Resumé

At first sight, Danish survey data seem to support Philip Converse's theory on (the absence of) coherent belief systems in mass publics. An alternative model is presented, however, accounting for the same data; the existence, within a larger population, of 'belief publics' holding differently structured belief systems, may distort measurement of belief coherence. The data support this model; different belief structures do exist.

Finally, there is no sign of 'elites' with more coherent belief systems within the sample, although such elites should exist according to Converse's theory. Therefore, the model of 'belief publics' is regarded as the more plausible explanation for the apparent absence of patterned beliefs. The article is primarily concerned with methodological aspects of belief systems analysis, and warns against taking national correlations of opinions at face value; especially the possible existence of 'split correlations' may seriously distort measurement of, and theorizing on, belief systems.

Belief systems have never surrendered easily to theory construction. Despite the availability of huge quantities of data on mass publics' opinions, controversies on their interpretation have persisted: the data seem to generate, rather than terminate, theoretical debates. Each of the 'schools' involved seems able to muster empirical evidence - though often quite different types of evidence.¹

This article contributes a methodological note to the debate, by showing that one single set of data apparently may support two quite different models of belief systems. The data stem from Danish survey research in the early 1970'5:

- interviews held in 1971 with a random sample of 1302 respondents; and
- interviews held in 1973 with 533 of these same respondents.

---

* Thanks are due to Prof. Ole Borre, Institute of Political Science, University of Aarhus, and to Dr. C. P. Middendorp, Steinmetz Archives, Amsterdam, for comments on earlier drafts; and to the Danish Ministry of Education, for a scholarship.

In Section 1, a first analysis suggests that the data fit well with Converse's well-known theory on the nature of belief systems. In Section 2, an alternative model is sketched, suggesting that the concept of belief publics holding differently structured belief systems may be a fruitful tool for analysis. A further review of the data in Section 3 seems to support this approach. This result necessitates a further confrontation of the two models (Section 4). In the final section, some methodological implications are hinted at.

1. The Data: A First Analysis

The 1971-interviews included 32 opinion items on a variety of political issues; regarding 32 Likert-type statements, respondents were asked to rate themselves on a five-point scale, ranging from 'agree completely' to 'disagree completely'. In the 1973-interviews, nine of these items were repeated in exactly the same form. These nine are a fairly representative selection from the original 32 items.²

In both years, a question on respondents' voting motivation was included.

These data allow some basic comparison with Converse's theory on mass publics' belief systems. Summarized briefly, this theory states that, except for some small ideological elites, the public at large lacks genuine, patterned political beliefs; at best, people may belong to narrower 'issue publics', centered around a single issue, virtually without connecting it to other beliefs (Converse 1964).

1.1. Checking Converse's model

Crucial variables in Converse's original presentation of the theory are: the proportion of'ideologues'; the (in) stability of beliefs; and the coherence or 'constraint' between beliefs. Let us inspect the Danish data at these points.

(a) Ideologues. As both interviews were held in connection with national elections, an (open-ended) question was included, reading: 'What was the main reason why you voted the way you did?' From the answers, the 'ideologues' can be identified: those who indicate any ideological perspective, either by using ideological labels ('Liberal', 'Conservative', etc.), or by referring generally to 'the ideas' of political parties.³ Despite these generous criteria, the ideologues comprise only 18 per cent of the voters, both in 1971 and in 1973.

(b) Stability. Stability is defined here as the correlation4 between item

scores in 1971 and 1973. The 1973-sample shows a low stability at the nine
items mentioned above: the average stability at these nine is .36. The
stabilities at the individual items range from .17 up to .53.

(c) Constraint. With belief constraint, I will denote the average
strength5 of correlations between the 32 opinion items mentioned above.
For the 1971-sample, this constraint shows the poor value of .103.

A more detailed picture can be given by classifying the items into
different 'issue areas'.6 Six such areas could be distinguished, which I
labeled:

(1) Tolerance (10 items), especially as to: who should be allowed the
right to vote; deviant political opinions (extremism); and rights of foreigners
in Denmark.

(2) Economic liberalism (6 items), concerning the role of the state in
economic life and in 'private economy' (e.g. level of taxes).

(3) Egalitarianism (6 items), centering around questions of equal opportunities
for everyone.

(4)National autonomy (4 items), concerning Denmark's membership of
Nato and the European Common Market.

(5) Alienation (3 items), centering on whether some (unspecified) 'big
organisations' have taken too much control over individuals and politics.

(6) Trust in politicians (3 items), including an item on the desirability of a
'strong man' to come to power.

Within these issue areas, the inter-item correlations are consistently positive ⁷ Between areas the picture is less consistent: between every pair of areas , there is a clear majority of either positive or negative correlations, but also a discordant minority. The average correlations within and between areas (direction of correlations taken into account) are given in Table 1.

The average strength of correlations (disregarding directions) is: within
areas, .177; between areas, .087.

This result might have been expected: within areas, correlations tend to
be stronger than the overall constraint of .103, though still moderate;
between areas, there seems to be hardly any coherence of beliefs.

1.2. Are constraint levels comparable?

The proportion of ideologues, stabilities, and constraint levels are strongly reminiscent of the data that inspired Converse to his theory. Especially with regard to constraint levels, however, this direct comparison may seem questionable at some points, which therefore require discussion.

(a) The choice of items might be guilty of the low constraint. Converse gives data on a political elite, in order to show that sophisticated observers, at the same items, indeed score a higher constraint than does the public at large. Lacking data on some Danish elite, we do not have such a standard for calling the Danish constraint 'low'. Maybe an elite would not do better at these same items.⁸ Some research abroad, however, suggests that. 103 really must be 'low constraint'. We may take the Dutch case for comparison - a European multiparty democracy not essentially different from Denmark. In 1971, a Dutch sample showed a constraint between 7 items of only .121 (computed from Pearson coefficients), not unlike the Danish level; a Dutch elite (the Members of Parliament) showed a constraint between the same items of .493 (Thomassen 1973 : 166). There is no reason for supposing the Danish situation to be radically different; we might reasonably expect Danish MP's to have a quite significant constraint on our 32 items.

(b) The wide scope of items, tapping a variety of dimensions, might account for the low constraint. However, Table 1 showed that even within more limited issue areas, correlations are far from impressive. Furthermore, USA data from 1972 show that a wide scope of items does not in itself preclude higher constraint levels. Miller et al. (1976) report Pearson correlations between 14 items on a variety of issues, also distinguishing several issue areas; they find an overall constraint of .24, the average correlation within areas being .36, between areas . 18.⁹ If a mass public can show this constraint on a wide scope of items, one might expect more sophisticated elites to reach even much higher levels.

(c) Our Pearson correlations cannot be compared directly to data as
Converse's, based on tau-gamma coefficients. Though this indeed precludes
detailed comparison of figures, the Dutch and USA research mentioned

above indicate that the use of Pearson's coefficient does not preclude
high constraint levels.

Though these remarks cannot give conclusive proof, they suggest that neither the choice of items, nor their scope, nor the correlation coefficient used, offer sufficient explanation for the low constraint level. Furthermore, they would not explain the poor opinion stability, nor the small proportion of ideologues. Consequently, we may conclude that the Danish data provide another case in favour of the Conversian model of belief systems. Indeed, the model would logically account perfectly for the data presented so far.

There is, however, another model logically fitting the same data. This
will be explored in the next sections.

2. An Alternative Model: Belief Publics and Attitudes

2.1. Belief structures and belief publics

With the term belief system I will denote the opinions of a specified
population on a specified set of issues. In this article, the issues involved
are the 32 items referred to above.

As any system, a belief system must have a structure; this belief structure I define as the network of relations between the individual issues of the belief system. Both the level and the direction of these relations, as expressed in correlation coefficients, are relevant.

In this network of relations, like in a spider's web, some issues will take a more central position, while others are more peripheral. A first characteristic of a belief structure, then, is the relative centrality of each issue in it, and this centrality can be defined as the intensity of relations of each item with all other items taken together. Operationally, I will define the centrality of an item as its average absolute correlation with all other items. (This implies that the average of centralities of all items is equal to the constraint of the belief system.) Note that this definition in no way refers to the stability of opinions, as other descriptions often do. Centrality is a static concept-of-structure, measured at one point in time; stability is a dynamic concept-of-process. This careful distinction enables us to empirically assess the relation between the two, instead of merging them into each other already by definition.

Belief structures may differ, not only with respect to the centrality of
items, but also with respect to the direction of relations between the same
items.

A specific belief structure may or may not be shared by larger groups of

people. Groups or categories sharing a belief structure, I will call belief publics. Within one population, several belief publics, with their differently structured belief systems, may coexist.¹⁰ Where this is the case, aggregating the constraint over that population will result in (a) a decrease in the differential centrality of items, resulting in a rather flat distribution of moderate centralities; (b) a decrease in the constraint level, insofar as belief structures with reverse correlations between issues are involved. Thus, aggregating different belief publics may give a distorted picture of a population's coherence of beliefs.

Of course, belief publics are a matter of degree: one would hardly expect to find some perfectly demarcated publics with perfectly coherent beliefs. Furthermore, such belief publics are not necessarily manifest, and people need not be aware of their 'membership'. Thus, belief publics are largely a construct, a model - but not merely artificial, as we shall see.

2.2. Beliefs and attitudes

One more possible source of distortion in constraint measurement deserves attention: the difference between beliefs and attitudes. Rather than supposing that those who are 'in the information flow' simply learn by heart which single opinions go with which, I would assume that there is a deeper level, a latent structure, guiding the patterning of single opinions on single issues. This latent structure I will call thcattitude structure. General attitudes may guide specific opinions, but do not necessarily prescribe them with coercive force. Thus, even a firm attitude structure may leave some space to variations in single-belief patterning; consequently, the belief structure possibly conceals a more coherent attitude structure.

Supposing that such attitudes can be measured, we can treat the resulting attitude scores in a way exactly parallel to single-item scores. That is, we can define the concepts 'attitude system, - constraint, - centrality, and - stability' parallel to the corresponding concepts of the belief system. Furthermore, the argument from the previous subsection also applies: measurement of overall attitude constraint may be distorted by the existence of different 'attitude publics' within a larger population.

Summarizing, low constraint need not indicate weak opinion formation, but may also be produced by the existence of more coherent but differently structured belief systems. Where structural uniformity in a population is low, overall constraint only reflects the net balance in the aggregate of structures, not the actual coherence of beliefs. Furthermore, belief structures may conceal more integrated attitude structures.

3. Some Empirical Implications

It will be clear that the model outlined above logically might account for
low constraint. In this section, some empirical implications will be
checked.

3.1. Attitudes versus beliefs

The above statements on the relationship between attitudes and beliefs imply that, if some latent attitude structure indeed exists, it should be more tightly patterned than the single opinions are; thus, the attitude constraint should be higher than the belief constraint.

This can be tested. In Section 1, six issue areas were distinguished and the inter-area correlations inspected. Instead, we may simply aggregate the item scores in each area into attitude indices, a rough kind of scales.¹¹ The correlations between these six attitudes are given in Table 2, which can be compared directly to the inter-area correlations (see Table 1).

This matrix differs drastically from Table 1; the attitude correlations are more than twice as strong as the corresponding inter-area correlations. So, even if the overall belief constraint would imply the absence of patterned opinions, we must conclude that the Danes do have patterned attitudes to a much higher degree. It is as yet an open question which of both is the more relevant for political behaviour.

In the following, I will report on attitudes and their constraint in several
sub-samples; it will be seen that attitude constraint always is considerably
higher than belief constraint.

3.2. The role of belief structures

In identifying possible belief publics with different belief structures, no
more than a first exploration can be given here. As yet, there are no

guidelines for where to find such publics; even respondents themselves may not be aware of which belief structures they share with whom. Also the size of such publics is uncertain; perhaps uniform patterns of beliefs only develop in rather small groups, having some form of regular interaction.

Instead of identifying some clear-cut belief publics, some more general implications of the model will be checked. The logic of the procedure is a simple deduction: if different belief structures are present in our sample, then any category of respondents sharing at least some structural feature in their belief systems should show a higher constraint than average; a parallel argument goes for attitude structures.

In the following, then, I will inspect some constraint levels for (a) categories of respondents sharing centrality of some items in one issue area, (b) categories sharing centrality of two issue areas, and (c) some groups which, in addition to shared centrality, share a specific direction of relations between two attitudes. As these three categories, in this order, share a progressive number of structural features, we should also expect progressive constraint levels.

(a) Shared centrality, one issue area. As centrality, being derived from correlations, is a group characteristic, it cannot be observed directly from individuals' item responses. Thus we cannot immediately select categories of respondents with shared centralities. Instead, I will assume that if a certain issue area has a high centrality for some respondents, these will be more disposed to give extreme and consistent responses to the items involved. Consequently they will end up at the (low or high) extremes of the attitude scale constructed from that issue area. The respondents at all scale extremes, then, should constitute categories sharing high centrality of the items involved.

This assumption, however crude, yields the intended results. As extremes' I arbitrarily defined the (around) 25% most extreme respondents on each scale. Each of these six groups indeed shows an unusually high centrality of at least some of the items involved; the result is indeed a higher constraint in the belief and attitude sytems generally. The belief and attitude constraints for each category are given in Table 3.

A description of one of these categories, namely the 'trust'-extremes, may serve to illustrate the procedure. The 'trust'-extremes together comprise 26.9% of the whole sample -15.5% on the low side of the scale, 11.4% on the high side. As intended, in this category the three 'trust'-items have an outstanding centrality: they are the three most central items in its belief

system, whereas they only rank 3rd, 9th, and 23rd in the sample as a
whole. As expected, the category shows quite higher constraint levels.

Of course there is a certain overlap between the six categories (i.e.,
some respondents belong to more than one); but together the six cover a
major portion of the sample: 78%.

(b) Shared centrality, two issue areas. The overlaps of the six categories have a special significance: they constitute categories which probably share centrality of items in two issue areas. This means a higher uniformity of belief structures, and should result in higher constraint levels. The six scales give 15 such overlaps. Indeed, their belief and attitude constraints are considerably higher than those of the previous six categories. The attitude constraints for all 15 are given in Table 4.

One may wonder if these overlaps do not contain 15 times almost the same respondents. Perhaps there is an elite-like stratum of consistent respondents, which we here analyze in 15 different ways? The data suggest that this is not the case: the 15 categories together cover no less than 47 % of the whole sample.

(c) Specific direction of relations between attitudes. All 15 categories, of course, consist of two types of respondents: those who show a positive relation between the two attitudes involved (that is, score either high or low on both scales); and those who show a negative relation, combining a high score on one scale with a low score on the other. By separating these two types, we identify another structural characteristic in the attitude system: a specific direction of a relation between two attitudes.

Space limitations preclude analysis of all 30 subgroups which thus can be identified. Instead, I will concentrate on one of the fifteen categories as a typical example: the overlap between the 'trust' and 'liberalism' scale extremes. This category contains 128 respondents. Of these, 85 (which I will call the 'majority subgroup') show a negative relation between 'liberalism' and 'trust'; the other 43 (the 'minority subgroup') show a positive relation. Let us now compare the pattern of attitude correlations of the whole category to that of both subgroups. The three matrices are given in Table 5.

The significant feature of Table 5 is not the correlation between 'trust' and 'liberalism' themselves, which forks into two strong but reverse correlations; this is simply the intended effect of the splitting-procedure. Most interesting, however, are the three correlations printed in italics in all three matrices. Here we see three insignificant correlations in the common matrix fork into two quite significant but opposite correlations in the subgroup matrices. The result is a higher attitude constraint in each subgroup than in the category as a whole. It might also be noted that the whole pattern of correlations in the subgroups looks rather different (notice, for instance, the different 'behavior' of the tolerance-scale in them). Neither of the patterns looks much less rational than the other. Indeed, the two subgroups come close to what I would understand by two different belief publics.

In short, Table 5 shows several instances of what Lipset et al. (1954 : 1160) call 'split correlations'. They all result in concealing relevant information behind seemingly low correlation. It is especially such 'split correlations' that can decrease and distort constraint levels of populations, if different belief- and attitude structures are hidden in it.

All 15 categories can be split up in the same manner; they all reveal such
different attitude structures.

4. The Models Confronted

The previous section suggests that a model of belief publics holding different belief systems might well account for the low constraint level found. Section 1, however, suggested strongly that the quite different Conversian model was supported by the data. Obviously, a further comparison of the models is needed.

4.1. Centrality and stability

In presenting his theory, Converse recognizes explicitly that difference of belief structures might account for low overall constraint; he calls them 'individually idiosyncratic' configurations of beliefs (though it is unclear to me why 'idiosyncrasies' always should be individual). But he rejects this possibility through his analysis of the (in)stability of beliefs. The peculiar

pattern of opinion change, he argues, is only compatible with a model of
largely random change; thus, a large part of the sample obviously gave
random answers to the items.

Unfortunately, this argument falls prey to a common logical fallacy: ascribing characteristics of a collectivity to its individual parts. To conclude individual random response from a collectively random pattern is, logically spoken, an analogue to saying that if the average height of the Danes is 1.73 meter, then every Dane must be 1.73 meter high.

On the contrary, the existence of different belief structures does provide an explanation for a collectively random-like model. The crucial fact is that stability of opinion on an issue is closely connected with the centrality of that issue in the belief structure. For Converse, this is true already by definition (Converse 1964:208). From my definition of centrality, this can be assessed empirically. For the nine issues included in both the 1971 and the 1973 interviews, I computed the correlation between centrality (in the 1971 sample) and stability. This was done both for the whole sample and for several subgroups. In all groups centrality and stability of the items were highly correlated - typically at a level around r = .80, for some subgroups even stronger than r = .90.¹²

Now, widely different belief structures in a population imply a wide dispersion of centralities of issues among that population. This gives a wide dispersion of the probability of opinion change through the population; and insofar as different belief structures with reverse correlations between issues are concerned, people may rationally change opinions in different directions under the same circumstances.

In short, idiosyncratic' belief systems will lead to 'idiosyncratic' opinion change, which may well result in random-like patterns for populations at large. 'Low' overall stability may thus be explained by the same models explaining 'low' constraint.

4.2. Belief publics versus elites

At one point, though, the models have different implications. Converse's model would imply that the main source of constraint should be some small elites within the population; outside the elites, correlations between beliefs would be virtually zero. A model of belief publics, however, would imply that the source of constraint cannot be located in one 'ideological stratum'.

Along three lines, possible elites in the Danish sample were identified.
First, I selected the party members; one would expect a rather well-developed
belief system among those interested and informed enough to be

paying members of a special party. Second, the ideologues were selected; as Converse suggests, the use of ideological labels '... betokens a contextual grasp of politics that permits a wide range of more specific ideaelements to be organized into more tightly constrained wholes' (Converse 1964). Finally, I used a special 'detector' for possible elites. Elsewhere, constraint levels have been found related to education, political involvement, general interest in politics, and media use (Converse 1964, Borre 1973, Nie and Andersen 1974, Thomassen 1976). Accordingly, a sophistication index"¹ was constructed from 18 relevant variables.¹³ The distribution of the sample on this index turned out to be skewed to the low side, with a thin tail to the high side. From that tail, I selected both a larger elite (the 17.2% highest respondents) and a smaller subgroup of it (the 8.6% highest).

Even if these possible elites are not identical to the real elites, one would expect at least a large part of the real elites somewhere within them, producing a substantial constraint level. The constraint levels of the four possible elites are given in Table 6.

Generally, these constraint levels are no higher than those from Table 3, which covered 78 % of the sample, and considerably lower than those from Table 4, covering 47%. And yet we would expect the categories from Table 6 to have a somewhat better developed belief system than the population at large. How to explain their moderate constraint?

After what we saw in the previous sub-section, we might answer that also these possible elites, far from being homogeneous groups, are aggregated from different belief publics. That is, it may not be meaningful to speak of elites as such.

4.3. Some implications

To a large measure, the two models can explain the same data; only

regarding the low elite constraint does a model of belief publics seem the
more reasonable explanation.

It should be stressed that the models are logically incompatible. Some of
their different implications will be discussed here.

(a) The concept of belief publics may seem reminiscent of Converse's final explanation: the segregation of mass publics into 'issue publics'. However, the two concepts are each other's opposite. The assumption behind 'issue publics' is that people take a separate stand on some issues, and no stand at all on most issues: 'Certain rather concrete issues may capture their respective individual attentions.... This engagement of attention remains narrow however: Other issue concerns that any sophisticated observer would see as 'ideologically' related to the initial concern, tend not to be thus associated ...' (Converse 1964:246). But belief publics show the opposite: as we saw above, groups sharing a centrality of some issues or attitudes, also share a._stronger association of other opinions and attitudes.

(b) The models imply different explanations for changes in constraint levels over time. For example, Nie and Andersen (1974) report a sudden upswing in constraint levels in the USA as from 1960. Their explanation is a modification of the Conversian model, namely that the more salient political context since 1960 made people develop real patterned opinions. This implies, of course, that such opinions indeed were virtually absent in the 1950'5. A model of belief publics, however, would ascribe a raise in overall constraint levels not mainly to formation of beliefs, but rather to a unification of belief structures.

(c) Thus, the model of belief publics ascribes importance to (changes in) belief structures, whereas the other model does not even mention the structure of belief systems. However, changes in belief structures have often been noted as substantial, apart from changes in belief content. Nie and Andersen themselves remark, unfortunately without elaborating on the issue, that'... our data suggest that not only specific political attitudes but the structure of mass attitudes may be affected by politics .. .' (italics in original). And on salience of issues - a structural characteristic similar to centrality - Lipset et al. note: 'Thus, the most immediate effect of political propaganda is upon the voters' feeling of saliency of issues. Short-term propaganda probably does not affect people's attitudes greatly, but it can well affect the way they see a concrete electoral situation, and, by changing its 'hook-up' with an existing attitude structure, might well affect their final decision.' (Lipset et al. 1954:1159; italics by the authors.)

5. Discussion

The core of the above argument is that national correlations of opinions should not be taken at face value. In belief system analysis, the national electorate is not necessarily the most relevant unit of analysis; it may conceal important facts, rather than reveal them. Different belief publics, hidden under its surface, may be more relevant units. One should especially be aware of possible 'split correlations' in national samples.

In much social research, there seems to be a tendency to distrust any strong correlations found - one tries to refute them by introducing controlling variables or computing partial correlations - whereas weak correlations are often taken for granted. To me it would seem more reasonable to distrust any correlations, whether strong or weak. If'split correlations' hide in the data, then calling a national constraint level 'low' would be an analogue to calling a country 'flat' just because its mountains compensate for its valleys.¹⁴

The Conversian model, by taking national correlations at face value, implies that a national population either must have uniformly patterned beliefs, or else has no patterned beliefs at all. The model of belief publics does not presuppose such uniformity. If I prefer the latter model for the Danish case, this is not because of some 'rationalistic' bias: quite the contrary, I think that on purely empirical grounds it is the more plausible model for the Danish situation of 1971.

The two models are based upon some different background assumptions, however, which are worth noticing. If Converse's model presupposes a national uniformity of belief structures, this implies that the source of belief structuring must be a national one. Indeed, the model seems to ascribe special importance to the mass media, which obviously are supposed to imprint their (or: the elite's) belief structure upon all individual receivers in the same way. This background model of belief structure formation is strongly reminiscent of a mass society model. The belief publics approach, on the other hand, conceives of society as existing of many different groups and communities that may develop specific belief structures, and in which the influence of mass media may be received in different ways. To me, it is hard to see why such different groups as blacks and whites, farmers and businessmen, the poor and the rich, should be expected to hold one and the same belief structure. The background model here may thus be called a society of publics.

In this article, only a first exploratory treatment of the model of belief
publics could be given. Many questions remain. First, better criteria for

defining relevant belief publics are needed; the method used in this article is neither the only nor the best possible. Second, their development over time is as yet an unknown matter. Third, their sociological backgrounds require further exploration. Fourth, their importance for actual voting behaviour is a crucial but unanswered question. Perhaps analyses along these lines could further our understanding of the formation and development of belief structures.

As yet, the questions outnumber the answers. But sometimes, asking
new questions may be as essential as answering old ones.

---

NOTES

1 Some extreme cases are Converse (1964), concluding people's ideological poverty from survey data, versus Lane (1962), concluding the 'hidden' existence of ideology from few but very deep interviews.

2 The nine items cover all six issue areas described in sub-section 1.1.

3 Thus, the category called 'ideologues' here is quite comparable to Converse's 'ideologues' and 'near-ideologues' combined, which cover \5Vz% of the voters in Converse's sample.

4 Throughout this article, Pearson's (product-moment) correlation coefficient is used.

5 1.e., the numerical value of correlations, disregarding their directions (positive or negative).

6 The distinction of these issue areas was partly inspired by Worre's attitude scales, which were constructed from the same set of items as is used in this article (Worre 1978).

7 Within almost all areas, some items were worded positively (i.e., tolerant, liberal, etc.), some negatively (intolerant, anti-liberal, etc.). Where appropriate, the item scales were reversed before computing correlations, so that high scores always point to the 'positive' side, and positive within-area correlations always indicate 'consistent' opinions.

8 It should be noticed that also Converse lacks elite data really comparable to his sample data, as a result of significantly different item formats for elite and sample. The sample was given one Likert-type statement at each issue and asked to react (agree/disagree), the elite was given a clear choice between two or three different opinions (Converse 1964:257). The latter format is suspected to produce substantially higher correlations; a very similar difference in item formats has often been held responsible for the sudden jump in USA constraint levels since 1964 (reported by Nie and Andersen 1974).

9 Difference in item formats precludes an exact comparison of these figures with the Danish. But the USA format is no guarantee for strong correlations: the low Dutch constraint reported above is basedon virtually the same format.

10 Note that the concepts of 'belief structure' and 'belief public' do not refer to any content of beliefs; a Communist and a Conservative might well belong to the same belief public, sharing a similar belief structure - in which, for example, the issue of 'private enterprise' may be highly central for both.

11 For this purpose, respondents' item scores within each area were added; these sum scores for each area were converted into nine-point scale scores. As noted in sub-section 1.1., all item correlations within areas were positive, which justifies this adding procedure. If a respondent's answer on an item was missing, his or her score on the relating attitude scale was also treated as 'missing', and ignored in the computation of correlations between attitudes.

12 Converse's data also allow inspection of this relation, as he reports correlations and stabilities of 7 policy issues in his Table VII and Figure 3. Unfortunately, the stabilities are given in imprecise form, but with my 'best guess' from his Figure 3,1 find a correlation between centrality and stability of r = .86.

---

12 Converse's data also allow inspection of this relation, as he reports correlations and stabilities of 7 policy issues in his Table VII and Figure 3. Unfortunately, the stabilities are given in imprecise form, but with my 'best guess' from his Figure 3,1 find a correlation between centrality and stability of r = .86.

13 The characteristics summarized in this index are: length of formal education; interest in politics; discussing politics in the family, with friends, and with colleagues; level of participation in such discussions; persuading others to vote for a specific party; party membership; reading on politics in newspapers; following political programmes on radio and TV; regularly following radio and TV news programmes.

14 Of course, insofar as national correlation matrices give a distorted picture, this also affects the meaning of techniques based upon them - for example, factor analysis.

REFERENCES

Borre, O. 1973. Party and Ideology in Denmark. Aarhus: Institute of Political Science.

Converse, P. E. 1964. 'The Nature of Belief Systems in Mass Publics', pp. 206-261 in Apter,
David E. (ed.), Ideology and Discontent. New York: Free Press.

Lane, R. 1962. Political Ideology. New York: Free Press.

Lipset, S. M., Lazarsfeld, P. F., Barton, A. H., and Linz, J. 1954. 'The Psychology of Voting: An Analysis of Political Behavior', Chapter 30 in Lindzey, Gardner (ed.), Handbook of Social Psychology, Volume 11. Reading, Mass.: Addison-Wesley Publishing Company.

Miller, A. H., Miller, W. E., Raine, A. S., and Brown, T. A. 1976. 'A Majority Party in Disarray: Policy Polarization in the 1972 Election', pp. 176-195 in Niemi, R. G., and Weisberg, H. F. (eds.), Controversies in American Voting Behavior. San Francisco: W. H. Freeman and Company.

Nie, N. H., with Andersen, K. 1974. 'Mass Belief Systems Revisited: Political Change and
Attitude Structure', pp. 540-587 in Journal of Politics 36.

Thomassen, J. J. A. 1976. Kiezers en gekozenen in een representatieve demokratie. Alphen
aan den Rijn: Samsom.

Worre, T. 1978. 'Partisystemets ideologiske dimensioner', pp. 176-210 in Borre, O. et al.,
Vælgere i 70'erne. København: Akademisk Forlag.