National Academies Press: OpenBook

Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers (2003)

Chapter: Identifying International Networks: Latent Spaces and Imputation

« Previous: Using Multi-Theoretical Multi-Level (MTML) Models to Study Adversarial Networks
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 345
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 346
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 347
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 348
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 349
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 350
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 351
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 352
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 353
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 354
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 355
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 356
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 357
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 358
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 359
Suggested Citation:"Identifying International Networks: Latent Spaces and Imputation." National Research Council. 2003. Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. Washington, DC: The National Academies Press. doi: 10.17226/10735.
×
Page 360

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

IDENTIFYING INTERNATIONAL NETWORKS: LATENT SPACES AND IMPUTATION Michael D. Ward Peter D. Hoffi Corey Lowell Lofdahl Political Science Statistics Simulation & Information University of Washington University of Washington Technology Operations Seattle, Washington Seattle, Washington SAIC 98195 98195 Burlington, MA 01803 Introduction Most analyses of world politics and studies of national security policies recognize the inter- dependence among the salient actors across the salient issues. ~aditionallv. internation?~.1 _1_~_ __ ~ _ _ 1~ 1 _ ~ 1 1 1 ~ . . ~ . ~ J ~ —~^ v _^ ~ ^~ v ^ ~ ^ ~ we id; nag Been aennea as tne scope and extent ot the relations among independent coun- tries, thought to be the most important elements in world politics This means that actors as well as their actions are strategically interdependent (Signorino 1999). Ignoring the in- terdependence among these phenomena would appear to be a serious oversight that plague attempts to understand, let alone predict, the course of national security policy and world politics more generally. With very few exceptions, quantitative, systematic studies of in- ternational relations and national security have assumed that the major actors and actions that comprise world politics consist of unconnected actions and actors. Game theoretic models are legion, but rarely deal with more than two actors at a time. Some beginning attempts to model the interdependency in international relations have appeared in the lit- erature (Ward and Kirby 1987; Gleditsch and Ward 2001; Gleditsch and Ward 2001; Ward and Gleditsch 2002; Gleditsch 2002; Lofdahl 2002), but as yet network models have yet be widely applied in scholarly or policy work on international politics. This is somewhat surprising, since it is evident at first blush that international politics is about the interde- pendence that appears around the world. Social network analysis is one technique that has been developed to map and measure the relationships and flows among agents. The nodes in the network are the individuals and groups and the links among them illustrate their interdependencies, both in terms of structure and in terms of the flows of information from one node to others. Since the development of the sociogram (Moreno 1934), sociologists among others have been _ ~ Aged ~ ~ 1 : _ ~ _ ~ 1 ~ 1 2 1 ~ 1 - · ~ ~ ~ ~;u ~~ ana~yz~,- cne linkages among 1nulvloua~s and groups. An interesting early example is found in the early work of Coleman, Katz and Menzel (1957). Most early theoretical advances were based on graph theory as developed and advanced by Ffank narary and nets students and collaborators (Harary 1959; Harary 1969; Harary, Norman and Cartwright 1965). The so-called "Columbia school" worked throughout the 1960s and beyond to further advance the substantive findings in this arena of sociology (White 1963; White, Boorman and Breiger 1976~. Broader dissemination of these ideas came much more recently with didactic writings (Knoke and Kuklinski 1982; Scott 1991) as well as early s ~ ~ ~ iPeter Hoff's research is supported by Office of Naval Research grant N0001~02-1-1011. DYNAMIC SOCKS NETWO=MODEL~G ED ANALYSIS 345

applications that are by now canonical (Padgett and AnseD 1993; HanseL 1983). Indeed, social network analysis has become even more fashionable as outside of sociology in technical fields, and has spread to the wider press as an important method for understanding an increasingly perplexing and complicated social environment (Garreau 2001). However, it is somewhat ironic that to date there are no published applications of network analysis to the study of international relations.2 Hoff, Raftery and Handcock (2002 in press) developed probabilistic models of links among actors based on latent positions of actors in an unobserved "social space." We apply such a model a large database on international relations that is typical for the national secu- rity and international politics literatures, and discuss making predictive inference on links that are missing at random. In particular, we analyze the interactions among important social actors in Central Asia, using data taken from the Kansas Event Data Survey an autos mated textually oriented data generating process (Schrodt, Davis and Weddle 1994; Gerner, Schrodt, Francisco and Weddle 1994), specifically the CASIA database, available from the KEDS Web site at http://www.ku.edu/ keds/data.html. This database captures the daily ebb and flow of cooperative and conflictua] events among important political and economic agents (typically ceded "actors" in the international relations literature) in the Central . . . Asian region. Our main purpose is to illustrate the value in using a latent space approach to under- standing network structure in an applied, international relations context. Event Data on International Relations among Central Asian Countries Event data are nominal or ordinal codings of the recorded interactions of international actors.3 Berelson (1952) introduced the concept of content analysis to the social sciences, but it was North, Holsti, Zaninovich and Zinnes (1963) that pioneered its use in studies of world politics . Event data have been widely used in quantitative international relations research aIld in policy research for four decades, following their introduction, event data in international relations were widely used (North 1967; McClelland and Hoggard 1969; Azar 1980~. Until the development of machine coding the World Event Interaction Survey (WEIS) and Conflict on Peace Databank (COPDAB) were the two dominant schema. The contemporary, state-of-th~art is found in the Kansas Event Data System (KEDS) which uses automated coding of English-language news reports to generate political event data (Schrodt 2000; Schrodt, Davis and Weddle 1994). According to Schrodt, there are three major steps involved in creating event data. 2Steven J. Brams (1966; 1968) and later Schofield (1972) tried to estimate linkages among countries, but this line of research was not pursued. Some work with elementary graph theory in the field of international relations has appeared more recently (Lad 1995), but like most of the early work, this applies to a small number of actors, typically three. 3This section is taken and adapted with permission from the KEDS Web site at http://www.ku.edu/ keds/intro.html. 346 DYNAMIC SOCIAL N~TWORKMODEL~G~D ^^YSIS

1. First, a source of news is identified. Typically a news summary is used, ideally one that is already available in a machine readable format. The two current de facto publically available standards are the Reuters news service lead paragraphs or the Foreign Broadcast Information Service (FBIS)4 2. Second, a coding system is developed, or one of the extant coding systems such as the World Events Interaction Survey (aka WElS), the Conflict and Peace Data Bank (COPDAB), or CAMEO (a KEDS coding schema (Gerner, SchroUt, Omur Yilmaz and Abu-Jabr 2002~) is chosen. This coding system must specify what types of interactions constitute an "event." This requires the specification of which actors will be coded, for example, whether nonstate actors such as NATO and the Uruted Nations or gueriLa movements or salient individuals wall be included. At the same time the coding rules must specify what basic issue areas will be included. The COPDAB data set includes a general "issue area" which describes whether an action is primarily military, economic, diplomatic or one of five other types of relationship. In contrast, WEIS also had a few specific "issue arenas" such as the Vietnam War, Arabs Israeli conflict, and SALT negotiations. . The coding rules themselves may be developed in terms of a manual that is given to human coders or more frequently is encapsulated in a computer program such as KEDS, which uses extensive dictionaries to identify actors and events and associate these with specific numerical codes. These dictionaries are developed theoretically by specification and tinned practically-by coding a large number of test sentences from the actual data and adding the appropriate vocabulary when the machines makes an observed error. Table ~ shows a sample of the lead sentences of reports on the Reuters newswire that preceded Iraq's invasion of Kuwait in August 1990. Generally each lead corresponds to a single event, though some sentences generate multiple events. For example, the lead sentence for July 23, 1990 is "Iraqi newspapers denounced Kuwait's foreign minister as a U.S. agent Monday." This corresponds to WElS category 122, defined as "Denounce; denigrate; abuse". In this event, Iraq is the source (actor) of the action and Kuwait is the target. Together, these generate the event record that corresponds to an event in which Iraq denounces Kuwait.5 The WEIS codes and associated Goldstein (1992) weights ale given in 2 for some of the WEIS categories. Goldstein scores are psychometrically determined weights' where a positive weight means that the event has positive affect; conversely' a negative Goldstein score indicates negative affect. Table 3 shows the Reuters stories converted to WElS events. Event data analysis relies on a large number of events to produce meaningful patterns of interaction. The information 4FBIS is available at http: //199 . 221 . 15 . 211/, while Reuters can be contacted via ~ . renters . cam. 5This gives "900723 IRQ KUW 122" where "900723" is the date of the event, IRQ is a standard code for Iraq, KUW is the code for Kuwait, and 122 is the WEIS category. DYNAMIC SOCIAL NETWORK MODELING AND ANALYSIS 347

Table I: Renters Chronology of 1990 fraq-Kuwait Crisis, adapted from Schro~t (http://www.ku.edu/ keds/intro.htmiJ, with permission. The Iraqi denunciation of Kuwati on July 23, 1990 is a typical confict event. Date Headline July 17, 1990 RESURGENT IRAQ SENDS SHOCK WAVES THROUGH GULF ARAB STATES Lead Sentence Iraq President Saddam Hussein launched an attack on Kuwait and the United Arab Emirates (UAE) Tues- day, charging they had conspired with the United States to depress world oil prices through overproduction. July 24, 1990 IRAQ WANTS GULF ARAB AID Debt-burdened Iraq's conflict with DONORS TO WRITE OFF WAR Kuwait is partly aimed at persuading CREDITS Gulf Arab creditors to write oh bil- lions of dollars lent during the war with Iran, Gulf-based bankers and diplo- mats said. July 24, 1990 IRAQ, TROOPS MASSED IN GULF, Iraq's oil minister hit the OPEC cartel DEMANDS $25 OPEC OIL PRICE Tuesday with a demand that it must choke supplies until petroleum prices soar to $25 a barrel. July 25, 1990 IRAQ TELLS EGYPT IT WILL NOT Iraq has given Egypt assurances that ATTACK KUWAIT it would not attack Kuwait in their current dispute over oil and territory, Arab diplomats said Wednesday. July 27, 1990 IRAQ WARNS IT WON'T BACK Iraq made clew Friday it would take DOWN IN TALKS WITH KUWAIT an uncompromising stand at concilia- tion talks with Kuwait, saying its Per- sian Gulf neighbor must respond to Baghdad's "legitimate rights" and re- pair the economic damage it caused. July 31, 1990 IRAQ INCREASES TROOP LEVELS Iraq has concentrated nearly 100,000 ON KUWAIT BORDER troops close to the Kuwaiti border, more than triple the number reported a week ago, the Washington Post said in its Tuesday editions. August 1, 1990 CRISIS TALKS IN JEDI)AH BY Talks on defusing an explosive cri- TWEEN IRAQ AND KUWAIT COG sis in the Gulf collapsed Wednesday LAPSE when Kuwait refused to give in to Iraqi demands for money and territory, a Kuwaiti official said. August 2,1990 IRAQ INVADES KUWAIT, OIL Iraq invaded Kuwait, ousted its lead- PRICES SOAR AS WAR HITS ers and set up a pro-Baghdad go~rern- PERSIAN GULF ment Thursday in a lightning pre-dawn strike that sent oil prices soaring and world leaders scrambling to douse the flanks of war in the strategic Persian Gulf. 348 DYNAMIC SOCIAL NETWORK MODELING AlID ANALYSIS

Table 2: Selected WETS action categories and Goldstein Scores. Yield (1.0) Retract (2.0) Comment (0.0) Neutral Comment (-0.2) Consult (1.0) Receive (2.8) Praise (3.4) Kidnap, Jail (-2.5) Non-Injury Destruction (-8.3) Riot, Violent Clash (-7.0) Surrender (0.6) Accommodate, Cease Fire (3.0) Decline Comment (-0.1) Optimist Comment (0.4) Meet (1.0) Vote, Elect (1.0) Endorse (3.6) Spy (-5.0) Non-Military Destruction (-~.7) Assassinate Torture Execute (-9.0) Retreat (0.6) Cede Power (5.0) Pessimist Comment (-O.a=) Explain Position (0.0) Visit (1.9) Approve (3.5) Rally (3.8) Force (-9.0) Military Engagement (-10.0) Coup Attempted (-8.0) provided by any single event is very limited; single events are also affected by erroneous reports and coding errors However, important events trigger other interactions throughout the system. For example while Iraq's invasion of Kuwait by itself generates only a single event with WElS code 223-military force-the invasion triggers an avalanche of additional activity throughout the international system as states and international organizations de- nounce, approve or comment, so the crisis is very prominent in the event record. This kind of implicit triggering is analyzed and described in Throat and Mintz (1988) and Ward and House (1988~. Table 3: Coding of the 1990 fraq-Kuwait Crisis, using WETS coding scheme. These events appear in textual format in Table 1, above. Date Actor Target WEIS Action Code Type of Action 900717 IRQ KUW 121 CHARGE 900717 IRQ UAE 121 CHARGE 00 ~ 3 ~Q ~ . DEMAND DEMAND ASSURE WARN MOBILIZATION REFUSE MILITARY FORCE .. . .. 900724 900724 900725 900727 900731 900801 900802 .... . . _. . IRQ IRQ IRQ IRQ IRQ KUW IRQ ~ ~ '. _ . ~ . ARE OPC EGY KUW KUW IRQ KUW 150 150 054 160 182 112 223 Data generated in this fashion are exactly the same kind of data that are used to repre- DYNAMIC SOCIAL NETWORK MODELING ED ISIS 349

350 sent social networks. Yet, to date, despite the widespread use of such data in international relations, there are no published studies which analyze these kind of data from a social network perspective. We turn to the application of latent space analysis of social networks using these ciata in the following sections. Latent Space Models of Network Structures Let pi j denote the value of a relationship between agent i and agent j; these relationships may be measured discretely or continuously. The matrix Y is variously called a transaction matrix, a sociomatrax, or a spatial weights matrix. Let X comprise observed characteristics (c~variates) that can be specific to the agents i or j, or specific to their interaction i, j 6 The observed network is assumed to be a function of all relevant c~va~iates, observed or not observed. The presence of important non-observed co-variates often leads to de- pendencies in the network Y. The models of (Hoff, Raftery and Handcock 2002 in press) assume the dependencies in the data can be represented via a latent, unrealized position or characteristic zi for each node i, and that the network responses are conditionally in- dependent given the set of latent positions. Given this assumption, we can express the probability of the given network conditional on the latent positions of the agents and their characteristics as P(Y~Z' X, §) — ~ P(Yi,j ski, Zj, Xi,j, 0, ~2) (~) it Unconditional on the zi's, the data are dependent. If the data are binary, Equation (~) can be parameterized as a logistic regression mode! in which the probability of linkage depends on some projected closeness between the agents (pi and zj) and covariates such that: '7i,j log oddsty:,j I Pi, zj, Xi,j, Ct. is, ~2)—~ + 3 Xi,j + Hi zj, (2) where zi and Zj represent the projected positions of actors i and j in the latent space. Suppose that each actor i has an associated vector Zi of characteristics. Each vector can be thought of as comprising a position on a k-dimensional sphere of unit radius (the direction of the Zi), as well as an "activity level" (the length of zig. In the mode! above, agents i and j are more likely to have linkages if they have similar locations on the sphere, and they are "active," that is, if Z~Zj is large and positive. This leads to a (Iog) probability of the sociomatr~x specified as: 10gP(Y~77) = ~(Yi,j(a + ~ xi,j + Zi Zj)—tog(l + em+ Hi j+Zi Zilch. (3) it 6Most of the notation (but not necessarily the terminology) herein follows (Hoff, Raftery and Handcock 2002 in press). DYNAMIC SOCIAL NETWORK MODELING AND ANALYSIS

Maximum likelihood and Bayesian estimates for the parameters in this model can be ob- tained in a straightforward way. First, the maximum likelihood estimates of the parameters are formed by direct maximization of Equation (3~. Then, with this estimate as a starting value and with diffuse prior distributions over the model parameters, a Markov chain is constructed to generate samples of the parameters from the posterior distribution Such a sample is generated by drawing proposal values from a symmetric proposal distribution, and accepting the proposal with an appropriate probability. For example, in sampling a new Z value at the kth stage of the chain, we sample a proposal value Z from a proposal distribution )(Z~Zk), where Zk is the most recently sampled value. The proposal is ac- cepted as the new value Zk+l with probability P(Y~ZC'k'pkk'j'X) Skit where (A) is the prior distribution for Z. If the proposal is not accepted, then Zk+i iS set equal to Zk- This approach has been used by Hoff, Raftery and Handcock (2002 in press) to estimate several of the classic social network analysis data sets.7 The basic setup is quite general and can be even more widely employed. Estimation of Network :Links in CASIA database We use this framework to estimate the network structure of the political interactions of the primary actors in Central Asian politics over the period from 1989 through 1999. This region has a great deal of conflict and spotty coverage in English language media, despite its contemporary salience. Based on the CASIA database, there are ll3 such actors which have been deemed by substantive experts to be significant. Of these, there are 51 country level actors that have interactions with one another during this eleven year period. We sum the paired interactions among these 51 countries across the eleven year period. A link is deemed to occur for any interaction between two countries during this period. Thus, our data is a 51 x 51 sociomatrix in which an entry is ~ if and only if there is an interaction between i and j in the CASIA database between 1989 and 1999; otherwise it is 0. We use a single covariate for this analysis: xi,j is the distance in thousands of kilometers between the capital city of each of the countries Distance is widely employed as an indicator of interaction in international relations: countries closer together have higher probabilities of having linkages. We have glossed over the important content of the interactions. Some will have been cooperative and others highly conflictual. There are many debates in the national security literature about reciprocity. It turns out that countries that have high levels of cooperative interactions also terry to have high levels of conflictual interactions. So this seems a rea- sonable approximation, though it is certainly possible to disaggregate these data by event type, issues, and time. Treating all years in one aggregation is not optimal perhaps, but it does reduce consid- erably the sparseness of the data. We also recognize that some pairs of countries will have 7These are the drosophila melarlogaster of social network analysis, the so-called Monk data (Sampson 1968), as well as data on Florentine marriage patterns among Medici families (Padgett and Ansell 1993), and data on classroom friendship networks (Hansel! 1983~. DYNAMIC SOCL4L NETWORK MOL)~:~G ED ISIS 351

many more interactions. Examination of the histograms of the actual data suggested to us that most of the information about the linkages was captured in the dichotomy; most of the responses were zero and the second most likely value was I. Equation 3 was estimated using direct optimization of the maximum likelihood to gen- erate starting values for the MCMC. One million iterations of the chain were run to obtain estimates of the parameters or and ,5 as well as the latent positions Z and their underlying variance ~2 The negative coefficient for the distance covariate indicates a lower probability of interaction at greater distances, consistent with many published results from different contexts. These estimates are presented in Table 4. Table 4: Maximum Likelihood and MCMC estimates of parameters for the sociogram of the 51 cov~ntraes involved in Central Asian politics over the period from 1989-1999. Quantile based confidence intervals are provided for the MCMC estimates. (J2 95~o Confidence Interval Parameter MLE MCMC 2.5~o 97.5'~o l or -4.32 -4.20 -4.69 -3.06 ,5 -0.26 -0.25 -0.40 -0.21 5.82 4.48 10.53 Figure 1 illustrates the trajectories of the log likelihood and the parameters or, ,d, alla ~2 over the 106 scans of the Markov chain. These plots suggest that the chain mixes reasonably well for all the estimated parameters. The density of these estimated distributions are Figure 1: MCMC Diagnostic plots of parameter estimation via 106 scans of Markov Chain. (a) MCMC ~ (b) a^2 (c) ~ (d) ~ presented in Figure 2. These represent the marginal posterior densities, with vertical lines representing the maximum likelihood estimate. Each of these densities presents a fairly narrow bandwidth. 352 DYNAMIC SOCIAL NETWORK MODELING AND ANALYSIS

Figure 2: Marginal Posterior Densities of the Estimated Parameters. Vertical lines present the MLE. (b) ci (C) ~ As interesting as these estimates and diagnostics are, the most interesting output of a latent space analysis is the position of the actors in the latent space. Figure 3 illustrates these positions for the 51 countries analyzed. Figure 4 displays the relative, latent positions of countries projected onto a circle. Countries that are close together on this circle have higher probability of sharing a fink. Since this set of countries shares many ties, many countries are close to one another in latent space. Imputation of Missing Network Linkages in CASIA Gauging whether a network is completely sampled is perhaps the holy grad! of network analysis. As yet there is no simple solution to this perplexing problem. We offer no complete solution here. However, practically, it may be useful to use imputation methods along with the latent space estimates to gauge whether or not a link that does not turn up actually may be missing at random. We conduct an experiment using the CASIA database through the following procedure: I. Randomly assign NA to 100 yi,j's, "missing" data; 2. Fit the mode! using the non missing data; ~ keeping track of the 100 actual values of these 3. For each missing y:,j, use the parameter estimates to the calculate predicted proba- bility Pi,j that each missing value Yi,j equals one, i.e. is a hidden links; 4. Compute the number of correct and incorrect predictions, using as a first cut a 0.5 threshold; DYNAMIC SOCIAL NETWORK MODFL~G ED ANALYSIS 353

Go To ~4 Latent Dimension 1 Figure 3: The Latent Positions of 51 Countries as a function of their interactions in Central Asia. The United States has links to many of the other countries, as do Pakistan, Afghanistan, China, Russia, and India. 5. Compute the Brier score (Brier 1950~: train—I Id—yi,j)2. 6. Repeat the above steps 200 times; and finally, 7. Compare these results to a standard, logistic framework with the same covariates. The results of this experiment are quite supportive of using latent space to predict non-sampled or hidden network linkages. For the 200 runs, the average Brier score was 0.087, which is quite low. The proportion of correctly predicted observations was 0.~. The original sociomatr~x has about four non-links for every link. This means that a modal guess of O would result in correctly predicting about 0.80 of the observations. Thus, the latent space approach improves significantly upon that result, garnering an additional ~ of the remaining 20 percent. Specifically, conditional on the true value being no linkage, the predicted value is O with probability of 0.95. Given that the true value is ~ (i.e., linkage), the predicted value is ~ with probability of 0.67. 354 DYNAMIC SOCIAL fJETWORK MODELING AND ANALYSIS

4~ <>I'm ~~§ it .~ ~ ~ /< - J .~N 1 G'~'$~°~: Figure 4: The latent positions of countries is projected on a unit circle. Because many of the positions are very close to one another, the positions have been loosely jittered so that the overlap of country labels is reduced. In comparison, a logistic mode! using the same geographic covariate, the average Brier score is almost twice as high (0.17~; higher Brier scores indicate poorer predictive perfor- mance. A logistic model, as is typical of this approach, will correctly predict at! the zeros and none of the links, because it, like the modal guess, always predicts 0. The upshot of this experiment is the important implication that if we sample network ties at random, then estimate the latent positions, this approach can be used to predict the Yi,j that were not sampled. Although quantitative models in international relations that make actual predictions are themselves rare (SchroUt 2000; King and Zeng 2001; Ward and Gleditsch 2002) for a variety of reasons (SchroUt 2002), these results are strong in comparison. More broadly this approach identifies an effective way to sample networks. DYNAMIC SOCIAL NETWORK AdODEL~G ED TRYSTS 355

Conclusion The latent space approach to social network analysis seems promising. It performs quite well in identifying observed, complete networks in the national security realm. It does so in a way that embraces the interdependence of the network data, rather than assuming that it is generated randomly. Moreover, the approach facilitates the presentation of network positions in an intuitively satisfying way, mapped into a small number of dimensions. These locations incorporate measures of uncertainty. Perhaps most importantly this approach is quite general, since it encapsulates a broader class of models. Specifically, a variety of discrete and continuous specifications can easily be adapted, depending upon the data generating process. Finally, our experiments on using the latent space positions to impute missing at random network links proved to be remarkably productive, especially given the absence of any substantive covariates. This leads to the exciting result that it may be possible to use this approach to sample network ties at random, then estimate the latent positions in order to predict network ties that were not initially sampled. References Azar, Edward. 1980. "The Conflict and Peace Data Bank (COPDAB) Project." Journal of Confict Resolution 24:143-152. Berelson, Bernard. 1952. Content Analysis in Communication Research. Glencoe, IL: Fiee Press. Brams, Steven ]. 1966. "Transaction Flows in the International System." American Political Science Review 60:~80-898. Brams, Steven J. 1968. "DECOMP: A Computer Program for the Condensation of a Directed Graph and the Hierarchical Ordering of Its Strong Components." Behavioral Science 13:344-345. Brier, G.W. 1950. "Verification of forecasts expressed in terms of probabilities." Bulletin American Meteorological Society 78:~-3. Coleman, James S., Elihu Katz and Herbert Menzel. 1957. "The Diffusion of an innovation among physicians." Sociometry 20:253-70. Ga~reau, Joel. 2001. "Disconnect the Dots: Maybe We Can't Cut Off Terror's Head, but We Can Take Out Its Nodes." Washington Post September 17:C-01. Gerner, Deborah J., Philip A. SchroUt, Omur Yilmaz and Rajaa Abu-Jabr. 2002. Conflict and Mediation Event Observations (CAMEO): A New Event Data Framework for a Post Cold War World. Annual Meetings of the American Political Science Association Boston, MA: . 356 DYNAMIC SOCKS NETWO~MODEL~G ED TRYSTS

Gerner, Deborah J., Philip A. SchroUt, Ronald Francisco and Judith L. Weddle. 1994. "The Analysis of Political Events using Machine Coded Data." International Studies Quarterly 38:91-~19. Gleditsch, Kristian S. and Michael D. Ward. 2001. "Measuring Space: A Minimum Dis- tance Database and Applications to International Studies." Journal of Peace Research 38~6~:749-768. Gleditsch, Kristian Skrede. 2002. All International Politics is Local: The Diffusion of Confict, Integration, and Democratization. Ann Arbor, MI: University of Michigan Press. Goldstein, Joshua. 1992. "A Conflict-Cooperation Scale for WElS International Events Data." ~Tournal of Confict Resolution 36~2~:369-385. Hansell, Stephen. 1983. "Cooperative Groups, Weak Ties, and the Integration of Peer Friendships." Social Psychology Quarterly 47~49:316-328. Harary, Frank. 1959. "Graph theoretic methods in the management sciences." Management Science 5:387-403. Harary, Frank. 1969. Graph Theory. Reading, MA: Addison-Westey. Harary, Frank, Robert Norman and Darwin Cartwright. 1965. Structural Models. New York: Wiley. Hoff, Peter D., Adrian E. Raft ery and Mark S. Handcock. 2002 in press. "Latent Space Ap- proaches to Social Network Analysis." Journal of the American Statistical Association tbattba) :tba. King, Gary and Langche Zeng. 2001. "Improving Forecasts of State Failure." World Politics 53~4~:623-658. Knoke, David and James H. Kuklinski. 1982. Network Analysis. Beverly Hills, CA: Sage. Cat, David. 1995. "A Structural Approach to Alignment: A Case Study of the China- Soviet-U.S. Strategic Triangle, 1971-1988." International Interactions 20~49 :349-374. Lofdahl, Corey L. 2002. Environmental Impacts of Globalization and Trade: A Systems Study. Cambridge, MA: The MIT Press. McClelIand, Charles A. and Gary Hoggard. 1969 Conflict patterns in the interactions among nations. In International Politics and Foreign Policy, ed. James N. Rosenau. New York: The Free Press pp. 711-724. Moreno, Jakob L. 1934. W71O Shall Survive? Washington, D.C.: Nervous and Mental Disease Publishing Company. DYNAMIC SOCKS N~TWO=MODEL~G ED ^4YSIS 357

North, Robert C. 1967. "Perception and Action in the 1914 Crisis." Journal of International Affairs 21:103-122. North, Robert C., Ole R. Holsti, M. George Zaninov~ch and Dina A Zinnes. 1963. Content Analysis: A Handbook with Applications for the Study of International Crisis. Chicago: Northwestern University Press. Padgett, John F. and Christopher K. Ansell. 1993. "Robust Action and the Rise of the Medici, 1400-1434." American Journal of Sociology 98:1259-1319. Sampson, S.F. 1968. A Novitiate in a Period of Change: An Experimental and Case Study of Relationships, unpublished Ph. D. Dissertation, Department of Sociology, Cornell University, Itaca, NY. Schofield, Norman J. 1972. "A Topological Mode! of International Relations." Peace Re- search Society, Papers I8:93-~12. SchroUt, Philip A. 2000. "Forecasting conflict in the Balkans using Hidden Markov Mod- els." Presented at the Annual Meetings of the American Political Science Association, Washington, D.C. SchroUt, Philip A. 2002. Forecasts and Contingencies: From Methodology to Policy. Annual Meetings of the American Political Science Association Boston, MA: SchroUt, Philip A. and Alex Mintz. 1988. "A Conditional Probability Analysis of Regional Interactions in the Middle East." American Journal of Political Science 32~:217-230. Schro~t, Philip A., Shannon G. Davis and Judith L. Weddie. 1994. "Political Science: KEDS-A Program for the Machine Coding of Event Data." Social Science Computer Review 12~3~:561-588. Scott, John. 1991. Social Network Analysis: a Handbook. Newbury Park, CA: Sage. Signorino, Curtis. 1999. "Strategic Interaction and the Statistical Analysis of International Conflict." American Political Science Review 92~2~:279-298. Ward, Michael D. and Andrew M. Kirby. 1987. "Re-examining Spatial Models of Interna- tional Conflict." Annals of the American Association of Geographers 77:86-105. Ward, Michael D. and Kristian Skrede Gleditsch. 2002. "Location, Location, Location: An MCMC Approach to Modeling the Spatial Context of War and Peace." Political Analysis 10~3~:244-260. Ward, Michael D. and Lewis L. House. 1988. "A Theory of Behavioral Power of Nations." 32~:3-36. White, Harrison. 1963. An Anatomy of Kinship: Mathematical Models for the Structure of Cumulated Roles. Engiewoods Cliffs, NJ: Prentice Hali. 358 DYNAMIC SOCIAL NETWORK MODEL~G~D ISIS

White, Harrison, Scott Boorman and Ronald Breiger. 1976. "Social Structure From Mul- tiple Networks; Blockmodels of Roles and Positions." American Journal of Sociology 81:730-799. DYNAMIC SOCIAL NETWO=MOD~L~G ED ISIS 359

Next: Summary: Themes, Issues, and ApplicationsLinking Capabilities to Needs »
Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers Get This Book
×
Buy Paperback | $73.00
MyNAP members save 10% online.
Login or Register to save!
Download Free PDF

In the summer of 2002, the Office of Naval Research asked the Committee on Human Factors to hold a workshop on dynamic social network and analysis. The primary purpose of the workshop was to bring together scientists who represent a diversity of views and approaches to share their insights, commentary, and critiques on the developing body of social network analysis research and application. The secondary purpose was to provide sound models and applications for current problems of national importance, with a particular focus on national security. This workshop is one of several activities undertaken by the National Research Council that bears on the contributions of various scientific disciplines to understanding and defending against terrorism. The presentations were grouped in four sessions – Social Network Theory Perspectives, Dynamic Social Networks, Metrics and Models, and Networked Worlds – each of which concluded with a discussant-led roundtable discussion among the presenters and workshop attendees on the themes and issues raised in the session.

  1. ×

    Welcome to OpenBook!

    You're looking at OpenBook, NAP.edu's online reading room since 1999. Based on feedback from you, our users, we've made some improvements that make it easier than ever to read thousands of publications on our website.

    Do you want to take a quick tour of the OpenBook's features?

    No Thanks Take a Tour »
  2. ×

    Show this book's table of contents, where you can jump to any chapter by name.

    « Back Next »
  3. ×

    ...or use these buttons to go back to the previous chapter or skip to the next one.

    « Back Next »
  4. ×

    Jump up to the previous page or down to the next one. Also, you can type in a page number and press Enter to go directly to that page in the book.

    « Back Next »
  5. ×

    To search the entire text of this book, type in your search term here and press Enter.

    « Back Next »
  6. ×

    Share a link to this book page on your preferred social network or via email.

    « Back Next »
  7. ×

    View our suggested citation for this chapter.

    « Back Next »
  8. ×

    Ready to take your reading offline? Click here to buy this book in print or download it as a free PDF, if available.

    « Back Next »
Stay Connected!