Best models go NEG.
Gallea and Rohner 21 – Quentin, PhD, postdoctoral researcher in political economy at the University of Zürich. Dominic, Faculty of Business and Economics, University of Lausanne. “Globalization mitigates the risk of conflict caused by strategic territory”, Proceedings of the National Academy of Sciences, Vol. 118, No. 39, https://doi.org/10.1073/pnas.2105624118, 09-28-2021
This contrasts with an intellectual tradition of arguing that globalization, business, and trade may—by fostering interdependence—curb the incentives for engaging in domestic and international conflicts. This argument has its roots ranging as far back as the thinking of De Montesquieu [1758 (2)] or Angell [1909 (3)], has been refined verbally by an array of “liberalist” scholars in international relations (see the survey in ref. 4), and has recently been scrutinized in formal game-theoretical models (see, e.g., refs. 5 and 6).
While a small body of empirical research has linked trade to interstate wars (4, 5, 7, 8), the arguably even more pressing question of how globalization and trade affect domestic conflicts has received even less attention.† This is a major gap in the literature, given that since World War II, roughly 80 to 90% of wars have been within rather than between states (11). Hence, the goal of the current article is to study the question of how places close to strategically important trade routes may be more or less subject to civil conflict and how their fate is affected by surges in globalization.‡
We have built what—to the best of our knowledge—is the most precise and fine-grained dataset of strategic location importance covering the entire globe. Our dataset allows us to investigate, using a regression analysis, how a location’s strategic centrality affects its risk of being drawn into an armed conflict and how globalization can influence this centrality-conflict nexus.
In order to develop an empirically testable hypothesis to guide our statistical investigation, we have built a game-theoretic model that systematically studies the incentives for engaging in conflict (SI Appendix). As discussed in detail in SI Appendix, our framework predicts that under mild conditions in years of low international trade openness, strategic territory tends to entail above-average levels of conflict (due to the strategic value of territory providing incentives for appropriation). In contrast, in years of roaring globalization, areas with strategic value are, on average, less combatted, as major international powers have incentives to intervene in local disputes to make sure that crucial trade routes remain open.§ In what follows, we will present the data and methods, before confronting these predictions to the data.
Data and Methods
To carry out our empirical analysis, we have constructed a panel dataset that consists of grid cells of size 0.5 × 0.5 decimal degrees (55 km × 55 km at the equator) covering the whole world from 1989 to 2018. We have designed an algorithm to detect strategic zones in the sea, so-called maritime “choke points” (e.g., straits or capes) that are points of “natural congestion along two wider and important navigable passages” that are typically of key strategic importance for international shipping. Crucially, we have built this measure purely based on geographical features, which has the advantage of addressing a series of potential statistical biases—such as reverse causation—that would arise if we were to focus on measures of actual water transport volumes.¶ In particular, drawing on a network model, we compute the betweenness centrality of any water location, allowing us to uncover where crucial strategic choke points lie. SI Appendix contains detailed variable definitions and sources and a full discussion of the construction of all variables.
It is widely accepted that maritime choke points are of crucial importance to world trade and global energy security. Our algorithm identifies real ship density and all famous maritime landmarks, such as the straits of Hormuz or Malacca and the canals of Suez or Panama. Furthermore, our measure provides a fine-grained scale of strategic importance for any water spot worldwide, including the great number of less well-known locations. Fig. 1A depicts for each water location how close it is to a choke point, as computed using our algorithm. Fig. 1B displays for illustration major marine traffic routes (observed density of ships in 2017 from https://www.marinetraffic.com/). Strikingly, the proximity to waterway choke points, as computed by our algorithm based on purely geographical features, matches remarkably well the actual marine trade routes, hence stressing the relevance of our measure.
After having computed strategic water choke points, we have then, in a second step, constructed, for all land locations, the distance to these waterway bottlenecks. The obtained values across the world are displayed in Fig. 1C. Darker colors indicate areas closer to choke points, which typically lie close to major straits and waterways, and brighter colors indicate zones that are further away from maritime choke points. This measure of the strategic importance of any land location worldwide is used as the main explanatory variable in our statistical analysis. We study its direct impact as well as how it interacts with the volume of world trade in a given year, which is measured by using world trade openness from the World Bank [trade in percentage of gross domestic product (GDP) (18)]. SI Appendix contains a graphical representation of the evolution of this variable, as well as of other key covariates.
In terms of the dependent outcome variable, for measuring conflict, we draw on fine-grained geolocalized information on conflict events from the Uppsala Conflict Data Program (UCDP) Georeferenced Event Dataset (GED) (19). This allows us to know for each cell and year whether at least one conflict event took place, as well as the types of events and their number. As mentioned, all data, variable construction, and methods are described in detail in SI Appendix.
Descriptive summary statistics of all variables of the analysis are provided in SI Appendix. In a nutshell, the final sample is composed of 64,818 cells covering the world from 1989 to 2018, resulting in a total of 1,944,540 observations. The unconditional likelihood for any type of violent events for the whole sample at the cell level is 1.5%, while for state-based it is 0.7%, for nonstate 0.2%, and 0.6% for one-sided events. The mean number of deaths is 1.126 per year per cell. The difference between the mean value of the outcome for the cells “close” to choke points (above median by proximity) and for those “far away” (below median by proximity) is also displayed. The difference is statistically significant and positive for any type of violence (using a t test with a bilateral null hypothesis; SI Appendix). These results highlight a positive association between the proximity to waterway choke points and violent events, which we shall investigate in more depth in what follows.
In terms of the methodology used, we carry out a multivariate regression analysis, focusing on Linear Probability Models (LPMs) when facing a binary dependent variable and on Ordinary Least Squares (OLS) estimators otherwise. We will include a battery of fixed effects, filtering out time-invariant location characteristics, as well as global shocks. Specifically, in some specifications, we go as far as including fixed effects at the cell level (i.e., separate constant terms for each cell), which control for all local, time-invariant potential confounders such as local climate, elevation, sea access, distance to capital, and historical population density, among others. We also control for annual time effects, which analogously capture all global shocks occurring in a given year, such as, for example, major geo-political shocks like the fall of the Soviet Union or 9/11, major recessions such as the subprime crisis, or health shocks such as a pandemic (e.g., severe acute respiratory syndrome or COVID-19). The various specifications, as well as additional estimation results, are reported in SI Appendix.
Results
We start by running a very simple regression specification before gradually increasing complexity. In particular, we focus first on comparing areas with high strategic importance scores (according to our measure) with cells for which our algorithm has found a lower strategic importance (i.e., that are further away from maritime choke points). Our main explanatory variable is “proximity” (to the nearest choke point), and the dependent variable of interest is the likelihood of experiencing at least one violent event in a given cell and year. The goal of this initial table being to report the parsimonious “raw” correlation, we limit ourselves to controlling for latitude fixed effects (i.e., a specific constant term for each latitude, filtering out climate zone effects and earth perimeter# ) and annual year dummies (which account for global shocks). All methodological details of this specification are provided in SI Appendix.
The regression analysis of Table 1 reveals that overall areas closer to maritime choke points face a greater risk of conflict, as shown by the fact that in all columns, the proximity variable has a positive, statistically significant coefficient. This holds when including a dummy for any violent event (column [col.] 1) and also for various subcategories of violent events (col. 2–4). It is imprecisely estimated for a violence intensity measure (col. 5). The effect is quantitatively sizable, as one SD greater proximity (i.e., 1,100 km closer to a choke point, corresponding to the straight-line distance from Paris to Rome or New York to Chicago) in the main specification (col. 1) corresponds to a 0.31-percentage-point increase in conflict risk, which is about a fifth of the baseline conflict risk for a given cell and year (1.5%). Note that the results of col. 2–4 show that the quantitatively largest effect emanates from state-based conflict (col. 2).
Next, we investigate the main prediction of our game-theoretic model (SI Appendix), namely, that while proximity to maritime choke points increases the conflict risk for moderate levels of trade openness, for peak levels of globalization, the prediction reverses, and locations of strategic importance are expected to benefit from a relatively low likelihood of conflict. We illustrate graphically how the locations of strategic choke points relate to conflict events—both for periods of high trade (Fig. 2 A and C) and low trade (Fig. 2 B and D). We zoom in on key strategic regions: Panama Canal and Cape of Good Hope (a full map of the world is depicted in SI Appendix). Visual inspection suggests—in line with our predictions discussed above—that strategic territory may bear a conflict potential, in particular, during periods of low trade, while in times of high trade volumes (when major powers are particularly keen to keep world trade routes open and secure), conflicts may be less concentrated around choke points. While these associations are interesting, they could be driven by various confounders, and, hence, we need to perform in what follows an in-depth regression analysis that allows us to control for confounding variables and statistical biases.
At present, we move to a regression analysis with this interactive effect. Note that SI Appendix presents a simplified regression specification (featuring the same controls as in Table 1) and provides all methodological details for the more demanding main specification that we shall now discuss. This main regression specification features, as before, as dependent variables several measures of violent events. As a main explanatory variable, we still focus on the proximity to maritime choke points, but now not only as a linear term, but also in interaction with a measure of world trade openness (imports plus exports) in percentage of world GDP. In this main baseline specification, we include a more stringent set of controls. As before, we control for annual time dummies (which account for global shocks) and latitude fixed effects (capturing, among others, climate zone effects, earth perimeter, and cell size), but now we also control for country fixed effects. These different constant terms for each country allow us to control for any time-invariant country characteristics (such as colonial heritage, tradition of autocracy, country size, geographical features, etc.), and, hence, our identifying statistical variation stems from comparing different locations of the same country (e.g., Medellin with Bogota or Miami with Nashville). Note that controlling for annual time dummies picks up the world trade openness measure (which takes the same value for each country and varies annually), which, hence, is dropped.
The results are displayed in Table 2. Consider the main specification of col. 1, where the linear effect of proximity has a statistically significant positive coefficient, whereas its interaction with world trade openness has the expected negative sign. This means that strategic territories face, on average, a higher conflict risk in periods of low trade openness, while with greater trade openness, they are relatively more shielded from armed conflict, which is fully consistent with our game-theoretic model in SI Appendix. This result carries over for subcategories of conflict (col. 2–4) and for a conflict intensity variable (col. 5). The results of Table 2 are represented graphically in Fig. 3.
The impact is quantitatively sizable, as moving one SD (1,100 km) closer to a choke point increases by 0.4 percentage points∥(24.8% of the unconditional baseline risk) the conflict likelihood in periods when trade openness is low (0.4), while reducing it by 0.2 percentage points** (12.1% of the conflict baseline risk) when trade openness is high (0.6).
In SI Appendix, we present the details of all specifications used in the main text, in addition to results for alternative specifications. In particular, we go one step further by running the same regression, but including controls for cell fixed effects. These constant terms are specific to each cell of 0.5 × 0.5 decimal degrees (55 km × 55 km at the equator) and, hence, filter out all time-invariant characteristics of this very fine-grained local area. In particular, this controls for the potentially confounding impact of elevation, microclimate, sea access, ruggedness of terrain, river proximity, and historical road network, to name a few. This specification is described in detail in SI Appendix. It is shown that all our results go through in this demanding specification and that the interaction term of interest between the proximity to maritime choke points and world trade openness continues to have a statistically significant negative sign in all specifications.