In the continuum of analytics—descriptive, diagnostic, predictive, and prescriptive—understanding cause and effect should be a core concept. Yet is often seems to be overlooked or treated as a tangential topic. Diagnostic analytics clearly depends on causal understanding. In a recent blog about prognostic analytics, Looking at the Future through Analytics: Predictive vs. Prognostic, I discussed the reality of predictive analytics operating without understanding why things happen. Prescriptive analytics goes a step beyond predictive to recommend and even automate actions that shape the future. It seems that three of the four common analytics types—diagnostic, predictive, and prescriptive—would get advantage from a healthy dose of causal analysis.

If we really want to understand cause and effect,
maybe it’s time to step away from the data.

Note that the term used here is causal analysis, not causal analytics. They are different but related things. Causal analysis seeks to find the reasons that things happen. Causal analytics explores data to explain why things happen. The two descriptions sound similar because the differences are subtle but important. Causal analysis seeks reasons—plural with the assumption that no effect is ever the result of a singular cause—and it doesn’t necessarily start with the data. Causal analytics begins with data exploration and extends to include data-driven experimentation. The point: If we really want to understand cause and effect, maybe it’s time to step away from the data and complement data-driven analytics with model-driven analytics. (See figure 1.)

 Figure 1. Data-Driven vs. Model-Driven Analytics

Okay, I know it’s weird for a lifelong data guy to say “step away from the data.” But I’m not saying walk away. Just step away initially; don’t start with data. Start with the questions—what are the reasons that things happen. Start exploring those questions with colleagues, peers, and domain subject experts using shared knowledge, beliefs, and experiences. Develop hypotheses about why things happen and represent those hypotheses as a causal model. Then turn to data and analytics to verify, validate, quantify, and refine the assertions of the model. If you apply the principles of systems thinking and the techniques of causal loop modeling, you’re sure to get better results from diagnostic, predictive, and prescriptive analytics.

Systems Theory

So, let’s look beyond analytics and think about systems. I don’t mean computer system here, although computer systems are one type to which systems theory can be applied. But it applies just as readily to human, organizational, and business systems.

Fundamental truths for all systems regardless of their type include the following assertions:

• A system is a collection of interacting parts.
• Behavior of any part is influenced by interaction with other parts.
• A system boundary defines the set of parts that comprise a system.
• A system may interact with things outside of its boundary.
• External interaction is less influential of system behavior than internal interaction.
• Behavior is understood by examining the entire system, not individual parts.

Systems Thinking – Applied Systems Theory

Systems thinking applies systems theory to create desired outcomes or change. It offers a unique approach to problem solving that views problems as part of an overall system. Traditional problem-solving approaches tend to focus on one or a few parts of a system, believing that changes to those parts offers a solution. The systems-thinking approach focuses less on the parts and more on interactions and influences among them as the core elements of solving problems.

Understanding of systems is achieved through identification, modeling, and analysis of relationships and interactions among the parts of a system – a distinctly different and more in-depth analysis than is possible with structural models of a system. Systems modeling is performed by representing the parts of a system and the interactions among those parts.

The most basic concept of systems theory is that a system is a collection of interacting things. I use the word thing to avoid the context-based connotations that might occur with terms such as entity, object, or component.

Things in a system are of many types. They may include (but are not limited to) entities that are familiar to data modelers, objects that are familiar to object-oriented systems analysts, and components as they are understood by software developers. Things in a business system encompass artifacts such as resources, capacities, limits, gaps, goals, desires, actions, results, plans, processes, rules, standards, and much more.

Influence is a behavioral characteristic of interaction. Interaction between two things in a system is directional – one thing has influence on another thing. System Behavior is important to understand why things happen in a system, and to predict what may happen in the future. Analysis of influences is the key to understanding system behavior.

Systems Thinking Models – Causal Loop Diagramming

Visually representing system behavior is widely practiced in systems thinking with a Causal Loop Diagram (CLD). Causal loop diagramming is a form of cause-and-effect modeling. The diagrams represent systems and their behaviors as a collection of nodes and links. Nodes represent the things in a system and links illustrate interactions and influences.

Influences are of two types – same direction and opposite direction. A same-direction influence means that the values of two things move in the same direction when change occurs: When employee morale goes up employee productivity goes up. When employee morale goes down employee productivity goes down. An opposite direction influence means that the values move in opposite directions: When employee stress increases employee productivity decreases. When employee stress decreases employee productivity increases. Figure 2 illustrates how these two examples are modeled. Note that a plus (+) indicates same direction and a minus (-) is used for opposite direction.

 Figure 2. Influence and Direction of Values

The diagramming techniques is called causal loop diagramming because real understanding comes from understanding the system as a whole. Cause-and-effect is typically not linear. It is circular with a sequence of influences producing a feedback loop. Loops are closed structures that represent a sequence of system interactions without a beginning or an end. A loop may contain any number of interactions greater than one. Feedback is a characteristic of loops in systems.

Feedback is a process by which the results of an activity or action are returned to the actor in a way that influences the behavior of that actor. Positive feedback occurs when the cumulative effect of all interactions in the loop is one of growth, amplification, or acceleration. Positive feedback loops are often called reinforcing loops. Negative feedback occurs when the cumulative effect of all of the interactions is stabilization or equilibrium. Negative feedback loops are also known as balancing loops or goal-seeking loops.

Figure 3 illustrates both kinds of feedback loops. Note that the kind of feedback loop – positive or negative – is indicated using a polarity symbol at the center of the loop. Polarity describes the positive or negative feedback property of a loop. Determining loop polarity is relatively easy. Simply count the number of subtractive interactions in the loop. An odd number indicates negative polarity, and an even number positive polarity.

 Figure 3. Positive and Negative Feedback Loops

Avoid the pitfalls of reading the arrows in causal models as process or flow. The arrows indicate influence, and only influence. Reading more into them leads to misunderstanding. The meaning (and all of the meaning) in the loop on the left is

• When performance bonuses increase employee productivity increases. When performance bonuses decrease employee productivity decreases.
• When employee productivity increases performance bonuses increase. When employee productivity decreases performance bonuses decrease.

This is a positive polarity loop, also known as a reinforcing loop. It reinforces a pattern of system behavior—either an upward spiral of growing productivity and bonuses or a downward spiral of shrinking productivity and bonuses.

The loop on the right is a negative polarity loop, also known as a balancing loop. It brings two opposing system components into balance. In this example, the opposing components are workload and workload capacity. The tension between the two is represented as capacity gap. This loop describes influences as:

• When workload increases capacity gap increases. When workload decreases capacity gap decreases.
• When capacity gap increases hiring increases. When capacity gap decreases hiring decreases.
• When hiring increases workforce capacity increases. When hiring decreases workforce capacity decreases.
• When workforce capacity increases capacity gap decreases. When workforce capacity decreases capacity gap increases.

Individual feedback loops are a step toward understanding cause and effect, but they only scratch the surface. It is often the interactions among loops that provide real insight into system behaviors by breaking down stovepipe views of the parts of a system. Figure 4 illustrates this principle with only one minor change to the diagrams shown in figure 3. The new model shows a connection between the two feedback loops. Finding these kinds of connections is the first step to developing a holistic view of a system.

 Figure 4. A System of Interacting Loops

In reality, a system consists of many loops and many interactions among those loops. It is that total system view that helps to achieve depth of understanding and real insight into the behaviors of complex systems. The intersection nodes – those that participate in two or more loops – are the core of system complexity, and they provide the greatest opportunity to discover side-effects, hidden influences, and unintended consequences.

Determining the boundaries of a system model can be challenging. Every system is a part of some larger system. For any node in a model, you can ask what influences it and create new nodes for the new influences. Then you can ask what influences those things, and so on endlessly. Therefore, it is possible to continue modeling infinitely. Stop modeling when you have acquired the knowledge and information that satisfies the purpose of the model. Stopping too quickly, however, brings the risk that you’ll overlook side effects and unintended consequences. Figure 5 illustrates the nature of this challenge.

 Figure 5. Seeking System Boundaries

Systems Thinking and Analytics

This article provides only a brief introduction to systems thinking, a subject that is deep, complex, and very much related to analytics. Only by understanding system dynamics can we really understand cause and effect, provide the most meaningful measures, and deliver analytics that are purposeful, insightful, and actionable. Sometimes that means measuring things, but more often it means measuring interactions and influences.

The discipline of systems thinking includes several archetypes – generic models that represent recurrent patterns in systems. The names of the archetypes are fascinating in themselves: accidental adversaries, fixes that fail, drifting goals, tragedy of the commons, etc. But even more interesting is the clear and certain relationship that exists between these archetypes and the patterns seen in time-series analysis.

The systems thinking approach also includes other modeling techniques. Causal loop diagrams illustrate influences. Another technique called stock-and-flow provides the means to quantify influences. Quantification enables simulation, and simulation is at the heart of “what if” analysis and an effective way to complement and enrich predictive analytics.

I will expand on these topics as parts two and three of a series of articles to describe a systems-thinking view of analytics.

Read - A Systems Thinking View of Business Analytics, Part II: Recurring Patterns in Systems