Beyond the micro/macro distinction

Introduction: The problem of discrepant data in electoral studies

The study of elections is bedevilled by the existence of two discrepant sets of empirical data about the same phenomena: microdata and macrodata (Scarbrough, 1987). It is the contention of this paper that microdata and macrodata or 'individual' and 'ecological' dataare in fact about different objects which, however, are involved in the same phenomena, elections. These objects, and hence data about them, are certainly related, but only through complex social and cultural systems which are the real context of electoral phenomena. This we shall call the structural context of elections.

Within the structural context, micro-and macro-data may be consistently related. But if the structural context is disregarded, no consistent relation will be discernable. In any case, no simple or direct translation of either type of data into the other is possible.

This view runs counter to an assumption about the relation between datalevels that underlies electoral studies and behavioural political science generally. The assumption is that macrodata must ultimately be reducible to some corresponding microdata. It is an assumption with a long and impressive pedigree, claiming among its progenitors both Popper and Hayek who pro-pounded it under the name of 'methodological individualism' (Popper, 1957: 136ff). To see what such a 'reductionist' project would actually entail, however, let us consider what was said on this by Heinz Eulau, one of the methodological founders of the modern discipline.

Eulau has written:

'. . . Taking it for granted that political science is interested not only in the behaviour of individuals but, above all, in the actions and policies of groups, institutions, and 'states', the problem arises as to how meaningful statements about large systems can be made on the basis of inquiry into the behaviour of individual political actors. This is the problem of the relationship between macro-and microanalysis. For, clearly, only if the relationship between macro-and microanalysis is satisfactorily settled can it be legitimate to say that . . . problems of interest to political science are potentially susceptible to behavioural treatment. In trying to solve this methodological problem, it is easy to fall into various errors. For instance, there is the fallacy of extrapolation from micro-to macrophenomena. Small systems are treated as analogues of large systems, and the findings on the microlevel are extended to the macrolevel. There is, secondly, the fallacy of personification: large-scale phenomena are 'reduced' to the individual level, as in the more grotesque descriptions of 'national character'. link. . . . I am not thinking of theoretical linkage alone but of empirical linkage also' (Eulau, 1967: 47).

As described by Eulau, the 'task of research' is incapable of attainment. If in order to accomplish our research tasks we really had first to 'build the . . . total chain of interrelationships' involved (and the 'horizontal patterns of relationship' besides), we should never have the time or resources to accomplish anything. Nor is the problem merely a practical one. Even if we could assemble all the relevant data on all the relevant levels, the question would still remain of how in principle to relate them.

A telling point made by Eulau is that micro and macro are in the eye of the beholder; what is micro and what macro is 'relative to the observer's standpoint'. Thus a family may be a microunit from the standpoint of the wider society, but a macrounit from that of an individual family member. Between any two adjacent levels, however, the relation is given, and it is not commutative. The macrounit is the 'higher-level' one; it contains and is composed of 'lower-level' microunits. Of course, that is partly a matter of definition: what is perceived as higher-level or containing is defined as macro, what lower-level or composing as micro. However, once a given level has been defined, an adjacent level (higher or lower) will necessarily be defined as the opposite. But as Eulau points out, any actual level (with one important exception) may be defined as either. The important exception is the individual level, which can only be defined as micro. And herein lies the source of both the 'ecological' and 'individualist' fallacies.

Individual-level variables

Inasmuch as 'level' is a matter of definition, data on any given level may be represented by either independent or dependent variables. This necessarily assigns the converse status to data on the adjacent level. It follows that data on any given level may be represented in any of four ways: as independent micro-variables; as dependent micro-variables; as independent macro-variables; or as dependent macro-variables. However, as in electoral studies one level is predefined as micro (the individual level), there are only two ways in which data on that level can be represented: as independent micro-variables; or as dependent micro-variables. These definitions are formal, not real; they define abstract, not concrete, objects. In empirical studies, however, it is common for the formal and abstract to be taken to represent the real and concrete. Thus, objects on the 'individual' level in electoral studies are assumed to be real individuals. (As we shall see later, this assumption is false.) Furthermore, formal (logical, mathematical) relations between variables are taken to represent material (physical, social) relations between things or events. In the micro/macro schema, the formal relation between independent and dependent variables is commonly taken to represent the material relation of cause-and-effect. Therefore, given that in electoral studies the individual level is necessarily the micro level, if individual-level data are expressed as dependent variables, this seems to carry the implication that they are effects of macro-level events.

In regard to individual voting behaviour, that implication is intuitively correct. But it runs counter to the equally intuitive presuppositions of methodological individualism, according to which 'superindividual' phenomena are (in Popper's words) 'macro-effects of micro-events'. Thus whatever position we adopt seems to involve us in a fallacy. From the standpoint of methodological individualism, drawing inferences about individuals and their behaviour from macro-level data involves us in an 'ecological' fallacy (Robinson, 1950). But eschewing that fallacy seems to commit us to the view that the voting behaviour of individuals is a cause and not an effect of macro-level phenomena. Leaving aside the purely formal effect of voting (ie, election results), this view is counter-intuitive. And if intuition is right in this case, then drawing inferences about macro-phenomena from micro-data also involves us in a fallacythe 'individualist' fallacy (Barton, 1968). Our intuitions seem to land us on the horns of a dilemma. This particular dilemma is by no means confined to electoral studies, or to political science, or even to the social sciences generally. It occurs in the natural sciences, also, and it is a recurrent problem in philosophy.

To pursue the larger issue would take us well beyond the scope of this paper. Less ambitiously, let us examine it in the terms in which it specifically confronts us: the problematic relation between the two types of data employed in electoral studies.

Eulau again:

'. . . Not unrelated to the macro-micro problem, is the problem of using both discrete and aggregate data in behavioural analysis. The difficulty arises out of the fact that what may be true of aggregates need not be true of the individuals who compose them. The reason for this is simple enough: moving from statements about the behaviour of aggregates, such as electoral districts, to the behaviour of any one individual within the aggregate involves an inference which may be wrong. The dubious procedure involved was pointed out sixteen years ago in a widely read article by the sociologist W.S. It seems that the extent to which certain problems of politics are susceptible to behavioural treatment depends on an answer to these questions' (1967: 4748).

In common with many later commentators, Eulau was of the view that the advent of the sample survey had overcome Robinson's caveat. Yet twentythree years further on the issue of the 'ecological fallacy' is stiIl with us, now counterposed with the 'individualist fallacy' supposedly inherent in the survey method itself. Eulau's questions remain unanswered.

It is a convention of discourse that when questions remain long unanswered, or seem unanswerable, the reason may lie in the terms in which the questions are posed. Let us therefore examine the terms in which these questions are posed: 'individual' and 'ecological', 'unit' and 'aggregate', 'micro' and 'macro'. To begin with, how individual is 'individual'?

Defining the individual

In much of the discussion of this issue the impression is given that the ultimate purpose of electoral studies is to account for the political behaviour of individuals. Indeed, that is not infrequently stated in so many words. However, while there may be some practitioners for whom it is true, it is not generally true. For the most part, accounting for the political behaviour of individuals is not the ultimate object of the exercise, if by 'individuals' we mean particular, nameable, 'real-live' persons. Of course, there are some 'real-live' persons who are of great interest to political scientists and whose political behaviour we should dearly like to explain -Mrs Thatcher or President Gorbachev, for example.

Political science has not so far been very successful at doing this, but it would still like to do it. It is different when it comes to the political behaviour of 'ordinary' people. Here, political science has an ulterior motive. Rather than doing it 'for its own sake', it does it in order thereby to explain something further, something superindividualpolitical change, for example. Explaining individual political behaviour is a means to explaining what that behaviour is in turn assumed to explain, namely, again in Eulau's words, 'the actions and policies of groups, institutions, and "states"'or, we might add, the outcome of elections and the fortunes of parties.

That political science is not ultimately concerned with the behaviour of particular individuals is evident from its preferred instrument of research at the individual level, the sample survey. One critic has described this instrument as a 'sociological meatgrinder' and compared its use to a biologist 'putting his experimental animals through a hamburger machine'; as a result, 'structure and function disappear, and one is left with cell biology' (Barton, 1968: 1). It may be apposite here to ask: Why not cell biology? Cell biology has contributed much to the understanding of structure and function on the larger, physiological scale. Perhaps political science might gain from developing its own 'cell biology' -that is (to draw a very imperfect analogy), from minutely analysing the political life of ordinary individuals, like the fifteen studied by Robert Lane (Lane, 1962). That, however, is precisely what the survey method does not do. When our 'experimental animals' are put through the hamburger machine even the cells are destroyed. All genuine individuality disappears, and in any case the individuals actually surveyed are only stand-ins for the anonymous rest.

But we may also ask: Why not hamburger? If the study of the political behaviour of individuals qua individuals is not the purpose of political science, does it matter if, as Eulau says, we 'do violence' to findings about it? These findings are treated in that waythat is, uggreguteedfor the good reason that they thereby serve the real purpose of political science, which is to account for superindividual political phenomena.

Individual, ecological and personal data

In the practice of electoral studies, whether individual or ecological, truly personal data scarcely figure at all. What we call 'individual' data are just as much aggregate data as what are actually so called, though they relate to different objects and are acquired by different methods. Individual data so-called are aggregate data relating to what we define as the micro-level, or the level of microunits. We know abstractly that these microunits are supposed to be individuals, but we do not know them 0s individuals. Furthermore, although the macrounits to which aggregate data so-called relate are parliamentary constituencies, congressional districts, and the like, these are in fact aggregations of individuals. Taking these points into account, we may conclude that, unlike truly personal data, individual and ecological data (or data on the individual and ecological levels) are two kinds of aggregate data. More pertinently, they are sturisticul data, which is essentially what aggregate data are.

The individual as a statistical unit

Statistical data that apparently relate to individuals are aggregate data expressed in unit terms, where the relevant unit is conventionally labelled 'the individual'. This entity is a pure abstraction, a statistical construct which, like 'the average man', has no physical existence in the real world. Mistaking it for an actual person is a category error which, not surprisingly, leads to fallacy, paradox and false inference. But it is not fallacious to draw inferences from aggregate data to 'the individual', so long as that construct is not mistaken for an actual person. Whether it is useful to do so is another matter.

The propensity to confuse this statistical unit with a real person is by no means confined to electoral studies or political science. Perhaps its most egregious manifestation is in economics, where the corresponding construct, 'economic man', is even endowed with personal qualities like 'rationality'. In fact, the 'behaviour' of this non-person is nothing more than the aggregate behaviour of markets (eg. aggregate demand) statistically expressed in unit terms. So little significance do economists attach to this fictitious being that they make no attempt to study it empirically or test their 'hypotheses' about it; and rightly so. Such an enterprise would be as profitable as empirical science fiction.

In political science we are fortunately not much troubled by Economic Man's alter ego, Political Man, although the 'rational voter' does put in an occasional appearance. The probable reason is the difficulty of constructing any plausible political equivalent of economic rationality. No set of presuppositions about people's political behaviour commands such a widespread consensus, and every branch of political science puts its own construction on 'the individual'. In addition, when the economic model is applied directly to voting behaviour, the conclusion seems to be that a truly rational person would not vote at all.

In the light of the discussion so far, it is tempting to suggest that after forty years it is time to stop worrying about the ecological fallacy, which may not be so fallacious after all, and forget about 'real live' individuals altogether. In the actual practice of electoral studies, individuals are generally forgotten. In survey research they are the source of data, but they are forgotten in the process of aggregating the data into a statistical form. In ecological research, insofar as they figure at all, it is only as 'intervening variables' linking two sets of statistical macrodata, political and non-political. This is not to say that individual persons are 'mere conduits' of social (macro) forces. However, while they are not mere conduits, their behaviourespecially institutionally defined behaviour like voting-consists, contrary to Popper, largely in 'microeffects of macro-events'. That not all individual behaviour is like this can readily be conceded. But a concession is not a principle. It cannot be extended to the great mass of the behaviour with which we deal. Politics does not begin on the individual level; it ends there.

Aggregate data and the problem of consistency

For the purposes of electoral studies, the significance of voting is manifested in the aggregate, whether the data are acquired on the individual or the ecological level. But if the data used in both forms of analysis are consistently of a statistical kind, the apparent inconsistency of their findings becomes superficially more puzzling. How is it, for example, that one kind of analysis can show a decline in class voting and the other its persistence?

The simplest explanation for their inconsistency is that it is an unavoidable consequence of using different operations for data-acquisition. The methodological doctrine of operationalism (from which the rather confused social science notion of 'operationalization' derives) holds that a concept is defined by the operations employed in its measurement. The principal exponent of this doctrine, the American physicist, Percy Bridgman, eventually abandoned it because he could not accept the implication that if different measurement operations were employed, it was logically impossible to say that the same thing was measured (Bridgman, 1960: 5;Popper, 1959: 440). The alternative is to accept that different measurement operations may unavoidably produce different results for the same thingsas may successive measurements with the same operations. The only thing to be done about this is to refine the operations. In the social sciences it is a problem to know whether we ever measure the same things, concurrently or successively. But if our problem is the result of using different operations to measure the same things, the answer may be to refine our operations.

That could mean using more sophisticated statistical techniques, although established techniques are perhaps already rather more sophisticated than the quality of the data warrants. A more promising approach might be to extend the scope of data acquisition by including variables representing factors known or thought to affect our objects of study but not previously tapped. For example, the effects of 'context' -the social setting of voting-may be included as either macro-or micro-variables. The relative density of the 'natural constituency' of a party in an area, or the relative frequency with which an individual who might on other grounds be expected to vote for that party interacts with others in the same category, may affect the party's success in securing its 'natural share' of the vote. There is, however, a danger in expanding our databases by including more and more variables representing more and more possibly relevant factors or effects.

How long is a piece of string depends on the refinement with which we measure it. The more refined the measure, the greater the measured lengthand the longer the time required to measure it. If every variation, however small, were allowed for, the length, and the time, would approach infinity. Similarly, in electoral studies, if we included every conceivable influence on voting behaviour, we could never predict the outcome of a forthcoming election since our prediction would not be completed before the election had come and gone. Nor could we 'postdict' an election, since the world and all elections would have ceased before our task was done. Like a Mandelbrot transformation, the process is potentially infinite. Even if we did not go that far perhaps no further than Eulau recommendswe must still bear it in mind that the more we include the longer our research must take, and the further back or ahead in time it must begin. And here we run into another difficulty.

For the further back or ahead in time we begin, the less reliable our predictions will be. Given a system of more than a certain degree of complexity, and especially one whose elements show multiple interdependence, the predictability of future events rapidly diminishes to the point where it is no better than if they were random occurrences. Chaos theory suggests that the time-horizon for reliable prediction in highly-complex systems is short, and beyond it things might as well be chaotic. Trying to improve matters by increasing the refinement (but also the complexity) of measurement operations may further shorten the horizon and so, paradoxically, reduce rather than enhance the reliability of predictions over a given term. Hence the abandonment of long-range weather forecasting and the poor showing of not-so-long-range economic and political forecasts. There is no virtue in the excessive refinement of operations.

Another kind of refinement might be applied, not to the operations themselves, but to the concepts they are supposed to operationalize. For example, the class concepts employed in electoral studies are notoriously crude. Furthermore, the official class categories generally used in ecological studies are not necessarily identical to those (if any) in the minds of individual voters. Again, the assumptions of electoral studies about what classes 'ought' to mean for voting may be naive. Why, for instance, should it be assumed that in Britain being working class 'ought' to mean voting Labour? It is not just that some voters whom political scientists classify as working class may not classify themselves as suchself-classification could take care of that. It is that for some, being working class may actually be a reason for voting Conservative. In the simplest case, they might calculate that the interests of working class people were better served by Conservative than by Labour governments because the Conservatives manage the economy better and thereby ensure higher real wages. If a working class individual generalized that view to workers as a class and voted accordingly, his vote would be as much a class vote as that of another who voted Labour on grounds of class. Hence, an increase in Conservative voting on that basis (and a corresponding decrease in Labour voting) would not mean a decrease in class voting.

In a more complex case, working class Toriesthose 'angels in marble' portrayed by Robert McKenzie and Allan Silver (1968)may vote as they do because they regard being working class as occupying a certain position in an established status hierarchy to which they subscribe and which they perceive to be upheld uniquely by the Conservative party. They need not even be content with their position within the hierarchy. Rather, they may consider it proper that working class people like themselves, whatever their aspirations, should support the established hierarchy by voting for the party that alone upholds it, the Conservatives. One should vote for (and generally show deference to) one's betters. That is as much class voting as voting Labour on the grounds that one should vote for (and generally show solidarity with) one's equals. Such voting behaviour is not at all aberrant in terms either of working class voting or Conservative voting.

Class voting is in reality an exceedingly complex phenomenon with many subtle economic, social, cultural and historical undertones, as well as political overtones. It is not surprising that the statistical treatment of it, which necessarily obscures its subtleties, should throw up inconsistent results -not least when undertaken by observers more versed in statistics than politics.

In the end, it may be that operating on different observational and analytical levels makes some degree of inconsistency inevitable, no matter how refined our concepts and operations. In that case, we shall just have to live with the problemas natural scientists apparently do with similar problems in their fields.

Discrepant data in the natural sciences

It may be instructive at this point to consider how the natural sciences deal with these problems. We may begin by observing that where macro/micro problems arise in the natural sciences, rather than being resolved they become the bases of separate but parallel branches of the same science. For example, in parallel with the historic corpus of biology, there exists the newer science or sub-science of micro-or molecular biology. Each even has its own characteristic theories. Darwin's theory of evolution by natural selection flourishes alongside micro-biological hypotheses deriving from Mendel's genetic theory of inheritance. Both theories, with their corresponding but different data, are used in the analysis of biological change. But they remain quite distinct; and though the idea of reducing the one to the other has been adumbrated, that is not how the research task of biological science has been defined. Biologists pursue their interests along both paths at once. And that is as well for biology, since it is difficult to conceive how the environmental (or as we should now say, ecological) data so essential to Darwinian explanations could be brought within the scope of Mendelian analysis.

Even where reduction is successfully carried out, parallel sub-sciences persistas in the case of classical inorganic chemistry and modern nuclear physics.

Again, in the field of 'celestial mechanics', Newton's theory of gravitation has been replaced, not only by Einstein's relativity theory but also, for some purposes, by quantum-mechanics. The consequences can be highly paradoxical. Not only do the different theories yield different results, but even the data themselves seem to be affected by which theory is involved in their acquisition.

Thus the objective status of the data becomes problematical, and that fact itself becomes a scientific principle -Heisenberg's principle of 'uncertainty' or 'indeterminacy'. If that is the case in the most exact of the natural sciences, should we expect anything different in political science? In all the social sciences, there are undoubtedly large areas of uncertainty in the ordinary sense of the word, and perhaps even indeterminacy in the technical sense, especially concerning individuals. For example, how confident can we be that electors' attitudes, opinions or intentions with regard to voting (other than in the context of an actual election) are 'really there' before being elicited by an observer asking questions about them? A significant degree of indeterminacy on this level would be sure to produce inconsistent findings on different data levels, and even on the same level between different observations and observers. Or would it?

The example of quantum theory suggests that it might not. For the indeterminacy at the level of the basic units of quantum analysis is not maintained at higher levels, where determinism still prevails. Even more relevant to our problem, the indeterminacy characteristic of individual units does not apply in aggregate. True, the laws governing aggregates in the natural sciences are 'probabilistic' rather than 'deterministic'. But these 'statistical' laws still have determinate probabilities (eg, 0.5) which are consistent between different observers and observations. In the social sciences there are, of course, no comparable statistical laws. That rather suggests that social science data are even more indeterminate than the corresponding natural science data. Before accepting that view, however, it would be as well to consider an important characteristic of the subject matter of the social or behavioural sciences that ought to make for less uncertainty or indeterminacy in the status of their data, and a less problematical relation between observer and observed, than appears to exist in the contemporary physical sciences.

The nexus of meaning

Among all our data, and between all observers and observed, there is the common nexus of rneaning(s) -not merely the constructed meanings (theories) in terms of which science makes the natural world intelligible to us, but the common meanings (usages) whereby we, observers and observed together, determine the nature and even the existence of our objects of study. Elections are an obvious instance of this.

Elections and electoral behaviour can be understood in a sense in which natural phenomena cannot. Hence it is possible to construct kinds of explanations of electoral behaviour (for example, 'rational choice' explanations) for which there are no counterparts in the natural sciences. Yet this nexus of meaning seems, if anything, to make it more difficult to connect different data levels than in the natural sciences where no such nexus exists. For the differences of meaning that characterize the different levels of data actually contribute to the difficulty of bringing them together in a single explanatory framework. To go back to an earlier example: consider what it means to an actual individual voter to be 'working class' and what 'working class' means to a political scientist observing that individual's (or an area's) voting behaviour. Because the term is meaningful on both levels, the likelihood of a mismatch between the macro/observer and micro/observed data levels is increased. Nor do we improve things by pretending that 'working class' does not mean anything to voters, or that it means the same to all of them, or the same to them as to those who observe them. The problem of levels of meaning may contribute more to the problem of the level dependence of data than we generally perceive (Sainsbury, 1987). Even so, while it may be fortunate for microphysics that 'electron' does not mean anything to electrons, it would be disastrous for electoral studies if 'election' did not mean anything to electors.

What this demonstrates is that the indeterminacy that exists in human affairs is of a different order from than in nature. Human indeterminacy is 'meaningful indeterminacy'. The behaviour thus characterized can be understood just as well as any human behaviour. But it cannot be predicted by any methods available to us. Nor can inferences be drawn from one instance of such behaviour (its occurrence in one person, for example) to any other. Nevertheless, it is as true in the human as in the physical sciences that there is less indeterminacy at aggregate than unit data levels. If it were otherwise it would not be possible to formulate valid laws or generalizations of any sort. The 'laws' of economics are a case in point. Contrary to text book accounts, aggregate demand is not derived by aggregating individual 'demand curves', which are indeterminate. Rather, its theoretical shape expresses a law-like statistical generalization, though not a quantum-like statistical law, about aggregate (market) behaviour which holds, if at all, regardless of the 'economic rationality' (or lack of it) of individuals. Such generalizations can be made and tested only by rising above the level of greatest indeterminacy, which is the unaggregated individual level. For that reason, we should not be too ready to follow the individualist line of disaggregating all data to the level of individuals. Aggregation may degrade data to some degree; disaggregation degrades rf l'outrance. Instead of explaining superindividual phenomena, it destroys the evidence for them.

This view directly contradicts the individualist conviction that all social research comes down to individuals. Though such analysis might contribute to understanding in the Weberian sense, it would not have much explanatory value if the microphenomena it addressedthe personal characteristics of individual voterswere themselves either determined by macro-events, which we wished to explain, or were indeterminate. This suggests that for electoral studies the final analysis cannot be microanalysis. Before accepting that conclusion, however, we need to examine the micro/macro distinction itself, since it is from that distinction that the entire issue arises.

The micro/macro distinction

One thing that can be learned form natural science examples is that there is more to the micro/macro distinction than just the matter of scale, or of level in a sense that equates with scale. That is to say, it is not simply a matter of having data relating to 'small scale units' (microunits) on one level (the microlevel), and other data relating to 'large scale units' (macrounits) on another level (the macrolevel). Nor is it just a matter of the 'larger scale' or 'higher level' units containing the 'smaller scale' or 'lower level' ones (or the former composing the latter). There seem also to be two different conceptions of the nature of 'units' as suchthe sense in which they are units at all-which is independent of their level or scale.

First, there is the conception of a unit as 'one whole thing'. An individual elector is an obvious example. A cell and an atom are others. Then there is the conception of a unit as a 'whole lot of things'. The electorate is an examplea whole lot of electors. The biomass or the atmosphere (whole lots of cells and atoms respectively) may be others. In point of fact, when conscious of this distinction we would probably not call examples of the second kind units at all, though we might still call them 'wholes'. That is because 'unit' suggests a kind and degree of integration that the former examples possess but the latter lack. The former sorts of units are integral (in politics we might say Corporate) entities, the latter merely aggregate ones. Clearly, either could be of any scale: small scale or large scale integrals; small scale or large scale aggregates. Equally, each could contain or be composed of others of the same kind but a different scale: large scale integrals contain or are composed of small scale ones (eg. parties and their branches); and the same with large scale and small scale aggregates (eg. national and local electorates).

This, up to a point, defines our 'levels'. But only up to a point. For we make different assumptions about the relations of the levels in the two cases. The simplest case is that of aggregates on successive levels which do just compose or contain (cf. 'mixture' in chemistry). National and local electorates are again an example of this. However, the two levels of analysis in electoral studiesthe individual and ecologicalare not so simply related. Although, as we have seen, the data on both levels are aggregate data, neither can be said to compose or contain the other. In fact, the two levels to which the two sets of data relate are not successive levels of aggregation at all. Rather, they are different levels of integration.

That fact may turn out to be the source of our problem. In order to elucidate it, however, we must first consider what we understand by 'levels' in relation to integrals and aggregates.

In the case of integrals, we assume that at least some of the lower level units that make up a higher level unit are more than just contained by it. Their containment involves some determinate set of relations, some specific form of integration among them, such as we express by the word 'organization' (cf. 'compound' in chemistry). Another word frequently used in this context is 'structure', although that word has different 'aggregate' and 'integral' usages: 'age structure' and 'wage structure' relate to aggregates; 'party structure' and 'power structure' to integrals. Despite these different usages, 'structure' is the term most widely employed to denote the micro/macro relationship between integrals on successive levelsthe relationship of integration. Some express this relationship in a phrase like 'the whole is more than the sum if its parts'. The important point, however, is that the 'whole' is itself an integral, just as the 'parts' are, though not necessarily as integrated. It can therefore properly be treated as a single entity. A political organization such as a party, or a political institution such as a parliament, is an integral. As such it can quite properly be treated as a single entity, for all that it is composed of other entities which are integrals 'in their own right'.

By contrast, an aggregate cannot be treated as a single entity. The electorate is clearly not a single entity despite the singular noun, and statements that treat it as if it werefor example 'the outcome of the election was indecisive because the electorate could not make up its mind'are nonsensical. In most respects an electoral area is an aggregate, and most data relating to it are, therefore, aggregate data. But in one important respectperhaps the most importantit is an integral. As a parliamentary constituency (or whatever the local equivalent is), it is an 'integral part' of a political system; and it is only in virtue of that fact that it (or rather its electorate) can be treated as an aggregate at all. Although there may be very little 'data' relevant to it as an integral, the 'level' to which the relevant aggregate data relate is a level of integration, not aggregation.

From the above it will be seen that, while aggregates cannot be treated us integrals, integrals can be treated us aggregates. However, in treating integrals as aggregates, some things about them are necessarily left out of account that may have to be brought back in if valid inferences are to be drawn about them or from them. These things can be regarded as 'contextual' and brought in (or not) as the requirements of analysis demand. But it is important to bear in mind that the relevant context is 'structural', in that it implies an ordered set of concurrent interrelationships among integrals through all (not just a pair) of the successive levels over which they are integrated. As some would put it, the context is a 'whole system'. It is within this context, the structural context, that different levels of aggregate data will be mutually consistent, if at all.

The structural perspective

From a structural point of view, the object world appears as a vertical continuum or hierarchy of entities, each composing and composed of others 'above' and 'below' it. Lower level entities are, therefore, seen as 'parts' or 'components' of higher level ones. But the 'parts' are also 'wholes'integralsin themselves, and can be studied as such; that is, as composed of 'parts' of their own, which in turn are 'wholes' having their own 'parts'. In this perspective, all levels are equally 'structural', and 'micro' and 'macro' do not equate with 'integral' and 'aggregate' respectively. Rather, higher level integrals are 'macro-structures' and the corresponding lower level ones (the 'components' of the first) are 'micro-structures'.

The structural continuum or hierarchy thus described is without determinate beginning or end. There is no definite lowest or highest level. However, different sciences, which cover different ranges of the continuum, tend to define a lowest and perhaps a highest level for themselves, if only provisionally. Thus in any science, the integrals on some structural level will be taken as 'basic units' for the purposes of that science, at least for the time being. This level represents a threshold, usually between two sciences. However, this will hold only so long as the 'higher level' science leaves the basic units-its 'atoms' component will be different from every other. This characteristic of the structure, which is generally overlooked by methodological individualists, makes it not particularly amenable to statistical treatment. That is why sciences that primarily treat of integrals -'structural' sciences like social anthropology, individual psychology and comparative anatomyare not particularly statistical. It is also why sciences or their branches that are statistical -'molecular' ones like social psychology and molecular biology -characteristically treat integrals as aggregates; it makes them more amenable to statistical treatment.

Structural and molecular approaches in the social sciences

Turning to the social sciences, we may first note that their common basic units are individual human beings, though these may not be the prime objects of study. Rather, aggregates of individuals and superindividual integrals are the prime objects of study of the several social sciences and their various branches.

Even when individuals are studied they are generally not studied integrally, as persons, but in aggregate, as populations.

In political science, neither the molecular nor the structural perspective makes individuals the prime objects of study, except as sources of data. However, in molecular approaches individuals are always the units of analysis whatever the level of research, while in structural approaches the units of analysis vary with the level, from whole societies or political systems down to families or party cells. The lowest level of molecular analysis is therefore an aggregate level, while the analytical unit of which aggregates on all levels are supposedly composedthe individual'is an abstract, statistical unit rather than a concrete integral. Structural analysis is not obliged to reach down to the level of individuals, though it sometimes does, if only for the sake of understanding. But when it does do so it is obliged to treat individuals as distinct persons and preserve their analytical integrity. It cannot aggregate data about them, and so cannot treat them statistically. The only alternative is to treat them as fictitious persons like Economic Man.

Applied to electoral studies, what these considerations indicate is that we are confronted, not just with two types of data, 'ecological' and 'individual', but with fourthat is, with aggregate and integral (or 'molecular' and 'structural') data, each on at least two levels. One level will be that of the basic units of all the social sciences, human individuals, and is therefore defined for us as the micro level. The corresponding macro level can be any superindividual level we choose to make it. But on any level, we shall potentially be faced with some data of an aggregate type and some of an integral type. Aggregate data on any level, including the micro level, relate to populations (electorates and other social groupings), and are statistical. Integral data relate on the micro level to persons (electors and other social actors), and on the macro level to the units appropriate to the level of integration chosen (constituencies, regions, political systems, or whatever). Such data are not statistical in nature, and drawing valid inferences between them and aggregate data is, if anything, an even greater problem than doing so between different levels of aggregate data. In fact, drawing inferences between any two of the four data types is a problem.

The situation can be illustrated by Figure 1.

The arrows in the diagram indicate the directions in which inferences may in principle be drawn. As data of any of the four types may be expressed as either independent or dependent variables, the arrows are double-headed. However, in practice, drawing inferences between any pair of data types is in each case attendant with problems and possible fallacies. Thus, the ecological fallacy is associated with inferences drawn from Type I to Type I1 data. One form of the individualist fallacy arises when inferences are drawn from Type IV to Type I1 data -Barton's 'meatgrinder' process. In its most general form, the individualist fallacy is the assumption that ultimately all other types of data must be inferrable from, because reducible to, Type IV data. A 'structuralist' fallacy may be involved in drawing inferences from Type I11 to Type IV and perhaps to Type I1 data. And even if fallacies are avoidable, the practical problems of drawing inferences validly between different data types are enormous. Yet there can be no doubting that data of all four types are relevant to electoral studies. Much of this is anticipated in Eulau's account of the 'micro-macro continuum'. But much is also obscured by his account because of his virtual exclusion of anything recognizably structural in favour of a one-sidedly molecularist perspective which he uncritically equates with behavioural analysis. This is clearly evident in another passage: 'It is necessary . . . to distinguish between the study of political behaviour and the behavioural study of politics . . . It is possible to do research on political behaviour without making use of the concepts and methods of the behavioural sciences. The only requirement for the study of political behaviour is that the individual political actor be the empirical unit of analysis whose behaviour is described -though probably not explained . . . In the behavioural study of politics . . . the individual remains the empirical unit of inquiry, but the theoretical units may be role, group, institution, organization, culture, or system, and so on, whatever conceptual tools may be most adequate for the purpose of a particular investigation'. (Eulau, 1%7: 36, emphasis supplied;cf. Eulau, 1963).

That this is a molecularist perspective is unmistakably indicated by the emphasized words. Their implication is that the behavioural study of politics cannot be other than molecular.

This exclusive perspective has characterized political sociology since its inception, despite the repeated pleas that it 'take structure into account' so strongly voiced from the 1950s to the 1970s. It was during that period, and in response to those demands, that methodologists (Paul Lazarsfeld most notably) formulated much of the nomenclature relating to the types of data since used in electoral studies, including such terms as 'contextual', 'global', and even 'structural'all, however, denoting kinds of aggregate data (Lazarsfeld and Rosenberg, 1955;Lazarsfeld, 1959;Lazarsfeld and Menzel, 1969). But despite increasing methodological sophistication, the treatment of structure as it relates to integrals has remained underdeveloped. Under the influence of methodological individualism, it has been entirely overshadowed by molecular analysis.

Methodological individualism may be defined as the method of treating all superindividual integrals as aggregates. Only individuals are allowed to be treated as integrals; but as we have seen, they end up being aggregated too. It is entirely legitimate to treat integrals as aggregates, though the reverse is not the case. But where structure is highly significant the results are likely to be problematical.

Taken literally, pure molecularism embodies a Laplacean view of the social world that would commit us to a research task even more stupendous than the one outlined by Eulau. Laplace held that if we could determine the current positions of all the bodies in the universe and all the forces acting on them we should be able to predict all future physical events (Laplace, 1820). The Laplacean political scientist would presumably hold that if we could determine the political positions of all individuals in society and the political forces acting on them we should be able to predict all future political events. Of course, such a project could not be carried out; for one thing, there are just too many individuals. But anyway, Laplace was wrong in principle, since, as we now know, the things he would have needed to determine are actually indeterminate. So obviously is that the case with the corresponding political phenomena that there are no Laplacean political scientists. And yet, as the extract from Eulau shows, some do in their more expansive moments come close to adopting such a standpoint, particularly when emboldened by the apparent promise in the survey method o € a cheap, mechanical way of acquiring data on indefinitely large numbers of individuals.

This neo-Laplacean illusion is the 'individualist fallacy' taken to its extreme. Even if it were not fallacious, however, it would still be infeasible, not least because its feasibility must depend on possessing a fully elaborated and validated molecular theory of society, which manifestly we do not have. That is why, though we have lots of statistics, we have no statistical laws. Nor does it seem likely we ever shall have. Yet absurd as the illusion is, it still seems to inhibit the development of structural perspectives, so much taken for granted in other sciences, and so much needed in ours.

Conclusion

We may conclude, therefore, that the problem of seemingly inconsistent sets of data in electoral studies does not arise only from the use of different methods of data-acquisition, or from indeterminacy at the individual level. It arises also, and inescapably, from the fact that, though the data are of the same statistical (ie. molecular) type, they relate to different structural levels of the political and social systems. These levels are not simply different levels of aggregation. Their relation is much more complex, and in certain respects more remote, than that. For example: how integrated is the average constituent in a constituency? Even to determine what the inconsistency 'means', let alone correct or control for it, it would be necessary to establish the structural context in which it occurred. That might be difficult. It would certainly require major new developments in political science. But it would not involve the sort of impossible project described by Eulau. To see what it would involve, let us review the fundamental difference between the structural and molecular perspectives on superindividual integrals. From the structural point of view, an integral, though having any number of components, is a single entity, and can be studied as such. Furthermore, though structural analysis can be carried all the way down to the relevant basic unitsin our case, individualsthat is not a requirement.

Consequently, the number of objects to be studied remains relatively small, certainly compared to the number of individuals, and in fact tends to diminish at higher structural levels.

This contrasts with the molecular perspective which treats all integrals as aggregates (ie. statistically). Because the empirical unit of inquiry remains the individual, the number of objects to be studied increases through progressively higher levels of aggregation until it embraces the entire population. Hence the necessity and appropriateness of using surveys and statistics. But in this process virtually all aspects of structure, including the internal and external integration and differentiation of all integrals, not least individuals, are unavoidably obscured. Some can be recovered as 'contextual' effects, but there are practical limits to that. The structures themselves of course remainthe class structure (a Type 111 variable), for example. It is just that they are no longer apparent from the data.

Assuming the data themselves are of value, the loss of structural infonnation need not be of decisive importance. Molecular analysis simply provides a different perspective on the same realityalthough here the lack of a molecular theory is important since that does seriously limit the value of the data. It is really only when inconsistencies occur that relate to the 'lost' structure that anything needs to be done about it. But what can be done?

One thing that cannot be done is simply to bundle structural information in with the molecular data and treat it as if it were all of the same type. Some can be treated 'contextually' in that way. But most is not sufficiently amenable to statistical representation to be treated thus, and attempts to do so are likely to overload the analysis with innumerable 'dummy variables' and other statistical spuria to no purpose. What could be done, would be to develop a parallel structural political scienceand to a considerable extent that has been done (since Aristotle, in fact). But it is a possibility that has been much neglected in recent decades, due to the unfounded belief, absent from the natural sciences, that only one kind of approach is legitimate. However, the two would remain distinct, each with its own limitations, both adding to our knowledge and understanding, but neither able to come to the rescue of the other.

Given the difficulties inherent in the subject matter of electoral studies, in the predominant mode of analysis, and in the relation between the two, it is not at all surprising that inconsistencies should have arisen in electoral findings. Nor is it likely that they will easily be resolved. As with the practitioners of other, more advanced and more exact sciences, we shall probably have to live with them.

Fig. 1 .

Robinson . . . 'The use of aggregate data, therefore, is likely to conceal a good deal of the variance in the behaviour of political actors which the use of discrete data reveals.