Basins at Risk: Predicting International River Basin Conflict and Cooperation

Abstract

The existing literature identifies several factors that may influence conflict and cooperation in international river basins. It offers theoretical arguments to that end, and it empirically tests these arguments by means of qualitative case studies and large-N statistical work.1 Empirical testing in this research is primarily of an ex-post nature, however. That is, it seeks to account for incidences or levels of international river basin conflict and cooperation observed in the past.In this paper, we seek to take research one step further by moving from ex-post empirical analysis to predictions and forecasts. Predictions are “conditional statements about a phenomenon for which the researcher actually has data, i.e., the outcome variable has been observed.”2 A forecast “is a conditional statement about how a phenomenon will develop in the future and/or whose values are truly unknown.”3 Our motivation is substantive and practical, but also methodological. As it will become evident in the course of this analysis, ex-post empirical results, for example in the form of regression coefficients in quantitative studies on river basin conflict and cooperation, may not tell us much about the actual influence of specific explanatory variables, such as political system characteristics, water scarcity, or river geography. As noted by Ward et al., “policy prescriptions cannot be based on statistical summaries of probabilistic models.”4Moving from ex-post empirical analysis to prediction/forecasting thus serves two purposes. First, it complements the former by allowing us to discriminate among explanatory factors according to their predictive power. Second, it offers a more solid scientific basis for forward-looking (ex-ante) analysis, which is highly relevant from a policy perspective.The most explicit forward-looking analysis of international river basin conflict and cooperation to date is the “basins at risk” study by Yoffe et al.5 This article, which was published about ten years ago, relied on the Transboundary Freshwater Dispute Database (TFDD).6 The TFDD codes water-related events (conflict and cooperation) in most international river basins on a continuum, the Basins at Risk (BAR) scale, which ranges from highly conflictual to highly cooperative events. Yoffe et al. correlated this scale with a large set of variables that might influence conflict risk or the chances of cooperation. Ultimately, and based on simple bivariate regressions and descriptive analysis, they identified those international river basins that appeared particularly risk-prone according to their analysis. On these grounds, Yoffe et al. placed river basins into three categories: (a) basins in which water conflict was already manifest; (b) basins in which conflict is possible in the future and for which there is evidence of existing tensions; and (c) basins in which conflict is possible in the future, but there is no present evidence of existing tensions. Table 1 aggregates these categories and summarizes the rivers Yoffe et al. predicted as “being at risk.”7Prediction and forecasting require a robust ex-post empirical foundation. The Yoffe et al. study was in this respect ahead of its time, since large-N research on international river basin conflict and cooperation was not well developed in 2003, and has made major progress only in recent years. Given current standards, Yoffe et al.’s risk-profiling aspect lacked a sufficiently robust ex-post explanation of river basin conflict and cooperation. Interestingly, however, subsequent quantitative studies on international basins have rarely used advanced prediction and forecasting techniques. While the literature frequently cites Yoffe et al.’s work, it has not advanced the most policy-relevant part of that research agenda, namely the forward-looking risk-profiling component.This paper systematically connects ex-post empirical analysis and the “basins at risk” agenda. Building on recent theoretical and empirical research, we first construct an ex-post explanatory model. We then use prediction and forecasting methods to evaluate the predictive power of particular explanatory factors and identify river basins that are prone to conflict or cooperation. We start by briefly reviewing the existing literature and then describe our move from ex-post statistical inference to prediction and forecasting. We discuss our results, and conclude by comparing our findings with the starting point for our work, the original “basins at risk” study.Whereas the earlier research identified twenty-nine basins at risk, our study classifies forty-four such river basins. We also arrive at different findings with respect to key determinants of river basin conflict and cooperation. Our analytical approach can increase the robustness of explanatory models in other areas of international environmental politics, and make their findings more policy-relevant by moving from ex-post analysis to in-sample prediction and to out-of-sample forecasting.The existing literature on international river basin conflict has been cumulative in nature, allowing us to be very selective as we focus on a set of recent quantitative studies.8 This approach allows us to develop our research on a broad theoretical footing and provides a sound empirical base for our predictions and forecasts. Existing explanatory models rely either on traditional models of conflict, such as the gravity model,9 or use variables that largely belong to three general clusters: (1) realist variables capturing state interests and the distribution of power; (2) liberal variables and transaction costs; and (3) river-specific and geographic variables.10Brochmann and Gleditsch, building on Toset et al., Furlong et al., Gleditsch et al., and Owen et al.,11 apply the gravity model. According to this model, conflict over river basins is likely to be driven by country size (population) and power (GDP per capita), and is inversely proportional to the distance between states (state contiguity and capital-to-capital distance).12 Most of these variables are also considered in Yoffe et al.13 Brochmann and Gleditsch find that the gravity variables affect conflict risk in the expected ways. They also find that other explanatory variables, including basin size, political regime type (democracy), and river geography (upstream/downstream configurations) have significant effects.Most scholars have identified conflictual interactions between states and inferred cooperation from the absence of conflict.14 Yet, there are some noteworthy exceptions that have focused directly on cooperation. Zawahri and Mitchell,15 for example, measure basin cooperation in terms of treaties between riparian country dyads. To explain cooperation, they use proxies for state interests, notably the share of a country’s surface area in an international river basin, the ratio of external to internal sources of freshwater, and precipitation levels. The distribution of power is measured by upstream and downstream countries’ economic and military capabilities. Liberal arguments are captured by the democratic form of government,16 and by the similarity of domestic legal traditions in a riparian dyad. The explanatory model also includes the total number of states in a basin and a measure of geographic contiguity. Zawahri and Mitchell find that greater dependence on freshwater resources makes cooperation more likely, while higher precipitation levels make it less likely. Furthermore, cooperation is more likely between democratic riparians, states with similar legal systems, and contiguous countries.Kalbhenn focuses on liberal arguments, conceptualized by political regime type and linkages between states.17 She theorizes that river cooperation is more likely among democracies, which increases trust and, therefore, promote cooperation; and among states with more extensive trade relations and joint international organization memberships. Her empirical analysis, which uses the same event data for river basin conflict and cooperation that we employ, supports these arguments. In line with earlier research, the analysis also includes river characteristics, realist variables, and variables for the gravity model.18Brochmann addresses a similar research question as Kalbhenn.19 Her research focuses on all international river basins between 1948 and 1999, using the TFDD data.20 In terms of the explanatory variables, Brochmann argues that realist variables, such as peace years, alliances, and state power are likely to affect river cooperation. Her model also includes variables mirroring Kalbhenn’s liberal approach as well as indicators that capture river-specific characteristics.21Based on the existing literature, as briefly reviewed here, we identify sixteen variables that are commonly used to explain international river basin conflict and cooperation. The online appendix22 summarizes these variables, points to the underlying theoretical rationales, and refers to the data sources.23 We construct an explanatory model that includes these variables as the starting point for predicting and forecasting basin conflict and cooperation.As noted in the preceding section, empirical studies have shown that the explanatory variables we discussed (see http://www.mitpressjournals.org/doi/pdf/10.1162/GLEP_a_00260) have significant effects on international river basin conflict and cooperation. Several scholars contributing to the more general literature on conflict and cooperation argue, however, that drawing inferences from statistically significant results can be misleading to the extent that those inferences are unlikely to tell us much about the predictive power of a specific covariate or an entire model: statistically significant results may improve our understanding of the relationship between variables in a given sample under study, but they may not provide information on the same relationship in another sample.24 Yoffe et al. are at least explicitly aware of this limitation when stating that “[c]ategorizing a basin at risk does not presume to identify basins in which acute conflict will occur, but to point to basins worth more detailed investigation.”25We submit that moving from ex-post statistical inference to in-sample prediction and out-of-sample forecasting can help improve the explanatory power of our models. Moreover, it makes research more policy relevant, as policy-makers are usually primarily driven by the desire to identify and mitigate the risks of future conflict. In more technical terms, as noted by Ward et al.,26 “it is quite possible to focus on statistically significant results that are artifacts in the sense that they do not generalize beyond the specific cases studied. This happens if we focus only on statistically significant relationships and may actually hinder our ability to generalize to out-of-sample situations, such as the future!”Accordingly, we evaluate (a) the predictive power of those explanatory variables we discussed (and as summarized in the appendix) using in-sample prediction techniques and (b) the ability of these variables to forecast basins at risk and those that are likely to see cooperation using out-of-sample approaches.27For our empirical work, we follow other studies28 and opt for the dyad-basin-year as the unit of analysis. The dependent variables measure international river basin conflict or cooperation, based on the International Rivers Cooperation and Conflict event data (IRCC) compiled by Kalbhenn and Bernauer for 1997–2007.29The rationale behind choosing these data is twofold. First, the IRCC data are coded from a uniform set of information sources. While the TFDD offers data for a longer time period than the IRCC, the main reason for using the IRCC instead of the TFDD is that major changes in the availability of news media texts over time (notably the advent of the digital revolution) make it problematic to use event data coded from partly changing sources for a very long period of time (as the TFDD does). Second, the IRCC data are publicly available in a format that is easier to use for an advanced statistical analysis of river basin conflict and cooperation than the TFDD.Ultimately, using the IRCC event data appears most appropriate in view of the outcome we want to explain: conflictual and cooperative water-related events in international river basins. Applying our analytical approach to other types of dependent variables that have been used in the literature, for instance river treaties, river claims, or militarized interstate disputes, would be straightforward and certainly useful. However, we could not implement our analysis for different types of dependent variables in this paper due to space constraints. Table 2 provides an overview of the IRCC scale.At this point, it should be noted that international river basin conflict can, according to the IRCC coding rules, include “water wars” that would receive a value of −6, e.g., interstate disputes over water resources. Yet, no events coded in the IRCC dataset received this coding. Thus, all observed and recorded events have the character of international political disputes, low-intensity violent actions, or international cooperative events. This assessment that such water wars are rare is supported by the fact that neither the TFDD nor the Issue Correlates of War Project (ICOW)30 include events that would qualify as a “water war.”We now examine the in-sample and out-of-sample predictive power (a) of two comprehensive models on river basin conflict and cooperation; (b) of the predictors in these models; and (c) by identifying whether basins are more likely to experience conflict or cooperation in the future. After estimating the ex-post statistical models that are based on the IRCC data and the explanatory factors introduced above, we move beyond such approaches, which rely exclusively on the statistical significance of explanatory variables. To this end, we aggregated the IRCC data, which record individual events, to the dyad-basin-year, and generated two binary variables for conflict and cooperation that serve as our dependent variables. The first binary variable (Conflict) receives a value of 1 if the median IRCC score for a dyad-basin-year is negative (0 otherwise); the second variable (Cooperation) receives a value of 1 if the median IRCC score is positive (0 otherwise).31 Consequently, we estimate two separate models: one for conflict (Model 1) and one for cooperation (Model 2). Next to the explanatory variables we discussed (and as summarized in the appendix), these models incorporate a conflict-years variable and a cooperation-years variable, respectively, as well as different sets of cubic splines to correct for temporal dependencies.32We considered, but did not use, two alternative empirical measures for the dependent variable. First, one could argue in favor of the yearly mean value for the dyad-basin-year before calculating the binary dependent variables. We decided not to use the mean, since it is more sensitive to extreme values than the median.33 Second, one could change the unit of analysis and compare individual events. While this would circumvent the issue of using either the mean or median for aggregation, it increases the number of observations substantially and, we believe, artificially, as all of our covariates are measured at the country-dyad or the river basin level. In other words, using the dyad-basin-year as the unit of analysis avoids inflating the number of observations, but requires data aggregation either in terms of the mean or median—and we believe that the median is likely to be the more accurate choice.Models 1 and 2 cover 1997 to 2004, the period for which data for our dependent variables (IRCC data) and our explanatory variables is available. Although this relatively short temporal span might initially appear as a shortcoming, it expands opportunities for out-of-sample forecasts to the period 2005–2007. We can then compare those out-of-sample predictions with the empirically observed values on the dependent variables to assess the models’ forecasting capabilities. Our results of the ex-post models are summarized in Table 3.Since our main interest is in prediction and forecasting, we discuss this table’s results only briefly. In Model 2, for example, more water dependence is associated with more cooperation. Conversely, we might expect a negative coefficient for the water dependence variable in Model 1. This coefficient is neither negative nor significant, however. In contrast to other studies, democracy appears to reduce the probability of cooperation (Model 2). The similarity of the legal system has a significantly negative effect on conflict risk, but also on cooperation. The findings for the number of riparian states and membership in international organizations are similar to the findings of other scholars.34 It is important to note, however, that dependent variables and time periods differ across studies. Hence, our initial findings should not be viewed as an empirical contest between models, data, or samples, but rather as a plausibility check of our empirical setup.The application of in-sample predictions is straightforward. First, we use the models in Table 3 to estimate the predicted probabilities of conflict and cooperation, respectively, for each dyad-basin-year. The predicted probabilities can vary between 0 (0 percent) and 1 (100 percent), and we group these into quintiles, which we compare with the actual instances of conflictual or cooperative dyad-basin-years in our data. For these calculations, we refer to the fifth, fourth, and third quintile as the “most-likely” group, i.e., those groups with predicted probabilities that are most likely to match with the actual instances of conflictual or cooperative dyad-basin-years. The first and the second quintile, in contrast, are the “least-likely” group, which is equivalent to those groups with predicted probabilities that are comparatively low and, thus, less likely to correctly predict observed onsets of conflict or cooperation. The findings are summarized in Table 4 and Figure 1.The most-likely group for Model 1 includes twenty-seven out of thirty-one conflictual dyad-basin-years in our data (87 percent). With regard to Model 2, 416 of 445 (93 percent) cooperative dyad-basin-years are placed in the most-likely group. In other words, only four dyad-basin-years that were de facto conflictual are characterized as least-likely cases. Similarly, only 29 of 445 cooperative dyad-basin-years are classified as non-cooperative, but were actually cooperative.Figure 2 sheds more light on the in-sample predictive power of Models 1–2. The Receiver Operator Characteristic (ROC) plot shows the extent to which models with more predictive power generate “true positives at the expense of fewer false positives.”35 Thus, a perfectly predictive model would correctly classify all empirically observed cooperative or conflictual dyad-basin-years and never generate false positives, i.e., dyad-basin-years that were not conflictual or cooperative, although our estimations predict the opposite. The importance of the ROC plot is highlighted by Gleditsch and Ward:36 any “threshold for considering an event as predicted could be seen as an arbitrary description of the continuous distribution of the probabilities.” Hence, despite our careful selection of the thresholds for the most-likely and the least-likely groups above, this argument by Gleditsch and Ward clarifies why the ROC curves approach is more precise in showing the predictive power of models. Figure 2 emphasizes that although our models do not perfectly predict either river conflict or cooperation, these models yield a higher predicted probability for a randomly chosen event than for a randomly chosen non-event. This finding is reflected in the ROC curve statistic (AUC), which theoretically varies between 0.50 (no predictive power) and 1.00 (perfect predictive power). As demonstrated by Figure 2, our models perform quite well in this regard: Model 1 has an AUC value of 0.78 and Model 2 has an AUC statistic of 0.82.The previous section prompts the question of how well our models and their covariates fare in the “harder” test of an out-of-sample forecast, i.e., what is the model’s predictive power for outcomes that are not “within the very same set of data that was used to generate the models in the first place.”37Our first step in this section is a so-called four-fold cross-validation setup,38 for the full models and for models that omit one or some of the explanatory variables from the estimation. This approach divides the existing data into four subsets, while the dyad-basin-year observations are randomly assigned to these four different sets. All except one of the four subsets are then pooled together and this pooled set of observations is used to estimate the models shown in Table 3. The remaining subset, also called the “test set,”39 which we do not use for the pooled set of observations and the initial model estimation, then serves to assess the predictive power of the model estimated for the pooled subsets. Put differently, we predict the outcome variables of the test set with models that are based on a pooled set of randomly-assigned observations. Then, we calculate the AUC for measuring the predictive power.40 We repeat this procedure ten times and then present the mean AUC for several model constellations in Figures 3 and 4.Figures 3 and 4 show the results (AUC values on the vertical axis) for the full models as specified in Table 3 and models in which certain explanatory variables are omitted (labeled as stages on the horizontal axis). The values presented in the first stage of either Figure 3 or 4 indicate the AUC of the full model that leaves out one covariate at a time. Lower values than the value of the “Full Model stage” indicate that a covariate contributes to the out-of-sample prediction power of the model. Higher values than the AUC from the full model indicate that including a specific covariate reduces the forecasting capability of the model. Hence, predictors associated with lower values have a higher power for forecasts. After completing the calculations for the first stage, we then identified the strongest predictor (label underlined) and left it out for the second and third stages, while repeating the calculations for all other covariates again. On this basis, we identified the three strongest single predictors for river conflict and cooperation. The reason for this is that demonstrating that an entire model or its alternative specifications perform above the AUC level of 0.50 does not allow for firm conclusions concerning the forecasting power of individual covariates. We thus follow Ward et al.41 and estimate the AUC for each model, dropping covariates sequentially.The forecasting power as measured by the AUC value both of the full models and the reduced models at the three stages, where we exclude an explanatory variable at a time, is generally lower for the four-fold cross-validation than for the in-sample predictions. Nevertheless, the forecasting power of the full models remains reasonably high: we obtain a score of 0.66 for conflictual dyad-basin-years and 0.80 for cooperative dyad-basin-years. These findings are driven by relatively few variables for which the average AUC is lower than the AUC of the full model. For example, Population, Small (log) is the strongest predictor in the first stage of the conflict model. When dropping this variable, the out-of-sample predictive power decreases from an AUC of about 0.66 to 0.63 in the conflict model. Similarly, Number of Riparian States (decrease in AUC from 0.80 to about 0.79) is the strongest predictor at the first stage of the cooperation model. We then discard both these strongest predictors from the respective model for the second stage and repeat the procedure of a four-fold cross-validation for all remaining explanatory variables. After identifying the strongest predictors in the second stage, we reiterate this procedure for a third stage. Our results show, therefore, that dropping some of the explanatory variables included in Table 3 from any model estimation would not only be misleading from the perspective of statistical significance, but also from the viewpoint of predictive power. However, this conclusion only applies to a small subset of covariates that contribute to the forecasting power of our models.Figure 3 documents that Legal System Similarity, Downstream Power, and Population, Small (log) contribute most to the out-of-sample power, i.e., they display a lower AUC than the full model when leaving these items out of the estimation for the first stage; for the second stage, the contributing covariates are Precipitation, Downstream Power, Democracy, External Water Dependence, and Population, Large (log); for the third stage, only Precipitation contributes to the forecasting power of the conflict model. With regard to cooperation (Figure 4), the variables Number of Riparian States, External Water Dependence, and Population, Small (log) are the three strongest predictors. Most other variables, including IGO Membership, contribute very little to the cooperation model’s predictive power. Figures 3 and 4 also suggest that more parsimonious models can perform better in forecasting river basin conflict or cooperation than more complex models. Already (and only) the three strongest predictors for either conflict or cooperation seem to work very well in this respect.Finally, in a last step to assess the forecasting power of our models, we again use the grouping of the predicted probabilities by quintiles and compare these with the empirically observed conflictual or cooperative dyad-basin-years. The crucial difference between the otherwise-similar approach underlying Table 4/Figure 1 above and the models (not reported here) leading to Table 5/Figure 5 below is that we now use the covariate values in 2004 to predict river conflict and cooperation between 2005 and 2007. That is, we employ data for the explanatory variables in the last observed year (2004) and impute these into the years 2005–2007 to predict conflict or cooperation, as measured by our dichotomous dependent variables. This approach mirrors a true forecast to the extent we can make a conditional statement about how conflict and cooperation will develop in the future as we treat the dependent variables’ values as unknown.42 Again, we refer to the fifth, fourth, and third quintile as the most-likely group, while the first and the second quintile are designated as the least-likely group.Based on this test, the most-likely group comprises thirteen out of thirteen conflictual dyad-basin-years in our data (100 percent). Slightly less accurately, 182 out of 188 dyad-basin-years (97 percent) are captured by the cooperative most-likely group. This means that, although we now use “new data,” i.e., dyad-basin-years for the dependent variable that were originally not covered by the models in Table 3, the forecasting power in this case is higher relative to the in-sample predictions and the four-fold cross-validation. This finding is upheld by the values of the ROC plots in Figure 6. The accuracy of the model for conflictual dyad-basin-years increases to 0.97, while the forecasting power of the cooperation model is 0.90.The foregoing analysis shows that our models can produce accurate predictions and forecasts for conflictual and cooperative dyad-basin-years. We can, therefore, return to our principal motivation and compare our findings with the basins that Yoffe et al. identified as “at risk.”43 To this end and in our setup, a basin is predicted to be at risk (or to be characterized by cooperation) if at least 90 percent of all dyad-basin-years in both the in-sample and out-of-sample estimations are classified under the “most-likely groups” of conflict (or cooperation) introduced above. Table 6 summarizes the results.The first column mirrors Table 1. The second column lists all basins that appear in our in-sample and out-of-sample predictions or forecasts of conflict and fulfill the “90 percent threshold.” Note that this second column lists forty-four basins as compared to twenty-nine in the first column. There is a rather limited overlap as only six basins appear in both lists: Asi/Orontes, Cross, Han, Indus, Ob, and Tigris-Euphrates. One of these six basins constitutes a particularly interesting case: the Cross River in Nigeria is predicted to be one of the most conflict-prone rivers according to our research and is also included in the list by Yoffe et al. However, the IRCC do not record any conflictual events in this basin, and only two neutral events involving consultations between riparian countries. Still, according to our work and Yoffe et al., conflict is likely here.With regard to the third column, we obtain evidence that it is worth studying river basin conflict and cooperation side-by-side.44 This column shows that five basins, which are categorized as basins at risk by Yoffe et al., are predicted to have at least 90 percent of dyad-basin-years with a cooperative median according to our estimations: Indus, Jordan, Mekong, Nile, and Senegal. In addition, our work suggests six more basins are likely to be cooperative in the future. The Indus River, which is shared by India and Pakistan, is worth examining in more detail. Yoffe et al. categorize this as a basin at risk and descriptive statistics for the IRCC data suggest the same: 14 (29 percent) of the forty-nine events coded for this basin between 1997 and 2007 are coded as conflictual. Note, however, that these descriptive statistics may be misleading. In fact, our in-sample and out-of-sample work suggests that we might observe more cooperative interactions in the future.In other words, our approach matches some of Yoffe et al.’s predictions but also identifies many differences. Assuming that the IRCC and the TFDD data cover the same underlying theoretical concepts,45 the dissimilarities are probably caused by different methodological approaches. We believe that our approach leads to more accurate predictions, however. As stated above, bivariate regression models, which were used for the 2003 study, can produce information on statistical significance; but they cannot generate information on relationships between variables in other samples or with regard to the predictive/forecasting power.Differences in methodology also result in differences concerning predictors. Yoffe et al. identified the following predictors to be most crucial: high population density, low GDP per capita, overall unfriendly relations between riparian countries, politically active minority groups, proposed large dams or other water development projects, and limited or no freshwater treaties. While some of these variables could not be included in our setup due to limited data availability (e.g., politically-active minority groups) or because they are already coded within the IRCC scale (e.g., water development projects and treaties), it is interesting that only the population variable – in our setup, Population, Small (log) – appears as a robust predictor for conflict (first stage in Figure 3) and cooperation (third stage in Figure 4). GDP per capita, for example, which is identified by Yoffe et al. as an important predictor and is frequently used in other studies of river basin interactions46 and armed conflict,47 is unlikely to help us in anticipating future river conflict or cooperation.These differences notwithstanding, the majority of variables classified as the strongest predictors in our models support other studies that emphasize river characteristics and water availability.48 Other determinants, which are also frequently used in explanatory models of river conflict and cooperation, such as IGO Membership or Democracy, are unlikely to play a major role. Note, however, that the lack of predictive power for some of these variables that are seen as strong predictors in studies on civil war, for example, may be related to the absence of violent conflict in our data. The two GDP per capita variables as well as regime type are arguably the most prominent cases in this regard. That being said, civil war onset is a different outcome than conflict and cooperation over international river basins. Moreover, when confronting the statistical (ex-post) evidence for democracy with its prediction/forecasting power, Ward et al., for example, find that regime type is actually one of the weaker predictors.49 Our point here, in line with Ward et al.,50 is that findings exclusively based on statistical (ex-post) evidence may be misleading, and that prediction and forecasting are more powerful tools in this respect.We have sought to put ex-post empirical models of international river basin conflict and cooperation to a harder test by examining their ability to predict and forecast variation on the outcome variable of interest. The empirical models we estimated perform well in the in-sample tests (1997–2004), the four-fold cross-validation setup, and the out-of-sample forecast that allows for a comparison of predicted conflictual as well as cooperative dyad-basin-years with empirically observed cases (2005–2007).We were motivated to reassess the predictions of Yoffe et al.’s seminal “basins at risk” study. This study received strong attention from policy and academic circles, but was ahead of its time because large-N research on international river basin conflict and cooperation and prediction and forecasting methods were still in their infancy. Our main contribution is to augment the most advanced quantitative research on river basin conflict and cooperation with a prediction/forecasting approach, and to revisit the basins at risk issue. Our approach has produced a set of predictors supported by more robust empirical evidence that generates a substantially revised list of basins at risk. Although none of the river basins identified appears likely to experience a “water war,” our approach and its results can help policy-makers by drawing their attention to basins that are likely to require greater effort in conflict prevention and/or resolution.We see several opportunities for further research. First, case studies should focus in greater depth on individual river basins to re-examine some of our findings that may appear counterintuitive. Examples include the Cross River in Nigeria (see above) and the Aral Sea basin. The latter is listed as a basin at risk by Yoffe et al., but is not identified as conflict-prone by our in-sample and out-of-sample predictions for 1997–2007. However, it only barely drops out of our list, being just below the “90 percent conflictual dyad-basin-years threshold” in the out-of-sample predictions. Moreover, while our results suggest that that forty-four basins are at risk, the international community will be unable to alleviate the conflict potential in all those basins. Fortunately, most of the conflicts are likely to be at a sufficiently low level that the riparian countries may be able to resolve them readily. Case studies could thus examine more thoroughly, which basins are likely to have conflicts that go beyond the conflict resolution capacity of riparian states, or why some basins experience both cooperation and conflict, within specific dyad-basin-years or over time, whereas others are dominated by one of the two interaction types.51 The Indus River provides a good example.Second, our approach reduces information on conflict and cooperation intensity to binary variables, whereas the IRCC scale ranges from −6 to +6. Since extremely conflictual or cooperative international water events are rare, this approach is defensible.52 Further research could explore alternatives to using the median value of the IRCC scale for dyad-basin-years as extreme conflictual or cooperative events might have been averaged out with our approach (in many cases potentially to 0).Finally, future research could employ the prediction and forecasting approach used here to strengthen explanatory models of river basin conflict and cooperation using other outcome variables, e.g., militarized interstate disputes, river treaties, or alternative event data such as the TFDD. Such work could be linked to the broader literature on conflict and cooperation. Our approach might also shed light on other questions of interest in research on international environmental politics, such as explaining variation in environmental performance across countries or participation rates in and compliance with global environmental agreements. Adding prediction and forecasting to standard empirical models would also make them more policy-relevant, because policy-makers are at least as curious as researchers with respect to future trends, developments, and outcomes.