Return to Index

"Failed States and International Security:
Causes, Prospects, and Consequences"

Purdue University, West Lafayette
February 25-27, 1998

The State Failure Project:
Early Warning Research for U.S. Foreign Policy Planning

by Daniel C. Esty, Jack Goldstone, Ted Robert Gurr,
Barbara Harff, Pamela T. Surko, Alan N. Unger, and Robert Chen

State failure and state collapse are new labels for a type of severe political crisis exemplified by events of the early 1990s in Somalia, Bosnia, Liberia, and Afghanistan. In these instances the institutions of the central state were so weakened that they could no longer maintain authority or political order beyond the capital city, and sometimes not even there. Such state failures usually occur in circumstances of widespread and violent civil conflict, often accompanied by severe humanitarian crises These conditions may precede or follow the institutional collapse of the state; sometimes they are instrumental in causing it. In a general sense they are all part of a syndrome of serious political crisis which, in the extreme case, leads to the collapse of governance.1

The international consequences of state failures are profound. Failed states often threaten regional security and require major inputs of humanitarian assistance. And they pose long-term and costly challenges of rebuilding shattered governments and societies. A vital policy question is whether failures can be diagnosed far enough in advance to facilitate effective international efforts at prevention or peaceful transformation. The State Failure Task Force was established in response to a 1994 request from the office of the Vice President of the United States to design and carry out a data-driven study on the correlates of state failure. The ultimate objective is to develop a methodology to identify key factors and critical thresholds signaling a high risk of political crisis in countries some two years in advance. This chapter describes the approach taken in the first phase of the Task Force's work, outlines some general findings, and identifies research issues to be dealt with in future work.2

Identifying State Failures and Control Cases

No more than a dozen complete collapses of state authority have occurred during the last 40 years, too few for meaningful generalization. Therefore the Task Force broadened its focus to include partial state failures. This is consistent with distinctions that other researchers have made between states that have collapsed completely and those in the process of failing. Zartman (1995:9) observes that "State collapse is a long-term degenerative disease . . . an extreme case of governance problems." Helman and Ratner (1992-93:5) identify states whose survival was threatened in the early 1990s, among them Ethiopia, Georgia, and Congo/Zaire. In Africa, states were said to be in serious danger of collapse in Mozambique, Angola, Rwanda, Burundi, Chad, and Sierra Leone (Nellier, 1993; cited in Zartman, 1995:3, 11). Whereas all these states experienced very serious problems of violent challenges and ineffective governance in the early 1990s, by early 1997 some had pulled back from the brink—examples are Georgia and Mozambique—whereas others had collapsed and then were reconstituted under new leadership—as in Ethiopia and Rwanda.

The first challenge facing the Task Force was to identify systematically all occurrences of partial or complete state failure that began between 1955 and 1994. We began from existing compilations of revolutionary and ethnic conflicts, regime crises, and massive human rights violations that are typically associated with state breakdown. The problem set of state failures thus includes four categories of serious political conflicts and crises that have been identified in specialized studies and data collections. Within each category, events are scaled by magnitude to permit analysis of the conditions associated with the extent of state failure.3

Revolutionary wars are sustained military conflicts between insurgents and central governments, aimed at displacing the regime4 (n=40). They include:

  • Large-scale and intense guerrilla wars with more than 250,000 deaths; for example, mujahidin warfare against the Khalq regime in Afghanistan, 1978-92.
  • Large-scale and intense guerrilla and civil wars with 10,000 to 250,000 deaths; for example the Sandinista guerrilla war against the Somoza regime in Nicaragua, 1978-79.
  • Small-scale guerrilla wars and rebellions that result in 1,000 to 10,000 deaths; for example, left-wing guerrilla warfare against the Colombian government, 1984 to the present.
Ethnic wars are secessionist civil wars, rebellions, protracted communal warfare, and sustained episodes of mass protest by politically organized communal groups5 (n=75). They include:
  • Large-scale and intense ethnic wars with more than 250,000 deaths; for example, the rebellion by southern Sudanese, 1983 to the present.
  • Large-scale and intense wars with between 10,000 and 250,000 deaths; for example, the Kurdish rebellion against the Khomeini regime in Iran, 1979-84.
  • Small-scale communal wars and rebellions with 1,000 to 10,000 deaths; for example, the Intifadah campaign in the Israeli-occupied territories, 1988-94.
  • Protracted episodes of violent communal rioting, clashes, and terrorism; for example, violent protest by Azerbaijanis against Soviet policies, 1987-91.
Genocides and politicides (geno/politicides) are defined as sustained policies by states or their agents and, in civil wars, by contending authorities that result in the deaths of a substantial portion of members of communal or political groups. In genocides the victimized groups are targeted primarily because of their communal (ethnic, religious) characteristics. In politicides, by contrast, victims are targeted mainly because of their political opposition to the state or dominant group6 (n=46). They include:
  • Episodes involving more than 250,000 deaths; for example, the Khmer Rouge killings and starvation deaths in Cambodia, 1975-79.
  • Episodes involving between 100,000 and 250,000 deaths; for example, the civilian death toll from massacres and starvation during RENAMO's rebellion against the government of Mozambique, 1976-92.
  • Events involving between 10,000 and 100,000 deaths; for example, Tutsi army massacres of Hutus in Burundi, 1993.
  • Events involving fewer than 10,000 deaths; for example, the victims of the military's “dirty war” against the left in Argentina, 1976-80.
Adverse or disruptive regime transitions are major, abrupt shifts in patterns of governance, including state collapse, periods of severe regime instability, and shifts toward authoritarian rule7 (n=82). They include:
  • Collapse of central state authority for two or more years; for example, the collapse of central government in Somalia, 1989 to the present.
  • Transition toward autocratic rule by revolution or coup; for example, the military coup against the Allende regime in Chile, 1973.
  • Abrupt transitions toward autocratic rule by nonviolent means; for example, the replacement of democratic institutions by one-party rule in Sierra Leone in 1978.
  • Violent regime instability accompanied by revolution or coup, with no increase in autocracy; for example, north-south rivalry and civil war after the attempted merger of North and South Yemen in 1990.
These 243 conflicts and crises are the basis of the problem set, that is, the study's dependent variable. It includes almost all serious events of these types that began between 1955 and mid-1994 in all states in the international system with populations greater than 500,000.8 Area specialists subsequently have proposed some additions to and deletions from the set, mostly events of low magnitude, which will be taken into account in a future revision of the problem set. The analyses reported below are based on the initial problem set.

The four kinds of conflicts and crises in the problem set often coincide or occur sequentially in the same country. The Iraqi Kurds, for example, fought four successive rebellions between 1961 and 1992; and in 1988-89 the Iraqi government responded with the murderous al Anfal campaign, a case of politicide against a politically-organized communal group. Where conflicts or crises overlapped or came in quick succession, these were combined into consolidated cases. The 113 consolidated cases include 62 single events of the four kinds plus 51 complex cases consisting of linked sequences of events (of any kind) in which less than five years elapsed between the beginning and end of successive events.

The next research task was to match the 113 consolidated cases of state failure with a random sample of control cases three times as large. For each year in which a consolidated case began, three cases were selected at random from countries in which no such events of any type or magnitude were under way or about to begin in the next two years. The question for all statistical analyses was to identify those independent variables, and sets of variables, that discriminated most significantly and efficiently between the problem set and the control set.

Explaining State Failure: The Independent Variables

A very large number of independent variables were considered for inclusion, ranging from indicators of democracy and ethnic cleavages to the size of the youth age bulge, and from income inequality to the presence of IMF standby agreements. Candidate variables were identified in Task Force brainstorming sessions based on theoretical arguments, observed empirical regularities, and observations from policy makers. Cross-national data on important variables often proved to be limited or missing, especially data for the global South and for the 1950s and 1960s. Nonetheless it was possible to acquire a great many existing data sets; and to code or collect some new data, especially on political variables. No classified data sources were used, though the Task Force did ask whether comprehensive data might be available from intelligence sources on such variables as leadership traits and traffic in small arms; no usable comparative data on these variables could be identified. A team from the Consortium for International Earth Science Information Network (CIESIN) assembled an initial matrix of 617 measures for each country for each year it was in the study. The maintenance and updating of this dataset was and continues to be a major task of the project.

Expert assessments of the potential significance, quality, and coverage of the 617 measures were made by the analytic team from Science Applications International Corporation and led to a Task Force decision to concentrate statistical analysis on 75 high priority measures. Of the high priority variables 21 were demographic and social, 24 political, and 30 environmental and economic. The first cut was to conduct t-tests or chi-square tests to determine their capacity to differentiate between the states that had a regime crisis and the control cases that did not. This univariate analysis identified 31 powerful variables, listed in Table 2.1, that discriminated at a statistically significant level between the problem and control set.

The next step was to subject these 31 variables to two kinds of multivariate analysis: logistic regression analysis and neural network clustering. Regression analysis is a technique familiar to early warning researchers, neural network analysis is not. Neural network clustering provides a non-linear modeling technique, and can be used on datasets where the values for some variables are missing, a property that can become important when additional variables are added to the analysis. Combinations of two, three, five and up to 14 variables contained in the 31-variable set were examined together in an inductive approach to specifying the most accurate analysis or model. Regression analysis was used first, then neural net analysis was applied to the variables found to be significant in regressions. Several models approached 70 percent accuracy in identifying which states would fail between 1955 and mid-1994. The two techniques used gave similar results.9 The single best model, according to both regression and neural net analyses, included three variables: openness to international trade, infant mortality, and democracy.

Table 2.1: Independent Variables that Discriminate between State Failures and Control Cases (statistically significant using a t or x2 test with P < .05).
Military personnel/physicians ratio
Civil liberties index
Infant mortality
Life expectancy
Extended longevity
Percent of children in primary school
Percent of teens in secondary school
Girls/boys ratio in secondary school
Youth bulge
Labor force/population
Party legitimacy
Party fractionalization
Executive dependence on legislature
Separatist activity
Years since major regime change
Ethnic character of ruling elite
Religious character of ruling elite
Political rights index
Maximum ethnic cleavage
Defense expenditures/total government expenditures
Government revenues/GDP
Investment share of GDP
Trade openness (imports plus exports/GDP)
Real GDP/capita
Cropland area
Land burden (farmers/cropland) x (farmers/labor force)
Reports of famine

1. Openness to international trade is the sum of imports + exports as a percent of GDP; high openness is associated with a low risk of state failure. There are several reinforcing theoretical explanations for this finding. Trade serves as a measure of a country's integration into the global economy and the international community. One consequence of economic and political interdependence is that regimes are more inclined to adhere to international norms of good governance, and more sensitive to external encouragement to observe those norms and to censure for violating them. In domestic politics, a large volume of trade is dependent on a stable rule of law and fair enforcement of contract and property rights. Where trade is less significant to the economy, regimes may have more latitude for arbitrary, unpredictable, or corrupt behavior that offends key political groups and prompts them to revolutionary challenges.

2. Infant mortality is the ratio of reported deaths of infants under one year old per thousand live births: infant mortality rates above the international median for a given year are associated with a high risk of state failure. Infant mortality is indicative of the quality of life in a society. It tends to be lower in countries with high GDP per capita, but also is relatively low in some poorer countries whose governments devote substantial public resources to health, education, and welfare services. Its importance as an indicator of stability is likely due to its inverse association with popular discontent. Regimes that are unwilling or unable to raise the quality of life to international standards are at risk of popularly-based challenges. This relationship is likely to be more pronounced in democratic countries where discontented publics have greater opportunities to organize opposition.

3. Democracy is indexed by reference to competitive political participation, election of chief executives, and institutionalized checks on executives' exercise of power. Regimes above a middling threshold of democracy (5 or above on the 10-point Polity III scale; Jaggers and Gurr 1995:472) have a low risk of state failure. The finding is consistent with a large body of theory and observation about the conflict-inhibiting effects of democratic governance. Since democratic elites are dependent on popular support to gain and retain office, they are more responsive to popular discontents and more likely to accommodate potentially dissident political and ethnic challengers. They also are likely to have strong normative and institutional inhibitions against committing or tolerating massive human rights violations.

Two of the three variables in the general model correlate with a number of others used in the study. Infant mortality is a marker indicator that represents a basket of interdependent conditions. It performs slightly better in most of the models estimated than do other quality-of-life indicators such as per capita caloric intake, access to clean water, and GDP per capita. Depending on the availability of data, one could substitute some of these other variables for infant mortality without a major reduction in the predictive power of the models. Democracy is not a marker or surrogate indicator but rather a summary measure of open political institutions. It correlates strongly with indices of political rights, civil liberties, party legitimacy, and more weakly with some quality of life indicators. The indicator of trade openness has few close correlates either conceptually or empirically. Of the 75 key variables analyzed in the study, it correlates closely only with the density of roads—generally accepted as an indicator of economic development—and population size.

An important interaction effect was observed between democracy and the other two variables in the basic model. Among more democratic countries, the risk of failure was greater when infant mortality—which we interpret as a broad measure of living standards and quality of life—was high and when trade openness was low. Among less democratic countries, by contrast, the risk of failure was greater when trade openness was low, regardless of the level of infant mortality. Reference to specific groups of countries helps clarify these findings.

  • Among less democratic countries with trade openness above the world median (countries whose total imports + exports exceeded three-quarters of their GDP), risks of state failure were very low, irrespective of levels of infant mortality. Gabon and Zimbabwe are countries that in 1994 fitted this profile.
  • Among less democratic countries with lower levels of trade openness, risks of failure were substantial but levels of infant mortality do not help differentiate among levels of risk. Cuba and North Korea are examples of non-democratic countries with low trade openness in the early 1990s.
  • Among more democratic countries, the study found that the higher the infant mortality rate, the greater the risk of state failure. All other things being equal, more democratic countries with high infant mortality faced a greater risk of failure than did less-democratic countries. In an effort to probe this relationship, a cluster analysis was done of cases in the study with high (5+) democracy, GDP per capita below the world mean (used here as a surrogate indicator of quality of life) and openness below 60 percent. About 40 percent of the democracies in this situation prior to 1992 failed. Democratic countries where infant mortality substantially exceeded the 1994 world median of 45 deaths per thousand live births include Bangladesh, Benin, Bolivia, Kyrghizia, Nepal, and Paraguay. Several countries in this group, including Kyrghizia, had state failures between 1994 and mid-1997.
  • A significant risk of state failure was observed in democracies where infant mortality rates ranged between the median and one-half the median. Democracies where infant mortality were in this range in 1994—between 10 and 45 deaths per thousand live births—include Armenia and Romania. Examples of quasi-democratic states in this range in 1994 include Jordan, Mexico, and Thailand.10
  • A very low risk of failure was observed in countries where infant mortality rates fell below one-quarter of the median. In 1994 this included the advanced industrial democracies and, among others, Greece, Portugal and South Korea.

The Task Force recognized the need for models that are specific to each of the four types of state failures included in the problem set. Such analyses were done for the two most numerous types: violent or abrupt regime change and ethnic war, with results that dovetailed closely with the general model.

1. Regime crisis: Stepwise logistic regression identified two strong variables, infant mortality and number of years since the previous major regime change. The model's predictive power was improved by the addition of trade openness. These three variables together correctly estimated violent or abrupt regime transitions in 69 percent of cases. Neural net analysis achieved 70 percent accuracy using infant mortality and trade openness alone.

2. Ethnic war: Stepwise regression identified three strong variables that jointly provided 78 percent accuracy. As in the general model, trade openness reduced the risk. Two other variables, not in the general model, increased the risk of ethnic war: the ethnic character of the ruling elite, if it represents only one group in an ethnically divided society; and a youth age bulge, i.e. a large proportion of the adult population concentrated in the young adult years. Neural net analysis achieved 72 percent accuracy with the three variables from the general model plus youth bulge and years since last major regime change.

Strengths and Limitations of the State Failure Project

The State Failure project is the most broadly conceived empirical effort we know of to identify the correlates of political crises globally and across a long span of time. It makes use of a wider range of data than previous studies in conflict and crisis analysis and is one of the few studies of its genre to employ two different analytic methods. The dataset compiled for the study is unique in depth and coverage, though many gaps remain. The initial models are parsimonious, internally consistent, and have important implications for researchers' theoretical understanding of the preconditions of state failure and for long-range policy planning.11 And the project has gained substantial visibility and credibility among those responsible for the analysis of global security and for planning U.S. foreign policy. The first phase of the State Failure project thus strengthens the case for a systematic approach to risk assessment and early warning of political crises.

Having said this, some important cautions are in order.

  • The study has been largely inductive. It was not designed to test existing theoretical arguments about the causes of specific kinds of political crises.
  • The models classify correctly about 70 percent of historical cases. A model with 70 percent accuracy two years in advance would correctly identify about two out of three failures and two out of three stable countries. Given historical failure rates of about three per year, this translates into approximately one missed regime crisis per year. The number of false positivesstates classified at high risk of failure that did not fail in a given yearis roughly 50 out of the 166 countries in the analysis.
  • The models are based on historical analysis. It remains to be demonstrated that they will be equally accurate in identifying prospective cases of state failures.
  • Models have not yet been identified that help to account for the type or degree of state failure or the sequential relations among them. Researchers and analysts also need to learn more about the conditions that keep partial state failures from escalating, and that contribute to the reconstruction of collapsed states.
  • Many variables of theoretical and policy interest were not included in the initial analyses because suitable broad-coverage indicators did not exist. Others were used despite significant data gaps and irregularities. The limits are particularly serious with respect to international and domestic political factors and environmental variables.
  • Most variables in the models refer to background or structural conditions that are relatively slow to change. In the terminology used elsewhere in this volume, the models are (at best) suitable for long-term risk assessment. If they are to be used for early warning, they must be complemented by the analysis of potential accelerators and triggers.


Most of the limitations enumerated above can be overcome, at least in part, through future work. Members of the Task Force expect that most of these issues will be addressed, if not by the Task Force itself then by other researchers and agencies.

Improving the Quality and Coverage of Data

The initial results suggest priorities for future data gathering and coding on background variables. High priority should be given to identifying more precise and time-sensitive indicators of quality of life, international economic linkages, and ethnic polarization especially at the leadership level. Longer-run efforts are needed to improve the coverage and reliability of ecological and environmental indicators. All indicators that contribute significantly to the models, including revised models that may be identified in future research, should be kept as close to current as possible.

Types, Magnitude, and Duration of State Failure

Policy planning and risk assessment will be improved once we are able to say something about which type of state failure is most likely to occur in high-risk countries, and with what magnitude. One way to do this is to use an expanded set of background variables to identify the probability of occurrence of each type of state failure. Regression analysis also can be used to estimate models of magnitudes and duration. A more innovative approach to the problem of duration also can be suggested: to model the conditions associated with the ending dates of state failure. Indicators of regime capacity, economic productivity, and international assistance are likely to figure in such models. The authors suggest that some such models should be estimated.

Accelerators and Triggers

Researchers want to understand the political and social dynamics that link the background conditions to the onset or avoidance of state failure. Policy makers and analysts need more than watch lists: they want to know about the signals that suggest a state failure is imminent. There are several research approaches to this problem. One is to use theory and comparative evidence to identify accelerators of various types of conflict and crisis. Barbara Harff (chapter five in this volume) lays out the rationale for this approach and sketches suggestive evidence from four cases. The approach only works if causal models have already been specified: “accelerators” are, in effect, variables outside the parameters of a static or systems model whose effects depend in whole or part on their interaction with conflict-disposing background conditions. This approach is recommended for future work by early warning researchers. An inductive alternative is to search for patterns in “events” that are observed to cluster prior to the onset of historical instances of state failure. Examples of this approach are represented in this volume by Philip Schrodt and Deborah Gerner (Chapter 7 in this volume) and Peter Brecke (Chapter 9 in this volume).

Conflict Management

Foreign policy makers have for a long time relied on foreign assistance and preventive diplomacy to reduce the risks of state failure, or alternatively to help regimes recover from crises (see Lund, 1996). An innovative complement to the analysis of accelerators is to specify and study the effects of “de-accelerators,” i.e. the events and interventions that contribute to conflict resolution and recovery. The inductive approach to this research task is to look for evidence of cooperative interactions in the flow of conflict events. A stronger argument can be made for a deductive approach: the impact of conflict-management interventions surely depends on the context in which they are employed. This implies that we should begin with empirically plausible theories or models about the conditions that increase or decrease different types of political crises and conflicts.

Model-building and Expert Assessments

The 70 percent accuracy of the Task Force's initial models in accounting for failure or non-failure of states suggests how wide the gap is between general risk assessment and early warning of specific crises. Econometric models of national and global economic processes have similar indeterminacy but nonetheless have proven to be extremely useful for policy makers. Moreover the indeterminacy is a spur to building more refined and accurate models. Early warning researchers should be able to close part of the gap by identifying and analyzing the accelerators and triggering events of political crises. At the end of the process, though, all evidence on a country situation—data on background conditions and the stream of events—needs to be interpreted through substantive as well as theoretical lenses. The quantitative modeling and theoretical analysis of the dynamics of political crises is not designed to replace expert assessments. The two kinds of analysis are complementary. If risk assessment and early warning are to be put on a systematic footing, it will be done by teams of generalists and country specialists working with shared information within a shared perspective.


1. The terms state failure and state collapse are analogous to but broader than well-established analytic terms like political instability and internal war. Helman and Ratner (1992-93) discuss implications of state failure for international policy. Zartman (1995) gives the term state collapse more precise analytic content and uses it in comparative case studies. Also see Kaplan (1994) and Ayoob (1996). The State Failure Task Force is the first effort to identify a comprehensive list of instances of state failures.

2. The study was commissioned by the Central Intelligence Agency's Directorate of Intelligence in response to Vice President Gore's request and was carried out by a Task Force consisting of academic experts, data collection and management specialists from the Consortium for International Earth Science Information Network (CIESIN), and analytic methods professionals from Science Applications International Corporation (SAIC). The co-authors of this chapter are senior consultants to the Task Force. The views expressed in this chapter are those of the authors and do not represent the official view of the U.S. government, the U.S. intelligence community, or the Central Intelligence Agency.

3. The magnitudes of state failure were not analyzed in the first phase of the Task Force's work.

4. The primary source is Small and Singer 1982, with an update through 1993 provided by Professor Singer. Civil wars by communal groups are included under ethnic wars.

5. From the Minorities at Risk project's profiles of conflicts involving all politically active communal groups from 1946 to 1989, updated and annotated for the State Failure Project. See Gurr 1993a, 1994, and chapter 1 in this volume.

6. The general definition and distinctions are developed by Harff (1992: 27-30). The primary data source is her inventory of episodes of gross human rights violations since 1945 (Table 3.1 in Harff 1992), updated and annotated by Barbara Harff and Michael Dravis for the State Failure project.

7. Interruptions and abrupt shifts in regime authority patterns were identified from the Polity III data set (Jaggers and Gurr 1995), which includes annual codings of the authority traits of all regimes since 1800. The data set was updated and annotated by Keith Jaggers for the State Failure project. Note that nonviolent transitions from autocracy to democracy are not considered state failures and thus are not included in the problem set.

8. The United States was excluded from the study. If it had been included, 1960s racial violence in the South and urban riots and rebellions in the North would have entered in the problem set as low-magnitude ethnic war.

9. The accuracy with which cases are classified can be adjusted depending on whether the analysts care more about the number of “false positives” or “missed failures.” For example, models can be estimated that accurately classify all failures, but at the cost of vastly increasing the number of false positives. The accuracy rates reported here are based on a study design that divides errors evenly between “false positives” and “missed failures.”

10. The governments of these countries have some democratic features but as of 1994 were below the 5 threshold on the Democracy indicator.

11. Some theoretical and policy implications of the study are dealt with in the State Failure Task Force Reportfor official U.S. government usedated 30 November 1995.