Predicting State Failure: A Classification Tree Approach

Transcription

1 Virginia Military Institute From the SelectedWorks of Atin Basu Choudhary Winter December 27, 2010 Predicting State Failure: A Classification Tree Approach Atin Basu Choudhary, Virginia Military Institute Jim T Bang, Virginia Military Institute William F Shughart II Available at:

2 Predicting State Failure: A Classi cation Tree Approach Atin Basuchoudhary y Department of Economics and Business Virginia Military Institute Lexington, VA James T. Bang z Department of Economics and Business Virginia Military Institute Lexington, VA William F. Shughart II Department of Economics School of Liberal Arts University of Mississippi University, MS December 27, 2010 Abstract We deviate from standard practice in the literature on state failure to use classi cation tree methods to predict state failure. We argue that the rarity of state failure and simplicity of use and interpretation makes this approach more attractive. We determine simple decision rules, based on observable and measurable variables, to determine whether a country is likely to fail or not. Do not quote without permission. y Author names appear in alphabetical order z Corresponding Author 1

3 1 Introduction In this paper we use a classi cation tree approach to predict state failure. We identify a hierarchy of several observable, measurable, and relatively perceptionbias- free variables that critically lead to state failure. Goldstone et al. (2010) and others (e.g., Fearon and Laitin, 2003) seek to predict political instability by focusing on the marginal e ect of variables plausibly related to it. This approach has a number of problems. First of all, states tend to be quite stable. These authors try to address the rarity of political instability by using complicated econometric techniques that do not lend themselves to easy interpretation. Or, they focus their attention onto geographical regions where political instability is not a rarity, thus reducing the generalizability of their models quite dramatically. We move away from these procedures by creating a global data set that is then subjected to classi cation tree analysis. This procedure provides simple decision rules forto decidinge whether a state is likely to fall apart or not. We also note that that most of the theoretical advances in understanding con ict and state failure are game theoretic in nature. 1 Game theoretic models develop equilibrium strategies. Classi cation trees are better suited for testing game theoretic models in general because of their ability to produce empirical "if- then" statements that retain the avor of equilibrium strategies. Thus we also suggest that classi cation trees should be the empirical method of choice for testing game theoretic models. Moreover, there is an inherent bias arising from missing observations in most cross- country or withincountry empirical analyses of con ict and state failure because con ict- torn countries are notoriously poor record keepers. The classi cation tree approach is better than parametric approaches in analyzing state failure because, rather than throwing out all observations of a variable because of missing observations for a few countries, it uses "surrogate" information from other variables to proxy for missing observations. The literature on predicting political instability tends to follow three major strands. One strand follows the view that ultimately the a lack of economic opportunity ultimately can increase political instability at the state level (Collier and Hoe er, 2002, 2004). Fearon and Laitin (2003) claim, on the other hand, that political factors incentivize rebels to take on the state. Goldstone et al. (2010) argue that political structure is what truly matters in any attempt to predict state failure. We combine all three approaches to identify decision rules that predictascertain whether a state will fail or not. 1 For a comprehensive review of the literature please see Blattman and Miguel (2010) 2

4 2 Data and Methodology In this paper we choose a classi cation tree approach to determine the political, institutional, and economic variables that are critical for predicting state failure. This approach provides both simple and atheoretical decision rules for predicting state failure. We describe the data used in Section 2.1 and the methodology in section Data. In this section we introduce the data used to measure our dependent variable, state failure, and the explanatory variables Measuring State Failure. Our total sample for estimating and predicting the probability of state failure includes 12,409 country-year observations for 237 countries from 1950 through Our dependent variable, state failure, seeks to capture the existence of a systematic collapse in the government s ability to function and meet the needs of its constituents. The speci c measure we use is the classi cation system employed by the State Failure Task Force (SFTF), which is in turn based upon the classi cations of state failure and con ict compiled by the Political Instability Task Force (PITF) (Esty et al. 1995). The PITF dataset classi es the four di erent types of major political crisis within a country as: (1) ethnic civil war, (2) revolutionary civil war, (3) genocides and politicides, and (4) adverse regime changes. The SFTF state failure classi cation incorporates all four of these types of government crisis into its de nition of whether or not a state is to be considered to be "failed." This approach is not without its shortcomings. As the SFTF acknowledges in its own report, "Narrowly de ned, state failures consist of instances in which central state authority collapses for several years [emphasis added]. The political crises indicated in the PITF dataset may not capture the level of collapse of authority described by this de nition, but "fewer than 20 such episodes have occurred during the last 40 years, too few for meaningful statistical analysis." Thus, the SFTF justi es its use of the PITF data for tractability reasons, as well as "for the reason that events that fell beneath such a threshold nonetheless often pose challenges to US foreign policy, the task force broadened the concept of state failure to include a wider range of civil con icts, political crises and massive human rights violations that are typically associated with state breakdown" (Esty et al. 1995) Predictors of State Failure. As predictors of state failure we include a variety of economic and demographic variables from the World Development Indicators (WDI), published by the 3

5 World Bank and institutional variables from several sources. As economic predictors we include: (1) GDP per capita measured in constant 2000 dollars; (2) gross capital formation as a percentage of GDP; (3) the total volume of trade (imports plus exports) as a percentage of GDP; (4) the unemployment rate as a percentage of the labor force; (5) foreign aid and development assistance as a percentage of central government expenditures; and (6) income inequality, measured by the Gini coe cient. To these we add the following demographic variables: (7) the total population, (8) the population density and (9) the dependency ratio (the number of children and the elderly relative to the working-age population). As mentioned above, our institutional variables come from a variety of sources, including the International Country Risk Guide (ICRG), published by the Political Risk Services (PRS) Group; the Database of Political Institutions (DPI), compiled by Beck et al. (1995) and published by the World Bank; the Polity IV Project, published by the Center for Systematic Peace (Marshall et al. 2010); and the Cross-National Time Series (CNTS), published by Databanks International (Banks 2010). The rst set of variables that we enter as institutional predictors of state failure include several qualitative characteristics of the political systems of the countries in our dataset from the CNTS. These variables are: (10) the competitiveness of legislative selection (no legislature, essentially non-competitive, partly competitive, or competitive); (11) party legitimacy (no parties or singleparty, signi cant exclusion, "extremist" party exclusion, or no exclusion); (12) the selection of an e ective executive (no election, indirect election, or direct election); and (13) the type of executive or regime (monarch, president, premier, military, or other). To these we add from the DPI (14) the type of political system (presidential, parliamentary, or parliament-elected president). Finally, we aim to include several quantitative measures of institutional quality and stability from the ICRG, DPI, Polity, and CNTS datasets. Simply including a subset of the constructed measures from these sources is problematic, however, for two reasons. First, although these variables purport to measure distinct aspects of institutional character, most of them are highly correlated. Second, institutional quality is likely to be multidimensional, and di erent dimensions of institutional quality or stability may have di erent impacts on economic and political outcomes (Bang and Mitra, forthcoming). The nonparametric classi cation methodology that we use to predict state failure partially avoids this issue in the sense that we do not need to worry about bias in our parameters. However, we do need to worry that our classi cation might end up being dominated by several institutional features that actually measure the same fundamental concept. Also, including a large number of institutional variables (some of which might contain redundant information) introduces the potential for over- tting the model as a consequence of excess dimensionality. To avoid these problems, we perform a factor analysis on 28 variables from the ICRG, DPI, Polity, and CNTS datasets to obtain eight distinct (and relatively uncorrelated) measures of institutional quality and stability. These factors are: (15) the democratic-ness of the country s institutions, (16) the transparency 4

6 of the government, (17) the government s credibility, (18) the security of civil society, (19) the amount of stability within the political regime, (20) the amount of social unrest and protest, (21) the frequency of political turnover in the regime and (22) the concentration of political power. We explain brie y how these factors are derived below. The Multidimensionality of Institutional Structure. In view of the conceptual and statistical problems associated with measuring the impact of institutional variables individually on the education of immigrants, it may help to get a more general sense of the overall impact of institutional structure. Following Alesina et al. (1996), Perotti (1996), Knack and Keefer (1995), Jong-a-Pin (2009) and others, we therefore try to identify combinations of variables that explain some aspect of institutions and can be interpreted more broadly than a single institutional variable, yet at the same time are uncorrelated with each other. One simple method for doing this is to perform a principal component analysis on the institutional variables, and interpret the rst component as institutions. This is the essence of what Alesina and Perotti (1996), Perotti (1996) and Keefer and Knack (1997) do in the context of investigating the in uence of institutional quality on economic growth. 2 Alternatively, following Alesina et al. (1996), one could construct a unidimensional index of institutional quality by using logit analysis. However, as argued by Jong-A-Pin (2009) in the context of measuring the impact of political instability, institutions have multiple dimensions, and it is our hypothesis that these dimensions may have di erent impacts on immigrant selection. Hence, a unidimensional index would fail to capture the true impact of institutional structure. This raises the question as to why we do not construct two separate indices for institutional quality and stability. The problem is that this would require a prior classi cation of available variables into ones that capture stability and ones that stand for quality. Having undertaken such a task in our preliminary exploration of the data, we are convinced that many of the available variables are not clear-cut measures of one dimension of institutional character as opposed to others. An alternative to principal component analysis, used by Jong-a-Pin (2009) among others, is factor analysis. Factor analysis is related to principal component analysis, but while principal component analysis aims to extract the maximum source of variation in the variables possible, factor analysis only seeks to capture the common sources of variation among the variables. Also, whereas in principal component analysis the components are linear combinations of the observed variables, in factor analysis the observed variables are linear combinations of the constructed factors. These features allow us to interpret the predicted factors and attach conceptual meaning to them. Principal component 2 The rst two papers investigate the impact of income inequality on economic growth via its role in fomenting social discontent. As such, they focus on constructing indices of sociopolitical instability rather than general institutional structure. Keefer and Knack (1997), on the other hand, consider more general measures of institutions. 5

7 analysis does not lend itself as well to such interpretation. As a result, factor analysis proves more useful in our investigation. The variables from the ICRG that we include in our factor analysis are: (a) government stability, which assesses "the government s ability to carry out its declared programs and... stay in o ce"; (b) the democratic accountability index; (c) the investment pro le index, which captures the enforcement of contractual agreements and expropriation risk; (d) the corruption index, which measures the absence of corruption in the government; (e) the index of bureaucratic quality, which assesses the e ciency of the bureaucracy; (f) the index of internal con ict; (g) the index of external con ict and (h) the ethnic tensions index. Next, we include nine variables from the DPI dataset. They are: (i) legislative party fractionalization; (j) political polarization, which takes a value of 0, 1, or 2 depending on the distance between the legislature and the executive on the left center right spectrum; (k) the number of years the current executive has been in o ce; (l) the number of changes in the veto players in the government; (m) legislative concentration, measured by the Her ndahl-hirschman index; (n) the number of veto points within the government; (o) whether or not there were allegations of electoral fraud in the last election; (p) the legislative index of electoral competition; and (q) the executive index of electoral competition. To these we add nine measures from the CNTS, which are: (r) the number of assassinations, (s) the number of labor strikes, (t) the number of major government crises, (u) the number of demonstrations, (v) the number of purges, (w) the number of riots, (x) the number of cabinet changes, (y) the number of changes in the e ective executive and (z) the legislative e ectiveness index. Finally, we include two variables from the Polity IV project: (aa) the polity 2 index of democracy and (bb) regime durability. The results of the factor analysis of these 28 variables are reported in Table 1 in the appendix. Variables with higher factor loadings for a particular factor contribute more to the construction of that measure, and so we have highlighted variables with factor loadings greater than 0.3. Here, eight distinctive measures of institutional quality emerge: democracy, which is primarily based on the indices of electoral competition and legislative fractionalization/concentration measures from the DPI, the polity index, the democratic accountability index from the ICRG, and the legislative e ectiveness measure from the CNTS; transparency, which is mostly comprised of the corruption, bureaucratic quality and democratic accountability indices from the ICRG; credibility, which is based on the government stability and investment pro le indices from the ICRG; security, which is largely based on the con ict and ethnic tension variables from the ICRG; within-regime instability, which is comprised of the number of cabinet changes and changes in the holders of executive o ce; protest, which is based on the number of demonstrations, riots and strikes in society; political turnover, based on the number of years the current executive has been in o ce and the number of changes in the veto players; and political concentration, which is based on legislative electoral competition, fractionalization and concentration 6

8 indices Methodology. Tree methods are non parametric. The results can be very simply classi ed into a series of if-then statements. Moreover, this method makes no a priori assumptions about the underlying nature of the relationships between the dependent and independent variables (Statistica 2010). 4 In addition, this process captures the e ect, on the dependant variable, of small perturbations at the extremes of the ranges spanned by each explanatory variable. Regression trees use an iterative algorithm to partition the data. This algorithm breaks up the data by using every possible binary split on every variable. It then uses this information to determine how well a split in the explanatory variables can di erentiate between observations on the dependent variable. Thus, the algorithm chooses the partitions that minimize the sum of the squared deviations from the mean in the separate parts. Each partition is split further using the same technique. The process continues until the sum of squared deviations from the mean at a node is zero. Each partition represents a decision rule that predicts the levels of the explanatory variables which are associated with the dependent variable. Below we provide a brief description of this process as well as a summary of the overall advantages of using a classi cation tree approach to predicting state failure. There are several ways of predicting a discrete classi cation model: logistic regression, discriminant analysis, cluster analysis, and classi cation tree analysis. A detailed reference for all of these techniques is available in Hand (1997) and a more accessible summary can be found in Shmueli et al. (2007). This paper compares logistic regression and discriminant analysis against classi cation tree analyses for predicting the likelihood that a country will be classi ed as a failed state in a given period. In this section, we will describe the classi cation tree technique, which is less familiar in the literature than are the other two Classi cation Tree Analysis. De ne the data to consist of a random pair, (X; y) where X is the set of explanatory variables or symptoms and y 2 C; C = f1; : : : ; Jg, is the outcome variable. The symptom variables can be real valued or categorical. De ne the learning sample, L = f(x 1 ; j 1 ); : : : ; (x N ; j N )g as the set of symptom-outcome pairs for each observation. Also, de ne a decision rule, d(x), that takes values on y. A classi cation tree consists of a set of nodes and branches that divide the sample into subsets. There are two types of nodes: terminal nodes, which are 3 The rst four factors, democracy, credibility, transparency, and security are similar to those derived by Bang and Mitra (2010); protest and within-regime instability are similar to factors derived by Jong-a-Pin (2009). 4 This methodology does not make any assumptions about whether the underlying relationship between the dependent and the independent variables are linear or non linear. In fact the relationships need not be monotonic either. 7

9 assigned a class in C, and non-terminal nodes, which lead to further splits. The full set of nodes is denoted T and the set of terminal nodes is denoted as T. The tree has three main elements: (1) Splits in the explanatory variables; (2) criteria for splitting at a node or declaring a node terminal; and (3) an assignment of a class to each terminal node. To begin, the de ning the splits is accomplished by a set of questions, that generates a set of splits of the data. At each node, a split divides a proportion of the sample, p L, to the left and a proportion, p R, to the right based on one or more of the explanatory variables. Next, a criteria for splitting at a node and a decision rule for when to call a node terminal is determined by depending on the extent to which splitting the node further can improve the classi cation. To achieve this, de ne the impurity of a node i(t) as a function of the probability of each class given it is in node t that satis es (1=J; ::::; 1=J) = maximum (1; 0; ::::; 0) = (0; 1; ::::; 0) = (0; 0; ::::; 1) = 0 5 The decrease in the impurity by splitting t using s is: i(s; t) = i(t) p L i(t L ) p R i(t R ). A node is terminal if either i(s; t) or the number of observations in either of the subnodes is less than some predetermined critical value. Finally, the assignment of a class to each node will depend on the proportion of the classes at that node and the cost associated with misclassifying the cases in any particular direction (i.e. whether it is more costly to predict a state to be failed when it is not or vice versa). If the costs are symmetric, then the class that is assigned to each node will simply be the class that represents the highest proportion of cases at that node. The test for the overall goodness of the tree is the tree impurity, I(T ), which is the sum of the node impurity functions, i(t), at the set of all terminal nodes, t 2 T. De ne: I(T ) = X i(t)p(t) (1) t2t where p(t) is the probability of reaching node t. To assess the quality of a decision rule, de ne R (d) as the risk estimate, or misclassi cation rate, of d. To estimate R, let h() be an indicator function that equals one if the statement in parentheses is true and zero otherwise. The simple estimate of R(d), R (d), is: R (d) = 1 N NX h(d(x i ) 6= j) (2) i=1 5 These conditions simply state that impurity is maximal when classes are uniformly distributed and zero when only the predicted class is contained in that node. An example of a typical impurity function is i(t) = j2c (p(jjt)log[p(jjt)]), which is a log entropy measure. In principle, the impurity function is any function that satis es the conditions given. Another common impurity function is the Gini index, i(t) = j6=i p(jjt)p(ijt); or1 j p 2 (jjt). 8

10 The problem with using R(d) is that it is based on the same sample as is used to construct the decision rule. An appropriate way to overcome this for small samples is to use a cross-validated estimate. Cross-validation is performed by dividing the learning sample, L, into V random subsets of equal size. For every v; v = 1; : : : ; V; construct a tree using the learning sample L L v and de ne d (v) (x) to be the classi cation rule. Use the remainder of the sample, L v to construct an estimate for R. The risk estimate for each random subsample, v, is: R ts (d (v) ) = 1 N V NX (x i;j i)2l V h(d V (x i ) 6= j) (3) Repeating this for all of the random subsamples, L v, then the resulting estimate for R is R CV (d) = 1 V VX R ts (d (v) ) (4) Finally, once a tree is generated, it can be improved by pruning. The two methods of pruning are minimum risk and standard error. The tree structure itself is unchanged by the pruning method, only the way that splits are upwardly removed to minimize an impurity measure that penalizes based on the number of nodes. V = Advantages of Classi cation Trees There are several advantages to the classi cation tree approach. First, classi cation tree analysis is useful for both very large as well as relatively small datasets. For small datasets, the nonparametric nature of the technique allows us to draw insight and understanding about the key variables that might in- uence the outcomes without relying on parametric tests. For large datasets, classi cation tree analysis can be useful in cases where the data are of high dimensionality (i.e., there are lots of potential explanatory variables that may in fact be correlated with one another), include a mixture of various types of data and scales of measurement (i.e., interval, ordinal, or nominal scaled), or for datasets where the relationships are non-homogenous (i.e., where there may be di erent relationships between the variables in di erent parts of the measurement space). Second, since classi cation trees are based on a simple series of binary (or, sometimes multinomial) "if-then" statements, they are relatively easy to interpret. From a policy and decision-making perspective this simplicity has two advantages. First, it makes it easy to predict the outcomes for new cases. Second, it automatically selects variables that are included in the splitting criteria based on their economic importance, rather than on their precision (as signi - cance testing does). 9

11 Finally, as a practical matter, classi cation trees make better use of all available data than other, traditional, parametric methods. Speci cally, in parametric methods, such as discriminant analysis and logistic regression, if any particular variable is missing for any particular element of the dataset, the entire set of values for that observation are omitted. Since classi cation trees are more exible to di erent types of variables, it is able to classify missing values by using "surrogate" information from other variables (i.e. classifying elements based on second-best splitting criteria) or, if that is not possible, creating a separate classi cation rule for missing values. In our case, this is a particularly useful feature of the technique because many countries that experience state failure are also countries for which several of the explanatory variables of interest are likely to be "missing." 3 Results We start with observations on state failure. The classi cation tree algorithm examines the values of each explanatory variable and splits the overall data space in two, such that each space is as homogenous or as pure as possible in the dependent variable. The process continues for each subsequent space the goal is to attain as much information as possible in each subsequent partition. In our dataset the classi cation tree algorithm examines each explanatory variable and ranks them in order of how well they are able to separate observations associated with failed states from observations associated with successful states. It then picks the split that has the most observations associated with failed states in one category and the least in the other. The rst partition of this dataset is determined by PROTEST. The classi cation algorithm splits the dataset into two parts in a way that minimizes heterogeneity in each part. Thus, in one partition there are 419 observation of which 56.3% are associated with failed states. In the other partition there are observations of which only 8.4% are associated with failed states. The split occurs at PROTEST < Note that the classi cation tree algorithm has processed all possible splits for values of each explanatory variable and chosen PROTEST at the level because that split, among all others, assigns the greatest percentage of observations associated with failed states to one category and the least percentage of observations associated with failed states tothe other. PROTEST is a composite variable developed by extracting the principal factors from DEMONSTRATIONS, RIOTS, and STRIKES (see Table 1). It is interesting that an inability to protest is the source of the rst, broad, and even counter-intuitive split in our dataset. Hirshleifer (1998) claims that con- ict arises when peaceful bargaining between two parties fails. In his model, the failure arises when the "Potential Settlement Region" (the set of possible resource allocations that are attainable so that both parties to the con ict gain) becomes smaller, thus reducing the opportunity of negotiated, peaceful, settlement. Information about these possible future, peaceful allocations of wealth is 10

12 of course essential for any bargaining solution to be stable. Thus information asymmetries may be a source of con ict. In addition, the paucity of institutional structures that would even allow such peaceful bargaining between potentially warring factions, by reducing information asymmetries and other transactions costs, would also lead to con ict. Protests are a way for citizens to communicate directly with the state. Thus the inability overtly to communicate with the state is emblematic of an asymmetric information problem between the rulers and the ruled. Indeed, Esteban and Ray (2001) show that in a multiplayer contest asymmetric information about the costs of con ict make con ict inevitable when there are more than four players. Thus, heterogeneous societies lacking a structure for information sharing can be prone to con ict and political instability that in turn impairs the state s ability to provide political goods. 6 Now, states where citizens cannot protest on the streets are likely to have repressive regimes. The only way the vox populi can be heard is through regime change or overt con ict. Within the "low protest" cohort of states, however, one may assume that in some countries protests are infrequent because citizens are satis ed or have other, more systematized means of communication provided by civic institutions like parliamentary debate and news media. Thus, within the cohort of countries where protests are rare we should expect to nd a lower likelihood of political instability rooted in asymmetric information than in countries with vibrant civic institutions. In fact our classi cation tree provides just such evidence. Recall that the rst level partition of our data suggests that if PROTEST < then states are likely to fail. However, at this decision level the partition of countries for which state failure also is likely to be correlated with the LEGIS06 categories {No Parties/Single Party, Signi cant Party Exclusion, "Extremist" Parties Excluded}. On the other hand, those countries for which PROTEST is > , i.e., the countries unlikely to be failed states are also likely to be in the LEGIS06 category {No Exclusions}. In other words, countries where civic institutions are inclusive are also less likely to fail. In fact, at this level, state failure is also associated with the LEGIS04 categories {No Legislature, Essentially Non-Competitive} while success is associated with {Partly Competitive, Competitive}. Once again, countries without civic institutions that allow communication between the rulers and the ruled are less likely to be successful in the provision of political goods and more likely to fail. Moreover, such repressive regimes will have less overt protests that vent the vox populi in the absence of institutions that support civic engagement. Thus, in such countries the vox populi shows itself in con ict and regime change. Viva La Revolucion. Other variables are also correlated with state failure in the rst partition though they are not the source of the partition itself which continues to be PROTEST. We nd that failed states are more likely to be in the DIRECT_ELECT categories {Direct Election, No Election} while successful states tend to be in the {Indirect Election} category. It is plausible to surmise that states that 6 These "political goods" are the structure around which citizens of a state form expectations, derive obligations, and develop a political culture that determines the social contract between the ruler and the ruled (Pennock, 1966). 11

13 hold indirect elections for their chief executive are more likely to have complex civic institutions that provide checks and balances to executive power. Thus, states that have direct elections or no elections are likely to have executives constrained by few checks and balances. Thus, it seems plausible that countries with low levels of PROTEST are also likely to be in the category where executives have few checks and balances on their power. In fact, countries in the low PROTEST category are also likely to have executive regimes that are in the {Military, Other} category rather than the {Monarch, Premier, President} categories. Military leaders can plausibly be surmised to have fewer constraints on their power than mere kings, presidents, or premiers. Tanks go a long way toward overcoming civic challenges to power while cultural norms (in the case of kings) and civic institutions (in the case of presidents and premiers) may provide credible checks to the power of the executive. In the absence of credible civic and cultural checks on power, con ict and extreme political instability may be the only way for aggrieved citizens to demonstrate their unhappiness with the executive. Thus the rst decision rule for predicting state failure seems to suggest that political structures that reduce communication between the rulers and the ruled lead to state failure, especially in the absence of civic institutions that limit on executive power. These results are consistent with some of the theoretical models of con ict which suggest that asymmetric information that leads to problems with commitment and contracting drive con ict and political instability. 7 Thus, our model seems to capture the avor of these studies by focusing on the importance of communication (PROTEST) in countries that lack the civic structures that could peacably mediate principal-agent problems between the rulers and the ruled. It is also consistent with the empirical evidence presented by, among others, Fearon and Laitin (2003), La Ferrara and Bates (2001) and Skaperdas (2008), who nd evidence countries with weak civic institutions and few checks and balances on executive power are more likely to be prone to con ict and political instability. The next decision rule, while once again highlighting the importance of legislative stucture also includes a number of non institutional variables. Among states likely to fail because PROTEST < failure is even more likely (Node 4) if they happen to be in the LEGIS04 categories {No Legislature, Essentially Non-Competitive, Partly Competitive}. In fact at this second level, of the 419 observations placed in the category where state failure is most likely in the rst level of decision making, 334 (69.2%) are placed in the category associated with state failure. Of the remaining 85 observations (Node 4), only 5.9% are associated with state failure and are therefore placed in a separate category. In fact this second category of successful states cannot be further homogenized by the classi cation process and is e ectively a terminal node. Notice that each decision making level provides a ner grain decision rule 7 Esteban and Ray (2001) show how social decision making is impossible when players have private information about the costs of con ict. Sylvain Chassang and Gerard Padro-i-Miquel (2008a, 2008b, and 2009) also show how economic shocks increase the incentive to ght in the current period relative to future, discounted, peace. 12

14 with more homogenous categories. Thus, at the rst level PROTEST < de ned the decision rule for state failure. These observations were associated with the LEGIS04 categories {No Legislature, Essentially Non-Competitive}. At the second decision making level notice the classi cation tree adds another category within LEGIS04, {Partly Competitive}, in addition to the {No Legislature, Essentially Non-Competitive} categories that observations into which failed states are likely to fall. Thus, if protests are suppressed enough then even partly competitive legislatures do not provide a su cient institutional bu er against state failure. Moreover, we can argue, at this second level of decision making, that state failure is more likely in countries with relatively closed economies (TRADE < ) and which receive a lot of aid (AID > 57.64). Moyo (2009: 47 & 124) and Calderisi (2006: 9) both argue very e ectively that trade has improved African economies while aid has not. Further, the connection between strong economies and low levels of con ict is one of the few regularities observed in the empirical cross-country literature on con ict (Hegre and Sambanis 2006). Thus, the connection between high levels of aid and state failure and low levels of trade and state failure are perhaps unsurprising. However, it is another bit of evidence in the aid versus trade debate. However, note that our empirical approach does not ascribe causality. We merely claim that countries are more likely to be in the state failure category if they tend to be closed economically and also receive high levels of aid. We also nd that larger countries (AREA > 7680) are associated with state failure. This is possibly a consequence of the di culty of governing larger tracts of land when states are weak. This argument is consistent with the notion that commitment problems and incomplete contracts increase the incentive to renege on contracts and increase the possibility of con ict (Powell 2006). To the extent that weak governments struggle to impose their writ over large tracts of land contract enforcement seems unlikely which in turn would make con ict more likely (e.g., Fearon and Laitin 2003; Skaperdas 2008). The classi cation algorithm places Recall that our rst decision rule placed observations in a category (with PROTEST ) where state failure was not likely (only 8.4% of the observations were associated with state failure). At the second level of decision making, though, the classi cation tree process divides these observations in two categories nodes 6 and 7. Now in both categories the probability of any observation being associated with state failure is less than 0.5. Thus in either category state failure is an unlikely phenomenon. Nevertheless in one of these categories the probability of nding observations associated with state failure is much greater (34.4% of the 855 observations in this category) than the other (6.4% of the 11,135 observations). The split at this level is determined by EXEC_REG. The category with a greater proportion of observations associated with state failure is thus associated with the EXEC_REG category {Military, Other}, while the other category is associated with the EXEC_REG categories {Monarch, President, Premier}. Note that we get similar results at the rst desicion rule level however, the categories are more homogenous at this level as a consequence of the classi cation tree algorithm. Similarly, we also nd that more observations associated with state failures are also associ- 13

15 ated with the LEGIS06 categories {No Parties/Single Party, Signi cant Party Exclusion, "Extremist" Parties Excluded}, the LEGIS04 categories {No Legislature, Essentially Non-Competitive}, GDPPC < 1, and the POLIT08 category {No Election}. All of this is consistent with our discussion above. Node 7 is also a terminal node the observations in this node cannot be partitioned further in any meaningful way by the classi cation process. Given our discussion of the important variables at the rst and second iterations of the recursive partitioning performed by the classi cation tree procedure there are no surprises with the third iteration. There were 334 observations classi ed as being more likely to be associated with state failure at the second iteration (node 4). 8 Of these, in node 8, 175 observations are much more likely to be associated with state failure (85.7% of these 175 observations are associated with state failure) while 159 are less likely (50.9% of these 159 observations are associated with state failure) though both partitions have more than 50% of the observations associated with state failure. Once again a higher proportion of observations are associated with state failure if PROTEST is low (this time < ). Moreover, low levels of TRADE ( < ) and high levels of AID (>6.31) are associated with state failure. All this is consistent with our discussion above. However, low dependency ratios ( DEP 1 < ) and high population densities are now additionally associated with a greater proportion of observations that can be categorized as belonging to failed states. This is quite interesting. A low dependency ratio suggests that the proportion of children and the elderly relative to the population of individuals in the productive age group is quite low. Typically this should transform into less of a burden on the productive segment of the population. However, in states with weak institutions it just means there are more people of ghting age with fewer ties to either the past or the future when con ict does arise. Further, state failure is also associated with higher population densities. This febrile combination suggests that more people particularly of a ghting age add fuel to the ames of con ict. Node 8 is also a terminal node. However, the 159 observations in node 9 can be further re ned by the classi cation algorithm into 111 observations (Node 12) classi ed as being associated with the failed state category (65.8% of the observations are associated with failed states) while 48 (node 13) are not. This time the split is predicated on the dependency ratio, thus reiterating the importance of a low dependency ratio in generating failed states. At Nodes 12 and 13 Democracy makes an appearance but is relatively unimportant. Note that Democracy is a variable constructed from the principal components of Legislative Fractionalization (DPI), Legislative Concentration (DPI), LIEC (DPI), EIEC (DPI), Polity, Checks (DPI), and Democratic Accountability (ICRG). Democracy di ers from the other variables like Polit08, LEGIS04, LEGIS06, SYSTEM, and EXEC_REG in a fundamentally subtle way. The latter variables capture institutional structure, the nuts and bolts as it were. DEMOCRACY tries to capture the less tangible e ects of open societies. It is 8 The remaining 85 of the rst 419 observations associated with state failure were part of a terminal node. 14

16 gratifying to see that open societies are associated with successful states. However, note that the partition between observations more likely to be associated with state failure is not determined by DEMOCRACY. Rather it is dependency ratios that determine the split. Other variables associated with state failure at this level are consistent with splits at earlier iterations of the partition process. The 855 observations partitioned into node 6 can be further partitioned into two groups of observations (both with the proportion of observations associated with state failure being less than 50%) where one group node 10 has a higher proportion of observations associated with state failure (48.9% of 370 observations) than the other (23.3% of 485 observations) node 11. The split at this level is predicated on population density, with higher population densities being associated with state failure. The only additional insight obtained at this level is that low levels of investment (GFCF < 8.7), low GDP per capita (GDPC < ) are also associated with observations that belong to failed states. This result is perhaps not striking given that Collier and Hoe er (2004) associated low current growth rates with more con ict a result subsequently shown to be quite robust by Hegre and Sambanis s (2006) test of the sensitivity of various statistical regularities reported in the empirical con ict literature to the conditioning set. Countries with low levels of investment also grow more slowly (Solow????). Poorer countries may also be more likely to fail since the state may not have the resources to stem rebellion (Fearon and Laitin 2004). Node 11 is a terminal node. Node 10, however, can be further puri ed. The 370 observations at node 10 can be further puri ed by placement in even more homogenous groupings of observations predicated by TRADE. Once again low levels of trade are associated with more state failure. No additional insight is available at this level of disaggregation of observations. However, some insight can be obtained from the variable particularly those from the factor analysis that did not seem to be useful in determining state success or failure. TRANSPARENCY (which is mostly comprised of the corruption, bureaucratic quality and democratic accountability indices from the ICRG), CRED- IBILITY (which is based on the government stability and investment pro le indices from the ICRG), SECURITY (which is largely based on the con ict and ethnic tension variables from the ICRG), WITHIN REGIME INSTABILITY (which is comprised of the number of cabinet changes and changes in executive), POLITICAL TURNOVER (based on the number of years the current executive has been in o ce and the number of changes in the veto players), and POLITICAL CONCENTRATION (which is based on legislative electoral competition, fractionalization and concentration indices) are not relevant for classifying variables into those that are likely to be associated with state failure. Interestingly all of these variables capture elements of the "usual suspects" in much of the literature on con ict and political instability. We of course are able to classify variables into those that are likely to be associated with state failure in terms of categories that capture actual state structure, e.g., whether it is ruled by a military regime or not or whether there are direct elections or not. First of all, possibly, even if transparency, for example, is important to the lives of people, the information contained in constructed variables like TRANS- 15

17 PARENCY can be obtained from variables that pertain to the actual structure of the state. Moreover, notice that none of the variables that use "soft" perceptions based indices was relevant in classifying variables into categories of state failure or success. All the variables relevant for decision making in our model are either purposefully quanti able PROTEST or TRADE for example or are variables like REGIME, or LEGIS08 that are patently observable. Thus we argue that not only is our methodology not subject to perception biases that plague many of the indices that are used in this literature but that it provides clear decision rules based on observable or clearly measurable variables. 4 Conclusion We use classi cation trees to provide simple decision rules, based on observable and measurable variables, to determine whether a country is likely to fail or not. References [1] Banks, Arthur S Cross-National Time-Series Data Archive. Databanks International. Jerusalem, Israel; see Last accessed summer [2] Bang and Mitra. (Forthcoming). "TITLE." Economic Systems. [3] Beck, Thorsten, George Clarke, Alberto Gro, Philip Keefer, and Patrick Walsh, "New tools in comparative political economy: The Database of Political Institutions." World Bank Economic Review. 15(1): [4] Blattman, Christopher and Edward Miguel Civil War. Journal of Economic Literature. 58(1): [5] Calderisi, Robert The Trouble with Africa: Why Foreign Aid Isn t Working. Palgrave Macmillan. New York. NY. [6] Collier, Paul, and Anke Hoe er "On the Incidence of Civil War in Africa." Journal of Con ict Resolution. 46: [7] Collier, Paul, and Anke Hoe er "Greed and Grievance in Civil Wars." Oxford Economic Papers. New Series 56(4): [8] Esteban, Joan and Debraj Ray "Social Decision Rules Are Not Immune To Con ict." Economics of Governance. 2(1): [9] Esty, Daniel C., Jack Goldstone, Ted Gurr, and Pamela Surko "State Failure Task Force Report." State Failure Task Force Report. McLean, VA: Science Applications International Corporation. 16

18 [10] Fearon, James, and David Laitin "Ethnicity, Insurgency, And Civil War." American Political Science Review. 97: [11] Goldstone, Jack, Robert Bates, David Epstein, Ted Gurr, Michael Lustick, Monty Marshall, Jay Ulfelder, Mark Woodward "A Global Model For Forecasting Political Instability." American Journal of Political Science. 54(1): [12] Hand, David J Construction and Assessment of Classi cation Rules New York, NY: John Wiley and Sons, 214. [13] Hegre, Havard and Nicholas Sambanis "Sensitivity Analysis of of Empirical Analysis of Civil War Onset." Journal of Con ict Resolution. 50(4): [14] Hirshleifer, Jack "The Bioeconomic Causes of War." Managerial and Decision Economics. 19: [15] La Ferrara, Eliana and Robert H. Bates "Political Competition in Weak States." Economics and Politics. 13(2): [16] Marshall, Monty G., Ted R. Gurr, and Keith Jaggers, Polity IV Project Political Regime Characteristics and Transitions, Center for Systemic Peace. [17] Moyo, Dambisa Dead Aid. Farrar, Strous, and Giroux. New York, NY. [18] Pennock, J. Roland "Political Development, Political Systems, and Political Goods." World Politics. 18: , 433. [19] Political Risk Services, Inc International Country Risk Guide. [20] Powell, Robert "War as a Commitment Problem." International Organization. 60(1): [21] Skaperdas, Stergios "An Economic Approach to Analyzing Civil Wars." Economics of Governance. 9(1): [22] Statistica, cation-andregression-trees/. Statsoft (last accessed Dec 7, 2010). [23] Shmueli, Galit, Nitin R. Patel, and Peter C. Bruce Data Mining for Business Intelligence: Concepts, Techniques, and Applications in Microsoft O ce Excel R with XLMiner R Hoboken, NJ: John Wiley & Sons. 17

19 A The First Appendix Figure 1 Node Observations F Node (56.3%) Protest< Legis06={Extreme, One Party, Significant Exclusion} Exec_Reg={Military, Other} Legis04={Non Competitive, No Legislature} Polit08={Direct Elections, No Elections} Node (8.4%) Protest Legis06={NA, No Exclusion} Exec_Reg={Monarchy, Premier, President} Legis04={Competitive, Partly Competitive, NA} Polit08={NA, Indirect Elections} F Node (69.2%) Legis04={Non Competitive, No Legislature, Partly Competitive} Trade < Aid Area 7680 Node 5 85 (5.9%) Legis04= {Competitive} Trade < Aid Area 7680 Node 6 855(34.4%) Exec_Reg={Mili tary,other} Legis06={Extre mist Exclusion, One Party, Significant Exclusion} Legis04{Non Competitive, No Legislature} Polit08={No Elections} Gdpc < Node 7 11,135 (6.4%) Exec_Reg={Mo narchy, Premier, President} Legis06={No Exclusion} Legis04{Compe tetive, Partly Competitive} Polit08={Direct Elections, No Elections} Gdpc Continued on next page