Journal of Theoretical Politics 13(1): 53 80 Copyright 2001 Sage Publications 0951 6928[2001/01]13:1; 53 80; 015303 London, Thousand Oaks, CA and New Delhi PERCEPTIONS, OPINIONS AND PARTY PREFERENCES IN THE FACE OF A REAL WORLD EVENT CHERNOBYL AS A NATURAL EXPERIMENT IN POLITICAL PSYCHOLOGY Wouter van der Brug ABSTRACT Subjective agreement between voters and the party they voted for can be produced by three separate processes: rational selection of parties, persuasion by parties, and distortion of perceptions. Rational choice theory, balance theory and social judgment theory make different predictions about the strength of each of these processes. In this article the strength of the three processes are estimated by modeling continuous longitudinal change in attitudes and perceptions of positions on nuclear energy in a period in 1986 during which the accident at Chernobyl occurred. The study uses a short-term panel study of Dutch respondents covering this period. The analyses demonstrate that changes in attitudes and perceptions conform largely to the Downsian model of democracy. Evidence is found for a weak persuasion effect, but hardly any distortion of perceptions occurred. These findings challenge the applicability of concepts from balance theory to the field of electoral research. It is argued that voters have little psychological motivation to distort their perceptions of party positions because they are normally not strongly involved with politics or political parties. KEY WORDS electoral behavior non-linear systems political perceptions political psychology public opinion Introduction Normative notions of democratic representation hold that voters should vote on the basis of their policy preferences. Downs (1957) spatial model of democracy conceptualizes ideological or policy agreement between voters and parties as a proximity relation. Empirical research has shown that subjectively perceived policy distances have a strong effect on Various people have commented on earlier versions of this paper and helped me to improve it. I would like to thank the following people in particular: Mark Franklin, Cees van der Eijk, Rob Mokken, Erik Oppenhuis and two anonymous referees of the Journal of Theoretical Politics.
54 JOURNAL OF THEORETICAL POLITICS 13(1) electoral choices (Enelow and Hinich, 1984; Franklin et al., 1992; Granberg and Holmberg, 1988; Krosnick, 1990a; Niemi and Bartels, 1985; Oppenhuis, 1995; Page and Brody, 1972; Page and Jones, 1979; Wright and Berkman, 1986; Van der Eijk and Franklin, 1996; Van der Eijk and Niemöller, 1983). The fact that voters tend to vote for the party they perceive as being close to them in some policy or ideological space suggests that these voters choose rationally, and consequently that they are represented meaningfully. However, subjective agreement between voters and the party or candidate they voted for could also result from persuasion by these parties or candidates, or from distorted perceptions (Brody and Page, 1972; Granberg and Holmberg, 1988; Judd et al., 1983; Markus and Converse, 1979). A general concern that underlies many studies of public opinion is whether parties are responsive to opinions of voters or whether parties manage to influence the opinions of voters. In order to save information costs it would be rational for voters to take cues from actors they trust on the basis of previous experience (e.g. Fiorina, 1981). To the extent that opinions of voters are guided by the leadership of parties, subjective agreement is not the result of rational selection but of persuasion. The notion of distorted perceptions is a crucial topic in empirical tests of proximity and directional models of issue voting. When using subjective perceptions of party positions the former model usually predicts party preferences better than the latter, whereas this is the other way around when replacing subjective perceptions by the sample means. MacDonald et al. (1997) argue that mean perceptions should be used since, to the extent that individual perceptions are affected by party preferences, proximity follows from prior affect, and we would hence overestimate the degree of issue voting. Others have argued that individual perceptions should be employed in tests of models of issue voting, because it is theoretically unrealistic to assume that voters utility functions would depend upon some aggregate view of the world (e.g. Gilljam, 1997; Merrill and Grofman, 1997; Westholm, 1997). I agree with this line of reasoning. If the purpose of an investigation is to model individual choice processes, we must use individual perceptions. Yet, MacDonald et al. are right to argue that, to the extent that distortion of perceptions dominates other processes, subjective agreement does not indicate meaningful representation. It is precisely for this reason that various scholars have set out to separate rational selection from distorted perceptions (e.g. Granberg and Holmberg, 1988; Judd et al., 1983; Markus and Converse, 1979; Merrill and Grofman, 1997). However, for our evaluation of the mass elite linkage that results from elections it is essential not only to separate rational selection from distorted perceptions, but to assess the relative strength of each of the three processes that produce subjective agreement (rational
VAN DER BRUG: PERCEPTIONS, OPINIONS AND PARTY PREFERENCES 55 selection, persuasion and distorted perceptions). This is the object of the present study. 1 This article does not aim to test various models of issue voting. Rather than estimating or predicting subjective agreement, this study estimates and thus disentangles the relative strength of the psychological processes that underlie electoral choices, and that may or may not produce subjective agreement. In order to estimate these processes this study takes advantage of a short-term panel study of Dutch voters. By chance, the two waves of interviews took place before and after an immensely salient event: the accident in the nuclear power generating plant near the Ukrainian city of Chernobyl in April 1986. The study can thus model longitudinal change in attitudes of individual voters on nuclear energy and their perceptions of parties positions in a period in which the accident occurred. Before setting out the research design I first describe in more detail the three processes by which subject agreement of voters with their preferred parties might be realized. Rational selection of parties is the first process that produces subjective agreement. Rational selection means that party preferences are based on a comparison of attitudes on policies and perceptions of parties stands on the same policies such that voters choose to support the party closest to them (Downs, 1957). Rational selection is unlikely to yield subjective agreement between voters and the party they voted for on each and every issue, however, because of the large number of issues to be considered. Citizens who choose a party on the basis of policy positions might still disagree with the party they voted for on particular issues. Rational selection of parties is not only predicted by rational choice theory, it is also compatible with cognitive dissonance theory (Festinger, 1957) and balance theory (Heider, 1958), which presume that people generally strive for consistency or balance among their perceptions, evaluations, and actions. Rational selection of parties leads to such consistency. Persuasion denotes the effect of party positions on attitudes of voters. When voters change their attitude on an issue in such a way that it subsequently corresponds to the perceived position of the party they prefer, this leads to consistency among perceptions, attitudes, and actions. Although persuasion is not predicted by rational choice theory, it is not incompatible with it either. Rational choice theory does not specify how people form their attitudes on specific issues. It merely requires that, after whatever persuasion processes have taken place, subsequent party preferences are dependent on policy attitudes. Empirically, persuasion effects have been found to 1. Proponents of the proximity model as well as proponents of the directional model agree that distortions of perceptions to the extent that these exist will follow the logic of the proximity model. This is how such effects are modeled in this study.
56 JOURNAL OF THEORETICAL POLITICS 13(1) occur (Petty and Cacioppo, 1986; Price, 1989). Markus and Converse (1979) point out that negative persuasion (i.e. changing one s attitude so that it no longer corresponds to the position of a party one dislikes) is equally possible. Distortion of perceptions is the third way in which preferences and perceptions may be balanced. Cognitive dissonance theory, balance theory, and social judgment theory (Eiser, 1992; Sherif and Hovland, 1961), all predict that actors will be motivated to perceive little difference between themselves and an object they prefer. This may result in perceiving a preferred party as being closer to one s own position than it really is, a process often referred to as assimilation. The opposite, that is exaggerating the difference between one s own position and that of a party one dislikes, is called contrast (see also Granberg, 1983; Granberg and Brent, 1974; Granberg and Holmberg, 1988; King, 1978). 2 Assimilation and contrast are not the only processes by which people may acquire perceptions that do not reflect reality. As rational choice theory predicts, most citizens do not invest much time and energy to become politically informed. Instead, they rely on information shortcuts, or cues, to make sense of the political world. When lacking information about the position of a preferred party on some issue, one might use one s own position as a relevant cue to guess the position of this party. Although this has the same effect as assimilation, it does not derive from a psychological tendency to distort perceptions. Empirical research has demonstrated the existence of persuasion effects (Petty and Cacioppo, 1986; Price, 1989), as well as distortion of perceptions (e.g. Granberg, 1983; King, 1978; Sherif and Hovland, 1961) and students of public opinion and voting behavior have long been intrigued by the question to what extent policy preferences of voters, persuasion by political parties, or perceptual distortion underlie electoral decisions. There seems to be consensus among scholars that all three processes occur in the context of elections, but there is disagreement about the importance of each of these processes, which naturally leads to differential appreciation of elections as instruments of mass elite linkage. Many scholars have attempted to assess the strength of some of these processes relative to one another (e.g. Granberg and Holmberg, 1988; Markus and Converse, 1979; Merrill and Grofman, 1997; Page and Jones, 1979). This tradition of research has particularly focused on separating the effects of assimilation on the one hand from persuasion and rational selection on the other. Because of various methodological problems attached to the estimation of these processes 2. The tendencies to assimilate or contrast perceptions of political stimuli are also referred to in the literature as projection, rationalization, or false consensus (Brody and Page, 1972; Conover and Feldman, 1982, 1984; Fabrigar and Krosnick, 1995; Feldman and Conover, 1983; Markus and Converse, 1979; Visser, 1994).
VAN DER BRUG: PERCEPTIONS, OPINIONS AND PARTY PREFERENCES 57 there is little consensus about the extent to which the psychological tendencies to assimilate and contrast affect citizens perceptions of political objects. Discussions of various methodological problems attached to the estimation of perceptual distortion are provided elsewhere (e.g. Judd et al., 1983; Krosnick, 1990b). 3 Krosnick (1990b: 177) proposes an alternative way to model perceptual distortion, and finds American voters perceptions of the positions of Mondale and Reagan on various issues not to be distorted. He argues that what is often interpreted as perceptual distortion is in fact a persuasion effect by politicians. None of these studies, however, attempt to estimate simultaneously the relative strength of rational selection, persuasion and distortions in perceptions. In the present article a model consisting of a set of interrelated differential equations will be developed to do just that. To estimate this model I rely heavily on the pioneering work of Przeworski and Sprague (1986) and Brown (1991, 1993, 1995a,b) who introduced these types of models to the social sciences. These models have so far been used almost exclusively on aggregate data (one of the few exceptions is Chapter 3 of Brown, 1995b). Although this study is substantively motivated, it also makes a methodological contribution by demonstrating the great potential of these novel methods when used with individual survey data. Theoretical Expectations Balance theory and cognitive dissonance theory argue that people have a motivation to create balance (or consistency) among their actions, cognitions, and attitudes, because inconsistency is psychologically uncomfortable. So, if one has just spent one s life-savings on a boat which subsequently turns out to be terrifyingly unreliable, one would be strongly inclined to rationalize one s decision and to argue that it only needs some maintenance work, etc. Similarly, if someone is a heavy smoker and knows that smoking might cause cancer, one will tend to make light of the dangers of smoking. 3. The most fundamental problem is that the part of the variance in perceptions that is interpreted as evidence of assimilation is, for a large part, correlated measurement error, or (to be more precise) idiosyncrasies in response functions. These idiosyncrasies imply that not all actors with the same attitude towards an issue will place themselves in the same category of a response scale for example because some are more likely than others to use the extreme categories. Since perceptions of parties and attitudes of respondents are normally measured with the same rating scales, these idiosyncrasies in response functions will yield positive correlations between the perceptions of preferred parties and respondents own positions on issues. Yet, these positive correlations are not caused by psychological tendencies to reduce cognitive dissonance, but are the simple results of differences in the use of response scales (see also Granberg and Holmberg, 1988; Judd et al., 1983; Krosnick, 1990b; Van der Brug, 1998b: 342).
58 JOURNAL OF THEORETICAL POLITICS 13(1) There is no reason to doubt the fact that these psychological processes occur. Various authors have therefore argued that these processes will take place also within the context of an interview in which a respondent is asked to locate him/herself as well as a number of political parties or candidates on some issue scale (e.g. Granberg and Holmberg, 1988; Markus and Converse, 1979; Page and Jones, 1979; Visser, 1994). From a theoretical perspective there are, however, good reasons to doubt whether these processes occur in the context of a social survey. In the examples given earlier spending one s life-savings or issues concerning one s personal health the issues are very relevant for people s day-to-day life and in which their personal involvement can be expected to be very high. So these topics are likely to evoke strong emotional reactions from individuals. Under such circumstances inconsistencies between behavior and attitudes will be uncomfortable, so we can expect people to rationalize. However, very few people in modern democratic societies will consider an electoral choice remotely as important as either of these topics. In fact, public opinion research shows that voters are generally not very interested in politics (e.g. Kinder and Sears, 1985; Sniderman, 1993). Because of this, the psychological discomfort arising from a vote for a party with which one disagrees on some issues will be minor as well. I therefore hypothesize that, since politics is quite remote to most voters, the large majority of citizens do not assimilate or contrast their perceptions of party positions (H1). The second and third hypotheses are that although positive persuasion effects will be found (H2), there will be no negative persuasion effects (H3). First of all, this is because citizens are selective in media consumption, and hence they are more likely to hear the arguments of parties with which they have ideological affinity rather than that of other parties. They are therefore likely to be positively persuaded, but not negatively. A second reason why there is no expectation of negative persuasion effects is the rather awkward character of this concept. If attitudes were solely formed by a motivation to preserve balance between perceptions, attitudes, and party preferences, negative persuasion would occur as often as positive persuasion. But persuasion involves more than balancing preferences, attitudes, and cognitions. Parties try to give convincing arguments in favor of their positions. Experimental research has shown, among other results, that the stronger these arguments are, the more likely it is that these will affect people s attitudes (Petty and Cacioppo, 1986; Price, 1989). For citizens to be negatively persuaded by a party s arguments, these must be unrealistically unconvincing. The concept of negative persuasion is thus quite implausible. The final hypothesis is that rational selection will be more important than other processes leading to subjective agreement. This hypothesis derives from accumulating evidence indicating the rationality of voters (Alvarez,
VAN DER BRUG: PERCEPTIONS, OPINIONS AND PARTY PREFERENCES 59 1997; Franklin and Wlezien, 1997; Krosnick, 1990b; Van der Eijk and Franklin, 1996; Van der Eijk and Niemöller, 1983). The Study: Design, Context, and Data The three processes that produce subjective agreement between individuals and a party they favor involve an intriguing web of interrelated causal relationships. Party preferences may depend upon combinations of issue preferences and perceived party positions, issue preferences may depend upon party preferences and perceived party positions, and perceptions of party positions may be dependent upon party and issue preferences. How does one estimate (and thus disentangle) the impact of each of these different effects? Even advanced methods such as structural equation models, cannot estimate all the different effects, because attempts to estimate such non-recursive relations results in under-identified, hence inestimable, models (Asher, 1983; Bollen, 1989). Existing approaches therefore focus on each of these effects in turn (e.g. Markus and Converse, 1988) or on separating rational selection and persuasion, on the one hand, from distorted perceptions on the other (e.g. Granberg and Holmberg, 1988). Apart from the fact that these approaches do not provide the opportunity to simultaneously estimate the three different processes, there are some serious problems involved in the estimation procedure (see note 3). In this study the matter is therefore approached differently. The starting point of this approach is that causality involves change (cf. Bollen, 1989: 41). 4 To estimate the magnitude of the various processes by which attitudes, perceptions of party positions, and party preferences are interrelated, we should therefore focus on empirical situations in which perceptions and attitudes were observed to change substantially. The different effects can then be estimated by modeling how attitudes and perceptions change over time as functions of attitudes, perceptions and party preferences. 5 The type of longitudinal model that will be developed here has been explored theoretically by social scientists for some 4. This assumption is also well-grounded in the philosophy that underlies experimental research. An experimental group is exposed to some experimental condition, and a control group is not. To the extent that a change occurs in the experimental group and not in the control group, we conclude that the experimental condition caused the change. 5. The approach is fundamentally different from most statistical techniques to estimate causality in a non-experimental dynamic design. Structural equation models, for instance, estimate effect parameters on the basis of correlations between different variables. Most of the interesting changes over time are not captured by these correlations, however. In fact, the effect parameter of a variable onto itself (between two points in time), referred to as the stability factor, can be very strong, even though large aggregate changes have taken place between these two time points.
60 JOURNAL OF THEORETICAL POLITICS 13(1) decades (e.g. Coleman, 1964). Recent advances in computer technology make it possible to empirically estimate these types of models (Brown, 1991, 1993, 1995a,b). Before developing the model I will first discuss the research setting and the data used in this study. The Empirical Context On the 26th of April 1986 a series of explosions took place in one of the four reactors in the nuclear power plant near Chernobyl in the former Soviet Union. Radiation killed 31 people in or immediately after this accident. An unknown number of people died later or are still suffering from the consequences of their exposure to radioactivity. Over 100,000 people were evacuated. The radioactive cloud that was released during this accident spread over large parts of Europe, rendering crops, dairy-products, and cattle unfit for consumption. Because the government of the Soviet Union did not immediately notify the authorities in neighboring countries, it was some days before news about the accident reached the media. The Dutch public did not hear about it until the 30 April, when the Dutch media reported it for the first time. All this happened during the campaign for the general elections to the Dutch parliament of the 21 May 1986. Even before the accident, nuclear energy was an important political issue because the then ruling coalition of CDA (Christian Democrats) and VVD (Liberals) proposed constructing three new nuclear plants, a proposal that had generated great controversy and fierce opposition from, amongst others, the two major opposition parties the PvdA (Social Democrats) and the D66 (Social Liberals). The four major Dutch parties reacted very differently to the Chernobyl accident. The parties that favored expansion of nuclear energy tried to remove the issue from the political agenda as soon as possible. The PvdA and D66 re-emphasized and hardened their position against the construction of nuclear plants. As shown graphically in Figure 1, the Chernobyl accident generated a large and lasting change in voters attitudes to nuclear energy. Moreover, in the perception of the public, the four major Dutch parties all became more critical toward nuclear energy immediately after the Chernobyl accident. The setting is obviously very unusual for a discussion about how a public issue would, in other circumstances, develop. For several reasons the psychological incentive to assimilate or to contrast, will be particularly strong in this context. The Chernobyl accident catapulted the nuclear power issue to the top of the public agenda. As the accident itself generated strong personal anxieties for many people, the issue was not only highly salient for elites but also for most citizens at the time of the 1986 elections. The incentive to
VAN DER BRUG: PERCEPTIONS, OPINIONS AND PARTY PREFERENCES 61 Figure 1. Positions on Nuclear Energy: Voters Positions and their Perceptions of Party Positions reduce whatever cognitive dissonance existed at the time can thus be assumed to be strong. Also, as a result of the Chernobyl accident, positions of voters and parties changed abruptly. Consequently, equilibria that existed between people s attitudes, perceptions and party preferences, were suddenly disturbed. This creates more cognitive dissonance than in situations in which attitudes, perceptions, and preferences are stable. In view of the fact that not only (perceptions of) party positions on nuclear energy changed, but voters own attitudes on this issue changed as well, we may fruitfully investigate to what extent these changes were caused by persuasion, assimilation, or contrast. However, it must be noted that, because parties had little time before the election in which to influence public opinion, this context may show relatively strong tendencies to assimilate or contrast, whereas the degree of persuasion may be relatively small. Data The Chernobyl accident occurred less than a month before the parliamentary elections of 1986. It just so happened that the first of two waves of interviews, which constitute the Dutch Parliamentary Election Studies 1986, had almost been completed by the time the media first reported the accident. By then 1293 respondents had been interviewed. In this study the data of this group of respondents will be used. The second wave of interviews started right after the elections of 21 May. In both waves of interviews respondents were asked to indicate their own position regarding nuclear
62 JOURNAL OF THEORETICAL POLITICS 13(1) energy on a scale ranging from one to seven, of which the extremes were labeled the number of nuclear plants should be quickly increased in The Netherlands, and no more nuclear plants should be built at all in The Netherlands. In both waves, respondents were also asked to indicate the position of the four main Dutch parties (PvdA, CDA, VVD, and D66) on the same issue scale. The dataset also contains information about the preferences of respondents for the various parties. Therefore, these data provide an excellent opportunity to test a model of interrelated causal effects of attitudes, perceptions, and party preferences. It was decided not to treat the perceived positions of the four major parties as four separate variables, but to rearrange the data in the stacked form that is illustrated by Figure 2 (see also Oppenhuis, 1995; Stimson, 1985; Tillie, 1995; Van der Eijk and Franklin, 1996). The arrangement of the data in a stacked matrix has two major advantages. First, if the model is any good, its explanatory power should be independent of the specific party involved (but not of the respondent s preference for this party). So, instead of four variables preference for the PvdA, preference for the CDA, etc., the stacked matrix contains one variable party preference. The same is true for the perceived positions of each of the parties. By abstracting preferences in this way, party-specific idiosyncrasies are less likely to affect the model. Second, by deflating the number of variables and consequently the number of parameters to be estimated the model becomes more parsimonious, which increases the power of the estimates. The variables that relate to the PvdA, as presented in Figure 2, include the perceptions of the PvdA s position on nuclear energy, and the preference Figure 2. Transformation of Original Datamatrix into a Stacked Datamatrix
VAN DER BRUG: PERCEPTIONS, OPINIONS AND PARTY PREFERENCES 63 for the PvdA in both waves of the study. For the CDA, VVD, and D66 similar variables are available. The stacked form of the datamatrix inflates the number of cases and reduces the number of variables. As our analyses involve four parties, each respondent is represented by four cases in the stacked datamatrix. 6 The stacked datamatrix has a grand total of 3095 cases. The four separate parties can still be identified in the stacked matrix because the model also includes a variable objective party position, which only varies between parties, and not between respondents. This variable is measured by the sample median. 7 Party preference was operationalized by respondents responses to a question on the probability of ever voting for a specific party. This measure has been used extensively in European election studies in the last decade, and the validity of its use thoroughly tested (e.g. Van der Eijk and Franklin, 1996; Tillie, 1995). Recently, the superior properties of a probabilistic measure of party choice was demonstrated in the North-American context (Burden, 1997). The Model In this section a model is developed that describes mathematically how attitudes on nuclear energy and perceptions of party positions on this issue changed in the period from just before Chernobyl to the general elections a few weeks after the accident. By modeling change in attitudes and perceptions as a function of attitudes, party preferences, and perceptions, the model estimates all causal relations involved in the three processes of rational selection, persuasion, and perceptual distortion, as will now be explained. 8 We start out with a description of the various components of the model. I will first derive an algebraic term that describes an independent (from party preferences and party positions) Chernobyl effect on attitudes to nuclear energy. Then I will develop terms that describe positive and negative persuasion. Following this, changes in perceptions of party positions as a 6. The only real consequence of artificially inflating the number of observations is that it renders significance tests invalid. As no such tests are performed here (as will be discussed later), this is of no further concern. Missing data were deleted listwise from the stacked (not from the original) matrix, which thus contains full information for all 3095 cases included in the analyses. 7. Students of issue voting have mostly relied on the sample mean as an indicator of the true position of political parties or candidates on an issue (Granberg and Holmberg, 1988; Markus and Converse, 1979; Page and Brody, 1972). The considerations that prompted the use of the (interpolated) median instead of the mean are discussed in detail by Van der Brug (1997: 18 20), and the validity of the use of sample medians as objectified measures of party positions was amply demonstrated in that volume (see also Van der Brug, 1998a; Page, 1978). 8. Rational selection cannot be estimated by the effect on party preferences from attitudes to nuclear energy and perceptions of party positions on this issue, because the weight that
64 JOURNAL OF THEORETICAL POLITICS 13(1) function of parties objective positions will be modeled. Finally, I will model assimilation and contrast effects. These and other components of the model are now described in a series of consecutive steps. Components of the Model The following notation is used for the components of the model: i = index of cases in the stacked matrix X i = Respondent s attitudes to nuclear energy Y i = Respondent s perceptions of party positions on nuclear energy S i = Respondent s party preferences YO = Objective party positions on nuclear energy The model consists of two differential equations one for X i and one for Y i. Their respective derivatives dx i /dt and dy i /dt describe how the values of X i and Y i (for each of the 3095 observations) change between the two waves of interviews, as a function of values of X i, Y i, S i, and YO. The differential equations describe algebraically changes in attitudes and in perceptions. For each case the model predicts a trajectory that describes longitudinal change in attitudes and perceptions of the position of the respective party. Differential equations rather than difference equations were used to model time as continuous, and thereby to make the model less vulnerable to the effects of misspecifications. 9 voters attach to various issues, ideological positions, and other considerations is unknown. So, even if the nuclear energy issue would have had little effect on party preferences, one could not conclude that rational selection did not occur. To estimate the rational selection component the matter will be approached from the opposite perspective. The relevant literature tells us that a strong correspondence normally exists in The Netherlands (as well as in many other democracies) between policy positions of voters and their perception of those of the party they voted for (Alvarez and Nagler, 1995 ; Van der Eijk and Franklin, 1996; Van der Eijk and Niemöller, 1983; Enelow and Hinich, 1984; Franklin et al., 1992; Granberg and Holmberg, 1988; Krosnick, 1990a; Oppenhuis, 1995; Page and Brody, 1972; Wright and Berkman, 1986). Unless this is caused by assimilation, contrast, or persuasion, it must be the result of rational selection. This does not mean that rational selection is deduced from residual variation, i.e., variation that is not explained by assimilation, contrast, or persuasion. Rather, the rational selection component of the model is captured by two effects: 1) a general (not idiosyncratic) change in attitudes caused by a new appreciation of the dangers of nuclear energy, independent of party preferences and perceptions, and 2) changes in perceptions of parties positions as a function of objective changes in those positions. 9. Brown (1995b) presents simulations of a study with only two time points and multiple observations. He demonstrates that if the social process that is being examined is of a continuous time nature... the discrete-time model over two time points will likely suggest too low a fit and seriously incorrect parameter estimates (p. 23). In view of this, it was decided to model continuous instead of discrete change over time.
VAN DER BRUG: PERCEPTIONS, OPINIONS AND PARTY PREFERENCES 65 To take advantage of the mathematical properties of values that are bounded between zero and one (Sprague, 1976), all variables were recoded to this range from their original ranges of 1 7 or 1 10. For the variables X i, Y i, and YO, a zero (0) indicates a position in favor of nuclear energy, and a one (1) indicates a position against it. For party preference, S i, a zero (0) score indicates no preference for that party while a one (1) indicates the highest possible preference for that party. Independent Chernobyl Effects on Attitudes As a first step in building the model, a term should be developed that describes an (overall) change in respondents attitudes on nuclear energy, independent of the positions of parties. Because the accident with the nuclear reactor near Chernobyl occurred between the two waves of interviews, citizens are expected to have become more critical of nuclear energy, independent of party preference and of other factors that may determine their attitudes. In terms of the variable X i in the model (respondents positions on nuclear energy) it means that its values will increase. The parameter indicating this effect is called b 1. As a score of one is at the extreme of the scale, the model must not predict scores with values higher than one. Therefore, the derivative that describes change in X i must be bound by a term that limits its growth as it approaches one. This is achieved through multiplying by (1 X i ). An independent Chernobyl effect can thus be described as follows: dx i /dt = b 1 (1 X i ) Positive Persuasion Persuasion, as defined in the introduction, entails that attitudes of respondents change as a function of their party preferences and their perceptions of party positions. The relevant literature suggests that persuasion may be positive or negative (e.g. Markus and Converse, 1979). Whether a person will be positively or negatively influenced by a party depends on his/her evaluation of it. In my model I assume that positive and negative persuasion can occur simultaneously for each respondent, just as someone s judgement on a party can contain a combination of positive and negative evaluations. If positive feelings are stronger than negative ones, positive persuasion dominates negative persuasion. The strength of positive persuasion is expected to be a positive function of the difference someone perceives between the position of a party and himself. A large difference provides a stronger motivation to alter one s position than a small one. X i (the respondent s position) is thus expected to
66 JOURNAL OF THEORETICAL POLITICS 13(1) move in the direction of Y i. Combining these perspectives results in modeling positive persuasion as follows: dx i /dt = b 2 (S i ) (Y i X i ). In this term b 2 is the parameter of the model to be estimated. If the respondent has a strong preference for a party (S i has a large value), and the respondent perceives a large difference between his/her own attitude and the position of this party (i.e. a large value for X i Y i ), the motivation to adapt his/her own opinion will be strongest. Parameter b 2 is expected to take on a positive value. This will happen if values of X i change over time so as to coincide more closely with those of Y i. Negative Persuasion Negative persuasion, i.e. changing one s attitude away from a disliked party or candidate, if this occurs, can be modeled quite similarly. Because no explicit measure of negative affect is available, it is measured by the inverse of S i : (1 S i ). Negative persuasion will be stronger as the perceived difference between one s own position and that of a disliked party is smaller. This is represented by (1 Y i X i ). A problem that remains is the direction in which respondents move in the case of negative persuasion. When respondents perceive a disliked party to the right of their own position, they are expected to move leftwards, and vice versa. To ensure that X i moves in the assumed direction a signum function is included, which makes the term look as follows: signum(x i Y i ) (1 Y i X i ). The signum function is assigned the value 1 when the difference between X i and Y i is negative, and the value +1 otherwise. 10 To insure that the predicted values of X i are bounded by 0 and 1, the entire term is multiplied by X i (1 X i ), which gives the following term for negative persuasion: dx i /dt = b 3 (1 S i ) signum(x i Y i ) (1 Y i X i ) X i (1 X i ). In this term, b 3 is the parameter that describes negative persuasion effects. 10. When (X i Y i ) = 0, (X i Y i ) = +1. So, when the position of a respondent and that of a disliked party coincide, the negative persuasion term predicts that respondents, as a result move to a more critical position on nuclear energy.
VAN DER BRUG: PERCEPTIONS, OPINIONS AND PARTY PREFERENCES 67 Effect of Objective Changes in Party Positions on Perceptions Having designed the equation for longitudinal changes in X i, let us now turn to similar changes in the values of Y i, respondents perceptions of party positions. Parties altered their positions on nuclear energy right after the Chernobyl accident. Perceptions of party positions are expected to change as a consequence. Moreover, due to the accident, nuclear energy became an even more salient issue than it already was. Perceptions of party positions may not only change as a consequence of objective change, but also because increased salience, and increased media attention to party policies during an election campaign, makes voters more aware of the positions of different parties. Under these circumstances, the likelihood increases that any discrepancy between a voter s perception and the real position of this party is noted and that perceptions are adjusted accordingly. 11 This aspect of changes in perceptions can be modeled in prospective terms. Discrepancies between voters perceptions of party positions and the objective positions of these parties at time t, are expected to decrease over time, independent of voters party preferences. This can be modeled as follows: dy i /dt = b 4 (YO t Y i ), in which b 4 is the parameter to be estimated. In the case of a large discrepancy between YO t and Y i the term predicts a large change in Y i, i.e. it predicts a strong tendency to adjust the perception of the party position. The parameter b 4 is expected to obtain a positive value. If the perceived position is above the true position, the value of (YO t Y i ) will be negative, which means that the perception is adjusted downward. Conversely, if Y i is below YO t, the term (YO t Y i ) will be positive, so that Y i increases over time. As Y i moves toward YO t, the term captures an effect of increasing awareness as well as an effect of objective change in YO. Assimilation Next, two terms are defined to model assimilation and contrast effects. Assimilation and contrast effects are modeled similarly to positive and negative persuasion. As was the case with positive and negative persuasion, each respondent is expected to simultaneously have a motivation to assimilate as well as to contrast the position of each party in relation to his/her 11. The increase in voters awareness of where parties stand depends also on the way parties themselves react to events such as Chernobyl. Van der Brug (1997: 169) presents coefficients of perceptual agreement among voters about the positions of Dutch parties before and after the accident. These show that, on the whole, voters awareness of party positions on nuclear energy did indeed increase.
68 JOURNAL OF THEORETICAL POLITICS 13(1) own position. This is because someone s preference for a party is determined by a mixture of positive and negative evaluations. To the extent that someone feels more positively towards a party, the tendency to assimilate dominates the tendency to contrast its position, and vice versa. The strength of motivations to assimilate is expected to be positively dependent on the magnitude of the difference between one s own position and the perceived position of a preferred party. Mathematically, the tendency to assimilate can be described with a term that is quite similar to the one that models persuasion: dy i /dt = b 5 (S i ) (X i Y i ) in which b 5 is the parameter to be estimated, capturing assimilation effects. Contrast The tendency to contrast the position of a disliked party from one s own position, is expected to be inversely related to the distance between these positions. The tendency to contrast is captured by a term that is highly similar to the one that describes negative persuasion: dy i /dt = b 6 (1 S i ) signum(y i X i ) (1 X i Y i ) Y i (1 Y i ) in which b 6 is the parameter to be estimated. The Model I have just specified various terms that describe how the attitudes of respondents and their perceptions of the positions of the four main Dutch political parties on nuclear energy evolved between two time points, before and after the Chernobyl accident. Contrary to other approaches to causal modeling, the purpose is not to develop a model that is parsimonious in terms of the number of effect parameters. Rather, these terms aim to provide a comprehensive description of the various relationships between attitudes and perceptions predicted by the relevant theories (rational choice, balance theory, and social judgment theory). While explicating the terms an effort was made to provide the most direct mathematical translation of verbal theories of the interrelated causal structure of attitudes, preferences, and perceptions. The different terms describe separate and independent processes by which attitudes, party preferences and perceptions affect each other, that may therefore be combined in additive form. The two interrelated differential equations that thus constitute the full model, are:
VAN DER BRUG: PERCEPTIONS, OPINIONS AND PARTY PREFERENCES 69 and: dx i /dt = b 1 (1 X i ) + b 2 (S i ) (Y i X i ) + b 3 (1 S i ) (1) signum(x i Y i ) (1 Y i X i ) X i (1 X i ) dy i /dt = b 4 (Y0 t Y i ) + b 5 (S i ) (X i Y i ) + b 6 (1 S i ) (2) signum(y i X i ) (1 X i Y i ) Y i (1 Y i ) The model contains six parameters to be estimated, b 1 through b 6, each capturing an aspect of the causal structure linking attitudes, party preferences, and perceptions. 12 Estimating the Model No ready-made statistical tools are available to estimate a nonlinear system, such as the model developed here. To estimate the values of the parameters a numerically intensive procedure is applied that was proposed by Brown (1991). 13 The model describes change over time in two variables which are measured at two different moments. Two differential equations describe the trajectory of each observation unit between the two moments. Rather than solving the differential equations, the longitudinal trajectories that they describe are estimated by a numerical procedure (Brown, 1991: 56). So, all 3095 cases move simultaneously from their initial condition (the values on X i and Y i at t 1) to their second condition (the values on X i and Y i at t). Longitudinal change is described in a two-dimensional space that is 12. Initially an even more fully specified model was estimated, which was designed to capture the same processes, but describe aspects of these in more detail. It allowed for the possibility of the Chernobyl accident having a stronger effect on voters who initially (before the accident) took an intermediate position on nuclear energy. Moreover, it contained an exponential term for the variable party preference and its inverse in the terms capturing assimilation, contrast, and positive and negative persuasion. By including the exponent, this model allowed for the possibility that the strength of persuasion and perceptual distortion would be especially strong at the highest and lowest levels of party preference. As it turned out, the richer model (with nine parameters to be estimated instead of six) explained less than 1 percent more of the overtime variance in attitudes and perceptions than the more parsimonious one presented here. Also, the value of the exponent was estimated at 1.15, which shows that the tendency to be persuaded by a party or to distort one s perception of its position is almost linearly related to one s attraction to this party. Because the more parsimonious model captures all processes of substantive interest, and as it describes over-time changes in attitudes and perceptions almost as adequately as the richer model, it was decided to present the former one here. The results of the analyses estimating the latter model are presented in Van der Brug (1997). 13. To estimate the model a routine was written in SAS/IML, which is based largely on the one described in the Appendix in Brown (1991). The set-up that estimates the model presented here can be obtained from the author.
70 JOURNAL OF THEORETICAL POLITICS 13(1) defined by the variables X i and Y i. For each variable the model produces a vector of predicted values at the second time point. Similarly, for each a vector of residuals (the differences between the predicted and the observed values at the second time point) is computed. An iterative procedure is used to find the combination of values for the six parameters for which the sum of squared residuals is minimized. Therefore, the procedure has also been called a nonlinear least-squares method (Brown, 1995a: 41). Two measures are computed, which are described in detail elsewhere (Van der Brug, 1997). A first measure is called fit, which is closely related to the R 2 in ordinary least-squares (OLS) regressions. It is a measure for the proportion of over-time variance explained by the model. Delta-fit, the second measure, indicates for each parameter how much it contributes to the fit of the model. 14 Delta-fit can be used as a diagnostic criterion, which assists in comparing the impact of the various parameters on the model s fit. It cannot be interpreted as a test of statistical significance of the values of the parameters. Results Table 1 reports the estimated parameters of the model, the delta-fit of each, as well as the fit of the entire model. The model fits the data well. It explains 47 percent of the variance over time in variable X i (the change in the respondents opinions on nuclear energy between the two time-points), and 53 percent of the variance over time in variable Y i (the change in the perceived positions on nuclear energy of the four major parties between the two time-points. 15 14. As a first step to compute the delta-fit for each parameter, optimum values for all parameters of the model are estimated. Following this, the parameter in question is set to zero, while the other parameters retain their estimated values. All 3095 trajectories are then computed again, using this new set of parameter values. The delta-fit of this parameter measures the deterioration of the model s fit after the parameter s value is set to zero. Both measures fit and delta-fit are normally positive, but can in principle also turn negative. Setting all parameters to zero, implying that both derivatives are zero (i.e. that the model predicts no longitudinal change), yields a fit of zero. Fit becomes negative when the model s predictions deviate more from observed longitudinal changes than a no-change prediction. When the delta-fit is negative, the fit improves by setting the respective parameter to zero. If the parameter estimates are good (i.e., yielding a sum of squared residuals that indeed is minimal), the delta-fit is by necessity positive. 15. Given the amount of measurement error normally present in survey data, these proportions of explained variance indicate a good fit of the model. The model was also evaluated in different terms, namely the correlations of the residuals with the initial values (at t 1). In a properly specified model these correlations must approach zero. The correlation between the initial positions of respondents (X i at t 1) and the residuals in X i (at t) is 0.069. The similar correlation for variable Y i is 0.023. Obviously, this is no proof that the model is properly specified. Nevertheless, the negligible magnitudes of these correlations are an encouraging finding.
VAN DER BRUG: PERCEPTIONS, OPINIONS AND PARTY PREFERENCES 71 Table 1. Parameter Estimates and Other Statistics of the Model Parameter Estimated Value Delta-fit b 1 = independent Chernobyl effect 0.625 0.454 b 2 = positive persuasion 0.526 0.028 b 3 = negative persuasion 1.123 0.009 b 4 = objective change and increasing awareness of party positions 1.578 0.356 b 5 = assimilation 0.603 0.014 b 6 = contrast 0.816 0.003 Variables Fit dx i /dt = Change of respondents positions 0.470 dy i /dt = Change in perceptions of party positions 0.532 Because the model is specified in its derivative form, the relative importance of the parameters cannot be assessed from their estimated values. To compare the parameters, their delta-fit values should be inspected. When comparing the parameters by their delta-fit values, it is obvious that b 1 and b 4 have by far the largest impact on the model s fit. These two parameters describing changes in attitudes and perceptions, independent of prior party preferences and prior party positions resemble closely the Downsian model of rational choice. So, these findings support the fourth hypothesis. Let us discuss the different parameters one by one. Parameter b 1 denotes the changes in attitudes to nuclear energy that are independent of the positions of the political parties. This parameter has the strongest impact on the fit of the model. Sudden shocks such as the Chernobyl accident have an immediate effect on public opinion, independent of the behavior of parties. Parameter b 2 shows that some persuasion took place, albeit that, contrary to the expectation expressed by the second hypothesis, this effect contributes rather little to the model s fit. So, in the brief period under study here, political parties had relatively little success in influencing the attitudes of their adherents. It is not unlikely, however, that this effect would be stronger in contexts characterized by more gradual change over a longer period of time. The impact of negative persuasion by non-preferred parties, estimated by parameter b 3, is negligible in magnitude, and it has the opposite sign from what was expected. This finding corresponds to the third hypothesis, which holds that voters are not negatively persuaded by parties they do not prefer. Parameter b 4, describing the adjustment of perceptions of party positions as a function of objective changes and increased salience, is very important for the goodness-of-fit of the model. A major part of the change