The endogenous perspective: a friendly amendment

By Jack Homer, VP of Professional Practice

I have developed lots of SD models over the years for both private and public organizations.  My models have all been built to answer strategic questions for which there was no single obvious answer, due to the presence of dynamic complexities including accumulations, delays, nonlinearities, and feedback loops.  These complexities cause the impacts of interventions to look different in the long term than in the short term, and different than a static or purely linear approach would suggest.  This is what sets SD models apart as my clients see it.

While my models all include these complexities, they may not always adhere to a traditional view of what an SD model should and should not contain.  The party line is that a model’s boundary should be broad enough so that the system’s main observed behaviors—such as S-shaped growth, oscillation, or overshoot and decline—are fully explained by the model’s endogenous structure.  One should avoid the use of exogenous time series drivers, because they undermine the ability of the model to explain and to anticipate change.

I mostly agree with this view but want to offer a friendly amendment here.  In my experience with real-world clients, I have often encountered situations in which it makes sense to employ exogenous time series for the sake of completeness and realism. 

For example, one model, done for the Gates Foundation, addressed policies to accelerate the decline of childhood malnutrition and stunting in the developing world.  This decline is primarily the result of rising GDP per capita, which I represented (mostly) exogenously.  The contribution of the model was not to explain the fundamental reason for the decline of childhood stunting (the general GDP effect was already known) but rather to identify its detailed mechanisms, show their natural limits over time due to nonlinearities, and quantify the further impacts across the aging chain as childhood stunting turns into adult infirmity and disability. 

A second recent example involves a model of the U.S. opioid epidemic done for a client interested in quantifying attributable cause and financial liability.  This model starts from prescriptions being written and wends its way through the dynamics of illicit use, addiction, and overdose.  Some of the characteristic growth behavior of the model is driven by exogenous time series inputs for prescription frequency and dose, as well as time series for the influx of deadly fentanyl illegally manufactured outside the country.  Yet, in this case, too, the use of the exogenous time series did not take away from the model’s ability to do the dynamic analysis that most mattered to the client.

My experience suggests that we should be less doctrinaire about the endogenous perspective and understand that “endogenous” is a relative thing.  No model can be all-encompassing and explain all observed behavior patterns.  That’s why we define a model relative to some subset of behaviors also known as the dynamic problem.  As long as the model adequately addresses the dynamic problem, it shouldn’t really matter if the model has some exogenous time series included to improve the model’s realism.

I would like academics to listen more closely to the experiences of practitioners as they think about how SD can make a contribution in the world.  They need not fear that modifying their traditional view of the fully-endogenous model is an invitation to poor quality.  It is possible for a model to be somewhat less comprehensively endogenous than that view might dictate, but still high quality and capable of addressing the dynamic problem.

4 thoughts on “The endogenous perspective: a friendly amendment

  1. I agree that “less endogenous” models can be very useful. Is there an example of doctrinaire academics opposing this idea?

    As a counterpoint, machine learning models are rapidly eating up the territory for open loop models; I’d argue that there’s an oversupply of that kind of thinking.


  2. Dear Jack, I find your note important, valid and useful. It is important to underline that endogenous perspective is not an absolute one. No model (‘real-world’ model) can be 100% endogenous, there are normally significant external inputs that influence model dynamics. Absolutely no problem here, given that the model dynamics are not entirely dictated by external inputs. So a better worded criterion is: the endogenous structure of the model must play significant role in producing its dynamics. The dynamics of the model is normally shaped by a combination of external forces and model structure. We then expect the modeler help us understand what components of model dynamics are primarily produced by external forces, and what components by the endogenous model structure. This is not just OK, but it is excellent SD. Our statements seem extreme when we emphasize ‘endogenous perspective’, but the main reason why we do it is to guard against increasingly exogenous prediction tools (data science, machine learning…) that seem to invade academic programs. It is a critical balancing act: it is foolish to expect absolute endogenous modeling, but it is important to emphasize that there is no SD left if one uses Stella and Vensim software like spreadheet software to build models that have no significant feedback loops, driven entirely by strong external inputs…
    Thanks for this important note! Yaman Barlas


  3. Hi Tom and Yaman, thanks for your good comments. IMO we shouldn’t be afraid of data analytics (incl. AI, data mining, etc.) and other statistical analysis, but should rather temper them with commonsense real-world causal modeling. (See another blog post, “First, look at the data”. Including the Ed Tufte quote: “Correlation isn’t causation, but is sure is a good hint.”) My own modeling practice is an interplay of data analysis and endogenous causal thinking, and analytics are welcome when they don’t violate laws of causation. My guiding light in modeling is not explaining “the dynamics” (Yaman’s term), but rather creating a robust tool for policy analysis, capable of anticipating all potentially relevant feedback and delayed effects. (This is why I urge academics to pay more attention to real-world practice.) My most influential models do this, even if they don’t necessarily generate cyclical or otherwise complex dynamics.
    That’s why I think it’s interesting to think about, and even model, political resistance to certain policies that could lead to difficulty in their implementation (the subject of my “Let’s consider policy feasibility” blog.) I think we should be primarily focused on modeling realistic policy impacts, including political resistance, if it is potentially significant and we know something about it already from past experience.


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.