REPLY Meaning of Stock/Level (SD6909)

[SDMAIL Ralph Levine]

Posted by  "Ralph Levine" <leviner at msu.edu>

We have been looking at this discussion about the use of soft variables 
in system dynamic models with intense interest.  As community 
 psychologists, we see many places where  the integration of soft, 
psychological variables with conserved variables that conform to the 
bathtub metaphor  makes sense.  The inclusion of soft variables can 
contribute to better understanding the dynamics underlying the problem 
at hand for both modelers and stakeholders.  System dynamics is 
primarily noted for its emphasis on feedback mechanisms that determine 
behavior.  Most system dynamic models deal with feedback associated with 
human actors, who perceive, make judgments, and take actions, at various 
parts of the system to control the flow and ebb of the material 
processes of interest, the levels or stocks, such as the amount of water 
in a reservoir,  the size of the inventory on hand, and the size of the 
population in an urban area.  All of those examples, are what we would 
call “energy/material” variables that are conserved.  On the other hand, 
feedback mechanisms, which are a very much part of system dynamics, are 
informational in nature.  As noted by Jay Forrester, in a number of 
places, e.g.  Principles of  a basic feedback loop is a circular path 
composed of a decision, action, a state variable, and information about 
the state variable. We have found in our modeling efforts that soft 
variables play a valuable role in elaborating the informational aspects 
of loop processes. 

A number of points dealing with whether or not one should represent soft 
variables as levels have already been given.  We are going to address 
other issues in this controversy, such as measuring soft variables, and 
the costs of not including soft variables in a model when relevant. 

Let’s address measuring soft variables. Some of the posts described soft 
variables as “intangibles.”  They may be intangible, unlike holding 
dollars in your hand, or putting your hand in a bathtub filled with 
water.  However, the notion of being intangible does not imply that soft 
variables cannot be quantified or measured.  Douglas Franco addressed 
the quantification issue in his example of quantifying trust.  We think 
he is right on target.  In our work, we frequently quantify soft 
variables by anchoring them alternatively  between 0.0 and 1.0, or 0.0 
and 100.  So, if we want to quantify the level of depression, 0.0 would 
be absolutely no degree of depression and 100 would be the situation 
maximum level of depression one can feel, which may or may not be 
attained. 

Although related, there is a difference between quantifying a soft 
variable and developing a operational measure of the variable.  We think 
that John Gunkler has a point in saying that we must specify how the 
variables are measured (operationalized).  Actually, the technology of 
measuring and scaling social and emotional variables is widely used by 
research and applied psychologists.  It has been around since the 
1920’s, if not before, and it is not a big deal, if you have the 
training.  The mathematical theory underlying psychological measurement 
is well known, and from a practical statistical point of view, there are 
a number of tools, like confirmatory factor analysis that can test for 
internal consistency and reliability of one’s scale.  Concerning the 
validity of one’s scale, it is a matter of using correlational analysis 
to correlate one’s scals with measures of other concepts that in theory 
should correlate with ones scale, if the scale actually measures what is 
was designed to measure.

In general, psychologist scales, based on Likert scaling technology, 
fall somewhere between ordinal and interval scales, not ratio scales.  
This is fine for most correlational analysis, but a bit of a problem for 
using such scales as data to be fit by the model.  If the soft variables 
in the model, like trust, are quantified as ratio scales, a perfectly 
good model could be rejected, because the data, based on interval 
scales, may not conform to trajectories that were generated by the 
model, assuming the soft variables would be measured on ratio scales.  
We have written about this problem before. Stevens, who developed 
methods for measuring psychological and social variable on ratio scales, 
also found a way to transform data that is measured on an interval scale 
to a ratio scale. 

Now let’s discuss some of the implications (and unintended consequences 
) of leaving out soft variables out of “quantitative” SD models.  Let’s 
take the suggestion that one should not include soft variables in one’s 
quantitative SD , simulation model.  One unintended consequence is that 
it may limit or eliminate the use of a coflow process that may be the 
core of one’s model. The problem being modeled may actually deal with 
the lowered quality of new hires, for example. The coflow process is a 
very important, helpful tool in SD modeling.  However, unfortunately 
 most of the attributes, such as “Total Effective Experience,” are soft 
variables.  The example of Total  Effective Experience was taken from 
John Sterman’s text, pages 497-212.  Would it be better for the field of 
system dynamics to encourage Dr. Sterman to cut out a good portion of 
his twelfth chapter, or at perhaps keep it as an abbreviated section on 
coflows as an example of what not to include in SD models.

Actually, in the classic coflow model, the attribute, as noted by Jim 
Hines, can be represented b y a first order  information delay, the 
smooth.  The smooth is not conserved, yet we believe that most system 
dynamicists would not want the smooth  to be eliminated in one’s  SD 
model.  Indeed, it is an important tool representing the dynamics of the 
stakeholder’s problems. Perhaps to get around Alan’s objections to 
including soft, nonconserved variables in SD models, one might continue 
to use practice of representing psychological processes  as smooth.  
Along the same line, one could also continue a very traditional way of 
representing soft variables by defining them as converters (auxiliaries) 
first, and then sending the value of the converter to an information 
delay.  An example of this can be found in Jay Forrester’s urban dynamic 
model. First, he created a soft variable, AMM, the Attractiveness for 
migration multiplier as a converter.  However, it takes time to perceive 
attractiveness, so he fed AMM into  a first order information delay, 
which he denoted as AMMP, the Attractiveness for migration multiplier 
perceived.  We also should note that AMMP had a significant dynamic 
impact on the stock of underemployed people. 

There is certainly more to say about going beyond the use of first order 
information delays to model soft variables. There is the decision 
whether to only include them in simulation models or only in CLD’s and 
stock and flow diagrams. Stock and flow diagrams convey more information 
than CLD’s, but operational SD models put you in a much better position. 
One can then gain a lot of information concerning  strength of various 
levers of change by performing a sensitivity analysis on the 
quantitative model. Such a model undoubtedly contains a number of 
nonlinearities, the effects of which may not be known without the 
ability of running the model. Also, one can learn  a lot about the 
functioning of individual loop processes, and in some cases, one can 
assess the dominance of particular loops with some software now being 
developed.

Moreover, a  model that includes the soft variables can be validated by 
performing numerous tests developed by Forrester and Senge, Barlas, 
etc.Errors in specifying the equations containing soft variables should 
be picked up this rigorous way of validating models. If one has time 
series data, one can also perform  Theil inequality statistics to 
ascertain the source of errors the model makes in fitting the data.  If 
those who feel that including soft variables in one’s model can lead to 
false conclusions, certainly these should be picked up by performing 
these tests. We also should mention the potential problems in using the 
qualitative model as a policy tool, even with the inclusion of soft 
variables.  In our opinion, insights gained by a qualitative SD model 
are more likely to be in error, because of the inability of validating 
the model, or CLD in a rigorous way and not  seeing the implications of 
nonlinear effects of delays, and loop structure over time.

Finally, a couple of years ago we started a Psychology Chapter for those 
who are interested in including psychological and social processes in 
their models. Undoubtedly, the issues raised by Allen McLucas should be 
brought up at our annual meeting in Athens this year.  It is a great 
opportunity to discuss and debate the role of soft variables in system 
dynamic models.

Respectfully,

Ralph Levine, Ph.D.
Posted by  "Ralph Levine" <leviner at msu.edu>
posting date  Fri, 11 Apr 2008 17:51:45 -0400