist can learn that the models’ results do not crucially depend on the
models’ tractability assumptions. The robustness of the models’ results rationally increases her
degree of belief that the result depends on the substantial assumptions that she believed to be
true, not on the tractability assumptions that she believed to be false.
Seeking robustness of models thus is a strategy to cope with the inevitability of using idealising
assumptions in models. We often cannot get rid of them through de-idealising, arriving at a true
model. But we can sometimes control for effects of inevitable errors, introduced through
necessary (but variable) tractability assumptions, by checking whether the relevant model results
depend on them.

6. Summary
Models are representations of targets. This representation is constituted by a similarity between
some model and target properties, to some degree. Targets can be objects or events of the real
worlds (either phenomena or data), or mere possibilities, figments in the imagination of
modellers. Models may be material objects or processes, or abstract entities described by
functions, diagrams or axiomatic systems.
Models are used for many different things. They are used to present model targets to us, but also
to manipulate, to demonstrate and to analyse. Models also play an important mediating role
between theory and reality. Furthermore, models are used to explore possibilities, and often also
as heuristic devises that help produce better theories.
The truth of models matters for the success of scientific endeavours, but it is not simple to
determine when a model is true. The naïve view, according to which models are true, only if they
are identical to their targets in all properties is obviously absurd. Instead, three alternatives were
discussed: de-idealisation, re-interpretation of assumptions, and isolation.

17
Finally, models can satisfy many epistemic virtues, simplicity, generality, accuracy, precision
and robustness amongst them. Interestingly, there are trade-offs between at least some of these
virtues, which an optimal model design for a specific problem will have to take into account.

7. References
Achinstein, Peter (1968), Concepts of Science. A Philosophical Analysis. Baltimore: Johns Hopkins Press.
Axtell, R. L., Epstein, J. M., Dean, J. S., Gumerman, G.J., Swedlund, A. C., Harburger, J., Chakravarty, S.,
Hammond, R., Parker, J., & Parker, M. (2002). Population growth and collapse in a multiagent model of the
Kayenta Anasazi in Long House Valley. Proceedings of the National Academy of Sciences, 99 (3), 7275-7279.
Bogen, James and James Woodward (1988), “Saving the Phenomena,” Philosophical Review 97: 303-352.
Boumans, M.J. (2005). How Economists Model the World into Numbers. London and New York: Routledge.
Cartwright, N. 1983 How the Laws of Physics Lie (Oxford: Clarendon Press)
Forster, M. and Sober, E. (1994): “How to Tell when Simpler, More Unified, or Less Ad Hoc Theories will Provide
More Accurate Predictions.” British Journal for the Philosophy of Science 45: 1-36
Giere, RN 1988. Explaining Science. A Cognitive Approach, xxi+ 321 pp. Chicago, London: University of Chicago
Press.
Grüne-Yanoff, Till (2009) Learning from Minimal Economic Models, Erkenntnis 70(1): 81-99.
Hughes, R.I.G., (1997). Models and Representation. Philosophy of Science, 64, Supplement, Proceedings of the
Biennial Meeting of the Philosophy of Science Association, Part II: Symposia and Invited Papers, S325-S336.
Koopmans, T. (1957) Three Essays on The State of Economic Science McGraw Hill, New
Kuorikoski, Jaakko, Lehtinen, Aki & Marchionni, Caterina (2010): "Economic Modelling as Robustness Analysis",
British Journal for the Philosophy of Science 61, 541-567.
Mäki, U. (1992) On the Method of Isolation in Economics, in: C. Dilworth (ed.) Idealization IV: Intelligibility in
Science (Poznan Studies in the Philosophy of the Sciences and the Humanities 26, pp. 319-354).
Mäki, Uskali (2011) "Models and the locus of their truth", Synthese, 180, 47-63.
Matthewson & Weisberg (2009) The Structure of Tradeoffs in Model Building. Synthese 170 (1):169 - 190.
Musgrave, Alan (1981). 'Unreal assumptions' in economic theory: The F-twist untwisted, Kyklos. 34: 377–387
Parker, W. “Does Matter Really Matter? Computer Simulations, Experiments and Materiality”, Synthese, 2009.
Schelling, Thomas C. 2006 Strategies of Commitment and Other Essays. Cambridge, MA: Harvard University
Press.
Sugden, Robert (2000) Credible Worlds: The Status of Theoretical Models in Economics. Journal of Economic
Methodology 7(1): 169-201
Suppes 1962 “Models of Data”, in Ernest Nagel, Patrick Suppes and Alfred Tarski (eds.), Logic, Methodology and
Philosophy of Science: Proceedings of the 1960 International Congress. Stanford: Stanford University Press, 252-
261.
Wimsatt, W. C., (2007) Re-Engineering Philosophy for Limited Beings: piecewise approximations to reality.
Harvard University Press.

18