Books Discussed in this Review Essay

Cederman, Lars-Erik, Kristian Skrede Gleditsch, and Halvard Buhaug. Reference Cederman, Gleditsch and Buhaug2013. Inequality, Grievances, and Civil War. New York: Cambridge University Press.

Starr, Harvey. Reference Starr2013a. On Geopolitics: Space, Place, and International Relations. Boulder, CO: Paradigm.

Stasavage, David. Reference Stasavage2011. States of Credit: Size, Power, and the Development of European Polities. Princeton, NJ: Princeton University Press.

Space has always been a central, if sometimes implicit, element in the study of international relations. Where things happen is a key factor in what happens and why. Although spatial relationships have often been analyzed in a general way, new tools have recently enabled more sophisticated studies. In particular, Geographic Information Systems (GIS) offer techniques for collecting and analyzing data in an explicitly spatial fashion, and these tools are being incorporated with increasing frequency into the study of international relations—without, unfortunately, adequate consideration of all of the implications of GIS as a method. This review essay considers how GIS has been applied and suggests how this set of tools can be made a more effective and appropriate technique for international relations, in particular by focusing on the assumptions of GIS, pointing out particular problems that arise with its use in this field, and then outlining possibilities for ameliorating them.Footnote ¹

The increased capabilities offered by GIS make possible studies that would otherwise be prohibitively difficult to carry out. International relations scholars have developed new data sets and variables, disaggregated and merged data in new ways, and tested novel spatial hypotheses. Yet the opportunities offered by GIS also entail potential pitfalls. GIS might be new in its application to international relations, but it should be subject to the same rigorous examination as other methods have been, such as statistical analysis or qualitative causal inference.Footnote ² In fields such as geography, history, and sociology that are more closely involved with the creation and use of GIS, scholars have critiqued GIS extensively.Footnote ³ Unfortunately, this discussion has been largely absent in the subfield of international relations even though the potential problems with GIS are especially relevant to our subject. Although GIS enables the collection and analysis of an extraordinary range of data tied to spatial position, that position is defined in one way only: as points on an abstract grid. Those points are used to demarcate lines, polygons, or pixels, and relationships are then analyzed among the resulting spatial entities. GIS, in short, offers a sophisticated set of tools for analysis built on geometric measures of spatial areas, distances, and relationships.Footnote ⁴

But what if the phenomena being studied are not effectively captured by the geometric tools of GIS? In other words, what if there is a fundamental mismatch between the assumptions built into GIS and the political ideas, practices, and institutions we in international relations seek to measure, analyze, or understand? In particular, two specific challenges emerge from the application of GIS to international relations: measurement validity and selection bias. Measurement validity may be undermined when institutions, practices, ideas, or behaviors are not accurately described by the points, polygons, or pixels used in the GIS data set. A selection bias problem appears when the focus on spatial entities or ideas prejudices unit selection, removing nonspatial units from the analysis.Footnote ⁵ If not addressed directly, both of these issues threaten the ability to make valid causal inferences using GIS.

These problems are not inherent in GIS as a method but instead have to do with the particular way in which GIS is applied: how data are coded in GIS data sets, the types of data available for those coding choices, and how data are then analyzed. Existing uses of GIS in international relations include both problematic applications and potential solutions. Because these two specific challenges have not been sufficiently highlighted, we do not yet have good means of distinguishing between more and less effective applications of GIS. In addition, there is a close fit between some of the deeply embedded assumptions of GIS and the implicit assumptions of international relations, creating a situation in which some applications appear to solve these problems without actually doing so.

In short, GIS is unquestionably a powerful tool, and its use is beginning to yield impressive results in international relations. Nevertheless, these particular challenges have not been sufficiently addressed, undermining our field's ability to use this tool to make valid causal inferences and thereby to accumulate knowledge.

Measurement Validity: Challenges

GIS offers a sophisticated set of tools for measuring and coding spatial variables. Yet, as I discussed, GIS encodes spatial information in a particular way: building on a coordinate grid and then either drawing points, lines, and polygons (vector data) or coding predefined pixels with data values (raster data). The question then becomes whether the relevant characteristics of political institutions, practices, or behaviors can be captured by data defined in this way. Of course, nearly every political action takes place somewhere, but that does not mean that the concepts being measured can always be described accurately in the geometric terms of GIS. Measurement validity is undermined when the features of the units, or the relationships between units, are not measured accurately by this particular data structure.

Many studies take political institutions as their unit of analysis, and those institutions sometimes fit the spatial characteristics of data in GIS, but they do not always do so. For example, the ideal-typical modern state can be coded as a polygonal territory defined by discrete, linear boundaries (although even this is not entirely unproblematic). Yet in some circumstances, both historical and contemporary, political organization has taken on forms other than territorial jurisdictions, which creates implications for the measurement validity of GIS data sets seeking to code those organizations.

For example, studies of the development of historical political institutions, including Stasavage's States of Credit and a number of articles,Footnote ³⁴ use GIS to build spatial data that allow for new forms of empirical testing of important theoretical questions. These studies make use of Euratlas, a GIS data set of political units and boundaries in Europe from antiquity through today.Footnote ³⁵ This data set essentially takes the traditional historical atlas and puts it in GIS form, coding political units and their boundaries as polygons. Unfortunately, this framework does not provide a valid measure of the spatial extent of political authority before the modern era. In brief, the depiction of political units in the fashion of today's maps—with clearly delineated boundaries separating homogeneous territories—misrepresents the actual character of rule during much of the premodern era, and it is particularly problematic for the European Middle Ages (a period included in many studies).Footnote ³⁶ The Euratlas data set has a layer for “sovereign states,” with each state defined by a linear boundary and coded as having exclusive and complete authority over the space within. The data set, in fact, explicitly does not allow for claims by multiple parties to the same territory: the only overlap incorporated into the data set is the hierarchical nesting of “dependent states” and substate “administrative” units within “sovereign states.”

Although the territorially demarcated state usefully describes political organization in many parts of the world today, it runs directly contrary to several decades of research on the complexity and nonterritorial nature of medieval political organization and authority.Footnote ³⁷ Claims to rule were made over persons, places, and nonterritorial jurisdictions, and were only very occasionally asserted over boundary-defined spatial expanses. In spite of the variety of political organizations in the Middle Ages, “in none of these was authority contingent on, or defined by, control over a specific territory with fixed boundaries.”Footnote ³⁸ At the broadest level, of course, the political depictions in Euratlas or in traditional printed historical atlases can give a sense of the spatial extent of some rulers' domains or of how different kingdoms related to each other geographically. Yet the data imply a deceptively high degree of certainty about boundaries and about the types of political units depicted. The coding used in the GIS data set applies to jurisdictions defined by polygonal units, but it fails to describe jurisdictions of a nonterritorial nature. This represents a fundamental problem of measurement validity: the indicators used in the GIS data set do not “adequately reflect the concept the researcher seeks to measure.”Footnote ³⁹

One study, for example, relates representative institutions to the overall size of a polity,Footnote ⁴⁰ which the Euratlas GIS data set captures on a broad level—the polities drawn as larger “states” were almost certainly larger. Scale, however, is a complex concept, as this study recognizes in a discussion of the difficulty of adjusting for travel and communication by roads, rivers, coastlines, and so on.Footnote ⁴¹ Yet there is also a more fundamental issue inherent in relying on this GIS database of boundaries. In Euratlas, the state boundaries are used to define who is a member of a particular polity—and thus who is potentially involved in the representative institutions being studied—but that membership often had little to do with the boundaries drawn by the Euratlas data. What about jurisdictions subject to overlapping claims? What about the inhabitants of polities that were not defined territorially? Who was involved in the creation or maintenance of a representative institution, therefore, is not validly measured by a GIS data set that focuses on the where of “state boundaries” that did not, in fact, exist.

The validity of linear boundaries as a measure of the extent of political organization changes over time, an evolution that is not captured by the Euratlas data set. By the eighteenth century, most polities within Europe were indeed defined by demarcated borders. In the late Middle Ages, however, political organization did not reflect Euratlas's linear boundaries between “sovereign states.” This systematic error, across time, could significantly bias results. Unfortunately, the problem is difficult to perceive because the potential bias in the data set matches a common underlying assumption in international relations—that polities are defined by territorial demarcation.

The potential issue of measurement validity when using GIS applies not only to historical work but also to research focused on contemporary international politics today. Although studies of contemporary international relations consider subjects that might be more likely to fit GIS's spatial assumptions, GIS can still “over-territorialize” the analysis. In other words, the danger remains that nonterritorial variables or processes are being measured as territorial. For example, the use of GIS to delineate ethnic groups is a major advance, seen in data sets such as GREG, discussed earlier.Footnote ⁴² The ethnic group is posited (convincingly) as a more appropriate level of analysis, between the overly narrow individual actor, on the one hand, and the overly aggregated state, on the other. Theories about ethnic conflict support this choice of an intermediate level of analysis.Footnote ⁴³ What those theories do not necessarily support, however, is the choice to treat ethnic groups and the relationships among them in terms of delineated spatial areas. These data sets, thus, still rely on the state form of territoriality to delineate units, even if the specific boundaries drawn are of ethnic groups, not states. Are ethnic groups—collections of people sharing an identity—always best measured by territorial delimitation? When approaching that question with GIS, the answer is, by default, positive. And for groups with a distinct settlement area, the measurement of spatial features using GIS can be effective. Yet for other types of groups, such as ethnic diasporas, polygonal data may not achieve valid measurements. These potential problems have been difficult to discern because they are often hidden in the assumptions and data structure of GIS. The appearance of universal coverage (such data sets often include polygons for groups worldwide) masks the way in which polygonal units may be ineffective measures of the concept in question.

There are, of course, political institutions that are unquestionably territorial. Data on contemporary international boundaries, for example, can code a polygonal institution (sovereign states) with polygonal data. (The “CShapes” data set, for example, covers state boundaries 1945–2012.)Footnote ⁴⁴ Fortunately, the use of a geometric form of territoriality is not conceptually problematic in the modern period in the same way it is for data sets covering earlier eras. At least in terms of the ideal of statehood toward which most governments aspire, today's international system is defined by discrete boundaries delineating exclusive claims to rule—in other words, by polygons. Yet problems can still arise. For example, in territories subject to violent rebellion, the situation can also be complex in a way that this GIS data set is unable to capture: even those rebel groups that exercise some degree of governance over territory may have no interest in building a “real” state with delineated boundaries or international recognition.Footnote ⁴⁵ In some regions, in other words, political rule can be validly measured by discrete boundaries between states, but in other regions those lines systematically fail to capture the consequential elements of political organization and action. Similar issues can arise when examining other institutions in contemporary international politics. International organizations, for example, often have a spatial component, in terms of their membership, activities, jurisdictional scope, and so on. Yet it is not clear what form of spatial data coding would be easily appropriate for measuring the features of these organizations.

The challenge to measurement validity goes beyond the issue of including nonspatial cases in a spatial data set (the obverse, in a sense, to the selection bias problem discussed later). Measurement validity is potentially threatened in analyses where both spatial and nonspatial units are included in a single data set and where all of them are measured using the same spatial variables, with analysis carried out based on those values. For example, a data set of “all political units in year X” should include both spatial and nonspatial units, but the same measures of spatial characteristics should not be applied to both types of units. That is the challenge faced by certain applications of GIS, and it is one that poses significant risks for making valid causal inferences—especially for studies in which the specific concepts that are vulnerable to mismeasurement (such as size or distance) are central to the arguments being tested.

As with any sophisticated method, potential challenges like these can be missed when they are related to basic assumptions and data structure. The problem is especially acute in this case: some GIS data sets, for example, are designed specifically to be accessible to scholars who have little familiarity with the technicalities of GIS and who, thus, are unlikely to see the mismatch between concept and measurement. In those analyses, therefore, the spatial assumptions of GIS, which were integral to the creation of the data set and its limits, are sometimes completely hidden.

Measurement Validity: Solutions

These challenges to valid measurement can be overcome. The first step should be to make an empirically and theoretically informed decision about the spatial structure of the units or processes in question, and then seek to create a GIS data set reflecting that decision. In other words, the determination of a theoretically appropriate level or scale of analysis should be matched by a careful selection of the proper spatial or nonspatial form of the units in question. In many cases this will be straightforward, since there are phenomena that can be effectively measured using the default spatial assumptions of GIS. Structures that are less institutional than physical, such as trade routes, urbanization, or land use, can be usefully analyzed in GIS, as can purely geographic variables. The hypothesized causal effect of mountainous terrain, for example, does not rely on what the actors involved think of the terrain; what matters is the terrain itself.Footnote ⁴⁶ In addition, polygonal measures of spatially defined political institutions—such as electoral districts, territorial subnational units like US states, and others—are unquestionably appropriate: those units are delineated as polygons both in law and in practice.Footnote ⁴⁷ As the Cederman, Gleditsch, and Buhaug book notes, the best practice is always to use “theoretically relevant data and measures, rather than the standard toolbox.”Footnote ⁴⁸

When incorporating complex or nonspatial political or social variables, however, specific attention needs to be paid to addressing the possibility of mismeasurement. Three potential solutions emerge from existing work: using points instead of polygons, expanding how boundaries between units are defined, and incorporating network analysis alongside spatial data.

Moving away from polygonal data as the default way to construct data sets in GIS is a potentially powerful solution that is being increasingly adopted in a variety of international relations studies. The main advantage of this approach is that it can sometimes match more closely the measurement to the underlying concept. For example, for the medieval political organizations I discussed (erroneously coded as polygonal spaces in Euratlas), maps and textual descriptions from the period reflect the idea of political authority claimed over towns as places, rather than over territory as a space.Footnote ⁴⁹ The distinction between place and space is essential to how historians and geographers confront measurement challenges in GIS, but it has remained unexamined in international studies, with a few notable exceptions.Footnote ⁵⁰ Places are locations held to be important for their particular characteristics rather than for their inclusion within a delineated spatial expanse. When space is the focus, ideas and practices are defined by bounded, expansive areas. Relying on textual knowledge and place-focused maps meant that medieval rulers' understanding of the territorial extent of their rule was fundamentally different from the modern, mapped ideal of the sovereign state. The interests, actions, and goals of those rulers were shaped by these ideas; after all, a polity is a collection of social institutions, not a geographic feature, and has no real-world existence outside of the ideas, practices, and behaviors of the individuals and groups that create and interact with it.Footnote ⁵¹

Thus, one effective approach could be to move away from polygonal data and use points to define locations, which can be described in a variety of quantitative and qualitative ways. A careful reading of historical texts can reveal exactly how political rule was conceptualized—and practiced—allowing the creation of a data set that incorporates point locations instead of polygonal units when point locations are more empirically valid. Historians using GIS have wrestled with exactly these difficulties, acknowledging that in some cases points may be a more appropriate form of coding than polygonal units. The author of one GIS data set project on China notes that, for much of the historical period covered by the data set, Chinese administrative units were defined more often by their administrative seat at the center than by their boundaries: “GIS is about geographic space, but in the historical record, place, not space, was the focus.”Footnote ⁵²

Likewise, today's complex political phenomena that are not captured by linear divisions and polygons—the distribution of diaspora ethnic groups, for example—might be better described by points. Several recent armed-conflict event data sets code conflicts and violent events by their point location rather than by aggregating total violence to the state (polygonal) level.Footnote ⁵³ Similar advances are being made in geocoding development data in terms of point locations, such as the sites of foreign aid projects.Footnote ⁵⁴ This could usefully apply to development outcomes as well—subnational variation in outcomes is important, and could be better captured by measurements at point locations (towns, neighborhoods, etc.). This approach, in short, could be applied to any event, process, or institution that has more of a point location than a spatial area. The fact that there are existing studies doing this shows how it can be done, but since the potential problems with polygonal coding have not been highlighted, studies have often continued to rely on polygonal data even when other options might well be more appropriate.

Another possibility for making GIS fit better with political institutions or phenomena that have varied spatial structures is to define the concept of a boundary between two units more broadly, that is, to include options other than lines. Thus the first question should not concern where a boundary between two units falls but rather what the nature of the boundary between those units is. One study of urban neighborhoods by sociologists illustrates this approach: “Our method … allows for the possibility that the boundaries between neighborhoods might be sharp edges or larger zones of transition.”Footnote ⁵⁵ The coding of such boundaries—their type as well as their location—thus builds on explicit theoretical assumptions or empirical findings, not on the implicit assumptions built into polygonal vector data sets.Footnote ⁵⁶

In international relations, one of Starr's projects aims to code international boundaries in terms explicitly intending to “go beyond simply observing the number of borders a state possesses, whether or not a border existed between two states, or the length of that border.”Footnote ⁵⁷ Instead, borders can be described by a number of variables, including, potentially, the very nature of the border itself: discrete versus fuzzy, porous versus impermeable, fortified versus undefended, and so on. Again, this is an existing practice, but without highlighting the potential challenges built into the polygonal data, it has not been made clear just how useful this approach could be.

Finally, some institutions or actors are better measured and coded as networks than as purely spatial entities, even when those institutions or actors do have spatial locations. As noted, recent work has combined spatial analysis using GIS with network analysis tools.Footnote ⁵⁸ Thus the spatial elements—locations, distances, surface areas—can be measured effectively by GIS, but complex nonspatial features or relationships—authority relations, cultural links—do not have to be coded in a spatial framework. This could offer one route to resolve, for example, the difficulties of coding premodern political institutions. Many of the problems that arise when using the Euratlas data set mentioned earlier result from the purely spatial (and polygonal) nature of the political units as they are coded. If, instead, one combined locational data (in some cases as points rather than as polygons) with network data on the connections and interactions among different places subject to a political claim, the combined data would more accurately capture the complex dynamics of rule before the advent of the territorial state and, thus, would enable more valid inferences to be made about the origins or effects of those institutions.

This would, of course, present a new issue: a huge amount of data would need to be gathered, versus the ready-made solution offered by Euratlas data. That is a significant challenge because it is far easier to include only the spatial element; the other information is much harder to come by. Yet a GIS data set can present the appearance of complete coverage (encompassing, for example, the entire land surface in question), thus obscuring the need for further, more complex, data. Nevertheless, scholars should aim for the kind of spatial and nonspatial coverage that would more effectively measure the phenomena in question, and combining network analysis with GIS may offer one means to do this. Similar solutions could be applied to political settings today in which diverse spatial and nonspatial units and relationships coexist: for example, areas of limited statehood, contested authority claims in civil war zones, or the complex overlapping authorities of international organizations.

Valid measurement using GIS, in short, can face significant challenges, but those challenges are by no means insurmountable. Solutions have appeared already in a number of studies, all building on the essential point that the measurement needs to match the underlying concept, not just in terms of scale but in terms of the spatial character—or nonspatial character—of the phenomena in question.

Selection Bias: Challenges

The use of GIS in international relations can also lead to selection bias, a well-known methodological problem for both quantitative and qualitative studies.Footnote ⁵⁹ In the case of GIS-based analysis, the potential bias is against any units or cases that are nonterritorial or nonspatial and thus are not as easily incorporated into a GIS data set. This can be particularly challenging for studies that take political institutions as the units of analysis: spatial units that fit the framework of GIS are incorporated, while nonspatial units are not. Although this point may seem so obvious that it is hardly worth stating (spatial data leave out nonspatial units), applications of GIS to complex institutional environments can create a situation in which that bias, while sometimes evident, is not always perceived to be a problem. The result is biased data that, while still presenting the appearance of complete coverage, end up excluding units that could prove useful for testing the hypotheses at hand. The excluded units, moreover, can differ not just in being nonspatial but also in terms of the key variables and concepts that are the focus of the theoretical arguments being tested, thereby potentially undermining causal inference.

For example, the use of polygonal data sets to code historical political units removes nonterritorial units from analysis and simultaneously gives the incorrect impression that the study relies on a complete and exhaustive data set on political organization within a given geographic area. One study explicitly uses Euratlas data to define the units of analysis (that is, those political entities designated in the data set as “sovereign states”), thus excluding nonterritorial polities.Footnote ⁶⁰ Likewise, “autonomous cities” are omitted from another analysis because there are “insufficient GIS data about their exact geographic scale.”Footnote ⁶¹ In the case of these cities, however, scholars lack accurate data about their geographic scale, not because their data are incomplete, but because those cities were not defined in terms of territorial delimitation. Instead, they were understood as places with some jurisdiction over surrounding areas but without discrete territorial borders.

The excluded nonterritorial units include political organizations that were anything but marginal. For example, the Hanseatic League was a confederation of independent trading cities spread across Northern Europe—it is the most well known of a number of city-leagues of the Middle Ages. These leagues performed many of the same functions that states did, and for centuries they coexisted with the earliest examples of territorial states.Footnote ⁶² Spatially, however, city-leagues were completely different from states: the several towns were geographically dispersed, and many had only limited control over their local hinterlands. Although these leagues often had representative institutions, the leagues are left out of the analysis because they were not territorially delimited and thus do not appear in a database containing only “state-like” polities. Their geographically dispersed nature might have provided additional leverage on the question of the link between spatial scale and representation, but only if the concept of spatial scale can incorporate units that are not territorially delimited. Likewise, the analysis of “ruler duration” could be altered by including these polities that do not fit our territorial expectations but that nonetheless served as an effective means of political organization.

This critique does not mean that findings generated using Euratlas will necessarily be incorrect. Ideally, it would be possible to construct an alternative data set incorporating all units and their relationships, both spatial and otherwise, thereby more effectively testing the same hypotheses. Unfortunately, that would be prohibitively difficult for the same reason that many studies rely on data sets such as Euratlas—because GIS and polygon-defined units are the shortcut that makes such studies feasible. Unfortunately that shortcut inadvertently also excludes a set of cases that could alter the results.Footnote ⁶³ With our tendency to assume that the most-important political organizations are spatial, this bias has been difficult to discern.

Similar problems can arise with regard to GIS data sets that focus on contemporary politics: once again, unit selection is biased in favor of “state-like” entities, undermining causal inference when other types of units could systematically differ in terms of independent or dependent variables. Thus, although the GREG data set of ethnic groups includes “stateless” groups that have a well-documented area of settlement, it leaves out groups without a spatially delimited presence in a territory. The authors of the GREG data acknowledge that this can lead to systematic omissions: “GREG only includes territorial groups, that is, those that have one or more settlement regions in a country. Migrant groups and foreign workers residing only in urban areas are not represented.”Footnote ⁶⁴

This absence can be difficult to keep in mind, given the appearance of complete and exhaustive coverage suggested by GIS data sets like this, particularly in the maps they generate that show a completely filled-in geographic area.

Efforts to improve on the GREG data have continued to make progress toward more complete coverage, but they also illustrate that the standard framework used to apply GIS to this topic may be problematic. The GeoEPR data set, for example, has more detail than GREG and represents a significant improvement.Footnote ⁶⁵ Geographically dispersed groups are included, as are urban groups that also have a regional settlement area, but urban-only groups and migrant groups are still left out. Depending on the outcome being studied, or the relative presence of these excluded types of groups in the area in question, this may not be a serious problem. In a number of conflicts, however, a significant role is played by groups without a “settlement region,” such as far-flung diaspora communities, which leads to a measureable impact on outcomes in terms of violence.Footnote ⁶⁶

The creation of new ethnic-identity GIS databases will undoubtedly focus on resolving some of these remaining absences—the authors of both GREG and GeoEPR explicitly note the need to incorporate other types of groups. I would argue, however, that so long as these data sets begin with polygonal areas, they are likely to continue both to refine the coding of certain types of groups (groups with a region of settlement) and to leave out, or to have significantly more difficulty in coding, other types of groups. What appears to be a data availability problem (that is, data on settlement patterns are hard to find for many groups) is actually a problem that originates with the polygonal data structure itself. If that data structure consistently biases selection against particular types of units—including units that, by their very nature, could have potentially different relationships to outcomes—valid causal inference could be undermined.

Like the measurement errors discussed earlier, the potential selection bias in GIS-based studies has sometimes been difficult to discern—GIS applications thus make it appear that there is no problem. There is a correspondence between some of the assumptions built into GIS and our own standard assumptions about territorial political organization, and those assumptions in GIS are embedded in a sophisticated set of technological tools. So, for example, when a data set declares that it incorporates “the geographic boundaries of all political entities covering the earth's surface in the box between the longitudes 15 west and 50 east and the latitudes 20 north to 60 north,”Footnote ⁶⁷ that sounds comprehensive and exhaustive. The resulting virtual map has been filled with directly abutting polygons—the entire land surface of the area in question has been spoken for. This feels accurate and appropriate because international relations scholars approach the study of political institutions with the assumption that the most important units, at the state level or below, are territorially delimited—an assumption that is, in general, well justified today.Footnote ⁶⁸ If that is not empirically accurate for other historical periods or for other types of contemporary phenomena, however, the GIS data set is less than comprehensive or exhaustive. Yet this is hidden from view by the data set's technical sophistication and the deeply embedded spatial assumptions in the method.

Selection Bias: Solutions

Addressing this selection bias issue is feasible, requiring first a recognition of the possibility that any analysis might need to include spatial and nonspatial institutions, groups, or processes. Several solutions then follow: using multiple coding methods to combine nonspatial and spatial units and variables, defining units analytically rather than empirically, and focusing research questions on those units and processes that are not being selected out.

First, efforts can be made to incorporate all relevant units, using coding methods that are theoretically appropriate rather than relying on the default polygonal structure of many GIS tools. Political units, for example, could be coded as polygons, points, or nonspatial network connections—or some combination thereof. This would reduce the tendency to assume that a filled-in spatial expanse indicates an exhaustive data set on all political units, rather than just all political units of a territorially expansive character. In some cases, GIS-based analysis can be used as merely one part of a research design, taking advantage of the benefits in terms of spatial analysis but avoiding too much reliance on GIS for the nonspatial aspects of the questions at hand. Cederman, Gleditsch, and Buhaug make a strong case for this approach, which may, by necessity, be more feasible in book-length studies than in articles. This is, of course, an obvious solution. But, as with many of the other solutions I have suggested here, if the specific nature of the challenge is not emphasized it can be difficult to see the need to make this effort.

Another potential solution is to define units analytically rather than empirically. In GIS, this could involve building raster data sets, where the pixels represent fixed units for which variables are coded. Alternatively, coding can delineate polygons that are explicitly not meant to reflect the boundaries or spatial extent of real-world institutions, social or economic features, or identity groups. Wimmer and Min's study of the effect of the shift from empires to nation states purposely fixes the spatial boundaries of the units at a single point in time to avoid having to line up the boundaries of the units of analysis with the constantly changing political units that are the focus of their study.Footnote ⁶⁹ Likewise, to study shifting territorial boundaries as a causal variable driving later territorial conflict, Abramson and Carter's study divides Europe into a fixed grid of square units for measuring variables and outcomes of interest.Footnote ⁷⁰

Finally, in what may be the most comprehensive effort to date along these lines, the PRIO-GRID data set uses a grid of square polygons covering land surfaces worldwide, with each polygon then being coded with data on both fixed and variable features, from geographic terrain to social or political variables.Footnote ⁷¹ As the authors of this data set note:

This reduces the likelihood of selecting units based on an assumption that certain units (that is, spatial ones) are more important than others. In data sets like these, the units are unquestionably artificial, thus making it easier to keep in mind the difference between the framework of analysis and the ideas and practices being studied. With clearly abstract units of analysis as a foundation, GIS can be a powerful tool for measuring and analyzing spatially distributed variables without biasing selection against nonspatial (and especially nonpolygonal) institutions or processes.

Finally, existing GIS data sets can be effectively applied without selection bias being a problem if the questions being asked work within the structure of the data. As the Cederman, Gleditsch, and Buhaug book notes, “selection bias only applies in relation to specific research questions and is not an inherent property of data sets.”Footnote ⁷³ Thus, in cases where it may be impossible to gather enough accurate information to make a data set incorporating all spatial and nonspatial units, the inherent selection bias against the nonspatial institutions or actors can be avoided by asking questions that focus on explicitly spatial issues. Abramson and Carter's grid-based approach to explaining the role of previous boundary change in territorial disputes represents a clear example of this, since the causal variables and the outcomes are all focused on the types of territorial factors that their GIS data effectively capture.Footnote ⁷⁴ Goemans and Schultz's study of the role of ethnicity in territorial disputes in Africa similarly uses the GREG data set on ethnic identities but does not run into the potential selection bias problem I discussed: the hypothesis being tested specifically concerns the types of geographically expansive ethnic groups that the GREG data set captures well.Footnote ⁷⁵

Moving Forward with GIS

International politics is inherently spatial, but not every phenomenon is spatial in the same way. Some actors or units are conceptually defined as polygons, others are not. Some relationships are understood in terms of geographic distance, and others are better conceptualized in terms of social networks. Some questions can be answered by data that exclusively cover spatially expansive units, while other questions may require a combination of spatial and nonspatial unit selection. The potential mismatch between the embedded assumptions of GIS and the conceptual framework of institutions, processes, and outcomes in international relations thus poses challenges to measurement validity and unit selection. Although the two are related, the challenge of validly measuring the features of units is distinct from the challenge of incorporating all analytically important units, spatial and nonspatial. All of the solutions discussed in this review essay are known to experienced GIS users and are, in fact, based on existing applications. Nevertheless, because these two specific challenges—which are particularly relevant to international relations—have not been well examined, mobilizing these solutions in direct response will help support valid causal inference with GIS in the future.