投票意向形成與變遷的代理人基模型：模擬與實證資料的比對分析
NSC 96-2414-H-110-001 研究成果報告(精簡版) 劉正山 中山大學政治學研究所 助理教授 前言 ... 2 文獻探討... 2 研究方法... 5 結果與討論... 11 參考文獻... 12 計畫成果自評 ... 15前言
代理人基模擬 (agent-based modeling，ABM，或 multi-agent modeling) 是個以電腦程式 （或稱作模型）模擬「代理人」之間長時間互動的結果，並以該結果進行理論分析的研究方 法。「代理人」所代理的對象，可以隨著學科及研究主題不同而指涉不同的個體，如蟻群、 鳥群、消費者、乃至於選民等等。在此所謂的「互動」，指的是代理人之間，以及代理人與 大環境之間的訊息交換。以民意的形成過程為例，這樣的互動指涉的是某選民與其他選民討 論政治之後，或是該選民接觸媒體之後，從訊息源取得訊號的動作。 目前社會科學界在評估「代理人基模擬」途徑時所在意的是電腦程式是否具備內部有效 性（internal validity）和外部有效性（external validity）。內部有效性指的是模型本身的設計 有多大程度呼應被模擬客體的特質，亦即，程式中的代理人是否合理地代表了被模擬的客 體。外部有效性則建基於內部有效性上，指的是模擬的結果多大程度回應了真實世界的現 象。從社會科學方法論的角度來看，模擬若不能達成以上兩個要求，則模擬產生的資訊不但 無法增進學術社群對政治現象的瞭解與推斷，甚至可能造成誤導。因此，使用此途徑的學者 首要的工作是拉近模型設計與實證資料的距離，亦即比對模擬產生的資料與實證的資料。唯 有如此，模型本身才可能得到學術社群的認可，學者也才可能以此模型為平台進行模型設計 與 模 擬 結 果 的 討 論 。 本 計 畫 創 造 出 一 個 可 以 依 不 同 地 區 和 選 區 的 資 訊 進 行 客 製 化 （customized）的模型”S-RAS”，盡量符合內在有效性的要求之後，再透過模擬結果與實證結 果之間的差異比較來判斷這個模擬途徑的外部有效性。本計畫所要探討的課題是：這個在物 理學、經濟學、生態學受到重視的研究途徑，在政治學中如何被適當地運用。 由於本計畫成果已以英文寫成，且成果已順利發表於美國政治學會今年（2007 年）於芝 加哥舉辦的年會（2007.8.31-9.2）。以下文獻探討、研究方法和結果討論等節將以英文作整篇 論文的重點摘要。成果自評的部分則以中文作會後心得整理。
文獻探討
Validation of ABM: Internal and External Validity
In general,the term “validity” meansa good correspondencebetween a real system and an artificial model. Laurent (2000) suggests that a clear understanding and description (or "the verbal model") of the real system through qualitative methods (e.g., interviews) is necessary before designing a mathematical, experimental, or computational model like ABM.
By a rigid definition, a validated ABM model is both internally and externally valid. An internally valid model is the one from which researchers can draw confident causal conclusions, so such design would yield robust and replicable results. McKelvey (2002) refers internal validity to "analyticaladequacy,” meaning that"the model(in an isolated idealized setting,such asalab or
computer) correctly produces effects predicted by the theory" (p.766; parentheses in original). A different but more relaxing definition of internal validity is having a consensus about the model design from members of the associated academic community (Becker et al., 2005; Schmid, 2005).
As long as the behavior rules and environmental settings of a model make sense, this model can be seen internally valid (see, Entman and Herbst, 2001; Graber and Smith, 2005; Lansing, 2003). An externally valid model allows a researcher to generalize conclusions to situations that prompted the research, including from a sample to a larger population, generalizing across settings or population, and both (Lucas, 2003; Schram, 2005). Especially when an assumption held by a model is supported by robust empirical evidence, this model will have a higher degree of methodological realism (Kagel and Roth, 1995). McKelvey (2002) refers external validity to "ontological adequacy,” meaning that a model passes the process of testing the model against evidence from the real world.
In effect, however, there exists no model that is constructed on the basis of complete information or the whole picture of a story. Details of a case told by field researchers may still not be complete enough to form an agent-based model; similarly, a model that is designed on the base of empirical findings still have many unspecified assumptions when it comes to realization in ABM. Laurent (2000), hence, argues that ABM modelers should not be worried about not being able to tell the whole story.
Looking from a logical positivist perspective, Lucas (2003) also argues that external validity of experiments should be rooted in theory instead of methodological procedures. "The major drawback to experimental investigations is that, at times, they cannot create or manipulate all theoretically meaningful variables. However, if an experiment does manipulate every theoretically relevant variable and finds an effect, then to say that the effect will not generalize to naturally occurring situations is not a criticism of the experiment as having low external validity; rather, it is a critique of the theory for not taking every factor influencing the phenomenon of interest into account" (Lucas, 2003, p.238).
Despitethediscrepancy,one would seethat“alignmentwith theory” isa shared commonality between the two conflicting perspectives about the role of theory in experiments. Laurent (2000), emphasizing the connection to the empirical world, suggests that a modeler describe (1) the characteristics of the real life system, such as retailing costs in economic analysis (2) the assumptions, and (3) empirical or factual information that "re-formalize the model in a more realistic manner,” such as realmerchandising costs in real stores Laurent (2000, p.181). Lucas (2003), emphasizing the role of theory, thinks that well-designed experiments (1) simplify naturally occurring situations and (2) incorporate only theoretically relevant elements. The key to evaluating any research design is "whether the situation adequately represents the concepts or relations specified in the theoretical hypotheses under test" (Lucas, 2003, p.246).
To sum up, an important step to construct the validity of the model, which appears a newly formed consensus among scholars using lab experiment, is careful evaluating the theoretical elements of a simulation model. Especially when the goal of research is not making an empirical
prediction, or when data for external validation are unavailable, researchers are expected to carefully choose parameters derived from theory (construct validity) and, if the theory does not explicate some assumptions or parameters, formulate assumptions based on empirical findings (analytical adequacy).
Research Methodology of Conducting External Validation
Beyond alignment of theory, Mehta and Bhattacharyya (2006) suggest that researchers need to consider two more stages to fulfill the requirement of external validation: the alignment of
observable processes and the alignment of outputs. Regarding how to do these deal with these two stages correctly, experimental economists have formed three approaches (Fagiolo et al., 2006): the indirect calibration approach (IC), the Werker-Brenner approach (WB), and the history-friendly approach (HF). In sum, the key elements drawn from the above approaches include (1) describe the choice of a limited number of parameters based on a theory, (2) construct the correspondence between these elements of the model to a targeted theory, (3) employ empirical data to calibrate initial conditions and parameters, (4) evaluate how robust and sensitive the values of key
parameters are, and (5) discover how much the statistical patterns of simulation corresponds those found from empirical data.
Comparison across patterns
As Janssen and Ostrom (2006) suggest, there are four areas to compare simulation results with stylized facts: First, if data is large in number and good in quality, and if one can draw stylized facts from the such empirical data, researchers can derive statistical distribution, such as the power law distribution, and other stylized facts from the empirical data (e.g., Cederman, 2003). Second, if the model is relatively uncomplicated, researchers could ask what the simple rules in the model that generate these stylized facts are. After finding out these rules, the next step is to investigate the modeled conditions that result in statistics similar to the observed styled facts. Third, researchers can use empirical findings from role games or field experiments to develop and test assumptions held in ABM. Fourth, use case study with rich information and data to parameterize the model (conjoint analysis) (e.g., Garcia et al., 2007). Besides comparing stylized facts or patterns, sensitivity analysis or robust tests (in the IC approach) is an important step that makes an ABM experiment communicable across disciplines. Robust tests make the findings more persuasive to readers who concerns with the influence of extreme conditions during simulation. Because parameter values are used to test if simulation results are stable in extreme scenarios, it opens the modelers “to the propositions that a model should be judged by the criteria that are used in mathematics:i.e.,precision,importance,soundnessand generality” (Fagiolo etal.,2006,p.22).
Robust tests provide a range of possible situations for empirical evidence to fit, but whether or not the model is external valid still depends on how much selected criteria of patterns, and the mechanismscausing thepatterns,match an empiricalcase,noton robustteststhemselves.“External
validity inferences do not have much bite unless one systematically investigates the degree of similarity and dissimilarity between laboratory and targetsystems” (Guala,2005,p.229).
Given all the above discussion drawn from the literature, the second half of this paper will present an example of applying the IC approach to a study of preference dynamics. As each approach has its limit and because no agent-based model is perfect at the time created, the example should not be regarded as a perfect one. Instead, through this practice, one will see more clearly the limits of ABM and what need to be done to advance an agent-based model and to validate simulation results.
研究方法
Alignment of Theory
The Swarm-RAS model, or S-RAS, is a “Swarm” version of John Zaller’s (1992) Receive-Acceptance-Sample (RAS) theory of voter preference.6It has been difficult to study the dynamics of preference formation with RAS where the axioms of information processing should be considered together. I designed S-RAS with a goal to revive this framework that has been more cited than directly applied.
S-RAS allows researchers to operationalize the original RAS theory. Using S-RAS, multiple agents being able to process information in R-A-S fashion will be put into an artificial society where they will interact with self-selected news media, political experts, and other fellow citizen agents. Moreover, S-RAS considers real-world complexity like individual differences by including three types of individuals—the politically aware, the politically unaware but with clear electoral preferences, and the politically unaware with no preferences about candidates.
Given these elements, hopefully, a researcher can use S-RAS to study long-term influence of a specific element of the model. The next subsection describes how the model design of S-RAS was associated with the RAS framework, particularly the four basic axioms of information processing.
Figure 1 summarizes how agents in S-RAS acquire information and process received messages. In each time step agents complete a loop of action, from whether or not accessing a news media object to update its voter preference. As Figure 1 on the following page shows, the following features of agents in S-RAS are initiated: a party identification (1 or 0), an opinion about candidates (0.00 to 1.00), a voter preference (1 or 0), a favorite media object (1 or 0), and eight political discussants with political experts on the top ofthe agent’scontactlist.Notethatitisassumed that agents habitually check out news media reports before finding someone to discuss politics (see, Mutz and Martin, 2001). For every time step or iteration during the simulation, every agent finishes its own loop of processing political information, including accessing the news media, discussing politics, or doing nothing if no discussant is available.
The value of Opinion can be seen as a “true”preferenceofan agentabouta candidate,which other people hardly catch accurately, while Voter Preference is better viewed as a general impression ofone’struepreferencethatotherscan tellfrom dyadicinteraction.IfagentX has Opinion value 0.62, for example, other agents who interact with X will obtain an impression that agentX favors“1”.Similarly, if it has Opinion value 0.37, its network members who interact with it at a given time will store 0 in their memory.
S-RAS maximizes individual differences among agents by considering general distinction between political experts and ordinary citizens (see, Bartels, 2004; Converse, 1964b, 1990; Fiorina et al., 2005) and interpersonal differences within each type (see, Oliver, 2002). As Table 1 summarizes, there are three types of agents in S-RAS and researchers can vary the proportion of each type of agents for calibration and robust tests: ordinary citizens with clear voter preference (C1 agents), political experts or the politically aware (C2 agents), and ordinary citizen having no preference about candidates or claiming independence (C3 agents).
A news media object in S-RAS refers to any source of information other than dyadic interpersonal discussion. By this definition, the news media object include political elites who usually appear on TV, newspapers, internet, and other kinds of news channels.11A news media object consistently holds a consistent voter preference. Unlike political discussant agents that may be unavailable sometimes, the news media objects can be accessed by any agent at any time. Hence, the two media objects in S-RAS—one favoring “1”and the otherfavoring “0”—can be seen as politically polarized media groups.
That media objects are assumed polarized does not mean that they always broadcast 1 or always broadcast 0. It is further assumed that about two-third of chance agents get news messages consistent with its partisan orientation. Therefore, agents, according to their propensity of selective perception, will get a variety of messages from the self-selected news source. An agent having voter preference“1” and isselective atatimestep,forexample,willstore 1 in itsmemory atthe time when it accesses a news medium. If it is not selective at another time step, it will get whatever the news medium gives, either 1 with a probability of 0.67 or 0 with a probability of 0.33.
Alignment of Observable Processes
One way to align a theoretical model with observable processes is customizing the model with specific parameter values. For this project, data were collected before and after the 2006 Kaohsiung mayoral election (Dec. 9, 2006) in Taiwan. The first wave was conducted between Dec. 4 and Dec. 7 (N=764) and the second wave was conducted between Jan. 20 and Jan. 27 (N=650). The data are weighted by population distribution across administrative districts, age, gender, and education level. This election has a number of empirical features that are consistent with the design of S-RAS. First, as the descriptive statistics of the first-wave survey shows, voters chose between two political parties—Democratic Progressive Party (DPP) and Kuomingtang (KMT). Candidates nominated by other marginalized political parties had almost no impact on the election results (only 1.3% of the total votes). Second, news media are generally polarized; voters can find channels supporting one side while some others favoring the other side. Third, most voters find like-minded people to discuss politics: 76.7 percent of respondents find that their discussants agree most of time; 78.8 percent find their favorite discussant support for the same candidate. Forth, Kaohsiung voters go to TV before discuss politics: 91.4 percent of respondents report that they obtain political news from TV and 52.9 percent of respondents will continue go to TV for further information.
For the calibration of S-RASA, Table 2 lists four criteria to draw information from the first wave of survey: the proportion of political experts, the proportion of political experts who favored DPP, the proportion of ordinary voters who favored DPP, and the proportion of ordinary voters who have no preferencesaboutpoliticalparties,atleastsay “itdepends”or“Idon’tknow”when they were asked about political party for which they tend to support. These four criteria are easy to be found in most surveys and can find their corresponding parameters in S-RAS. First, 16.7 percent of respondents who did not think themselves as politically knowledgeable (probExpertYES) favor red DPP; the proportion of C1 agents favoring “1” (probYES) is set to .167. Second, 34.3% of respondents subjectively thought that they know more about politics than their political discussants; the proportion of C2 agents in S-RAS (probExperts) is hence set to .343. Third, 52.1% of those who thought themselves politically knowledgeable favor red DPP ; hence, in S-RAS I set the proportion ofC2 agentsholding voterpreference“1” (probExpertYES)to .521.Forth,58.8% ofrespondents who felt less politically knowledgeable have no clear voter preference. Hence, the parameter value for probNOIDEA is set to be .588.
Alignment of Outputs
When we are not sure about if the models derived from a theory do not have the internal resourcesto specify theirown domain ofapplication,“modelsmustbe putin correspondencewith the real world by means of a theoretical hypothesis stating what kind of relation holds between a given model (or set of models) and a given real-world system (or set of systems)" (Guala, 2005, p. 156).
Not every theory or theoretical framework gives clear guidance for creating theoretical hypotheses. The RAS framework in this project is an example. RAS gives explanation instead of ways to prediction about the formation of voter preference. Hence, in this project I try to employ a naïve approach to alignment the outputs: present the results before and after the simulation and compare them with descriptive statistics of the second-wave survey. The alignment of outputs is presented the order of (1) simulation results based on calibrated parameters, (2) robust test or sensitivity test results derived from a series of simulations on the calibrated model, (3) comparison across the two surveys and simulation.
Kaohsiung election) is run for 150 time steps to simulate the short-term period of opinion dynamics. Figure 2 presents four snapshots of the 40 x40 opinion grid at the 3rd, 20th, 50th, and 150th time step.Thegray scale ofeach ofthe1600 cellsreflectsthecorrespondentagent’sopinion atthe given time step.Whitedenotesfavoring DPP (“1”or“YES” in the originaldesign)and black denotes favoringKMT (or“0” or“NO” in the originaldesign).The formation ofwhite and black clusters suggests a decrease in the number of independent voters and an increase in the number of opinion clusters.
[ Robust Tests and Table 3 detailed in the conference paper are skipped here]
Comparison between survey results and simulation results
Table 4 presents the comparison by five parameters that can be find in both simulation and survey questionnaire: probYES, probExperts, probExpertYES, probNOIDEA, and Diversity. The first four are those used for calibration (recall Table 2 on page 25). As discussed in the first half of the paper, there is no much reason to compare simulation results and empirical findings directly. It is, therefore, not a big surprise to see that the simulation results resemble the second-wave survey results. What we can inspect, however, is if there is any trend found in empirical data caught in simulation. Both empirical survey data and simulation results suggest that during the six weeks (150 time steps) (1) the proportion of individuals supporting for DPP increases, (2) the proportion of the politically aware decreases, and (3) the level of Diversity remains stable.
Two discrepancies require attention. First, the proportion of (self-claimed) political experts decreases during the election, while in S-RAS this proportion is assumed to be stable over time. This discrepancy can be a result of using self-evaluation of political knowledge, instead of objective measure of political knowledge, in the surveys. Therefore, it will be important to explore if the assumption “the levelofpoliticalexpertise remainsstable”in S-RAS pass more empirical testing.
Second, the proportion of respondents saying no preferences in the second wave of survey soars from 52 to 92 percent. S-RAS was not able to catch this trend and, instead, suggests that the proportion will decrease dramatically. Similar puzzle is found regarding the level of Diversity. Although parameter values remain stable, S-RAS suggest that the Diversity level would decrease
結果與討論
This paper first summarizes how scholars across disciplines deal with the issue of empirical validation for computational simulations and then gives an example of applying the principles to a study of the dynamics of voter preference. One should find thatitisimportantto consider“how to communicate with colleagues distrusting computer simulations” on the one hand; but it is a complex processto “find and presentasuccessfulempirically valid modeland itsresults”.
S-RAS as an agent-based model built on the R-A-S framework of opinion origination has to deal with theoretical vacancy that the framework does not address, such as how an individual form his or her discussion network (S-RAS assumes everyone has 8 discussants located in a Moore neighborhood), how and when one changes its voter preference (S-RAS assumes that one will be seen as changed mind if his or her opinion go below or above 0.5), and how those independent individuals different from ordinary voters (S-RAS assumes no difference), etc.
A dilemma for an ABM model emerges: Listing all those assumptions will make a paper more technical and even more difficult to read, but without explicating assumptions as much as possible, it is also impossible to fully communicate with the scholarly community.
Recall what Janssen and Ostrom (2006) suggest, if data is large in number and good in quality, and if one can draw stylized facts from the such empirical data, researchers can derive statistical distribution and other stylized facts from the empirical data; if the model is relatively uncomplicated, researchers can ask what are the simple rules that generate these stylized facts and then investigate the modeled conditions under which they can derive similar statistics as the observed styled facts. Indeed, comparing statistical relationships based on empirical data and those found in a simulation is a more sophisticated way than an attempt to plug some empirical values and to make prediction. The discrepancy shown in this paper discourages the later. What they do not say, however, is when an agent-based model is ready for delivering hypotheses for empirical tests? The more a researcher digs into the world of ABM, the more challenge and uncertainty he or she will encounter, such as (1) what parameters to choose when there is few hypotheses to be drawn from a theory like RAS and (2) how much can we say that a created agent-based model is completely a based on a theory when a modeler needs to incorporate a good number of untested assumptions to make the model function? Without fully understand a crystal ball or a black box created in the laboratory, how could empirical validation be done?
ABM modelers may also usually find themselves standing at the crossroad of either making the practice of modeling theoretical or predicable. Although a researcher should choose a goal of research before start modeling, the lesson I learned from this project is that a project aiming at a theoretical goal can be discouraged when taking into account empirical validation—a feeling of lacking sufficient elements to make the paper persuasive. I also can expect that a prediction-oriented project in political science will face even more challenges about how the
internal validity (a theory) of the model is preserved. Looking into the future, ABM modelers will continue to think about and be challenged regarding the issue of external validity.
To make S-RAS more applicable for further test with empirical data, two things need to be done. First, assumptions held in S-RAS need to be relaxed and calibrated, if not empirically tested. Calibration by changing parameter values, as demonstrated in this paper, is not enough because this type of calibration should be based on a solid ground that assumptions held in the model are checked. For example, S-RAS assumes that all agents think about checking out news media before discussing politics. What if this assumption is relaxed or what if all agents reverse this process? To what extent will this assumption change affects the overall result? Second, there is a further step beyond robust test: to meet “generative standard” or "proximal similarity"—an assumption of generalization from concrete particular treatment operations (McQuarrie, 2004).
As McQuarrie (2004) suggests, one important approach to promise proximal similarity is using difference combination of parameters/variables chosen from the set of parameter/variables drawn from the theory. Specifically, the goal of proximal similarity assumes that "if one member of a set of parameters {t1...tk.} causes a certain outcome, a new sampling from that set will cause a similar outcome" (p.145). "We have to warrant to suppose that any other treatment operation drawn from the set {t1...tk.} would also have a similar effect" (McQuarrie, 2004, p. 144). If more than one candidate can explain the phenomena of interest, as is often the case in computational modeling, "further work is required at the micro-level to determine which [model specification] is the most tenable explanation empirically" (Epstein, 1999, p.43).
參考文獻
1. Axelrod, R. (1997a). Advancing the Art of Simulation in the Social Sciences, chapter Simulating Social Phenomena, pages 21–40. Simulating Social Phenomena. Berlin: Springer, Berlin.
2. Axelrod, R. (1997b). The complexity of cooperation : Agent-based models of competition and collaboration. Princeton University Press, Princeton, N.J.
3. Bartels, L. (2004). Unenlightened self-interest. The American Prospect, 15(6):A17–A19. 4. Becker, J., Niehaves, B., and Klose, K. (2005). A framework for epistemological perspectives
on simulation. Journal of Artificial Societies and Social Simulation, 8(4).
5. Brown, D. G. and Robinson, D. T. (2006). Effects of heterogeneity in residential preferences on an agent-based model of urban sprawl. Ecology and Society, 11(1):46.
6. Cederman, L.-E. (2003). Modeling the size of wars: From billiard balls to sandpiles. American Political Science Review, 97(1):135–150.
7. Converse, P. (1964a). Ideology and Discontent, chapter The nature of belief systems in mass publics, pages 206–261. Ideology and discontent. New York: Free Press.
Discontent, pages 206–61. New York: Free Press, New York.
9. Converse, P. (1990). Popular representation and the distribution of information, chapter Information and Democratic Processes, pages 369–90. Urbana, IL: University of Illinois Press. 10. Druckman, J. N., Green, D. P., Kuklinski, J. H., and Lupia, A. (2006). The growth and
development of experimental research in political science. American Political Science Review, 100(4):627–635.
11. Druckman, J. N. and Lupia, A. (2000). Preference formation. Annual Review of Political Science, 3:1–24.
12. Entman, R. M. and Herbst, S. (2001). Reframing public opinion as we have known it, chapter Mediated Politics: Communication in the Future of Democracy, pages 203–225. New York : Cambridge University Press.
13. Epstein, J. M. (1999). Agent-based computational models and generative social science. Complexity, 4(5):41–60.
14. Epstein, J. M. and Axtell, R. (1996). Growing Artificial Societies: Social Science from the Bottom Up. Complex adaptive systems. Cambridge, MA: The MIT Press, Washington, D.C. 15. Fagiolo, G., Windrum, P., and Moneta, A. (2006). Empirical validation of agent based models:
A critical survey. A working paper, pageshttp://www.lem.sssup.it/WPLem/files/2006–14.pdf. 16. Fiorina, M. P., Abrams, S. J., and Pope, J. (2005). Culture War? The Myth of a Polarized
America. New York: Pearson Longman.
17. Garcia, R., Rummelb, P., and Hauserc, J. (2007). Validating agent-based marketing models through conjoint analysis. Journal of Business Research, 60(8):848–857.
18. Gilbert, G. N. and Troitzsch, K. G. (2005). Simulation for the social scientist. New York: Open University Press, Buckingham ; Philadelphia, Pa., 2nd edition.
19. Graber, D. A. and Smith, J. M. (2005). Political communication faces the 21st century. Journal of Communication, 55(3):479–507.
20. Guala, F. (2005). The Methodology of Experimental Economics. New York: Cambridge University Press.
21. Huckfeldt, R. (2001). The social communication of political expertise. American Journal of Political Science, 45(2):425–438.
22. Huckfeldt, R., Johnson, P. E., and Sprague, J. D. (2004). Political Disagreement: The Survival of Diverse Opinions within Communication Networks. New York: Cambridge University Press.
23. Huckfeldt, R. and Mendez, J. M. (2004). Managing political heterogeneity: Perceptions of disagreement
24. within communication networks. In the Midwest Political Science Association 62nd Annual National Conference, April 15-8, 2004, Chicago, IL.
25. Huckfeldt, R. and Sprague, J. D. (1995). Citizens, politics, and social communication: Information and influence in an election campaign. Cambridge University Press, Cambridge, England; New York.
26. Janssen, M. A. and Ostrom, E. (2006). Empirically based, agent-based models. Ecology and Society, 11(2):37.
27. Jerit, J. and andToby Bolsen, J. B. (2006). Citizens, knowledge, and the information environment. American Journal of Political Science, 50(2):266–282.
28. Johnson, P. (1999). Simulation modeling in political science. American Behavioral Scientist, 42(10):1509–1530.
29. Kagel, J. H. and Roth, A. E., editors (1995). The Handbook of Experimental Economics. New Jersey: Princeton University Press.
30. Lansing, J. S. (2003). Complex adaptive systems. Annual Reviews of Anthropology, 32:183–204.
31. Laurent, G. (2000). Improving the external validity of marketing models: A plea for more qualitative input. International Journal of Research in Marketing, 17(2-3):177–182.
32. Laver, M. (2005). Policy and the dynamics of political competition. American Political Science Review, 99(2):263–281.
33. Liu, F. C. (2007). Constrained opinion leader influence in an electoral campaign season: Revisiting the two-step flow theory with multi-agent simulation. Advances in Complex Systems, 10(2):233–250.
34. Lucas, J. W. (2003). Theory-testing, generalization, and the problem of external validity. Sociological Theory, 21(3):236–253.
35. Lustick, Ian S, D. M. and Eidelson, R. J. (2004). Secessionism in multicultural states: Does sharing power prevent or encourage it? American Political Science Review, 98(2):209–229. 36. McKelvey, B. (2002). Model-centered organization science epistemology, chapter The
Blackwell Companion to Organizations, pages 752–780. Malden, Mass. : Blackwell Publishers.
37. McQuarrie, E. F. (2004). Integration of construct and external validity by means of proximal similarity: Implications for laboratory experiments in marketing. Journal of Business Research, 57(2 ):142–153.
38. Mehta, K. and Bhattacharyya, S. (2006). Design, development and validation of an agent-based model of electronic auctions. Information Technology and Management, 7(3):191–212.
39. Midgley, D. F., Marks, R. E., and Kunchamwar, D. D. (2007). The building and assurance of agent-based models: An example and challenge to the field. Journal of Business Research, 60(8):884–893.
40. Mutz, D. C. and Martin, P. S. (2001). Facilitating communication across lines of political difference: The role of mass media. American Political Science Review, 95(1):97–114.
41. Neuman, W. R., Just, M. R., and Crigler, A. N. (1992). Common knowledge: News and the construction of political meaning. Chicago: University of Chicago Press.
42. Nowak, A., Szamrej, J., and Latane, B. (1990). From Private Attitude to Public Opinion: A Dynamic Theory of Social Impact. Psychological Review, 97(3):362–376.
43. Oliver, M. B. (2002). Individual differences in media effects, chapter Media Effects: Advances in Theory and Research, pages 507–524. Mahwah, NJ: L. Elbaum Associates, 2nd edition. 44. Rousseau, D. and Van der Veen, A. M. (2005). The emergence of a shared identity: An
agent-based computer simulation of idea diffusion. Journal of Conflict Resolution, 49(5):686–712.
45. Schmid, A. (2005). What is the truth of simulation? Journal of Artificial Societies and Social Simulation, 8(4).
46. Schram, A. (2005). Artificiality: The tension between internal and external validity in economic experiments. Journal of Economic Methodology, 12(2):225–237.
47. Sun, R., editor (2006). Cognition and Multi-Agent Interaction: From Cognitive Modeling to Social Simulation. New York : Cambridge University Press.
48. Waldrop, M. M. (1992). Complexity: The Emerging Science at the Edge of Order and Chaos. New York: Simon & Schuster, New York. M. Mitchell Waldrop.
49. Zaller, J. (1992). The Nature and Origins of Mass Opinion. New York, NY: Cambridge University Press.
計畫成果自評
本研究與原計畫相符程度超過九成。唯一的出入在於民調所收集到的資料不如一開 始預期的多（主要受限於每一通電話受訪者的訪談時間） ，導致在比對的時候只能使用 四組參數。原則上，本計畫達成了預期的目標：針對「該怎麼用 ABM 做民意研究」作了 詳實的文獻綜覽、建立了一個初步可用的 ABM 模型、以及經過與民調資料知道這個模 型未來可以再調整的地方。另外，本人在發表這篇文章之後，亦得到預期之外的想法和 收穫。以下簡述就本人在美國政治學年會的心得作為結尾。本次在芝加哥舉行的美國政治學年會以”Political Science and Beyond”為大會主題， 廣邀在研究議題和研究方法上進行跨學門融合的論文。本人論文題目為"Simulating the Dynamics of Voter Preference: Contrasting Agent-Based Modeling Results with Empirical Data"，於八月三十一日下午在以”They've Given You a Number and Taken Away Your Name: Agent-based Modeling in Political Science”為題的座談會（panel）上報告如何運用從 經濟學、生態學和物理學發展出來的代理人基研究途徑到政治學研究上，以及有興趣利 用此一途徑的政治學者應該注意的事項。出席聆聽的學者約有十五人。本人的評論人為 紐約大學政治系教授 Michael Laver。 就本人發表的論文發表情形和與會者對論文的評價，以及評論人對其他發表者的評 價來看，我覺得代理人基模擬在政治學門中要被廣為接納還有很長一段路要走。大多數 會對代理人基擬感興趣的政治學者，多是從賽局理論（game theory）角度出發來看電腦 模擬如何用到賽局的實驗上。由於賽局是個由象的理論，推算時所用到的假設十分的精 簡，因此運用 ABM 時並不會有太大的障礙。會議上的其他三篇論文多是這樣的應用，
而三篇論文的作者也都大力稱揚這個研究途徑的潛力。 本人同意 ABM 可以應用在賽局理論上，但是 ABM 的應用程度不只於如此。我在論 文中強調，這個途徑還可以用在民意形成的領域，但是學者在進行這樣的研究時，必須 考量到模型設計是否過於簡化，以及有多大程度我們可以將模擬的結果推論到現實生活 之中。我在論文整理了不同領域學者提出的具體做法，並用我在高雄搜到的民調資料與 我的模擬結果作了初步的比對。 我的作法得到台下幾位博士生的共鳴，他們在會後向我表示，他們認為我做的題目 很重要，但「很可能無法用它找到工作」。意思是，我試圖將模擬結果與現實資料作一 個比對的做法是 ABM 應用到政治學上很重要的一步，但是第一，不好做，第二，會挑 戰到很多既有發表的論文，因為大多已發表的、用 ABM 研究途徑寫的論文未處理到我 提出的問題，因此我的呼顩會影響這些文章的地位。我也發現 Laver 教授對我的評論並 未切中我這文的核心論點。他說了兩個模型設計上的問題，如實驗室中虛擬的時間如何 與實際生活作比較等。這些都是容易解釋的問題，但是當我在回應他的評論時反問「那 麼既有的文章中對模擬的時間沒有設限難道不個問題嗎？」他表示他的確沒想過這個問 題。 由此，我發現即美國政治學者在應用 ABM 這個途徑時，有避重就輕的傾向，亦即 有意識地回避外部有效性問題（external validity），而直接呈現模擬的結果，再以模擬的 結果印證自己的觀點。畢竟學門內真的能夠看懂並挑戰模擬的程式碼的人還不夠多，所 以這麼做所遇到的挑戰最少。但是，這麼作會打擊了年輕學者在評估是否要採用 ABM 這個途來作研究的信心。誰會選用一個客觀性不夠強的研究途徑來做研究呢？因此，建 議國內學界使用 ABM 作為發展理論的研究方法（而非用來預測的工具），並強調使用這 個途徑的學者公開程式的原始碼，以期未來有充份溝通的語言。
