System Effects:
Complexity in Political and Social Life

Robert Jervis (Princeton University Press, 1997)

System Effects, by the distinguished political scientist Robert Jervis, is a fascinating exploration of the complications that are intrinsic to the understanding of socioeconomic phenomena, both in terms of explaining past events and predicting future ones. Jervis takes the reader on a remarkable and broad-ranging tour of political, ecological and biological examples where system-wide thinking, defined as considering the social and/or physical environment that connects the elements under study, is essential to the social science enterprise. For example, the effect of a treaty between two nations cannot be assessed without understanding the effects on other countries. Similarly, the effects of a change in government regulations cannot be understood without considering how those who are regulated will respond. Jervis makes a persuasive case that at an abstract level, system-wide thinking is essential in the social sciences.

Even the most technically oriented researcher in complex and related systems will find the book delightfully written and very suggestive. That being said, the book is far less successful in its efforts to link its rich historical discussions to the formal literatures on complexity, nonlinearity and the like. To the extent that the book is designed to introduce complex systems and related thinking into social science, it cannot be judged a success.

One problem for Jervis is that in a number of cases he exhibits some confusion in the discussion of the formal concepts with which he is concerned. For example, he makes several errors concerning the relationship between various properties of systems and nonlinearity. On p. 37 he claims that when a system that experiences one shock followed by the opposite shock exhibits a net effect, the system is nonlinear. This is mathematically false. A standard object in probability theory is the so-called moving average representation of a time series, which is the decomposition of the series (for example, the unemployment rate) into a linear weighted average of current and lagged shocks to the series. Unless the weights in this representation are all equal, which is generally not the case, the cancellation to which Jervis refers will not hold by definition.

Similarly, the presence of positive feedback effects does not logically entail, as is claimed (p. 146), that there are nonlinearities in a system. This is not to say that the examples that Jervis gives of nonlinear systems are actually linear systems, but the arguments made in support of nonlinearity are frequently incorrect. In other cases, Jervis, in adopting systems- wide metaphors, is really placing new labels on old bottles.

Many of the examples in System Effects are cases where government policies, because they ignore interdependences between social, economic and political actors, have proven to be counterproductive. To an economist, this is old hat. In a system of actors whose behaviors are linearly connected, it is possible for the direct effect of the change in a variable on one actor’s behavior to have the opposite sign as the equilibrium effect of the change, due to indirect effects. In linear economic models, this is known as the difference between structural and reduced form equations, where only the latter, when derived from the former, allow one to compute the full effect of a change in an exogenous variable on the equilibrium of a system. Indeed, the many cases of unintended consequences described by Jervis seem to result from a failure to consider the full range of causes of individual and group behavior, not because of interactions, emergence, nonlinearity, etc.

In other cases, there is a lack of clarity in the definitions of terms such as nonlinearity. Statistics provides a simple example. A standard statistical problem is the modeling of the probability of a binary outcome (e.g., go to war or remain at peace) as a function of a set of causal factors. Suppose that the probability of going to war depends on the linear combination of a set of factors. Since probabilities are bounded between zero and one, it is generally the case that the effect of this combination on the probability must be nonlinear. So, is this a nonlinear model, in that the sum of the factors has a nonlinear effect on the probability of the outcome of interest, or is it linear, because the different factors can be traded off at fixed rates with no effect on the net probability? (Models of this type are sometimes referred to as generalized linear models!)

More broadly, the problem is that one can construct models that from one perspective are linear and from another are not. In such cases, how does one determine what it is about nonlinearity that matters in explaining the phenomenon of interest? While this determination can be made through a careful consideration of the structure of a system (or a historical episode of interest), the analyses in System Effects are far too cursory to be persuasive.

My most serious concern is that it is unclear how complexity, nonlinearity, and related formal ideas enrich the sorts of analyses that Jervis conducts. Consider Jervis’s discussion of Vietnam, where he argues that the success of the US in conventional fighting had the unintended consequence of causing the North Vietnamese to choose guerrilla tactics and thereby win the war, providing a clear example of unintended consequences. What makes such an example persuasive is a deep examination of the history of the case in question. It is not made more persuasive by the ex post determination that aspects of the example are similar to some features of certain formal systems. Nor is it obvious that ex ante knowledge of such systems would have led to any differences in the analysis or interpretation of the historical episode in question.

In general, I am strongly skeptical that the wisdom required to develop careful historical arguments can be substantially augmented by a knowledge of the basic ideas of complexity theory. Indeed, there is a certain sense in which the use of mathematical tools conventionally associated with nonlinear systems can retard the understanding of social science phenomena. When utilizing these tools, mathematical tractability can frequently require one to make assumptions about human behavior that fail to reflect the cognitive power and purposeful nature of individuals (who after all, behave differently than the particles, atoms or species whose behavior nonlinear systems were generally desugned to explain).

Without strong cognitive foundations, the conclusions of complex interactive system models can be as misleading as non-system-based thinking is shown to be. This is, in my judgment, the main failure in the current use of various complex or agent-based systems to study social phenomena, namely, the failure of many implementations of such systems to reflect the cognitive strengths of human actors. I do not mean that formal social science models need to assume the complete rationality of some neoclassical economic models, but that irrationality is no substitute for a properly modeled bounded rationality.

Now, Jervis himself is certainly not guilty of this failing (except indirectly through his insouciant enthusiasm for these methods); much of the value of the narrative in System Effects is that it avoids making the sorts of unrealistic assumptions often required when using the various formal methods to which Jervis frequently refers. In my view, Jervis underestimates the extent to which the careful, albeit on mathematical, reasoning in which he engages can successfully explain patterns in social phenomena. In short, System Effects is fascinating both for the imagination and erudition it presents, as well as for its demonstration of how formal systems methods have yet to contribute much beyond metaphors to certain aspects of social and historical science.