REPLY Definition of root cause (SD6827)

[SDMAIL Jack Harich]

Posted by  Jack Harich <register at thwink.org>

Jean-Jacques Laublé wrote:
> Hi Jack
> You write :
> <This is because the field still clings to phrases
> <like "dynamic hypothesis" of what is causing the problem instead of root
> <cause and feels that if a model can reproduce the symptoms, then it must
> <contain the root cause.
>
> I do not think that anybody pretends that reproducing the symptoms does
> guarantee anything.

Perhaps you came in late on this thread. One reply, from Bill Braun, was:

"Jack Harich asks about the definition of root cause. Broadly stated, it 
would be the structure that produces the reference mode up to the 
present and, if policies were left as is, would result in the "feared" 
mode in the future."

Another reply, from Jack Homer, said: "We call 'root cause' the dynamic 
hypothesis."

Also, Sterman writes that a key SD process step is: "Formulate a dynamic 
hypothesis that explains the dynamics as endogenous consequences of the 
feedback structure."


> The notion of dynamic hypothesis is not part of the SD paradigm but 
> belongs
> to the most known and taught method of building models.

Thanks. I think the concept of a dynamic hypothesis has been central to 
SD from the start.

Without using the term "dynamic hypothesis" Forrester wrote in Urban 
Dynamics, 1969, page 113 that: "The first step in modeling is to 
generate a model that creates the problem. Only if we understand the 
processes leading to the difficulties can we hope to restructure the 
system so that the internal processes lead in a different direction. If 
the model is to create the difficulties, it must contain all the 
interacting relationships necessary to lead the system into trouble. The 
troubles are not imposed on the system from outside the structure being 
modeled. The model will be a closed model which is not dependent for its 
inherent characteristic behavior on any variables transmitted across its 
boundary from the external world."

This says the model contains "all the interacting relationships 
necessary to lead the system into trouble" and "The troubles are not 
imposed on the system from outside." In other words, the model contains 
the underlying causes, which I'm calling the root causes. If the 
troubles are not imposed from the outside, then it follows that what 
imposes the troubles must be inside the model.

Unless I'm missing something, "The first step in modeling is to generate 
a model that creates the problem" seems to be the same as "The first 
step in modeling is to generate a model that reflects your dynamic 
hypothesis of what it is that is causing the problem."

Ergo, it follows that a dynamics hypothesis model is assumed to contain 
the root cause. My point is that particularly in difficult problems, it 
may not. Ability of a model to reproduce a system's behavior does not 
guarantee it contains the root causes. But far too often, modelers and 
model users assume otherwise.

This is a contentious point, so here's an example: Consider a model with 
a constant, such as the infectivity rate for a virus. If the problem 
deals with spread of a disease, and the virus is the disease, then the 
infectivity rate is one of several root causes. The solution may center 
on how to lower human infection rates to the point where an epidemic 
cannot occur. This would resolve the root cause by reducing the power of 
the virus to cause an epidemic. So in this example, the model appears to 
contain the root cause, and easily mimics system behavior in epidemics.

But along comes a creative thinker. She asks "Why is the infectivity 
rate for that virus so high? That's a 20th century virus. It appeared 
due to favorable evolutionary circumstances." Subsequent investigation 
shows that standard treatments were not knocking out the original virus 
fast enough. This gave it time to mutate in patients. This led to new 
more virulent strains. One of these was the one in the model. Here the 
root cause is the standard treatments. They were insufficient to prevent 
mutation. A second model could be build to include this. Or you could go 
deeper in the root cause analysis, and ask "Why did the social system 
allow this to happen?"

The point is that the first model reproduced the problem's symptoms, but 
did not contain the "true" root cause.

One might quibble that they didn't take the starting time of the model 
back far enough. If they did, they would have had to include treatment 
and mutation, and the model would have contained the true root cause. 
But if the problem being modeled was an epidemic, there is no apparent 
need to start the model long ago, when the virus was born. There is only 
the need to model the current epidemic under investigation.

Every constant, every equation in a model is a candidate for probing 
deeper. This can possibly lead to a bonafide root cause.

As another example, in Forrester's Urban Decay model, a reasonable root 
cause would have been related to the question: Why did America's urban 
managers and politicians allow the crisis to spiral out of control? 
There was plenty of early warning. Why were decision makers unable to 
solve the problem proactively? A possible answer, of course, is SD was 
not yet one of their problem solving tools. This root cause was not in 
the model, but yet the concept model produced excellent symptom behavior.

> Some authors do not use dynamic hypotheses while respecting strictly 
> the SD
> paradigm. For instance R.G. Coyle never mentions dynamic hypotheses and
> builds his models only from the principle of causality, not 
> necessarily root
> one because of the 'heap of tortoises'. He does not use either reference
> modes to build models but only as a way of comparing reality to the model
> behaviour once it is built.
Let's assume that the definition of "dynamics hypothesis" is a theory of 
what is causing a problem's symptoms, or more broadly, a theory of what 
is causing a system's behavior of interest. Then even though Coyle 
doesn't use the term, his models are each a dynamic hypothesis, because 
they are based on causes of system behavior.

The terminology doesn't matter to me, except for the fact that to 
converse efficiently, we need standard terms with standard definitions.

Jean-Jacques - Thanks for pointing these things out. For me these were 
very focusing observations.
Posted by  Jack Harich <register at thwink.org>
posting date  Wed, 19 Mar 2008 18:13:56 -0400