Richardson, George, "Concept Models", 2006 July 23-2006 July 27

Online content

Fullscreen
Concept Models

George P. Richardson’
Rockefeller College of Public Affairs and Policy
University at Albany, State University of New York

Abstract

Working with groups unfamiliar with system dynamics, modelers need a quick way to introduce the
iconography of the approach and some of its framing assumptions. In the early exploratory days
of group model building interventions at the University at Albany, we settled on the use of
sequences of tiny models for this purpose, which we call “concept” models. The intent is to begin
with a sequence of simulatable pictures so simple and self-explanatory, in the domain and language
of the group’s problem, that the group is quickly and naturally drawn into the system dynamics
approach.

Previous papers have sketched in passing the notion of concept models as we have used them.

Here we provide a number of illustrative examples and describe in detail the ways we use these little
models, the assumptions behind them, some design principles that have matured over time as our
experience has grown, and a discussion of possible problems with the approach.

Background

Working with groups unfamiliar with system dynamics, modelers need a quick way to introduce the
iconography of the approach and some of its framing assumptions. In the early exploratory days
of group model building interventions at the University at Albany, we settled on the use of
sequences of tiny models for this purpose, which we call “concept models.” The term reflects the
conceptual nature of these little models in two senses. The models introduce concepts,
iconography, and points of view of the system dynamics approach. In addition, the models are
designed to try to approach the group’s own concepts of its problem in its systemic context.

Our intent is to begin with a sequence of simulatable pictures so simple and self-explanatory, in the
domain and language of the group’s problem, that the group is quickly and naturally drawn into the
system dynamics approach. Within thirty minutes or less, we'd like to working with the group on

' Although this paper happens to be the work of a single author, who was the author of the concept models described
in it, it is essential to understand that projects in which these models were used were highly integrated team efforts.
The welfare reform work in which the first concept model discussed below appears was initiated by John Rohrbaugh,
facilitated by David Andersen, and supported significantly by Tsuey Ping Lee and Aldo Zagonel. The second concept
model discussed in this paper appeared in two workshops organized by Jose Gonzales, led and facilitated by David
Andersen, and supported extensively by Eliot Rich. The third concept model appeared in an intervention led and
facilitated by David Andersen in which Charles Finn contributed significantly with stakeholder analyses and other
facilitation support. I am grateful to these colleagues for the many ways they contributed in the background to this
paper and to the ideas it contains.

Richardson, Concept Models Page 1
their problem in their terms, listening hard to what they have to say, facilitating their conversations,
and structuring their views of the problem.

Previous papers have sketched in passing our notions and use of concept models (Richardson et al.
1992, Richardson and Andersen 1995, Andersen and Richardson 1997). Here we present a number
of examples illustrating a range of possibilities and describe in detail the assumptions behind the
idea of concept models, some design principles that have matured over time as our experience has
grown, and a discussion of possible problems with the approach.*

An Archetypical Example from Welfare Reform

In the late 1990s, the United States Congress passed sweeping welfare reform legislation that
changed the Federal government’s role and responsibilities in providing support for the nation’s
poor. The legislation put an end to programs that began in the depths of the Great Depression
providing the potential of lifetime federal support for an indigent family, and replaced it with
Temporary Assistance to Needy Families. A needy family now has up to five years of Federal
TANF support, cumulative over the lifetime of the parent. The rest, if any, would have to come from
states and counties. The frightening prospect that families could begin timing out of Federal
support as early as 2002 caused a number of New York State counties to want a systems view over
time of likely or potential scenarios (Zagonel et al. 2004).

We began the group modeling phase of the project with a sequence of three concept models, which
we will develop here in detail to show the process. The sequence is chosen
* to introduce the iconography of stocks, flows and feedback loops,
* to initiate understandings that dynamic behavior is a consequence of such system structure,
* and to motivate discussion on the group’s problem.
We began with the structure shown in Figure 1.

On At risk
. ————
assistance Job finding employed

rate
Figure 1: First stage of the Welfare Reform concept model sequence.

The picture was hand-drawn on a white board (we think that is important), first the labeled boxes,
then the flow. In explaining what the boxes and the pipe mean, we usually draw an old-fashioned
tub showing wavy water inside, with a faucet pouring water in and a drain pouring water out,
spilling out over the floor in concentric ripples. People in these boxes, we explain, are being

* This work falls within the growing body of work on group model building in the system dynamics community.
The system dynamics bibliography (version 2006a) lists at least 64 works referring to group model building. None
focus specifically on models and methods of the sort ed here. The general works in the field include Vennix
(1996), Vennix et al. (1997), and Zagonel (2002, 2004).

Richardson, Concept Models Page 2
thought of as flows and pools of water, rather than as individuals.

Acknowledging that the picture is inadequate because it leaves out the possible flow back to needing
assistance, we added that recidivism flow to the picture, as in Figure 2.

iv

7
Job loss rate

On At risk
assistance] toh finding employed

rate
Figure 2: Second stage of the first Welfare Reform concept model.

We then argue that we'd like to create some simple algebra for the two flows so that we can move
toward quantifying this picture. That motivates drawing, one at a time, the structures determining
the two flows showing in Figure 3. We take pains to be sure that the group sees there is no magic
here, just simple arithmetic, with the diagram showing the quantities that come together to determine
each flow. We don’t write the equations, just state them verbally to keep the tone conversational;
we do, however, note that arithmetic guarantees that the units of both flows are “people per year,”
as units appear to be helpful to everyone’s understanding.

vd

5
Job loss rate *—
~y Fraction losing job

r year
On At risk perv!
assistance] to, finding employed

rate

Avg length of stay
on assistance

Figure 3: Final stage in the first Welfare Reform concept model, showing the structure of the
algebra determining the flows.

At this point, we turn on the LCD projector and display the model in its modeling environment (we
use Vensim, but began years ago with STELLA/iThink). It is crucial at this moment for everyone to
see that the software picture is identical to the hand-drawn picture. We then simulate the model in a
carefully chosen scenario designed to connect to the group’s problem. In this case, we begin the
simulation in equilibrium and step down the Average Length of Stay on Assistance, motivating that
scenario as a likely or intended effect of the welfare reform legislation. This result is show in

Richardson, Concept Models Page 3
Figure 4.

At Risk Populations
4,000
3,000
2,000
1,000
0
0 2 4 6 8
Time (Year)

On assistance : welfare!
At risk employed : welfare!
Unemployed and unassisted : welfare!

Figure 4: Dynamics of the structure in Figure 3 disturbed from equilibrium by a step down in the
Average Length of Stay on Assistance.

The result is portrayed as reasonable to the group, probably not requiring simulation to see. With
families spending on average less time on assistance, more are finding jobs. The On Assistance
population declines over time, and the Employed (though still at risk) population rises.

At this point the group has seen an example of the diagrams we will use and has some
understanding of the icons, and the use of simulation to play out the implications of the model’s
assumptions. But they would see little need here for simulation, as they could probably predict the
model’s behavior from simply simulating their mental model of this structure.

The next step is crucially important. We return to the hand-drawn picture on the white board and
modify it, ostensibly to make it more realistic, closer to elements in the group’s problem. In this
case we added a ratio of the At Risk Employed to a fixed number of Jobs available, and an influence
of that saturation ratio on the fraction of At Risk Employed losing their jobs per year. That
influence converts the Fraction Losing Job per Year from a constant in the first model, to a variable.

Richardson, Concept Models Page 4
Job loss rate*—J

Fraction losing job

er year
On rs ; At risk P

assistance} toh finding employed

rate

Employment
Avg length of stay ratio
on assistance
Jobs

Figure 5: The second Welfare Reform concept model, building on the structure of the first by
adding job availability.

Turning again to the computer, this second model is displayed, and we make certain the group sees
it is the same as the hand-drawn version. Simulation produces the behavior shown in Figure 6.

At Risk Populations

0 2 4 6 8
Time (Year)

On assistance : welfare2
Atrisk employed : welfare2
Unemployed and unassisted : welfare2

Figure 6: Behavior of the second Welfare Reform concept model, showing that the change in
structure has alternated the behavior of the model somewhat

Participants notice, and we emphasize, that the dynamic behavior is somewhat different. The At
Risk Employed still increase after the simulated initiation of welfare reform (the step decrease in
Average Length of Stay on Assistance), but they then decline somewhat, while the At Risk
Employed begin to increase. The difference is slight, but significant; it paves the way for the
remarkable difference the next structural change brings.

Returning to the hand-drawn model for the final time, we observe that the model does not include
the crucial element of the welfare reform legislation, that families have a limited eligibility for

Federal assistance. So we draw in the structure shown in Figure 7, explaining as we go.

Richardson, Concept Models Page 5
iv

ry
Job loss rate.
v Fraction losing job

per year

On x At risk
assistance | Job finding Lemployed

_ rate a
Loss of AF 4

assistance rate Employment
v Avg length of stay ratio
on assistance
Unemployed and
unassisted Jobs

Avg total time on
Fraction losing assistance
assistance per year

Figure 7: The final Welfare Reform concept model, showing the addition of structure for families
timing out of the system and becoming ineligible for Federal assistance

Once again returning to the computer to simulate the model, participants see the dramatic results
shown in Figure 8.

At Risk Populations

0 2 4 6 8
Time (Year)

On assistance : welfare3
At risk employed : welfare3
‘Unemployed and unassisted : welfare3

Figure 8: Dynamic behavior of the third Welfare Reform concept model, sowing the rise of an
Unemployed and Unassisted population and a return of the At Risk Employed population to its
original value.

In this scenario, the On Assistance population declines as before, and as intended by the welfare
reform legislation, but the At Risk Employed population experiences only a temporary rise,
eventually declining to something like its original value. Now the new population in the model, the
Unemployed and Unassisted, begins to rise after the initiation of simulated welfare reform (the step
change in Average Length of Stay on Assistance).

Richardson, Concept Models Page 6
This sequence of concept models shows just the sort of behavior with which a modeling team
would love to start a group discussion. The behavior of the first model is the expressed intent of
the welfare reform legislation: get people off assistance and into jobs. The behavior of the second
model starts to cast a little doubt on that scenario, and the third completely contradicts it. The third
model looks like a nightmare of advocates for the poor: the legislation, this third scenario suggests,
will not improve people’s employment prospects, but rather will simply reduce the Federal
government’s financial responsibility and shift it to the states and counties. Needless to say, the
group was ready to talk about the structure and future of welfare as they saw it!

The sequence of concepts models here did its job extremely well.

¢ The iconography of system dynamics structure was introduced and became familiar;

* The use of simulation to play out behavior over time became clear;

¢ The dynamic behavior of the system, initially obvious, became less so with each structural
addition, making simulation seem more and more necessary;

* The behavioral changes were so dramatic as to move from one end of the political advocacy
spectrum to the other, creating irresistible motivations within the group to talk about what’s
wrong with these little pictures of what they know to be a very complex welfare support
system;

¢ The entire development, from white board drawings to the final simulation, took less than
thirty minutes.

Observations on the Archetypical Example

The concept models used in the welfare reform group modeling sessions were very simple, starting
with just two stocks (actually a first-order system since the stocks are not independent), and ending
up eventually with just three stocks and three flows. Three stocks is probably an optimal goal.
Creating such a compelling a sequence of concept models as this one with no more than three
stocks gets the conversation off to an informed and highly motivated start with a minimum chance
of confusing or losing people.

To all the experts in the room, and to those of us on the modeling team, these concept models were
wrong. Reflecting a long line of wise reflection in science in general and the system dynamics field
in particular, Sterman assures us that all models are wrong (Sterman 2000, pp. 846ff)’ , so we
should not be surprised that the welfare reform concept models are wrong. But these models are
wronger than most! And they are for some very good reasons.
¢ First, the models are tiny -- little imitations only hinting at the structural and dynamic
complexity of the Federal, state and local welfare system(s). When we begin a group

* The often-repeated statement to the effect that “All models are wrong, but some models are useful” appears as a
heading in a chapter of Box (1979). It seems quitely likely to me that there are earlier sources of this famous
shibboleth, but I am not aware of them.

Richardson, Concept Models Page7
modeling intervention, we know more than we put into the concept models that get the
discussions started. The models have to be tiny to serve our introductory purposes.

Second, the models are constructed and sequenced to get across very particular aspects of the
system dynamics approach, including the idea (initially obvious, then questionable, finally

repeatedly affirmed in practice) that dynamic behavior is a consequence of system structure.
Paying attention to getting those ideas across often necessarily draws attention away from the
real system and toward a representation of it that will serve our concept model purposes. In
particular, illustrating dramatically that behavior is a consequence of structure may require
less robust or unrealistic structural formulations such as open integrators or other omitted
feedback loops.

Thirdly, concept models should have simple, even obvious, algebraic structure; that
requirement of simplicity may force leaving out structure essential for coming close to
realism. One reason for requiring simple algebraic (or nonlinear graphical) equations is in
case a participant asks to see an actual equation. In our experience, that almost never happens
-- people seem content to get a gradual introduction and are anxious to get to the complexities
they came to discuss -- but the concept model should be ready to be investigated a bit,
nonetheless.

In this spirit of avoiding mathematical ugliness and striving for simple algebraic structure,
concept models should always embed graphical functions within a variable (as
STELLA/iThink always does and as Vensim can do). The focus should be as much as
possible not on software or mathematical elements, but rather on the structure and dynamics
of the group’s problem. It is expecially true for concept models that “mathematics should
be on tap, never on top.”

All these design principles can push in varying degrees toward formulations that one might not put
forward as best practice in the field.

A Recent Example
To illustrate variations and extensions on these themes, we turn now to a recent intervention about

cybersecurity. An energy client group was dealing with potential security risks on ocean oil
platforms as day-to-day operations were shifted to remote (onshore) internet-based electronic
control. There were two workshops, separated by a summer and involving a somewhat different
cast of characters each time. We needed concept models to start both group modeling building
workshops, but could not really start from scratch in the second workshop. Here’s the sequence of

three concept models that began the first workshop.

* Attributed to General Motors engineer Charlie Wilson in MacCorduck (1979, 388).

Richardson, Concept Models Page 8
Potential capacity to
transition

Speed of
transition
Capacity in —
Traditional Benapasiy
ops Transition to a ls

Integrated ops

Risk
Figure 9: The first concept model in group model building to study a transition to integrated,
remote e-operations on ocean oil platforms.

Following the procedures laid out above, the model was first drawn on a white board, one element at
a time, explaining the meaning of icons and quantities as we went. Turning to simulation, we
showed the model

Summary
100
0.2
50
oO.
0
0

0 12 24 36 a8 60

Time (Month)

Capacity in Traditional ops : transition
New capacity in Integrated ops : transition
Risk : transition!

Figure 10: Behavior of the first concept model in the oil platform study

The second concept model added feedback structure contributing to the speed of the transition to
new work processes in integrated e-operations.

Richardson, Concept Models Page 9
Synergies betw
potential and

experience Experience with
Potential capacity to \ ge transition
transition

Speed of
rf transition

sett New capacity in

Inte ited
ops Transition to HiEgrare? OPS

Integrated ops

4

Risk Cost per bbl

Figure 11: The second concept model in the oil platform study, showing an epidemic-like
reinforcing loop helping to speed the transition to integrated e-operations

Summary

300. kr/bbl

150. kr/bbl

0 kr/bbl

0 12 24 36 48 60
Time (Month)

Capacity in Traditional ops : transition
New capacity in Integrated ops : transition2
Risk : transi
Cost per bbl : transition2.§ ————_——— inibbl

Figure 12: Behavior of the second oil platform concept model, showing a dramatic shift from
simple goal-seeking behavior of the first model to s-shaped growth and decline.

This second model shows the declining production cost per barrel of oil, a variable of major interest
and the reason for migrating oil platforms to cheaper integrated remote e-operations. But it appears
to show that the risk in this transformation rises throughout the run. That observation leads to the
addition of a third stock in the system, which we labeled Mature Capacity in Integrated Ops, as
shown in Figures 13, 14, and 15.

Richardson, Concept Models Page 10
Synergies betw
potential and

SSRERISNCE: Experience with
Potential capacity transition
to transition
Speed of Time to fully
transition integrate ops

Capacity in 7 Mature
Traditional Neii-capacity Th capacity in
ops Transition to Tategeated ops Maturing Integrated ops
Integrated ops

tegrated ops

co

Risk Cost per bbl

Figure 13: Structure of the third and final concept model in the oil platform sequence showing
more complexity in the assumptions about Cost and Risk and the speed of the transition.

Summary

200 kr/bbl

100 kr/bbl

0. kr/bbl

o

10 20 30 40 50 60
Time (Month)

Capacity in Traditional ops : transition3
New capacity in Integrated ops : transition3

Mature capacity in Integrated ops : transition3
Risk : transition3
Cost per bbl : transition3 kr/bbl

Figure 14: Behavior of the third oil platform concept model, showing new behavior in New
Capacity (rising to a peak and declining). Risk follows the same pattern.

To facilitate comparisons of the behavior of these models, we showed comparative graphs of Risks
and Costs per Barrel shown in Figure 15. The comparison shows vividly that each structural
change brought about changes in the behavior of the system. One could argue, as we did when we
added bits of structure, that we were getting closer to something that looked like the actual nature of
this dynamic system, but no one was fooled into thinking this was an adequate representation of the

Richardson, Concept Models Page II
complexities of the transition the group needed to understand.

Risk comparison Cost comparison
o4 200
ors 150
0.05 100
0.025 50
6 o
° io 20 rT) ry 3 oo 0 io 30 3 ry 3 oo
Time (Month) Time (Month)
Risk: transition! Cost per bbl: tansitiont bb

Risk : transition? Cost per bbl : transition? Aeibb|
Risk : Transition3, Cost per bbl : transition}, <i

Figure 15: Comparisons of Risks and Costs per Barrel in the three oil platform concept models,
illustrating the principle that behavior is to some degree a consequence of structure.

Observations on this Recent Example

Modelers will observe that there must be some unusual things going on in these models focusing
on a transition to integrated e-operations on oil platforms. There appear to be some omitted
elements and unusual naming conventions. How, for example, is the variable Potential Capacity to
Transition formulated in the three models? How is the variable Experience with the Transition
formulated in the second and third models? And mathematically, what do we mean by the Speed of
Transition or Synergies Between Potential and Experience?

First, the modeler’s unvarnished truth:
¢ In all three models the variable Potential Capacity to transition was simply set equal to the
stock of Capacity in Traditional Ops.
¢ In the first model, the Speed of the Transition was a constant, the fraction of traditional
capacity transferring each month (0.03).
¢ In the second and third models, the Speed of the Transition was the product of Synergies
times Experience.
In the second and third models Synergies was now the constant, the fraction of Traditional
Capacity transferring each month (0.002, smaller now because Experience is in the product).
¢ Finally, in the second model Experience with Transition was simply set equal to New
Capacity in Integrated Ops, and in the third model Experience was set equal to the sum of the
new and mature capacities.
What was the modeler thinking about here?! Why were these decisions made?

Taking the bullet points in order, the first question can be interpreted as asking why the variable

Potential Capacity to Transition is in the model at all if it is just equal to the Traditional Capacity.
We felt including that wording told a story about making changes in traditional operations that the

Richardson, Concept Models Page 12
client group would identify with. Some operations could be turned into integrated e-operations and
maybe some could not. One might think a better way to formulate this story could be to set
Potential Capacity to Transition equal to a fraction of the traditional capacity, so as to leave some
traditional capacity as permanently not transformable. But this sort of formulation does not do
what it intends, and the proper formulation is a good deal more complex and not particularly
intuitive’ so we chose to leave it as we wrote it.

The second bullet could be interpreted as asking why we didn’t name the constant as the “Fraction
of Traditional Operations Transferred per month.” Why the “Speed of Transition”? We thought
that “speed of transition” communicated the real-world significance of the parameter, while a name
with “fraction per month” in it would focus attention on the mathematics rather than the real-world
significance. It’s a debatable point: Richmond held that variables should always be named with the
most descriptive mathematical terminology, to avoid any ambiguity in the meaning of the
formulation. But these models are concept models, so conceptual names seem more appropriate.

The same issue pops up with the constant called Synergies Between Potential and Experience. The
name we chose hides the fact that this variable is a fraction (0.002). But naming it as we did
communicates the concept that if there are more synergies between traditional operations and our
experience with the transition to new integrated e-operations, the transition should happen faster,
and we thought that to be the desirable focus. We also stayed away from naming the Synergies
parameter a fraction of some sort because, like all such parameters in this Bass model-like structure,
it has dreadfully complicated units (something like 1/capacity units/experience units/month) which
are not at all intuitive. We thought it better to leave the name at the intuitive level.

These observations lead to the disturbing modeler’s truth revealed above, that Experience was
simply set equal to capacities. Why not some converter constant times capacity? In fact, it could
have been formulated like that, and would make a modeler feel better if it had been so because the
units could be made to make sense. But then to keep the structure clean and simple looking, we
probably would have hidden the constant anyway. So we chose to simply leave it out. It appears
that in explaining these structures as they are developed on the board, the audience puts the missing
constant in mentally, or ignores it, or doesn’t know it should be there -- and in any of these cases
we do better leaving it out rather than putting it in and raising the ugly specter (in this particular
structure) of dimensional consistency.

These thoughts suggest the following principles about concept models used for group model
building:

* The “fraction” formulation does not leave something in the traditional stock; all would be transferred eventually to
capacity in new integrated operations, just at a lower pace. The formulation such thoughts really aim toward is a
fraction of the total of capacity that is transformalble, with the potential capacity to transition set equal to that total
transferable capcity minus the amount already transferred. The complexity of this paragraph illustrates why we chose
to sidestep all of these thoughts and just set the Potential Capacity to Transtion equal to the Traditional Capacity.

Richardson, Concept Models Page 13
* Concept models and the ways we use them with a group should minimize attention to the
software and model mathematics. Instead, our work with concept models should maximize
the group’s attention to concepts in the real system and the model as a representation of the
real system. That can mean using variable names that sound less like mathematics and more
like real system structure. It can sometimes even mean hiding or embedding some constants
if they aren’t going to be focused on or changed.

In this spirit of hiding unnecessary software or mathematical ugliness, concept models
should always embed graphical functions within a variable (as STELLA/iThink always does
and as Vensim can do).

Finally, it is a reasonable extension of this principle about minimizing attention to the

software and model mathematics to suggest that mouse moves on the computer screen should
be kept to a minimum. Pull down menus should be avoided if keyboard shortcuts exist.

The theme of these observations is that participants’ thoughts about concept models should not be
about mathematics or software, but about the system structure and behavior of their problem.

An Extended Example that Breaks More Rules and Suggests More Principles

Our final example (Griffen et al. 2000) comes from work done a number of years ago with a state
agency that spends its time and resources each a year equalizing tax burdens across school district
and county lines. The agency would be for the most part unnecessary if localities assessed all
properties every year and set tax rates based on full assessed value. Unfortunately, New York State
constitutionally can not require localities to assess and tax like that. Furthermore, localities are
under a wide variety of forces that tend to push them away from assessing every year, or even every
five or ten years, and to set tax rates based on some fraction of full assessed value. So the state
agency steps in to assure that people who own similar houses or business properties in different
taxing regions have essential the same tax burden.

We were brought in to get a systems view of this problem, working with people from the state tax
equalizing agency and its visionary leader, together with various tax and assessment professionals at
the county and local levels: a disparate group with multiple agendas. After beginning with a series
of strategic planning exercises to get the problem well-defined and the group ready to tackle it
(Griffen et al. 2000), we introduced, in the fashion described above for welfare reform, the concept
model shown in Figure 16.

The reader will note immediately that this first concept model looks more complex than the first
concept models for welfare reform and cybersecurity on oil platforms. There are five stocks and
flows, two more than past experience suggests is wise practice. But we reasoned that the concepts
in the structure would look very familiar to the participants and that familiarity would carry over to
make the entire diagram seem reasonable. Both assumptions proved correct. Also enabling this
structure to be the first system dynamics map and model the participants had every seen is the

Richardson, Concept Models Page 14
simple arithmetic governing the flows. The first-order formulations are obvious from the picture,
even for non-modelers. One does not need to see the equations to see that each flow is some
number of towns moving per year from one category to the next. Even the units (towns/year) are
discernible from the arrows pointing to the flows and made sense to the participants.

iva

Revaluing
towns far from

il reval

frowns recently| y Towns not too x frowns far from|

revalued 2 FH far trom reval a at

Slipping from Slipping far
recent reval from reval A
Moving towns Moving
fp continous B Dey ecatctant
reval ' towns to reval
Towne Lrowns seldom|
continuously

revaluin
revaluing s

Figure 16: Initial view of the local tax assessment concept model, drawn at first by hand, one stock
and flow at a time, showing pools of towns in five assessment categories and how they can move
between categories. [“Reval” is the state’s shorthand for reassessment.]

fraction revaluing per year

Fraction Fraction
moving tof = becoming
continuous Revaluing compliant
bares ot es orp
reval
frowns recently os Frowns not too = frowns far from|
revalued LPP fa: from reval sy eval

Slipping from Slipping far
recent reval from reval A
Moving
fl DAI recalcitrant
towns to reval
Vv time to slip from time to slip to
Towns recent reval far from reval
Frowns seldom
continuously
revaluing
revaluing e

<Total cost> <Average time since reval>

Figure 17: Second view of the local tax assessment concept model, showing the simple arithmetic
governing the movement of towns among the categories.

Hidden in Figure 17 is the fact that all the constants in the model were formulated as graphical
functions of time, as shown in Figure 18. That technique is extremely useful if, as in this case, the
modeling team wants to simulate different scenarios. The base simulation of the model was in
equilibrium. Simulating different scenarios, such as increasing the frequency of revaluations shown
in Figure 18, enables the model to begin in its original equilibrium and transition through the
simulated scenario. Participants would be puzzled to see the variable Time involved in each of these

Richardson, Concept Models Page 15
‘constants,’ so Time is hidden from their view.

cere tootup - tration revatung peryeur Eee

moun. vats xmin' [1980 Fx); sro sae gma Pod Ee) eee Seeirg

ck Gye Home | Cem Alfork | Ca seul | Sher Keleeenos | HelCur teres

Figure 18: The ‘constant’ Fraction Revaluing per Year formulated as a graphical function of time
to facilitate simulating different scenarios, in this case almost doubling the frequency of
revaluations.

However, noticeable in Figure 17 are two quantities disconnected in this view from the chain of
towns: Total Cost and Average Time Since Reval. These two quantities are computed in three other
screens in the concept model (see Figure 19 for costs, for example), which we alluded to in the
workshop but did not bother to show. Participants were content that the concept model could
compute from average values the total annual cost of equalizing tax burdens among towns with
different assessment policies and the cost of moving towns from one assessment category to
another, that is, the cost of assessments. Costs figured prominently in what we did with this
concept model, as we shall see below, and the average time since revaluation generated an
unanticipated insight among the participants, so we know that these concepts were easy for them.
Had we shown the arithmetic in Figure 18 to the participants we probably would have obfuscated
much more than we clarified, so we purposely hid these “spreadsheet” calculations.’

° The computation of the Average Time Since Reval should have involved a coflow structure to keep track of time as
towns aged through the system or reassessed. However, such a computation would have looked terrible to
participants if someone asked to see it, so we opted for a less accurate but adequate algebraic formulation involving
weighted averages of time constants.

Richardson, Concept Models Page 16
<Towns <Towns not seria <Towns <Towns

recently too far from We eas seldom continuously
revalued> reval> revaluing> revaluing>
Annual cost of ‘AauGa Gost ee HET ‘Annual cost of Annual cost of Annual cost of
recent towns foo towne: Yar hewn: seldom towns continuous

towns

f

Unit cost of
recent towns

Unit cost of Unit cost of
not too far

towns

Unit cost of

Unit cost of
seldom towns

continuous
towns

far towns

~

Total annual
cost

<Moving towns to <Moving
continuous reval> recalcitrant

<Revaluing
towns far from

reval>

towns to rev:

Unit cost of Unit cost of Unit revaluing,
moving to moving seldom costs
continuous towns -
Cost of moving Cost of moving Revaluing costs
to continuous seldom towns

reval to reval

<Total annual cost

-_

Total transition costs ——age Total cost

Figure 19: Computing costs (dollars per year) in the local tax assessment concept model, from the
annual costs of equalizing tax burdens across assessment categories (top of the figure) to the
costs of assessing towns and thereby changing their assessment categories (bottom of the figure).

Several things may be observed already with this concept model:
* It contains more than the usually recommended three stocks.
It contains a lot of ‘spreadsheet calculations’ that must be hidden from view or risk bowling
over participants with model complexity.
¢ It starts with the modelers’ view of the problem of this agency wished to talk about, at a level
of aggregation selected by the modelers, not the clients.
¢ It is apparently not designed to show changes in behavior as structure is added, but rather as
parameters (like the Fraction Revaluing per Year) are changed.
* By implication, the model is not intended to help build the insight that “behavior is a
consequence of structure.”
« A second implication is that participants would not get a sense of changing model
structure as a part of the concept model phase of the group model building process.
¢ A third implication is that this concept model could anchor, perhaps inappropriately,
the group’s perceptions and discussions. (More about that below.)

This concept model was designed as a tool for engaging participants in thinking about different

Richardson, Concept Models Page 17
revaluation scenarios and to track their costs. It may come as something of a surprise that the
process we used to get there began with simulating the model with deliberately and outrageously
fictitious cost parameters. The annual costs of towns in the various revaluation categories were (left
to right in the top of Figure 19): $.50, $2, $3, $5, and $.50 per year. The costs of moving towns
were (left to right at the bottom of Figure 19) $50, $30, and $15 per town.

The process we followed with the model was to develop it slowly with the group, stock by stock and
flow by flow. We then projected the model and simulated it, showing equilibrium behavior. Then
we illustrated a series of simulations in which we increased the Fraction Revaluing per Year, the
Fraction Becoming Compliant per Year, and the Fraction moving to Continuous Reval per Year.

We showed the graphs of various quantities, e.g., Figure 20.

Graph for Towns recently revalued

400

300 Ta
Le Na

200

100

0

1990 1994 1998 2002 2006 2010
Time (Year)

‘Towns recently revalued : Revall nirvana towns
‘Towns recently revalued : Revall contin reval towns
‘Towns recently revalued : Revall runs | & 2 towns
‘Towns recently revalued : Revall becoming compliant. ———_______________ towns
‘Towns recently revalued : Reval! faster reval towns
‘Towns recently revalued : Reval! base towns

Figure 20: Towns Recently Revalued under six different scenarios in the concept model with
fictitious cost parameters.

Although the cost parameters were fictitious (silly, actually), participants saw the logic of the
concept model and these scenario experiments. They were motivated to engage in the exercise we
were working toward, eliciting realistic values for all the parameters in the model, particularly the
cost figures. We had prepared a worksheet containing all the parameters and gave the group the
task of estimating and reporting realistic parameters. The director of the agency became intensely
interested and took a major role in getting the group to settle on consensus values.

The model was then revised in real time as the workshop turned to another task, with structure
added at the request of the participants to capture the breakdown of costs in the state share and the

local share, costs. Then we repeated the scenario simulations, this time asking the participants for

Richardson, Concept Models Page 18
the parameter changes they wanted to test (for more frequent revaluations and the like) and the
names of the scenarios they invented. Some of the results are shown in Figure 21.

Graph for Total cost

200M
175M a —
——>
125M [J
100M
1990 1995 2000 2005 2010
Time (Year)
Graph for Average time since reval
6 T
4.5
3
Ls =
0)
1990 1995 2000 2005 2010

Time (Year)

Figure 21: Total Cost and Average Time Since Reval shown in five difference scenarios suggested
by the participants using accurate cost and share parameters in the concept model.

The impact of these results were, at the same time, expected and stunning. The simulations showed
that attempts to improve the state/local assessment process did indeed shorten the average times
since revaluation, but always cost more, except in one simulation. If towns could be moved to a
process of “continuous revaluation” (actually, annual evaluations of statistical samples of
property), the cost would rise in the short run, as in all the other simulations, but eventually decline:
classic worse-before-better behavior. The concept model behavior reinforced what they had sensed,
that the only way to improve the system and eventually lower costs would be to move toward a
system of continuous revaluations. The very simple concept model, introduced initially with

obviously fictitious parameters became the medium for the disparate group of property assessment

professionals to come together around a substantive conclusion that formed the focus of their work
for the next several years.

There was an unexpected insight here, which the director of the agency continued to talk about
years after this intervention. We needed a representation in the model for the fairness or equity of
assessments. The extreme high level of aggregation of the concept model and the need to keep it
extremely simple for this gathering did not allow for the formulation of anything like a real measure

Richardson, Concept Models Page 19
of fairness or equity. So the modeler settled on the Average Time Since Revaluation as a stand-in
or proxy for that concept, formulated as a weighted average of the various times since revaluation of
the town categories in the model (see footnote 2). That approximation came to the participants,
particularly the director of the agency, as a great insight, which apparently gave them a handle or
perspective on equity that they had not had before.

That’s the good news. There was some bad news in this concept model and the ways it was used in
the workshop. Most prominently, it did not lead to the focused grouped model building effort that
we thought it would. The group essentially stopped with the concept model with accurate
parameters, when we thought that would have been just the beginning.

Our plan was to use the rich conversations around the concept model to build motivation for a
careful study of what keeps towns from revaluing properties frequently. What keeps some towns
from revaluing at all, maintaining housing assessed values around $8,000 for houses worth
$300,000 in the current housing market? We thought that the ease of changing the Fraction
Becoming Compliant (the fraction of towns seldom revaluing which decide to move each year to at
least occasional revaluation) would stand in stark contrast in their minds with the extreme difficulty
of bringing about that change in reality. Participants, we thought, would see this stark contrast and
be highly motivated to move to a careful systems view of the forces and dynamics of “recalcitrant
towns.” We were unable to make that move in the workshop.

The story contains hints of what we believe in retrospect to be the reasons behind our failure to
make the transition from these experiences with this concept model to a study focused on
recalcitrant towns:

First, like all concept models, the model here was “owned” by the modeling team, not by the
participants. But unlike our usual work with concept models, the model structure was never
changed in front of the group. The group never got the idea that they could structure the model as
they saw the world. We never gave them “permission,” and they never asked for it. Instead, the
group got the idea that their job was to make the parameters in a fixed model accurate, and the
model would be have accurately. The group owned the parameters, not the model.

Second, the model was “too good.” Certainly, from the point of view of a professional modeler,
the assessment model was trivially simple — linear, with exogenous time series for scenario
parameters, essentially open loop, no rich feedback structure, no compensating feedback for policy
initiatives, and so on. But the model looked right and behaved right to the participants, particularly
after good parameter estimates were introduced. Unlike our usual concept models, it did not have
rather glaring simplifications or inaccuracies. It did not cry out to be fixed. It did not provide the
drive toward rich give-and-take conversation among the participants, facilitated by the modeling
team, trying to get a systems view of the tough assessment problem. In fact, it may have looked like

Richardson, Concept Models Page 20
it “solved” the problems.

Third, the model became an anchor. The group had no interest in developing a rich dynamic
systems view of the structure recalcitrant towns and forces that keep them from revaluing. The
concept model told the story they wanted to focus on. It turned out to be a pretty good anchor, but
it was an anchor nonetheless, which inhibited the group’s ability and willingness to move beyond it.

Fourth, the model tended to reinforce an exogenous view of system behavior rather than an
endogenous view. The scenario parameters were captured as functions of time. While we still
believe that that modeling choice for the scenario paraters was just right here, there is no denying
the fact that the model’s behavior traces directly to changes in these scenario parameters. It
contained no interesting stock-and-flow / feedback structure generating intriguing, unexpected
behavior. One could say, perhaps, that it exhibited “fixes that fail,” since speeding the frequency
of revaluations only got at part of the problem and probably would help to mask the real problem of
uncompliant towns. But there were no shifts in loop dominance from nonlinearities, no real
compensating feedback to puzzle policy makers. The concept model itself, and the ways we used it,
inadvertantly reinforced the idea that change comes from outside the system.

We went on in this intervention to be of substantial help to this agency, so we overcame some of the
shortcomings listed here. Most significantly, the work helped the organization move toward its
director’s vision to try to change from being a watchdog agency keeping towns in line with state
procedures to being a consulting agency helping towns acquire the capabilities to move to
continuous revaluation. The concept model undoubtedly helped. But it did not lead in the exact
path we thought it would.

Summary and Conclusions

The preceding discussions of concept models in three widely differing interventions contain lessons
for their use in group model building, and perhaps lessons for other kinds of modeling with clients
as well. The most crucial lessons learned from our experiences with concept models include:
* Choose very carefully the scenario to simulate in a sequence of concept models. It must...
* be realistically related to the group’s problem,
* be simple (and obvious) to implement in the simulation model,
* not be “too good,” so plausible as to fail to stimulate conversation,
* and not weaken the endogenous point of view.
Use simple algebra or obvious nonlinear effects that are as close as possible to being
understandable or intuitable from the picture.
¢ If there must be a tradeoff between good modeling practice and understandable simplicity,
always choose the latter. The point is not a good model, but a good start to a conversation.

Richardson, Concept Models Page 21
Bibliography

Andersen, D.F. and G.P. Richardson, 1997. Scripts for Group Model Building. System Dynamics
Review 13,2.

Box, G.E.P., 1979. Robustness in the strategy of scientific model building, in Robustness in
Statistics, R.L. Launer and G.N. Wilkinson, Editors. 1979, Academic Press: New York.

Griffen, T., D. Andersen, G.P. Richardson & C. Finn, 2000. Using Group Model Building To
Support Strategic Reform Of The Real Property Tax System In New York State: A Case
Study. Paper presented at the 18th International Conference of the System Dynamics Society,
Bergen, Norway.

MacCorduck, P. 1979. Machines Who Think: A Personal Inquiry Into the History and Prospects
of Artificial Intelligence. San Francisco: W.H. Freeman.

Richardson, G.P. and D.F. Andersen, 1995. Teamwork in Group Modeling Building. System
Dynamics Review 11,2 (summer 1995).

Richardson, G.P., D.F. Andersen, J. Rohrbaugh, and W. Steinhurst, 1992. Group Model Building.
Plenary presentation, Proceedings of the 1992 International System Dynamics Conference,
Utrecht, The Netherlands.

Vennix, J. A. M. (1996). Group Model Building : Facilitating Team Learning Using System
Dynamics. Chichester, Wiley.

Vennix, J. A. M., G. P. Richardson & D.F. Andersen, Eds., 1997. Group Model Building. System
Dynamics Review.

Zagonel do Santos, A. A. (2002). Model Conceptualization in Group Model Building : A Review of
the Literature Exploring the Tension Between Representing Reality and Negotiating a Social
Order. Proceedings of the 20th International Conference of the System Dynamics Society,
Palermo, Italy, The System Dynamics Society.

Zagonel, A.A., 2004. Reflecting on Group Model Building Used to Support Welfare Reform in
New York State. PhD dissertation, Rockfeller College of Public Affairs and Policy,
University at Albany, State University of New York.

Zagonel, A.A., J. W. Rohrbaugh, G.P. Richardson, and D.F. Andersen, 2004. Using simulation models
to address ‘What If’ questions about welfare reform. Journal of Policy Analysis and
Management 23,4 : 890-901.

Richardson, Concept Models Page 22

Metadata

Resource Type:
Document
Description:
Working with groups unfamiliar with system dynamics, modelers need a quick way to introduce the iconography of the approach and some of its framing assumptions. In the early exploratory days of group model building interventions at the University at Albany, we settled on the use of sequences of tiny models for this purpose, which we call “concept” models. The intent is to begin with a sequence of simulatable pictures so simple and self-explanatory, in the domain and language of the group’s problem, that the group is quickly and naturally drawn into the system dynamics approach. Previous papers have sketched in passing the notion of concept models as we have used them. Here we provide a number of illustrative examples and describe in detail the ways we use these little models, the assumptions behind them, some design principles that have matured over time as our experience has grown, and a discussion of possible problems with the approach.
Rights:
Date Uploaded:
December 31, 2019

Using these materials

Access:
The archives are open to the public and anyone is welcome to visit and view the collections.
Collection restrictions:
Access to this collection is unrestricted unless otherwide denoted.
Collection terms of access:
https://creativecommons.org/licenses/by/4.0/

Access options

Ask an Archivist

Ask a question or schedule an individualized meeting to discuss archival materials and potential research needs.

Schedule a Visit

Archival materials can be viewed in-person in our reading room. We recommend making an appointment to ensure materials are available when you arrive.