Journal of Environment and Development 5, No. 2 (Spring 1996): 132-48
by Thomas Homer-Dixon

Summary

This paper shows that some commonly advocated methodological principles of modern political science are inappropriate for the study of complex ecological-political systems. It also provides conceptual tools for thinking about the causal roles of environmental and demographic factors, and it discusses various strategies for hypothesis and inference testing

Recent research has focused on the effects of demographic and environmental change on political stability.¹ This research suggests that rapid population growth and “environmental scarcities” – scarcities of renewable resources such as cropland, forests, and freshwater – can contribute to widespread violence and social conflict. The research also highlights important ontological and methodological issues that arise when studying highly interactive and complex ecological-political systems, and it shows that some commonly advocated methodological principles of modern political science are inappropriate for the study of such systems.

Contemporary North American political scientists often advocate a quasi-experimental method of hypothesis testing and causal inference modeled after the natural sciences. By this method, researchers ideally use broad theories of political behavior to generate hypotheses about causal relations between variables that interest them. These should be key or “critical” hypotheses that are both testable and linked directly to core concepts and laws within the more general theories. The researchers then test the hypotheses – and, in turn, the more general theory – against empirical data.² Of key importance, according to this method, is the choice of data. Data should provide for variation of both the hypothesized independent and dependent variables while allowing for control of all other potentially confounding variables.

The problem of data choice is particularly acute in the field of comparative politics. Researchers must often rely upon selected case studies – of specific countries, for example – to test their hypotheses, but the procedures they use to choose cases can be contentious. Recently, some commentators have focused criticism on the procedure of selecting “on the dependent variable,” in which cases are chosen that exhibit a particular value, or range of values, of the dependent variable. It is generally thought that this procedure gives biased estimates of the effect of the independent variable and cannot therefore be used to draw causal inferences or test hypotheses.³

However, other analysts have shown that selection of the cases on the dependent variable is the best testing procedure when the independent variable is hypothesized to be a necessary cause of the dependent variable.⁴ I show here that there are additional circumstances where selection on both the dependent and independent variables is warranted. Specifically, research on the links between environmental scarcity and social conflict is often aided by explicit selection of cases in which environmental scarcity and conflict both occur. This is so because the subject matter is extraordinarily complex: the systems under study are characterized by an immense number of unknown variables and unknown causal connections among these variables; by interactions, feedbacks, and nonlinear relationships; and by high sensitivity to small perturbations. Such complexities and uncertainties make it virtually impossible to choose cases that control for potentially confounding variables.

There is another characteristic of the environment-conflict research program worth noting. The program does not aim to determine the range of factors that explains the current value of the dependent variable (incidence of violent conflict); rather, it seeks to determine if a specific independent variable (environmental scarcity) can be an important cause of changes in the dependent variable.

This is not a goal generally thought to guide social-scientific inquiry. Usually, researchers want to explain or understand the current causes of certain types of social events. They are interested in the factors that currently influence the value of a specific dependent variable, let’s say Y. They therefore ask: What factors cause or explain changes in the value of Y? But researchers studying the links between environmental scarcity and conflict have a different goal. They are not interested in the whole range of factors that currently causes changes in the value of the dependent variable (conflict); instead, they want to know whether, and how, a hypothesized independent variable in particular (environmental scarcity) can cause conflict. Their key question is therefore different: Can variable X, in particular, cause changes in the value of variable Y? Their emphasis consequently shifts from explaining the current incidence of the dependent variable (Y) to understanding the current and potential causal role of a specific hypothesized independent variable (X) and to understanding the nature of the causal relationship between the two variables.⁵

This shift in focus is not uncommon. It is reasonable, for example, when two conditions hold: first, the value of a variable in a complex system is changing significantly, or is thought likely to change significantly in the future; and, second, researchers want to know if this change will affect other variables that interest them. In cases where the value of the hypothesized independent variable has changed little or only slowly in the past, standard statistical procedures will reveal little relationship between this variable and other variables. If researchers are interested in the potential causal role of the independent variable, they will need to study past cases selected specifically to accentuate its variance.

The above two conditions apply in environment-conflict research: evidence suggests that environmental scarcity is getting worse rapidly in many parts of the world, and the incidence of violent conflict around the world is of concern to many political science researchers. Therefore, these researchers might reasonably ask the following questions:

1. Can environmental scarcity contribute to violent conflict?
2. If yes, how can it contribute to conflict?
3. Is this contribution interesting?

Identifying “how” environmental scarcity can contribute to conflict – that is, answering question 2 – means identifying scarcity’s possible causal roles. Identifying its causal roles will help answer the third question; for example, environmental scarcity’s contribution to a given conflict will be interesting if it is identified as a powerful and independent cause.

Clearly, all three of the above questions can only be properly answered with empirical evidence derived from careful research. But two difficult and interrelated questions immediately arise. The first is ontological: What are environmental scarcity’s possible causal roles? For example, researchers commonly distinguish among three: environmental scarcity might be a trigger that releases accumulated nonenvironmental social pressures; an aggravator of already existing conflicts; or an underlying stressor that is causally distant yet powerful. The second question returns us to the methodological issue introduced above: How do we test our hypotheses about the current and potential causal role of environmental scarcity as a contributor to violent conflict? Researchers argue over the relative merits of quantitative and case-study analysis, over how many and what types of case studies should be used, and over the circumstances that require rejection or modification of hypotheses. I address these two questions below.

Identification of Causal Role

Debate about whether and how environmental scarcity contributes to conflict often centers on the specific causal role of this factor.⁶ There are two useful ways of thinking about this issue: we can focus on how environmental scarcity influences rational actors; or we can focus on the nature of the hypothesized relationship between the cause (environmental scarcity) and its effect (conflict). Although not mutually exclusive, I discuss these two approaches separately.

The Rational Actor Approach

David Dessler provides an account of the rational actor model.⁷ Figure 1 shows that environmental scarcity influences the decision-making unit (DMU), which might be an individual, a group, or an organization, and the DMU in turn chooses to act in such a way as to produce a social effect. The DMU might choose, for example, to migrate, to change its resource consumption behavior, or to attack another group to obtain more resources.

Dessler dissaggregates the DMU and its environment into four components, each of which influences the ultimate choice that the DMU makes. First, the DMU confronts an “opportunity structure,” which is an external and objective set of constraints and pressures that determines its set of feasible actions. Second, the DMU’s cognitive processes and circumstantial factors in the DMU’s environment will influence the perceptual salience of opportunities and obstacles in its environment. Third, the DMU has certain relevant beliefs about the causal consequences of its various possible actions. And fourth, the DMU has preferences regarding the various outcomes that it believes will arise from its actions.

Dessler argues that environmental factors can affect all of these components, either singly or in combination. Environmental scarcity can obviously influence the set of feasible actions: severe land degradation, for example, might close off certain agricultural and economic options for the DMU. Environmental scarcity can also influence the salience of perceived options or obstacles: a sudden drought might make obvious the long-term deterioration of the agricultural economy. Dessler argues that environmental factors can also have multiple and complex influences on the DMU’s beliefs and preferences. In Dessler’s opinion, attention to these various effects of environmental factors on decision makers allows nuanced analysis of the causal role that the environment plays in contributing to conflict.

The Causal Relationship Approach

The other approach to analyzing the causal role of environmental factors focuses on the nature of the relationship between the cause and its effect. I discuss here the following seven variables that can be used to characterize this causal relationship: necessity, strength, proximity, exogeneity, multicausality, interactivity, and nonlinearity.

Necessity is a dichotomous variable: something is either a necessary cause of a given type of event, or it is not. Environmental scarcity is clearly not a necessary cause of violent conflict, since much violence occurs in situations of resource abundance. Unlike necessity, the strength of a cause can vary along a continuum or scale, from weak to “sufficient.”⁸ Causal proximity can similarly vary along a scale from distant to proximate. We commonly think of proximity in terms of causal distance in time or space. But proximity is really a function of the number of intervening causal steps or variables between the cause and its effect; the larger the number of intervening variables, the lower the causal proximity.⁹ The characteristics of proximity and causal strength are sometimes conflated, since a distant cause is often assumed to be weak. But intervening variables do not necessarily weaken the link between a cause and its effect.

The causal independence of a variable, or its exogeneity, can also vary along a scale from fully endogenous to fully exogenous. Many analysts assume that environmental scarcity is no more than a fully endogenous intervening variable linking political, economic, and social factors to conflict (see the first diagram in Figure 2). By this view, environmental scarcity may be an important indicator that political and economic development has gone awry, but it does not merit, in and of itself, intensive research and policy attention. Instead, we should devote our resources to the more fundamental political and economic factors.

But there are three reasons why this view is not entirely correct (as illustrated in the second and third diagrams in Figure 2). First, environmental scarcity can be an important force behind changes in the politics and economics governing resource use. Scarcity can cause powerful actors to strengthen in their favor an inequitable distribution of resources. Second, ecosystem vulnerability is often an important variable contributing to environmental scarcity, and this vulnerability is, at least in part, an external physical factor that is not a function of human social institutions or behavior. Third, in many parts of the world environmental degradation has crossed a threshold of irreversibility. Even if enlightened social change removes the original political, economic, and cultural causes of the degradation, it will be a continuing burden on society. Once irreversible, in other words, environmental degradation becomes an exogenous variable.¹⁰

The degree of multicausality of the processes producing social conflict also varies. If environmental scarcity contributes to conflict, it almost always operates with other political, economic, and cultural causes. Analysts who are skeptical about environmental scarcity as a cause of conflict often conflate the characteristics of multicausality and causal strength by assuming that if many factors are involved, each must be relatively weak.¹¹

Interactivity is a dichotomous variable: the relationship between two causes of an event can be either interactive or additive. Interaction is a common feature of environmental-social systems. In an interactive system of causes of a specific social event, none of the causes is sufficient, but all are necessary; thus causal strength and interactivity are linked, because no single cause can produce the event itself. But beyond this statement, it is meaningless to claim that a given cause in an interactive system is stronger – or should be given more weight in the analysis – than another.¹²

Finally, the degree of nonlinearity of the mathematical function describing the relation between a cause and an effect can vary from high to low. A system with highly nonlinear functions can exhibit unanticipated “threshold effects” and chaotic behavior in response to small perturbations. This is a key characteristic of many environmental-social systems.

Academic and lay discussions of environment-conflict linkages are usually larded with imprecise causal verbs such as “aggravate,” “amplify,” and “trigger.” These fuzzy, “folk” concepts are useful in everyday explanations of physical and social events, but they are not always helpful for research. We can clarify them a bit, however, using the above distinctions.

A claim that an environmental factor “amplifies” the effect of other causes of conflict implies that the factor interacts with the other causes to multiply their impact. In contrast, a claim that the factor “aggravates” the impact of the other causes seems to suggest that the factor’s effect is added to that of others. A “trigger” of conflict is always a proximate cause, and usually an unnecessary and insufficient one too. The term also implies that the system responds nonlinearly to the factor in question: that is, the factor triggers a disproportionately large response by pushing the system beyond a critical threshold. Stochastic and extreme environmental events – such as cyclones, floods, and droughts – can be important triggers of conflict. They can provide challenger groups with opportunities for action against a state whose buffering capacity has been gradually eroded by civil war, corruption, economic mismanagement, rapid population growth, or deteriorating stocks of renewable resources.¹³

“Aggravator,” “amplifier,” and “trigger” models are popular with skeptics, because they seem to relegate environmental scarcities to the status of secondary causes of conflict. While these models are often valuable, they offer inaccurate and incomplete explanations of interesting cases. Research shows that environmental stresses can be important contributors to conflict even if causally distant and even if the system is interactive and highly complex.¹⁴

Hypothesis Testing

All empirical research must begin with hypotheses. These often take the form of if-then statements about causal relations, or, at least, correlations between types of events.¹⁵ The “if” in the if-then statement identifies the independent variable, while the “then” identifies the dependent variable. The “if” also states any “scope conditions,” which are additional circumstances, perhaps intervening between the independent and dependent variables, that must be true if the whole if-then statement is to be valid.

Development of hypotheses is not a simple process. Researchers usually start with very simple causal or correlational hypotheses, perhaps – but not necessarily – derived from a general theory. They use each hypothesis to interrogate available evidence by asking the question: What does the evidence say about the hypothesized correlation or causal process? Evidence that flatly contradicts the hypothesis – often called a “null” finding – is valuable, but other kinds of evidence are valuable too, including evidence that supports the hypothesis and evidence that is equivocal but that suggests alteration of category boundaries, introduction of intervening variables, or addition of new causal linkages. On the basis of all this evidence, not just of “null” findings, researchers refine their hypotheses as their work progresses.

Over time, the boundaries of the independent and dependent variables are more precisely defined (which often involves generating additional categories of these variables), and the understanding of the scope conditions becomes more textured. This process is neither purely deductive (from hypotheses to evidence) nor inductive (from evidence to hypotheses), but rather an iterative cycle between increasingly sophisticated hypotheses and an increasingly comprehended empirical world.¹⁶ If, eventually, the hypotheses become sufficiently refined, and if they are linked by a definable set of binding assumptions or concepts, then we can reasonably speak of a “theory” that explains the set of events under study.

In environment-conflict research, a number of methods are available to test hypotheses against empirical evidence. Three deserve close attention, of which the first two are conventional, quasi-experimental methods in political science. First, researchers can undertake a correlational analysis of large amounts of quantitative data on the relative frequencies of environmental scarcity and conflict across many societies and over time. Such an approach involves statistical estimation of the probability of obtaining a given correlation observed in the data if, in actuality, there is no correlation in the real world between the variables in question.¹⁷ Second, researchers can undertake a controlled-case comparison, in which cases are selected that vary on the independent variable, environmental scarcity, but that are essentially the same for all other variables that might affect the incidence of conflict. Researchers aim to select cases that control for all variables except environmental scarcity so that scarcity’s effect on conflict can be isolated.¹⁸ If sufficiently similar cases are not available, researchers can instead undertake thought experiments using counterfactual analysis in which researchers ask: What would have happened if the independent variable changed its value but all other factors remained constant?¹⁹

Finally, researchers can undertake process tracing of the causal processes in a selection of cases where environmental scarcity apparently contributes to conflict. Here, in violation of the strict canons of conventional political science, cases are selected explicitly on both the independent and the dependent variables. The aim is to determine if the independent and dependent variables are causally linked and, if they are, to induce from a close study of many such cases the common patterns of causality and the key intermediate variables that characterize these links.²⁰ Process tracing often involves dropping down one or more levels of analysis to develop a more finely textured and detailed understanding of the causal steps between the independent and the dependent variables.²¹ In process tracing, George and McKeown write:

. . . the process of constructing an explanation is much like the construction of a web or network. The researcher assembles bits and pieces of evidence into a pattern; whether a piece is to be changed or added depends on whether the change fits with what already has been constructed, and whether it strengthens the web’s structure. Does the modification of the explanation create internal inconsistencies in the theory? Does the modification of the explanation create more new puzzles than it solves? If yes is the answer to these questions, the modification is rejected. Modifications that are consistent and produce smaller, more localized, and less frequent research puzzles are to be valued. The growth of the web orients the search for new pieces, just as the growth of a jig-saw puzzle guides the search for pieces that will fit together with what is already assembled.²²

A central claim of this paper is that the stage of research strongly influences the method of hypothesis testing a researcher can use to best advantage. During early research in a new field, especially if the subject matter is highly complex, hypotheses are liable to be too crude to support testing that involves quantitative analysis of large numbers of cases. Similarly, it may be inefficient for a researcher to spend a great deal of time examining cases in which the cause of interest does not occur, as would be required by a methodology of controlled comparison. Initially, at least, the researcher can often use research resources to best advantage by examining cases that appear, prima facie, to demonstrate the causal relations hypothesized – that is, by selecting on the independent and dependent variables. This narrow focus will allow the researcher to efficiently identify conceptual errors and basic empirical weaknesses in the early hypotheses. Later, as the hypotheses become more refined and as understanding of scope conditions becomes more textured, the hypotheses can be subjected to much more rigorous analysis.

The issue of selecting on the independent and dependent variables is contentious within environment-conflict research. Early, pioneering work focused explicitly and intentionally on cases where the hypothesized causal link between environmental scarcity and conflict appeared to exist.²³ Given prevailing methodological thought within political science, it could be claimed that this approach biased the work’s results in favor of positive findings.

The criticism would take the following form. If environmental scarcity is the independent variable and violent conflict is the dependent variable, and if each variable – crudely – has two possible values, then we have four possible outcomes, as illustrated in the matrix in Figure 3. All cases (say, countries) will be located in one of the four quadrants of the matrix. If environmental scarcity is a necessary cause of conflict, there will be no cases in quadrant 3, but there may be cases in any of the other three. If scarcity is a sufficient cause of conflict, there will be no cases in quadrant 2, but, again, there may be cases in the others. If scarcity is both necessary and sufficient, there will be cases only in quadrants 1 and 4.

A correlational analysis will attempt to determine if the distribution of cases across the four quadrants is significantly different from a distribution that could be expected by chance alone. A distribution that is significantly different provides evidence that environmental scarcity and conflict are correlated. A controlled case comparison will vary cases on the independent variable, environmental scarcity, without regard to values of the dependent variable; of particular interest are any “null” cases in quadrant 2 in which all the preconditions of the hypotheses connecting environmental scarcity with conflict hold, yet conflict does not occur. Finally, process tracing will focus on cases selected just from quadrant 1. This is the method used in much environment-conflict research to date.

Critics might contend that process tracing somehow avoids a fair test of hypotheses. However, in the early stages of research, process tracing is often the best, and sometimes the only, way to begin. It can show, for particular cases, whether or not the proposed independent variable is a cause of the dependent variable. It answers the important questions identified above: Are there any cases in quadrant 1 in which the independent variable is causally linked, in a significant and interesting way, to the dependent variable? If so, how does this causation work?

More importantly, in highly complex systems such as ecological-political systems, it is likely that the proposed independent variable will not be a sufficient cause of the dependent variable. Rather, as noted, multiple factors, including the hypothesized independent variable, will interact to produce the effect in question. If a hypothesis is to be valid for such a system, therefore, it will have to be more than a simple statement of “X causes Y” or “X is correlated with Y.” The hypothesis will require, in addition, numerous and detailed scope conditions; it will take the form, for example, of “X causes Y, when A, B, and C are true.”²⁴ Adding the right scope conditions should increase the causal strength of the whole set of independent variables and scope conditions taken together. As the causal strength of the whole set increases, and if the whole set is taken as the independent variable, cases in quadrant 2 should move to quadrant 4. If enough conditions are specified, it might be possible to identify a set that is a “jointly sufficient” cause of Y. In this case, quadrant 2 should be empty.

Without including adequate scope conditions, a statistical analysis of the distribution of cases across the quadrants in Figure 3 will probably reveal little correlation, even though there might be important and interesting causal links between environmental scarcity and conflict. Yet careful process tracing, involving close examination of the causal process operating in the cases in quadrant 1, will help identify the relevant scope conditions.

Highly complex systems also present problems for controlled case comparisons. Such an approach, which, ideally, varies cases on the independent variable, is appropriate only if the researchers can be sure that all other variables that might affect the incidence of conflict are controlled. Then they can see what happens with the sole difference of variation in the independent variable.

Unfortunately, however, with ecological-political systems, researchers can never be sure that everything relevant is controlled. As indicated above, these systems include countless unknown variables and causal connections; analysts may not even be aware of the existence of many variables and causal linkages, let alone how they operate. Moreover, the relationships among their variables are often nonlinear, reciprocal, and interactive,²⁵ which makes the systems highly sensitive to small perturbations by relatively peripheral variables. It is therefore often impossible to identify cases similar enough that key independent variables can be isolated. Unknown and ill-understood differences between cases selected to vary only on a specified independent variable may have a great influence on the occurrence of conflict. Consequently, a close study of the incidence of conflict in cases that do not exhibit severe environmental scarcity may not reveal anything about whether and how environmental scarcity contributes to conflict.²⁶

One possible response to this problem of lack of adequate control is to carefully compare “positive” cases in quadrant 1 with any “null” cases that appear in quadrant 2.²⁷Such a comparison might help identify hidden factors and processes that influence links between environmental scarcity and conflict. However, once again, the high uncertainty about the character of the systems under investigation means that it is not sensible for researchers to conduct such a comparison before they have a good idea of how environment-conflict linkages work. Early in a research program, a focus on cases in quadrant 1 using process tracing is an efficient use of resources. If, instead, researchers spend much of their time examining null cases in quadrant 2, they will probably waste resources following red herrings and bad leads.

Close study of quadrant 1 cases using process tracing allows the researcher to determine key scope conditions and intermediate processes and variables. Eventually, on examining cases in quadrant 2, the researcher can ask whether these scope conditions and intermediate variables were present, and if not, why not. If these factors were present, the researcher can then determine what other factors prevented environmental scarcity from causing conflict. This staged approach permits progressive refinement of hypotheses and their scope conditions.

In sum, when researchers investigate highly complex causal systems, such as ecological-political systems, the choice of methodology to test hypotheses should be partly determined by the stage of research. In early stages, more attention should be given to the process tracing of causal links in cases where the hypothesized causal links appear to exist. As hypotheses are refined, an expanded range of methodologies can be used, including correlational analysis and controlled case comparison. Truly robust hypotheses – that is, hypotheses that reflect the complexity of the system under examination and that have a high probability of validity – are necessarily a product of later stages of research. They are the product of an iterative process of engagement with empirical data, using a range of quantitative and case-based tests.

Conclusions

Experts on methodology in political science often advocate an approach to hypothesis testing and causal inference that is modeled after the natural sciences. This orthodoxy stresses quasi-experimental research designs – including large-scale statistical analysis and controlled-case comparisons – that supposedly permit control of confounding variables, allow for variance on selected dependent and independent variables, and permit the disaggregation of the relative causal “weight” of different independent variables. The study of ecological-political systems highlights some weaknesses of this approach.

First, political scientists, and social scientists in general, tend to use “folk” concepts of causation that often do not further our understanding of complex social systems. A more precise and differentiated grasp of the possible causal roles of environmental scarcity as a contributor to conflict should inform methodological decisions about how to test environment-conflict hypotheses. In particular, researchers must be aware of the multivariate and highly interactive nature of ecological-political systems; these characteristics often render moot questions about the weighting, or relative strength, of specific causal variables.

Second, these systems are very “opaque” to researchers, in that they are extremely complex, ill-understood, and sensitive to small perturbations – characteristics that can together overwhelm both statistical and controlled-comparison methods. Consequently, in early stages of research involving ecological-political systems, process tracing is the best method to develop, refine, and test hypotheses. As research progresses and hypotheses become more sophisticated, researchers can fruitfully use a broader range of methodologies. In general, analysts should remember that political science is, at best, a soft science; a pluralistic approach to research methodologies is therefore often justified.