Lessons from the Virtual Agora Project: The Effects of Agency, Identity, Information, and Deliberation on Political Knowledge, by and

experimental conditions including discussion and information-only conditions. The model is tested with data from pre- and post-test surveys of a representative sample of 568 Pittsburgh residents who came to a one-day deliberation experiment. We find little evidence that post-discussion knowledge responded to discussion, suggesting that learning occurred through reading materials. The paper discusses deliberative design implications. We also find that socioeconomic factors do predict inequality in learning, but are counteracted to a degree by agency. The agency variables suggest points of intervention to mitigate inequality. Agency theory may be a valuable theoretical framework for deliberation research more generally.


INTRODUCTION
Even if deliberation can enhance citizens' political knowledge and sophistication, it also faces important equality concerns. Deliberative knowledge benefits might accrue inequitably, much as in traditional political participation (Rosenstone andHansen 1993, Verba et al. 1995). Of course deliberative democratic theorists recognize this problem and consequently some advocate for economic and institutional reforms to accompany the expansion of deliberative practices (e.g., Benhabib 1996, Young 1996. Deliberative practices have expanded absent such reforms. In addition, even if deliberation does not aggravate inequality, inequality nevertheless presents a challenge because deliberation would ideally proceed with all citizens equally able to represent their viewpoints (Sanders 1997, 349).
The following questions guide this paper: (1) Can citizens' involvement in deliberative contexts improve their knowledge of political decisions? (2) If it does, what aspect of these contexts leads to learning-discussion, informative readings, or some combination of these? (3) What implications does the mechanism of learning have for the design of deliberative contexts? For example, if readings lead to learning and the promise of deliberation motivates doing the readings, how might deliberations best be structured to increase learning? (4) What factors enhance or hinder learning in deliberative contexts? (5) What do these factors and other results indicate about the degree of learning inequality in deliberative contexts? (6) Do these factors suggest interventions to reduce learning inequality?
Research on deliberation has found that people who participate in deliberative contexts do learn. However, this research has not adequately determined what aspects of these contexts lead to learning. Knowledge of the exact mechanism of learning in deliberative contexts is crucial to determine how best to design these contexts to foster learning. In addition, the research has not carefully examined the individual-level factors that can affect deliberative learning. Knowledge of these factors may help address concerns about the possible inequitable results of deliberation. This paper introduces a theory of political agency that suggests factors for explaining learning in deliberative contexts, including political reflectiveness and conceptions of citizen identity. The paper tests a model, involving these agency variables, demographic factors, and experimental conditions including discussion and information-only conditions. The model is tested with data from pre-and post-test surveys of a representative sample of 568 Pittsburgh residents who came to a one-day deliberation experiment. We find little evidence that post-discussion knowledge responded to discussion, suggesting that learning occurred through reading materials. The paper discusses deliberative design implications. We also find that socioeconomic factors do predict inequality in learning, but are counteracted to a degree by agency. The agency variables suggest points of intervention to mitigate inequality. Agency theory may be a valuable theoretical framework for deliberation research more generally.

DELIBERATIVE LEARNING
According the theorists, deliberation approximates an ideal speech environment (Habermas 1989) whereby one of the main procedural conditions is the cooperative search for agreement.
Opinions are changed in a deliberation because the participants persuade by giving reasons (see Chambers 1996). As participants gain knowledge in the process of discussion, this should enable them to better provide reasons, or to recognize the reasons of others and alter their opinions accordingly. Even if participants in a deliberation do not actually change their opinions, most deliberative theorists would agree they should gain knowledge. Young ( , 128, cited in Barabas 2004 says that deliberation, "…adds to the social knowledge of all the participants." Gutmann and Thompson (1996) write, "The moral promise of deliberative democracy depends on the political learning that reiterated deliberation makes possible." Recent research has provided consistent evidence for an increase in participants' knowledge and sophistication during deliberative experiences, but this research does not clarify the exact source of these improvements. Most of these studies were not suited to untangling the effects of deliberative discussion from the effects of information acquired through reading materials and other sources. In studies by Barabas (2004), Gastil and Dillard (1999), and Luskin et al. (2002), participants were, respectively, administered pre-tests, provided with reading materials, engaged in deliberation, and administered post-tests. Participants in the widely renowned Deliberative Poll receive policy briefings by experts as well as readings (Luskin et al. 2002). Given research designs with information and discussion sandwiched between pre-and post-tests, it is impossible to determine whether learning occurred because of discussion, because of reading materials or other information exposure, or some combination of these factors. Price's (2002) Electronic Dialogue Project, in which participants spent months discussing election-related issues in the 2000 Presidential run, constitutes an experimental study that could determine whether discussions increased knowledge. Participants in the discussion condition could be rigorously compared with control group members who could not participate in these discussions. Regrettably, we are unaware of any published findings from this study that employed the experimental findings to determine whether discussion led to knowledge gains. One paper from the project (Price, Cappella, and Nir 2002) shows that those who discuss more frequently and with people who disagree with them are also able to generate more extended argument repertoires-that is, they are better able to recall pro and con arguments regarding a policy. This finding was not from discussions in the experiment but from surveys in which people were asked about their everyday discussion habits with acquaintances, family, and friends. While suggestive, these results face certain limitations. They do not establish direction of causality-greater argument repertoire might have led people to more discussions. This is especially plausible in light of another finding-those with greater political knowledge and interest were more likely to come to research discussions (Price and Cappella 2002). Also, argument repertoire is a recall measure that may be an imperfect indicator of political knowledge. As for factual knowledge gains, an overview of study findings (Price and Cappella 2002) states there were few "large" knowledge gains and reports on significant gains exclusively on one knowledge item. It is unclear whether these reports concerned experimental discussions or self-reports of everyday discussions.
A study of an Australian citizens jury found jurors' opinions changed in the information rather than in the discussion phase of the jury proceedings (Goodin and Niemeyer 2003). Questionnaires indicated knowledge gains after the 12 jurors received background briefings and questioned experts, but not after the subsequent discussion. The sample size was very small, so these changes were not statistically reliable. But perhaps information alone could provoke what Goodin (2005) calls an "internal reflective deliberation" that enhances knowledge.
More rigorous studies that separate the effects of informative readings from discussion can successfully demonstrate the value of deliberative discussion. Morrell (2005), for example, shows that internal political efficacy rises in response to discussion, not readings. This research compared people who read information with those who both read information and participated in discussion.
Overall, while the empirical literature on deliberation has built a case for learning in deliberative contexts, the findings have been unclear about the mechanism-discussion or information sources. Establishing a mechanism has important implications for inequality as well. If the "informative readings" matter most, this finding could suggest reforms that will more equitably distribute political knowledge.
Also important for addressing inequality is a better understanding of what individual-level factors contribute to this inequality. We next motivate and sketch a theory of human agency that suggests factors that help explain inequality. The theory may prove to be a beneficial framework for deliberation research more broadly.

HUMAN AGENCY AND DELIBERATION RESEARCH
Human agency-loosely, the skills and inclinations that make people good decision makers and actors-should play a role in how readily and thoroughly people learn in deliberative contexts. For example, someone who is inclined to be an agent with respect to political matters should put greater effort into learning political information and may have better-developed skills for political learning. This section and the next will explore the relationship between agency and deliberation research, sketch a theory of human agency, and use that theory to identify novel variables that could be helpful in explaining learning in deliberative contexts.
Assumptions about human agency play a critical role in deliberation research. Deliberation research generally proceeds informed by one of two broad theoretical tapestries with crucially different understandings of agency-liberal democratic political theory and deliberative democratic political theory. In the archetypal core of liberal democratic theory, people are agents with respect to their means, but not their ends (Warren 1992). Natural outgrowths of liberal democratic theory are economic and rational choice theory, in which the sole role of agency is to select the most efficient means to maximize achievement of preferences that are typically assumed to be fixed or predetermined (Elster 1986;Smith 1990). Much political science research, which rests on economic and rational choice approaches, displays deep roots in liberal democratic assumptions. A standard and important assumption of liberal democratic theory is that people come to the political arena knowing their ends and thus have the exclusive objective of maximizing their achievement of these ends (Warren 1992). In short, politics is about horse-trading and arm-twisting among goal-focused, instrumentally rational agents. With such a reading of politics, liberal democratic thinkers can only imagine a quite limited role for deliberation: the passing of information that could help the public better achieve their preferences. Much mainstream political science research on deliberation focuses on its capacity to convey information (Gastil and Dillard 1999;Lupia 2005;Lupia and McCubbins 1998;Luskin, Fishkin, and Jowell 2002;McLean et al. 2000).
Deliberative democratic theory, in contrast, maintains that people can be agents not only over their means, but also their preferences and values (Benhabib 1986;Bowles and Gintis 1986;Warren 1992;Warren 1995). In the course of democratic deliberation, for example, criticism from others might problematize a person's values and lead to introspection and change. Reflective evaluation of personal values and preferences, many absorbed unthinkingly from the social milieu, should lead to greater personal autonomy-an important value in deliberative theory. Deliberation could also be an occasion for people to contemplate their relationship with the community and thereby recognize new responsibilities, develop improved conceptions of the common good, and feel a stronger identification with the community (Muhlberger 2005a). Within deliberative democratic theory, politics need not be purely about the strategic game of maximizing predetermined preferences but also about coming together as a community to define and pursue the common good. And, it can be about selfdevelopment-challenging personal assumptions with consequent growth in autonomy or, to use another word, agency. Research findings such as the effects of deliberation on social trust (Price and Cappella 2002) and citizen identity (Muhlberger 2005a) involve an exploration of topics pertinent to a deliberative democratic approach to deliberation. Research in this vein, however, is comparatively rare.
Given these options for how to approach the study of deliberation, researchers find themselves in a difficult position. Taking a liberal democratic approach to studying deliberation narrows the range of possible research topics and approaches deliberation from a viewpoint that is somewhat at odds with deliberative political theory. The liberal democratic approach also has special difficulty explaining political engagement or sophistication, issues highly pertinent to understanding deliberation. While the depiction in liberal democratic theory of people as rational agents might suggest that these theories view people as sophisticated political agents these theories say quite the opposite. Voluminous evidence suggests that much of the public is jaw-droppingly ignorant and unsophisticated (Converse 1964;Delli Carpini and Keeter 1996;Kinder 1983;Kinder 2002;Neuman 1986). The main rational choice explanation of such findings is that it is rational to be politically ignorant and apathetic (Downs 1957). Such a theory should be troubling for deliberation researchers because it raises the question of why anyone should be expected to engage seriously in deliberation. Also troubling is that in some cases people do, in fact, seem to act contrary to rational apathy and ignorance. Substantial numbers do participate in wellorganized citizen deliberations and observers typically conclude that they do so in part out of a desire to contribute to the community, a desire not readily explained by rational apathy. More generally, political knowledge and engagement is substantially higher among the educated-precisely the intelligent and goaloriented public that should be expected to be rationally ignorant and apathetic with respect to politics. While rational choice proponents have proven quite adept at explaining problematic empirical findings, sophisticated observers such as Green and Shapiro (1994) question whether on the whole rational choice approaches have succeeded in explaining anything or simply accumulated ad hoc explanations that do not amount to a productive theory. It would be difficult to say what non-obvious insights into deliberation have been afforded by liberal democratic approaches to its study.
On the other hand, researchers who take a deliberative democratic approach to studying deliberation encounter two other difficulties. First, with its assumption that people are socially constituted, deliberative theory may more readily explain political engagement, but perhaps have difficulty explaining widespread apathy. Second, it is not readily apparent what research tradition might be brought to bear to pursue a deliberative democratic approach. Often, researchers with a broader focus than that afforded by liberal democratic approaches have tapped mainstream social and cognitive psychology. Regrettably, as Bandura (2001) observes, much of the psychological research tradition does not afford space for human agency. This research typically assumes that behavior and attitudes can be explained by a variety of non-rational and deterministic psychological factors such as social conformity or cognitive dissonance. Deliberation research conducted through a lens that sees primarily non-rational processes, not surprisingly, finds manifold problems with deliberation and little systematic grounds to hope for better (Mendelberg 2002). Deliberation researchers need an approach that allows for human agency but also permits people to have an absence of agency. Ideally, such an approach would explain how people could develop greater agency, particularly with respect to politics. Bandura (2001) suggests that there are new approaches in psychology that do create elbow room for agency and hopes these approaches can be stitched into a comprehensive theory of agency. Ironically, this paper does focus on a mainstay of liberal democratic research-learning effects in deliberation-but it seeks to show the value of an agency approach for even this body of research.

AGENCY THEORY AND DELIBERATIVE LEARNING
This paper will sketch some important aspects of a theory of agency and apply these aspects to learning in deliberative contexts. An exposition of other aspects of agency theory can be found in Muhlberger (2005b). This agency theory is indebted to Mead's (1962) philosophical psychology, Carver and Scheier's (1981;1999) self-regulation theory, Vallacher and Wegner's (2000) action identification theory, Koestner's (1996) and Ryan's (1989) work on self-determination and internalization, and Rosenberg's (2002) developmental psychology.
Most mental processes are parallel, unconscious, and "self-organizing" (Carver and Scheier 1999). For instance, when people learn a complex dance, they do not consciously learn which muscle groups to fire and when-something that could not be mastered by the slow, serial processes of consciousness. Rather, they practice until they have built a non-verbal, unconscious, and self-organized mental structure that can execute the necessary responses. Self-organizing processes do not, however, possess high-level unity of purpose or the capacity for reasoned correction. A massively parallel processing device like the brain requires a serial guiding force, and that force is conscious attention and verbal mental structures. Learning to dance involves both. Conscious attention directs the dancer to practice dancing, and it is also involved in building consciously accessible verbal structures that seek to model the dancing. These verbal structures are accessible to conscious, symbolic manipulation, which allows dancers to reason about and thereby consciously correct specific problems in their dance execution. The correction does not take the form of directly altering the self-organized structure, which is not accessible, but in consciously intervening at a specific point in action execution, thereby slowly reprogramming the self-organized structure. Such structures need not be confined simply to movement, but can include cognitive processes as well.
Objectives and identity play crucial roles in agency. An efficient and key way in which overburdened conscious attention can direct behavior is by posing "objectives" for behavior to meet and then activating ensembles of pertinent nonverbal structures in an effort to meet these objectives. Objectives can be at varying levels of abstraction, such as: keep the car between those white lines on the road, drive to the Sierra Club meeting, and be an environmentalist. As the example suggests, objectives can come as a hierarchy that applies simultaneously to the same activity. At the precipice of this hierarchy of objectives rest conceptions of identity that broadly steer a person's activity. The coherence of a person's activities will depend critically on the continuity and unity of the self-a composite of both conscious concepts and unconscious routines captured by the "I" and the "Me", respectively, in Figure 1. The self helps create coherent behavior despite varying activities and contexts. More abstract and consistent selves will better accommodate fluctuations in environmental and bodily needs and demands, charting a consistent course through life. The self consists of a multitude of identities, held together in some people by a more general identity that constitutes the unity of the self. Identities can also be construed as "perspectives"-taking the attitude (Mead 1962) of a given role or type of person.  People find themselves enmeshed in a structure of interaction-patterns of input, routines, rituals, and demands shaped by social structure-that can limit or enhance their agency. Environments that engage a person in repetitive actions "program" self-organized routines in the person in much the same way as a dancer programs dance steps-through redundancy.

Material
Also, an environment that persistently focuses scarce attention resources on some matters and not others can deeply shape all aspects of a person-though some people may have greater resistance than others. For many, in the absence of effortful conscious intervention much of the self is colonized by social concepts and identities absorbed unthinkingly from the social milieu. Rich social environments that place contradictory demands upon the individual or point out contradictions in background assumptions can stimulate conscious reflection that challenges what has been passively absorbed from the social order (Habermas 1984). Democratic deliberation in particular may be a context in which people have an incentive to question the fundamental values and beliefs of others, stimulating selfdevelopment (Warren 1995).
The notion of agency derived from this theory meets the desiderata of a theory of agency that is both consistent with deliberative democratic theory and potentially useful to deliberative researchers. Within the conceptual landscape of agency theory, agency is the capacity to choose and successfully execute actions consistent with a coherent and reflexively determined self. This conception of agency makes sense of the claim in deliberative theory that people can have agency over their preferences or values. It is only by reflexively considering their values and preferences that people exercise agency-that is, only by making uncritically absorbed values and preferences subject to conscious and thoughtful reflection. Such inquiry must also pose the question of what preferences and values would best allow the person to be internally consistent and pursue a coherent life history. This notion of agency should also meet the desiderata of deliberative researchers-it allows for agency but also suggests that such agency is quite difficult to achieve and perhaps many adults have not. Also, deliberation, in particular, may be pertinent to achieving such agency. The inspirations of this agency theory involve a number of tested research programs that can inspire further work, including self-regulation theory, self-determination theory, action identification theory, and Rosenberg's developmental psychology.
Importantly, the proposed notion of agency allows us to speak of agency with respect to a given domain-such as politics-and suggests factors that may matter for agency within the political domain. Someone who is a political agent would take politics as a target of conscious reflection. Researchers should not assume that those interested in politics or engaged in political activity reflect on their beliefs. Even political discussion need not be reflective. For example, the Dutch Election Study of 1970-1973 found that 68% of respondents reported never or rarely giving their own opinions or listening during political discussions (Brady 1999, p. 772). In past decades, voting was heavily determined by ethnic, religious and other group memberships (Dalton and Wattenberg 1993), while today they may be more determined by a variety of less group-based but nevertheless unthoughtful considerations.
We introduce a measure in this paper-"political reflection"-to capture whether people take politics as an object of their reflection. The measure involves asking people whether they feel personally responsible for deciding their own political views. Responsibilities are key "objectives" in self-regulation. We anticipate that those with high levels of political reflection will learn more in deliberative contexts, whether from discussion or informative reading. We also anticipate that political reflection, as well as the other agency variables described below, should prove to better predict learning than standard political attitudes such as political interest or internal or external political efficacy.
Agency theory, furthermore, indicates that identities serve crucial selfregulative roles, which implies additional measures pertinent to deliberative learning. A key identity with respect to political engagement is seeing oneself as a "citizen" (Muhlberger 2005a). Identification by itself does not indicate whether an identity is reflective, but the responsibilities people associate with an identity can bear directly on how reflective a person might be. Two types of citizen responsibility may matter for political learning-active and authoritarian conceptions of citizen responsibilities. Active conceptions involve believing that a citizen should be actively involved in contacting officials and writing letters to the editor. To anticipate a finding from confirmatory analysis below, however, even active conceptions of citizenship are subdivided into two types: nondeliberative forms of responsibilities like contacting officials and deliberative forms such as engaging people who disagree on the issues. The deliberative notion of active citizenship implies a responsibility to be cognitively engaged and should therefore be most pertinent to learning about politics. Authoritarian conceptions of citizenship involve believing that good citizens have a responsibility to be obedient and reverent toward political authorities and punitive toward those who are not. Authoritarian beliefs are connected to agency theory via their relation to how people conceptualize agency-with authoritarians limited to non-systemic, non-dynamic conceptions of agency (Muhlberger 2006). People who believe in subservience to political authority have reason to be less attentive to political information because they believe they should count on authorities and do not wish to usurp the role of authorities.
We will also included false consensus beliefs (Hibbing and Theiss-Morse 2002) as an explanation of political learning. This is the false belief that there is a broad consensus regarding the most important problem facing the country and what to do about it. False consensus beliefs are related to authoritarian thinking about government and, thereby, conceptions of agency (Muhlberger 2006). We expect that believing in a broad consensus on problems and solutions will leave people unmotivated to learn new political information.
A fifth variable suggested by agency theory captures whether a person's structure of interaction encourages political learning. We stipulate that someone who has a political social network-close friends with whom they discuss politics-will be more motivated and perhaps better at picking up political information that might become grist for conversation. A final variable that might be housed under the rubric of agency theory is a norm of social cooperativenessbelieving that people should contribute to society. Someone who believes they should contribute to society and sees, as many do, political engagement as such a contribution should be more motivated to learn political information, especially in a deliberative context. Such a norm might emerge from a number of identities, but whatever its exact origin, it should help regulate action toward learning.

METHOD PARTICIPANTS
Knowledge Networks, a survey research firm noted for its sampling work on academic deliberation projects, conducted the recruitment for this study Virtual Agora Project (VAProject). Of a sample of 6,935 Pittsburgh city residents (defined by zip code area) who could be reached via random digit dialing (RDD), 22% agreed to participate in this research and took a phone survey. Sampling differed from the typical methodology on other substantial deliberation projects conducted by Knowledge Networks in that it did not utilize quota sampling to make demographic statistics more representative of the population as a whole. Thus, the sample accurately reflects who would come to this deliberation without demographic oversampling. This has two advantages. First, the sample better reflects what it would be if deliberations were a more widely used process in government because in this case quota sampling would likely be too expensive and contrary to legal equality requirements. Also, although quota sampling may result in demographics matching the population in certain crude categories, those who come to a deliberation after extensive oversampling of their demographic are most likely not typical of their demographic.
Of recruits who agreed to participate, 37% or 568 people showed for the Phase 1 on-campus deliberation. This is the only part of the study pertinent to this paper and will be termed "the experiment." A modest response rate was expected because recruits were asked to participate in a series of online deliberations that would take most participants eight-months to complete and which they could join only by coming to an initial on-campus, all-day deliberation experiment. The final participation percentages are not, however, incomparable to that of another substantial long-term deliberation study, Vincent Price's Electronic Dialogue Project (EDP) at the Annenberg School of Communication (Price and Cappella 2002;Price and David 2005). This EDP project started with an effective sample of the population from which its discussants were drawn of about 3,686 (Price and David 2005). The number of people who ever participated in any discussion over the course of the year is 543, and the average number of people who participated in a given discussion was 305 (Price and Cappella 2002). Ultimately, the response rates for both the EDP and VAProject are modest. For the VAProject, comfort can be drawn from several considerations: a fair similarity to population demographics, the fact that the sample represents people who might be expected to participate in longer-term deliberations, and the objective of this research which is experimental and focused on psychological processes that should be universal.
Despite a strict RDD sample and modest response rate, the participants in this project reasonably matched the Pittsburgh city population on most demographic criteria. The sample was 77% Caucasian and 18% African-American, compared with CPS population benchmarks for the relevant zip codes of 75% and 20%, respectively. Fifty-six percent of the sample was female, compared with 53% for the population. Twelve percent of the sample was 18-29 years old, 22% 30-44 years old, 26% 45-59, and 27% 60+. This compares with population values of 26%, 20%, 26%, and 27%. The elderly and thirtysomethings are accurately represented, the young are underrepresented, while mid-life adults are overrepresented. Average age, however, is the same as for the population. Perhaps the greatest departure from population values is for education, which, as expected, is greater than for the population. Median education is "Some College" for both the sample and the population. Lower educational categories, however, are underrepresented, with 10% of the sample having less than a high school education and 14% having just a high school education, compared with 16% and 31% for the population. Nevertheless, the sample does contain the full range of educational levels.
Pittsburgh is an ethnically and class diverse community with a city population of 334,583 and over one million including surrounding areas, according to the 2000 Census. Neighborhoods range from suburb-like residential areas to areas of urban poverty. Although Pittsburgh is known to have a moderately high quality of life for a city its size, people intimately involved with public life in the city do not believe this leads to either an especially high level of political involvement or non-contentious public dialogue.

MATERIALS AND PROCEDURES
Knowledge Networks obtained phone numbers for households in the City of Pittsburgh from a random digit dial (RDD) sample. Where numbers appeared in a reverse directory, the household was sent an advance letter on Carnegie Mellon University stationery describing the study and indicating that the household would be contacted shortly. A Knowledge Networks phone center called households in the RDD sample and requested the household member with the most recent birth date. Both the letter and the call center indicated that in exchange for participation in the study, participants would have a four out of five chance of receiving a Windows computer and eight months of ISP service. The remainder would receive $100. Those who received a computer would be expected to participate in a longer-term online deliberation from home. People who agreed to participate in the VAProject were given a short phone-based survey including questions about their pre-deliberation policy attitudes, and they were scheduled for a one-day, eight-hour on-campus deliberation. Participants were asked to come to a randomly chosen day from the deliberation schedule, which spanned three weeks in July 2004, including many weekends and weekdays.
Deliberations were held with up to 60 participants daily. After informed consent and a brief training session, participants took a web-based pre-survey. Next, they were given a 40-minute "library session" to learn more about the four policy topics, a break, 90 minutes for deliberation or contemplation (face-to-face, online, or individual contemplation, depending on condition), and lunch. Deliberation was conducted by trained moderators. The library session, break, and deliberation / contemplation (same condition as before) were repeated in the afternoon, and this was followed by the second survey. In addition to the experiment with type of deliberation, another experimental condition involved either receiving or not receiving reminders of citizenship. In the citizenship condition, participants were reminded to think like citizens in a brief "talkinghead" ahead of their deliberations (the non-citizen condition involved a different talking-head), their rooms had an American flag, and they were given name tags with American flags and the word "Citizen" preceding their names.

MEASURES
All question responses were measured on 7-point Likert scales, unless otherwise noted. Generally, one or two sample questions per scale are provided below. Please contact the author for full question scale items.
Deliberative Citizenship: Part of a series of 70 True / False reaction time questions measuring conceptions of citizenship. This includes such questions as: "A good citizen should discuss politics with those who disagree with them." and "A good citizen should be willing to justify their political views." Authoritarian Citizenship: "A good citizen should respect the President." and "A good citizen should condemn people who are un-American." Political Reflexivity: "I feel personally responsible for my own political views.  (Koestner et al. 1996): respectively, such questions as "I follow politics because I think it's important." and "I follow politics because I will feel bad about myself if I don't."

RESULTS
The results are divided into six sections. The first section presents factor analyses that establish what factors exist in the data and shows that such variables as political reflection and political interest are indeed distinct. The second section briefly examines the issue of whether the data here show that people learned during the on-campus experiment or merely revealed pre-existing knowledge differences. The third section considers whether decision knowledge has consequences such as affecting policy attitudes or making the knowledgeable person more persuasive to others. The fourth section examines what factors affect decision knowledge, including the agency variables, political attitudes, and demographics. The fifth section explores the degree of inequality in learning revealed in the data. The final section presents some results that indicate how malleable various factors might be-indicating which might be manipulated to improve learning in deliberative contexts.

EXAMINATION OF FACTORS IN THE DATA
The surveys conducted for this study contain multiple questions for each of several conceptual factors, including such novel constructs as political reflection. This raises the issue of whether the questions we believe go together as a factor do. Exploratory and confirmatory factor analyses were conducted to address this issue. In addition, confirmatory analyses help establish that different factors do not measure the same thing. The results in this section firmly establish that the separate concepts identified in this paper are statistically distinct-for instance, political reflectiveness is not the same as political interest. They also help identify subsets of variables that cleanly estimate underlying constructs, increasing confidence in subsequent regressions.
Confirmatory analysis indicates that decision knowledge consists of two factors. A two-factor model has good indicators of model fit: Goodness of Fit Index, GFI=.98 (above .95 considered very good); Adjusted Goodness of Fit Index, AGFI=.95; Root Mean Square Error of Approximation, RMSEA=.06 (.05 and below considered good) with 90% confidence interval (CI) of .04 to .08; Bayes Information Criteria (BIC) of -89.8 (below zero indicates a model better than the saturation model); Hoelter's N of 292 (indicates the N at which the 2 test is significant; values above 200 considered good); and N=562. All confirmatory and exploratory analyses were conducted in the R statistical package. The two factors identified appear to be policy knowledge, tapping such questions as what various policy options entail, and statistical knowledge, tapping such questions as the percentage of excess capacity in Pittsburgh public schools. The questions measuring knowledge are multiple choice and therefore have two values, correct and incorrect. Analyses of all such dichotomous variables were conducted using tetrachoric correlations. Tetrachoric correlations create some difficulties for model estimation, particularly for dichotomous variables that have very low variance. In the two-factor model estimated here, four of the 12 knowledge variables were removed to permit estimation.
The presence of two knowledge dimensions suggests that research on deliberative learning should be more sensitive to the possibility of different types of learning-past studies have assumed unidimensionality. Such sensitivity, however, will be only partially feasible here. Only two dichotomous variables are related to the statistical knowledge dimension, which does not leave enough variation for separate regression analyses in which we might have much confidence. We choose to focus mostly on a single 12-variable decision knowledge scale that collapses the different dimensions of learning, though we will report any apparently robust results from separate analyses of the two dimensions. The decision to focus on a single all-encompassing decision knowledge scale follows from a couple considerations. There is a strong presumption in favor of the view that anyone who can correctly answer a factual question about the topic of deliberation has in some general sense learned something. Also, the .61 estimated correlation between the policy and statistical knowledge factors indicates that the two types of learning are moderately related and are reflections of an underlying second-order learning factor.
An exploratory factor analysis suggests that 70 dichotomous citizen responsibility questions can be roughly divided into four categories: authoritarian notions of citizenship, active citizenship, inclusive citizenship, and nonsense questions. Only the first two of these are pertinent here. Confirmatory analyses of the authoritarian citizenship variables suggest a three-variable model. The fit of the three-variable authoritarian citizenship model is less than ideal, though perhaps not unexpected for tetrachoric correlations with low variation variables (GFI=.95; AGFI=.88; RMSEA=.11, CI of .9 to .13; BIC=-12.4; Hoelter's N=126; N=557). The fit for the same model with a standard covariance matrix is superb (GFI=.99; AGFI=.99; RMSEA=.001; BIC=-105.0; Hoelter's N=1040). The actual fit of continuous variables would likely lie between these extremes.
The three-factor authoritarian citizenship model parallels findings in research on Right-Wing Authoritarianism (RWA). The three-factors appear to be: obedience to and respect for national symbols and leaders, condemnation of those who are not similarly respectful (e.g., flag desecrators), and religiousness. These parallel the three dimensions of Right-Wing Authoritarianism (Altemeyer 1981): obedience to authority, punitiveness toward out-groups, and traditionalism. Altemeyer combines all three of these correlated components into a single RWA scale. In prior research in the data underlying this paper, it was found that only the obedience and punitiveness dimensions of RWA correlated with authoritarian conceptions of government, suggesting that these dimensions are political relevant (Muhlberger 2005c). Consequently, for purposes of this paper, the average of the variables composing the obedience and punitiveness notions of citizenship were used to construct a single authoritarian citizenship variable.
A two-factor confirmatory model fits the active notions of citizenship variables well (GFI=.97; AGFI=.94; RMSEA=.07, CI of .057 to .088; BIC=-106.47; Hoelter's N=219; N=558). One factor appears to be a non-deliberative notion of active citizenship that includes such activities as joining public interest groups, writing letters to Congress, and following the news. The second factor was deliberative: justifying personal views in discussion and discussing politics with those who disagree. Two additional variables, that could not be added to the confirmatory analysis because of very low variation prove to correlate cleanly with the deliberative citizenship variables and not with the non-deliberative citizenship variables and were added to the scale. The cognitive and deliberative foci of these variables suggest they will be especially pertinent to deliberative learning.
The combined model can be used to test whether the scales under consideration are statistically distinct. Indeed, the correlation patterns create a strong presumption that they are distinct. The correlation between political interest and internal efficacy, two standard political attitudes, is greater than the correlation between either of these variables and political reflectiveness or deliberative citizenship. Indeed, 2 tests of the difference of fit between the full model and models in which pairs of factors are collapsed indicate that the probability of political interest being the same as either political reflection or deliberative citizenship is vanishingly small (p<.0001 for both; 2 =118.7 and 146.3 with 1 d.f., respectively). Similar results hold for internal efficacy and the two agency variables ( 2 =130.4 and 256.5), or between the agency variables ( 2 =146.3). Factor correlations between the novel variables and external efficacy are so low they do not warrant further examination.

KNOWLEDGE OR LEARNING?
We contend that most of the decision-relevant knowledge of participants at the end of the on-campus experiment is due to learning during that experiment, not to prior knowledge. This cannot be rigorously and directly established because no pre-experiment knowledge questions were asked-these would likely have been reactive. It can be rigorously established that some learning took place during the experiment: The number of reading materials whose URLs were clicked by participants is significantly correlated with post-experiment knowledge ( =.24, cluster-robust p<.001, N=563). The effect is not large, but the measure is crude. Perhaps the best evidence of learning will be presented in the next section, which shows that post-experiment decision knowledge has a substantial effect on opinion change from pre-to post-experiment. Without learning, there should have been no opinion change.
More generally, it is well known that most of the public has little or no policy knowledge, making the high levels of decision-relevant knowledge recorded post-experiment unlikely without learning. The policy questions we asked did not, to the best of our knowledge, after having examined local newspaper archives, have answers in the mainstream news media,. The questions and answers were derived from government and think tank reports and expert testimony to which the public would had difficult access. Several participants commented that they were astonished by how poor and incomplete the information they received from the media was, compared with the policy briefs available in the experiment. Table 1 shows that post-experiment factual knowledge is related to key outcomes of deliberation, including attitude change and the effect of participants' opinions on other participants. The table indicates that those with high levels of decision knowledge after the experiment-those who presumably learned the most-also showed significantly greater policy attitude change on three of the five policies under consideration, with a trend for a fourth policy. These changes are among the most substantial effects found. Decision knowledge and other continuous variables in Table 1 are on a seven-point scale to insure comparability of raw coefficients. The finding that post-experiment decision knowledge substantially changed opinions on three of the five policies suggests that appreciable learning took place during the study. The smaller sample size of the regressions in Table 1 is due to the fact that Knowledge Networks was unable to interview 105 participants prior to their being in the experiment. Analysis indicates no significant differences in demographics or opinion among those who were interviewed prior to the experiment and those who were not. .39; 1.7 Notes: To promote comparability of coefficients, continuous variables were put on 7-point scales. "Change" is post-minus pre-deliberation attitudes. *** p<=.001; **p<=01; * p<=.05; †p<=.10; p-values are reported as one-sided for directional hypotheses. All others are two-sided. p-values are cluster robust and take into account possible error covariation between discussion group members and heteroskedasticity.

CONSEQUENCES OF DECISION KNOWLEDGE
Table 1 also hints that decision knowledge may be related to greater influence over other participants' views, though this is not statistically firm. The row labeled "Converge to High Know." indicates the opinion change effect of the gap between a respondent's pre-deliberation opinion and the mean postdeliberation opinion of highly knowledgeable participant's in the respondent's discussion group, omitting the participant. A coefficient of .40, for example, indicates that participants closed 40% of the gap between themselves and the mean post-deliberation opinion of highly knowledgeable others in their group. The row labeled "Converge to Low Know." indicates convergence to the mean opinion of low-knowledge others in each respondent's group. High and low knowledge were defined by a median split on decision knowledge. Table 1 shows that in three of five policy attitudes, this "gap-closing" coefficient appears appreciably larger and more significant for the high knowledge group than for the low knowledge group. While this is suggestive, post-hoc tests to determine the significance of the difference between the low and high gap coefficients show trends but no significant differences.

EXPLAINING DECISION KNOWLEDGE
One concern in this analysis is that discussion groups may help explain some of the variation in participants' decision knowledge. In cases of such nonindependence, hierarchical linear modeling ("mixed effects" or "multilevel") modeling would be ideal. The data here were tested to determine whether such a modeling effort would be desirable. One measure is the Intraclass Correlation Coefficient (ICC)-the percentage of dependent variable variation explicable by a random-effects intercept only model. The ICC here is 5 X 10 -4 , or much less than 1% of variation, suggesting that there is no group-level variation that needs to be taken into account by HLM. Adding the full model in Table 2 and allowing random effects for powerful values does not change the picture. In short, an HLM approach is not warranted. To take into account the possibility of some covariation of error terms within group, as well as possible heteroskedasticity, the analyses presented here will be OLS with cluster-robust standard errors. Table 2 shows a regression of decision knowledge, measured on a 0 to 100 scale, on a variety of explanatory variables. The first batch of variables incorporates a complete analysis of the effects of experimental conditions. Neither face-to-face nor online deliberation significantly influences factual decision knowledge. Reminders of citizenship during the deliberation did lead to a significant, though modest, three point increase in decision knowledge. By comparison, a move from lowest (0) to highest (6) value of general political knowledge results in a 18-point change in decision knowledge (6*3.03). The significant negative effect of the F2F X citizen condition simply negates the positive main effects. Variables were scaled to a seven-point scale (0 to 6) for comparability of raw coefficients, with the exception of dichotomous variables such as those measuring the experimental conditions, which are 0 and 1. Readers can divide dichotomous variable coefficients by six to get a sense of their relative effect sizes. .50; 13.8 Notes: To promote comparability of coefficients, continuous variables were put on 7-point scales. Deliberative citizenship in the 2nd regression has been centered to avoid confusion regarding experimental effects, which otherwise become large and negative to compensate for the large citizenship X discussion coefficient for a variable, citizenship, that has most of its values at the upper extreme. *** p<=.001; **p<=01; * p<=.05; †p<=.10; p-values are reported as one-sided for directional hypotheses. All others are two-sided. p-values are cluster robust and take into account possible error covariation between discussion group members and heteroskedasticity.
Education has the most substantial effect on decision knowledge, followed by general political knowledge and then age in years, which has a negative effect. The deliberative conception of citizenship has nearly the effect of age or general political knowledge, with a coefficient of 2.79. Political reflection is also significant and has about half the effect of deliberative citizenship, while authoritarian citizenship has a significant negative effect, as predicted. The absolute sum of the coefficients of the three core agency variables (political reflection and conceptions of citizenship) is 5.1, which appreciably outstrips even the effects of education. The beta coefficients of the variables provide another way of viewing their relative effects. Having a political talk network (close friends with whom the respondent talks about politics) and strong social cooperativeness norm also help to a degree, while falsely believing in a political consensus reduces decision knowledge. Income, being black, and having children also affect knowledge.
Standard political attitudes-political interest and internal and external efficacy-do not significantly and positively contribute to decision knowledge. Oddly, political interest does have a significant negative effect. Regressing decision knowledge on just political interest reveals a highly significant positive effect (t=3.74). Controlling for just education, however, makes political interest insignificant (t=1.33), which suggests that the effect of political interest could be spurious. Political interest may merely be a proxy for some aspect of education, such as cognitive skills. Perhaps the negative coefficient in Table 2 reflects people who exaggerate their political interest. If the more specific motivational variables are harder to exaggerate, then controlling for these will leave behind a residue of exaggerators, resulting in negative coefficients.
Perhaps no effects were found for discussion because only discussion in particular groups succeeded in educating their members. This possibility was tested in two ways, but no evidence for such effects were found. The two possibilities tested were: Perhaps discussion groups in which members other than the respondent had high levels of decision knowledge succeeded in raising respondents' knowledge. Alternatively, perhaps a group-level effect might be seen by examining a variable that indicates how other members of a group scored in decision knowledge relative to what would be expected from the first regression in Table 2-that is, if their knowledge gain outstripped model expectations. This test should be highly sensitive to any possibility that certain groups educate their members. Variables were constructed to test both possibilities and inserted in the full regression of Table 2, but no evidence was found for group-specific effects (p=.99 and p=.89, respectively). Alternatively, perhaps no effects were found for discussion because decision knowledge hit a ceiling from just reading alone. This seems improbable because only 43 of 554 participants in the Table 2 regression had a perfect decision knowledge score.
An examination of the two components of decision knowledge-policy knowledge and statistical knowledge-indicates that there may be differences in which factors contribute most to each of these types of knowledge. Regressions for these two types of knowledge (not depicted in Table 2) naturally have weaker standard errors because the dependent variables are means of fewer dichotomous variables than decision knowledge. In particular, statistical knowledge consists of only two such variables, while policy knowledge consists of eight. Remarkably, however, several factors have larger and more significant coefficients for statistical knowledge than for policy knowledge. The citizenship reminder experimental condition appreciably affects statistical knowledge (ß=9.66, p=.01), but not policy knowledge (ß=1.54, p=.21). Apparently, then, the citizenship reminder results in a careful attention to detail that helps with learning statistical facts. Political reflection also seems to matter more for statistical than policy knowledge (ß=2.77, p=.05 vs. ß=1.1, p=.08). There are also signs that authoritarian citizenship may have a larger negative effect on statistical knowledge (ß=-1.60, p=.08 vs. ß=-.79, p=.06), while deliberative citizenship concepts may have a larger (positive) effect on policy knowledge (ß=2.71, p<.001 vs. ß=.89, p=.22). These differences make sense in light of the greater effort involved in learning statistical details.
An important result is depicted in the second regression reported in Table  2. The coefficient of deliberative citizenship has a larger and more significant effect on decision knowledge in the discussion conditions than in the control conditions. Indeed, a post-hoc contrast indicates that the discussion and control condition coefficients are indeed significantly different (p=.03). The coefficient for deliberative conceptions of citizenship is 2.5 times larger in the discussion conditions than in the control condition (discussion conditions were collapsed because there was no significant difference between them). This suggests that deliberation, despite the null main effects for the f2f and online conditions, does have an effect-it amplifies the impact of deliberative citizenship.
It would, then, be tempting to conclude that discussion does increase learning-at least for those with high levels of deliberative citizenship. But, regrettably, it is insufficient to look exclusively at the coefficients of the two deliberative citizenship variables to conclude whether the overall effect on decision knowledge is significant. The main effects of a given condition, such as f2f or online, must also be included. When they are, only one contrast comes close to significant: those with the highest level of deliberative citizenship in the f2f, no citizen reminder condition prove to have almost significantly greater decision knowledge than those with the highest level of deliberative citizenship in the control, no citizen reminder condition (p=.059). Of course, the cutoff p-value of .05 is arbitrary, and a p-value of .059 is very nearly as reassuring as one of .05. On the other hand, the difference between the f2f-high deliberative citizenship and the control-high deliberative citizenship conditions is hardly large in absolute terms-a difference of 2.6 points on the 100-point decision knowledge scale.

INEQUALITY
Inequality in political knowledge after a deliberation must be considered relative to inequality in such knowledge in the broader population that does not get invited to a deliberation. Typically, at most a few percent of the population are sufficiently involved in a particular policy issue to be able to answer specific questions about the issue, resulting in extreme inequality in decision-pertinent knowledge. Sometimes only leaders have policy knowledge, itself an extreme type of inequality. The observed mean value of decision knowledge for this project is 73 (73% correct answers), suggesting that considerable learning took place and that the participants may well be far more equal after this deliberative experience and more equal with leaders and decision makers.
The Gini coefficient for decision knowledge is .14, which indicates low overall inequality. This compares favorably, for example, with the U.S. income Gini of .41-also a politically pertinent number.
Of course, the first regression of Table 2 also indicates disparities among participants. To get a better grasp of these disparities, expected decision knowledge was calculated for four high and low combinations of demographics and agency variables. High demographics or agency was defined as the mean value of each variable plus one standard deviation for positive coefficients (or minus one s.d. for negative coefficients). Low values were defined as the mean minus one standard deviation (or plus for negative coefficients). The expected decision knowledge values were: 41 for low demographics, low agency; 62 for low demographics, high agency; 73 for high demographics, low agency; and 94 for high demographics, high agency. Real learning took place even among those with poor demographics and agency. These people have a projected score of 41 on the decision knowledge test, while the expected value for random guessing on the test was 25, Only 7% of participants actually scored 41 or less on the test. On the other hand, there is a considerable expected discrepancy between those with poor demographics and agency and those with good demographics and agency-53 points. Again, the number of participants with scores of 41 or less was exceedingly low, suggesting that it is rare to have the worst combination of demographics and agency. On the other hand, low decision knowledge scores for disadvantaged groups is not merely a hypothetical: the 20 African-Americans in the sample with income and education one standard deviation below the mean or worse had an actual mean score of 42. There were 104 African-Americans in the sample. Also, the 42 people of any ethnicity with income and education a standard deviation below the mean or worse had an actual mean score of 44. The effects of demographics and agency are moderately comparable-with low demographics, high agency giving a score of 62 with the reverse giving a score of 73. This opens the hope that the agency variables, which should be more susceptible to change than demographics, could help close some of the inequality between people. Table 2 shows that demographics clearly play an important role in deliberative decision knowledge, which should be of some concern. It suggests that some people will benefit more from deliberation and may be more likely to adjust their views as a consequence of deliberation. The cumulative absolute raw coefficient for demographics in Table 2 is 10.2, while the value for non-demographic factors is 11.08 (excluding experimental condition terms, non-significant variables and political interest, which has a counterintuitive effect). Cumulative absolute beta coefficients are .58 and .65, respectively. While demographics play an important role, suggesting that deliberative learning is demographically unequal, the overall impact of non-demographic factors is somewhat greater. Concerns about unequal capacities for learning in deliberative contexts might perhaps be addressable if various factors that affect learning could be influenced. Demographic variables cannot be readily changed, but it may be worth examining the extent to which various non-demographic factors are contextually modifiable. Two indicators may provide some insight into how contextually modifiable determinants of deliberative learning could be. First, if a nondemographic variable is appreciably explained by demographic variables, it is not likely to be a useful lever for addressing inequalities in deliberative learning. Second, variables that are significantly related to underlying motivational factors that have been shown in psychological research to be contextually changeable should be more promising as useful levers. Table 3 shows the proportion of explained variance (R 2 ) obtained from regressing each of the non-demographic variables on demographics and on two underlying motivational factors that have been successfully experimentally manipulated by psychologists using instructions to subjects about how they should approach an activity-introjected and internalized motivation (Koestner et al. 1996;Plant and Ryan 1985) .

POSSIBLE AVENUES FOR INFLUENCING DECISION KNOWLEDGE
With the exception of general political knowledge, the variables are minimally influenced by demographics. These variables have an appreciable effect on the outcome-with a cumulative absolute raw coefficient of 8.06, or 79% of the total raw effect of all demographics. Cumulative absolute beta coefficients is .44, or 77% of the total cumulative absolute coefficients of demographics. One of the variables, political reflectiveness, is appreciably influenced by the two motivational factors, introjected and internalized motivation. Political reflection plays an especially large role in learning statistical facts. Generalized political knowledge is also modestly affected. With the exception of false consensus, all the variables are significantly related to the combination of introjected and identified motivation (regression F-statistics p<.001). Of course, a relationship with introjected and identified motivation is not the only avenue by which these variables could be malleable. Indeed, online deliberation directly ameliorates false consensus beliefs (Muhlberger 2006), which should help improve decision knowledge in the long-term.

DISCUSSION AND CONCLUSION
This paper examines the sources of gains in decision-relevant knowledge during a deliberation experiment. Prior research has not generally been adequate to settle the question of whether learning (or, for that matter, attitude change) in deliberative contexts is due to discussion or information sources such as readings. The current study separates discussion and information by having an informationonly control group. It also explores whether decision knowledge matters by testing its relationship to attitude change. Finally, the paper investigates what individual-level factors account for learning in deliberative contexts. To do so, it introduces a theory of agency that suggests novel factors. The theory may more broadly benefit deliberation research.
Findings indicate that decision knowledge matters substantially for attitude change during the experiment (Table 1), and findings hint that more knowledgeable participants may have more of an effect on the policy attitudes of others (Table 1). Regrettably, the study finds no evidence that online or face-toface discussion significantly increased overall decision knowledge over just reading informative materials (Table 2, first regression). A counter-explanation that growth in knowledge may have encountered a ceiling with the study's measure of knowledge does not stand up to scrutiny. Likewise, two tests to determine whether particular discussion groups might have succeeded in educating their participants finds no support for this counterhypothesis.
The findings do not demonstrate that deliberation does not affect knowledge, only that there is no evidence of an effect on factual knowledge above and beyond that of informative readings and individual contemplation of these readings. Perhaps, if this study had contrasted people who neither deliberated nor received readings with people who deliberated and also received no readings, deliberation would have shown an effect. Nevertheless, the absence of additional learning from discussion, above the effects of readings, in the current study may indicate that readings are a more efficient way of conveying information than is discussion. Also, the findings here do not address whether discussion led to forms of learning other than factual knowledge. Perhaps discussion cements factual knowledge better in memory, helps build conceptual sophistication, or builds awareness of countervailing viewpoints. These possibilities should be considered in future research. Nevertheless, the findings here dispute the oftencited conclusion that deliberation promotes factual knowledge. It also suggests that those who wish to establish claims about deliberative learning need to do so with greater attention to methodology than has past research.
Given that readings and not discussion led to learning, how might those interested in advancing public political knowledge best do so? An important issue is how to motivate people to read the materials in the first place, and indeed deliberation may prove to be an important carrot, even if it is not the focal treatment.
The current study does suggest the value of deliberation as a motivational tool.
Anecdotal evidence and open-ended survey comments by a number of control condition participants, who read and contemplated but did not discuss, indicated that they were very disappointed and bored. A disgruntled control participant asked one of the authors why she had been "punished." Far fewer control participants would likely have come to the experiment had they known in advance they would not be discussing. In contrast to control participants, discussion participants in the on-campus portion of this study indicated significantly greater motivation (p=.046) to participate in online discussions in the next phase of the study, the at-home phase.
The findings here yield some advice for practitioners focused on advancing political knowledge. Deliberations should come with readings and there should be a strong expectation that people will in fact read them. Participants should not be given readings to do only at home, as occurs in such standard deliberation approaches as Deliberative Polling. Instead, time should be set aside for reading during the deliberation experience, perhaps a substantial portion of time-the present study had 44% as much reading as discussion time. Not everyone will take the time to do the readings at home, perhaps especially those who are economically stressed or are poorly educated. Making time for reading should maximize the learning effects of the deliberative experience and may reduce inequalities. On the other hand, putting standard deliberation approaches aside, it might be possible to use at-home readings to leverage deliberation on behalf of widespread increases in political knowledge. This would involve sending policy briefing materials to a large sample of participants and then accepting participants into a deliberation based on their knowledge of the material. Only a few need participate, but a much larger public might be motivated to learn about the issues.
The liberal democratic underpinnings of much deliberation research, such as research on Deliberative Polling™, focuses this research on the pedagogical effects of deliberation in an effort to address a key problem of liberal democracy-the apathetic public. But with such a focus, deliberation should be considered relative to other pedagogical methods, including simply reading and contemplating. The current study finds that it is the reading (and contemplating) that matters. Prior research does not dispute this result because it was not designed to separate the effects of readings and discussion. That reading begets learning should hardly astonish. Perhaps the liberal democratic focus on the pedagogical effects of deliberation has been excessive and researchers should consider a wider range of effects-deliberation may have value beyond addressing problems of liberal democracy. Agency theory may begin to suggest some avenues by which deliberation may affect deeper changes in people. Research in an agency theory vein has revealed that deliberation enhances citizen identity (Muhlberger 2005a) and reduces stealth democracy beliefs and vertical collectivism (Muhlberger 2006).
The agency theory approach in the current paper suggested that political reflectiveness and deliberative and authoritarian conceptions of citizenship would matter for deliberative learning. Not only is this the case, but also these agency variables prove significant while standard political attitudes either are insignificant or in the wrong direction ( Table 2). The fact that the agency variables affect not merely other attitudes but the accuracy of answers to knowledge questions provides a powerful demonstration of the validity of the agency variables. Deliberative conceptions of the responsibilities of good citizens proved particularly powerful in explaining decision knowledge.
These conceptions deserve more exploration in future research. The measure of deliberative citizenship in this paper is based on four dichotomous variables, and 71% of the observations are at the highest possible value. Using continuous measures deliberative citizenship may yield even more powerful effects and might firmly establish that at least those with deliberative conceptions of citizenship do learn to a degree in discussions-a possibility suggested but not quite proved in the current findings. The current dataset also contain measures of the speed of response to citizenship questions that might yield additional explanatory power by revealing the accessibility in memory of these citizenship constructs.
The overall inequality of decision knowledge of study participants was quite low. The Gini coefficient of decision knowledge was miniscule and compared very favorably with the U.S. income distribution. Of course, the comparison must be taken with a grain of salt because income is an essentially unbounded number while knowledge is on a bounded scale and different Gini values would likely hold for different sets of knowledge questions. Nevertheless, we believe the knowledge test was a good one, tapping the core understandings necessary to making the policy decisions. The low Gini probably does indicate that decision knowledge was reasonably well distributed, particularly in comparison with other inequalities pertinent to politics, such as income. In addition, inequality in decision knowledge after the experiment should be compared with inequality before the experiment. For reasons discussed in the paper, it is likely that very few participants would have been able to correctly answer the knowledge questions prior to the experiment. If so, there were considerable gains in factual knowledge by almost all study participants. Such gains should have brought participants closer to equality with anyone who was knowledgeable about these issues prior to the experiment-whether other participants or policy makers.
Nevertheless, the findings in Table 2 indicate that socioeconomic characteristics powerfully affect decision knowledge and thereby do create real inequality in knowledge among participants, albeit the agency variables serve as a countervailing force. In practice, at least some smaller subgroups did poorly with respect to acquiring knowledge-the 20 African-Americans and 42 people of any ethnicity in the study with low levels of income and education. These 42 people constitute 7.4% of the sample. While socioeconomic characteristics have the most powerful effect, agency variables also have substantial effects that come close to counterbalancing socioeconomic disadvantages. Importantly, the agency variables are largely unrelated to socioeconomic characteristics (Table 3), and the political reflectiveness agency variable is strongly related to motivational variables that psychologists have successfully altered through participant instructions (Table 3). Reflectiveness powerfully affects statistical learning.
The findings suggest a number of bits of advice for increasing factual learning and reducing factual learning inequality in deliberative contexts. First, in online deliberations participants will learn slightly more if they are reminded of their citizen role, as in the current study. This effect does not work in face-to-face discussion, but does work online or in no-discussion conditions. False consensus beliefs, which suppress learning, are reduced in online discussions (Muhlberger 2006), perhaps enhancing learning in the longer-run. Perhaps learning could be enhanced by challenging participants' false consensus beliefs prior to discussion by providing them with information on just how diverse views are about a policy and creating norms of attending to the views of others. Political reflectiveness, which matters especially for statistical learning, might be enhanced by instructions to participants about how they should learn. Psychologists have altered variables important for reflectiveness by instructing participants to focus on the informative quality of their readings, not on whether they are personally right or wrong about the issue. More generally, it may be possible to enhance learning and reduce inequality by suggesting norms to participants that bolster deliberative conceptions of citizenship, the desirability of reflectiveness, and norms of social cooperativeness while challenging authoritarian notions of citizenship. In longer-run deliberative efforts, learning could be enhanced by assigning participants political discussion partners outside formal discussions and encouraging and facilitating the acquisition of general political knowledgewhich has one of the most powerful effects on learning. This paper has provided some initial evidence for the utility of agency theory. The theory suggests variables that greatly outperform standard political attitudes in explaining political learning, help ameliorate the appearance of socioeconomic inequality in learning, and suggest avenues by which learning can be improved-empowering deliberation participants as political agents. One agency theory variable, deliberative conceptions of citizenship, has nearly the impact of general political knowledge and more of an effect than income. We hope that agency theory will prove to be a fruitful framework for research based in deliberative democratic concerns rather than the liberal democratic ones that guide much of today's research.