The Logic of Comparative Social Inquiry

Washington University
HENRY TEUNE University of Pennsylvania
CHAPTER TWO
Research Designs
"Most Similar Systems" Designs. "Most Different Systems" Designs. Univariale Comparisons. Comparing Relationships.
Most comparative studies take as their point of departure the known differences among social systems and examine the impact of these differences on some other social phenomena observed within these systems. An alternative strategy, however, is available. With this strategy, differences among systems are taken into account as they are encountered in the process of explaining social phenomena observed within these systems. Although emphasis wili be placed on the latter strategy, the assumptions and implications of both strategies will be the subject of this chapter.
As discussed in the previous chapter, a general theory is composed of propositions formulated in terms of variables observed either within social systems or at the level of systems, but devoid of the names of social systems. Since the number of the relevant determinants of any kind of social behavior is likely to exceed the number of accessible social systems, the objective of a theory free of all proper names will not be easily reached, and thus procedures must be formulated to maximize this objective. // "'
All research involves denning the population for which the study is to be conducted and selecting a sample from this population. Sampling methods vary greatly, depending upon the problems of the research and the nature of the population. Sometimes the sample is a random selection from the entire universe; sometimes it is selected in several steps in which some larger social units are chosen first and other social units within them are sampled subsequently; in other cases the sample is "stratified"-individuals are selected on the basis of their p o'sition on some variable, such as income or education. The common and ubvious procedure in cross-systemic re-
31

32
Ri'xtwch Designs

search is to first select systems and then to sample individuals or groups
within them.
For practical reasons (lie selection of countries can rarely be random. Even though the universe of social systems-countries, nation-states, cultures, and so forth-is fairly limited, the costs of conducting a study within random samples taken within each system will for a long time remain prohibitive. Therefore cross-national studies often have a quasi-experimental form, and the tactical choices are limited to the question of the "best" combination of countries, given the overwhelming limitations of money, access, and social scientists.
"Most Similar Systems" Design
The currently predominant view among social scientists seems to opt for the strategy that Naroll calls studies of "concomitant variation.1'1 Such studies are based on the belief that systems as similar as possible with respect to as many features as possible constitute the optimal samples for comparative inquiry. For example, Scandinavian countries or the two-party systems of the Anglo-Saxon countries are seen as good samples because these countries share many economic, cultural, and political characteristics; therefore the number of "experimental" variables, although unknown and still large, is minimized. This type of design is a "maximim" strategy. It is anticipated that if some important differences are found among these otherwise similar countries, then the number of factors attributable to these differences will be sufficiently small to warrant explanation in terms of those differences alone. A difference in the intensity of political partisanship between Sweden and Finland can be attributed to a smaller number of intersystem differences than between Sweden and Japan.
Alford's study of social determinants of voting was based on this kind of perspective. Describing the choice of countries, Alford noted:
"The Anglo-American countries-Great Britain. Australia, New Zealand, the United States, and Canada-are alike in the important respect that they may be termed "pluralist" political systems. . - . Each of the Anglo-American countries tends toward a two-party system. - . . The electorate is not fragmented into supporters of one or another small party hoping to gain a few seats and a voice in a coalition government."2
' Raoul Naroll, "Some Thoughts on Comparative Melhod in Cultural Anthropology." in H. M. Blal(M:k and Ann Bhitock. eds. Methodology in Social Research. McGr.iw-Hill, New York. 196R.
* R. R. Alford, "Parly and Society," in F. J. Munger, ed., Studies in Comparative Politics, Thomsis Crowd!, New York. 1967, pp. 66-67.

"Most Similar Sy-vienis" Design 33
He then discussed the differences between this set of countries and the multi-party systems of continental Europe, such as the relatively minor importance of religion as a determinant of voting among the Anglo-American countries. Finally, Alford specified the factors that differentiate the Anglo-American countries and (hat might explain tlie differences in the extent of class-voting. Allardt considered in similar terms the differences in class-voting among the Scandinavian countries and attributed the relatively high extent of such voting in Finland to the comparatively lower mobility rates in that country.3 In their study of civic culture Almond and Verba chose countries that have a "democratic political system" but differ with regard to their level of development.4 Studies of social mobilityR and suicide8 in Scandinavia followed this strategy. Cantril's 7 and Dogan's8 studies of Communist voting in France and Italy took as their point of departure the similarities between these political systems. This is also the perspective of the "area study" approaches in the social sciences, whether the area is defined in cultural or political terms.
Intersystemic similarities and intersystemic differences are the focus of the "most similar systems" designs. Systems constitute the original level of analysis, and within-system variations are explained in terms of systemic factors, Although these designs rarely have been formulated rigorously, their logic is fairly clear. Common systemic characteristics are conceived of as "controlled for," whereas intersystemic differences are viewed as explanatory variables. The number of common characteristics sought is maximal and the number of not shared characteristics sought, minimal. The resulting statements will take the following form: "Among the Anglo-American countries, which share the following characteristics. . . , differences with regard to class voting can be attributed to the following factors. . . ." There is no reason why these statements have to be formulated exclusively at the systemic level. One might find, for example, that among democratic countries that are economically developed, church at-
'Erik Allardt. "Patterns of Class Conflict and Working Class Consciousness in Finnish Politics," Publications of the Inslilute of Sociology, University of Helsinki, No.30,1964.
*G. A, Almond and Sidney Verba. The Civic Culture, Princclon University Press, Princeton, N.J., 1963.
Kaare Svalasloga. Prestige, Class, and Mobility, Gyldenal Scandinavian University Books. Copenhagen, 1959.
Habat Hendin, Suicide in Scandinavia: A Psychoanalytic Study of Culture and Character. Grune & Sirallon. New York, 1964.
T Hadley Caniril. Tin- I'fililir'i of Despair. Basic Books, New York, 1958.
Mattei Dogan, "Political Clea'.'age and Social Stratification in France and Italy," in S. M. l.ipsct and Stein Rokkan. cds-, Pa'ly Sysfe/nx mid Voivr A ligaments: Cross-National Perspectives. Free Press. New York, 1967.

34 Research Designs
tendance is either positively or not at all related to party identification, whereas among the less-developed democratic countries the relationship is
negative.9
If such a difference is found among the systems studied, the following theoretical implications follow: (1) The factors that are common to the countries are irrelevant in determining the behavior being explained since different patterns of behavior are observed among systems sharing these factors. (2) Any set of variables that differentiates these systems in a manner corresponding to the observed differences of behavior (or any interaction among these differences) can be considered as explaining these patterns of behavior. The second implication is particularly important. Although the number of differences among simitar countries is limited, it will almost invariably be sufficiently large to "overdetermine" the dependent phenomenon. Although "most similar .systems" designs focus on concomitant variation, the experimental variables cannot be singled out. There is more than one factor that ranks Great Britain, Australia, the United States, and Canada in the same order; there is more than one difference between the United States, Great Britain, and West Germany, on the one hand, and Italy and Mexico on the other. But even if we assume that some differences can be identified as determinants, the efficiency of this strategy in providing knowledge that can be generalized is relatively limited-
"Most Different Systems" Design
The alternative strategy takes as the starting point the variation of the observed behavior at a level lower than that of systems. Most often this will be the level of individual actors, but it can be the level of groups, local communities, social classes, or occupations. Although the goal of this strategy is the same as in the "similar systems" design, systemic factors are not given any special place among the possible predictors of behavior. For example, we may be interested in explaining variations in college student attitudes toward personal adjustment,10 perceptional illusion of movement," values of youth,12 or values of local leaders.1-1 The initial assumption is that
11 G. A. Almond and Sidney Verbil, op. rii.
10 J. M. Gillespie and G. W. Allport. Youth's Ouilook on the Future: A Cross-Ncilionol Smdy, Doubleday, New York. 1955.
"G. W. Allport and Thomas Pctligrew. "CuHural Influence on the Perception of Movement: The Trapezoidal Illusion among Zulus." Journal of Abnormal and Social PsychotoRv. 55. 1957.
" H- H. Hyman. Arif Payaslioglu. and F. W. Frey, "The Values of Turkish College Youth." Public Opinion Quarterly, 22. 1958.
'* P. E. Jacob, Henry Teune, and T. M. Watts. "Values. Leadership, anil Development," Social Science Infonnalwn, 7, 1968.

"Mosf Similar Sysfeim" Design 35
individuals were drawn from the same population; in other words, that systemic factors do not play any role in explaining the observed behavior. Further investigation consists of testing, step by step, this assumption in the course of cross-systemic research. As long as this assumption is not rejected, the analysis remains at the intrasystemic level; whenever the assumption is rejected, systemic factors must be considered.
The first step in this design is to identify those independent variables, observed within systems, that do not violate the assumption of the homogeneity of the total population. Although the samples are derived from different systems, they are initially treated as if the population from which they are drawn is homogeneous. If the subgroups of the population derived from different systems do not differ with regard to the dependent variable, the differences among these systems are not important in explaining this variable. If the relationship between an independent and the dependent variable is the same within the subgroups of the population, then again the systemic differences need not be taken into consideration.
To the extent that general statements can be validly formulated without regard to the social systems from which the samples were drawn, systemic factors can be disregarded. If rates of suicide are the same among the Zuni, the Swedes, and the Russians, those factors that distinguish these three societies are irrelevant for the explanation of suicide. If education is positively related to altitudes of internationalism in India, Ireland, and Italy, the differences among these countries are unimportant in explaining internationalist attitudes. Whereas studies of concomitant variation require positive identification of relevant systemic factors, the "most different systems" design centers on eliminating irrelevant systemic factors.
The difference between the two strategies should not be overemphasized. Both strategies can result in the confirmation of theoretical statements and both can combine intrasystemic and intersystemic levels of analysis. In the most different systems design, the level of analysis is shifted to systemic factors when the formulation of valid general statements is no longer possible for all of the subpopulations. If it is found that attitudes of internationalism in India and Iran depend upon exposure to mass media but do not in Ireland and Italy, then the differences between the two sets of systems become relevant and reference must be made to the systemic level. When this is necessary, concomitant variation is studied ex post facto, and intersystemic differences are attributed to the observed variations within systems.
Concomitant variation studies are focused almost exclusively at the level of systems. Certain systemic traits ?JC held constant, and others arc allowed to vary. Dcnumeration in terms ol national social or political systems or

36 Research Designs
cultures is only one of the many possible ways of conceptualizing social systems as the units of analysis in any theory. One could design research at the level of the American states, Finnish regions, Peruvian villages, Northern CaliforniEin tribes, and so forth. Similar systems designs, however, require an a priori assumption about the level of social systems at which the important factors operate. Once a particular design is formulated, assumptions concerning alternative levels of systems cannot be considered. The original assumption can be tested only in its entirety-either the systemic factors of the specified level of social systems are or are not relevant.
In the most different systems design, the question of at which level the relevant factors operate remains open throughout the process of inquiry. The point of departure of fins design is the population of units at the lowest level observed in the study, most often individuals. The design calls for testing whether this population is homogeneous. If subgroups of this population that correspond to some identifiable levels of social systems can be distinguished empirically, then factors operating at this level of systems will be considered. If a population of individuals is sampled from several communities within several countries, then differences among individuals will be tested both within and across communities and within and across countries. If communities differ, systemic factors operating at the level of local communities will be considered; if nations differ, national factors will be examined; if neither countries nor communities differ, the entire analysis will remain at the individual level and no systemic factors will be considered. The level that reduces to the greatest extent the within-group variance will be considered.
Although the subsequent technical discussion is based on a multiple regression model, it is also possible to visualize this design as one in which the patterns of interaction are being systematically examined for alternative ways of grouping individuals, whether based on a classification of various levels of social systems or some attributes measured at the individual level.14 Whenever classification into some level of systems results in the greatest reduction in variance and therefore yields the greatest gain in prediction, the level of analysis is shifted to factors operating at this level,
In the context of this design, the definition of comparative research becomes clear- Comparative research is inquiry in which more than one level of analysis is possible and {he units of observation are identifiable by name
"Computer programs lhal operate in a "iree" fashion and study interaction independently for each "branch" (e.g.. Ihe Automatic Interaction Detector) may be moat suitable for this purpose-

al each of these levels^ Thus a study of local leaders sampled from local communities in a single country is comparative, since research can proceed at both the individual and at the community levels. But if supranational regions are not identifiable, according to this definition a study conducted exclusively at the level of countries is not comparative.
Since the goal of research is to confirm general statements about human behavior, the process of sampling, even if it is not random, should be oriented toward this goal- No research based on a design other than a random multistep sample of all social systems will allow universal generalizations. The validity of generalizations and the guidelines for further research provided by the two research strategies will depend upon the nature of the findings that they respectively bring. Findings desirable in the most similar systems design are highly undesirable in the most different systems design and vice versa. Let us discuss this statement.
In the most similar systems design, systems with as many similar characteristics as possible are sought, Without attempting to provide a list, let the characteristics shared by the Scandinavian countries be denoted as Xi, X-i, . . ., X>,, and the characteristics that are not shared,as A^+i, Y^a, . . ., X,,. A dependent variable, whether it is a frequency distribution of one variable or a relationship between two variables, is found to vary among these highly similar countries. For example, according to Allardt the amount of class voting varies among the Scandinavian countries16 A data matrix for five countries in this kind of a design would assume the following form (all variables are dichotomized):
Country Variables Controlled "Experimental" Variables Dependent
The dependent phenomenon can either be a single aggregated attribute or a within-system relationship. Of course there are other factors that differentiate these systems in ways not associated with the variations of the dependent variable. The resulting finding, if stated carefully, may take the following form; "When the observed systems share characteristics X\^ A"a,
"It should be noted that this is Ihe meaning of the term "comparative" as used in psychology. Compunitive psychology is a study of organisms at different levels of structural differentiation.
" Erik Atl;"rdt, up. oil.

38 Research Designs
. . ., Xk. the variations of the dependent variable V (or of the relationship between an independent variable -V, and the dependent variable ~Y", both measured within systems) are associated with the variable X^^i (according to the hypothesis) or the alternative variables A'k+t, . . . , Xn (alternative hypotheses)."
What further implications follow from this finding? We obtain a positive, although overdetermined, explanation of the dependent variable Y-it either depends upon A'|.-+ i, as hypothesized, or the variables --Vk+2, . o o , Xn, which are not controlled, The original hypothesis is confirmed, although alternative hypotheses are not rejected. This certainly strengthens our confidence in the explanatory power of factor A\+i, and, although no rigorous inferences are possible, further research is directed toward testing the influence of A^+i in other settings.17 Thus if we find some other social system that shares with these systems all of the characteristics, X\, . .., X^, it is likely that a similar explanatory pattern will be found. If, however, any one of these characteristics is different, no inferences are possible since it is likely that this particular trait interacts with the dependent variable.
If a hypothesis is confirmed as a result of the most similar systems design, we gain some encouragement about the generality of the hypothesis. For example, if we find that among Scandinavian countries frequency of social mobility is associated with the frequency of class voting, we will be prompted to test whether mobility is also associated with class voting among the Anglo-Saxon countries. Moreover if we find that among the Anglo-Saxon countries, which share characteristics other than those shared by Scandinavian countries, mobility is also associated with class voting, the confidence in the explanatory power of mobility will be further strengthened. If, however, mobility is not related to class voting among the Anglo-Saxon countries, we are back where we started. All we now know is that class voting depends upon mobility, which in turn depends upon other factors that cannot be isolated.
The logic of the most similar systems design is based on the assumption that characteristics shared by one group of systems, such as Scandinavian countries, can be removed one-by-one in quasi-experimental manner. But this is an unrealistic assumption. As we argued previously, social phenomena vary in syndromes and it is difficult to isolate experimental factors.
"Let us note that we are talking here in psychological and not in logical terms. Within the present logic of inference, one cannot make any generalization beyond the population from which the sample has been drawn. However, i( is apparent (hat such theory of induction is not appropriate for social science and that, in their practical activities, social scientists arc actually willing lo take the risk of false generalizations rather than salisfy themselves with rigorous inferences about accidental populations.

Univaridie Comparisons 39
The most different systems designs eliminate factors differentiating social systems by formulating statements that are valid regardless of the systems within which observations are made- As long as these statements continue to be true in all systems, no reference to systemic characteristics is made. As soon as additional statements cannot be validly formulated across systems, however, the hypothesis concerning no difference among systems has to be rejected and the level of analysis is shifted to systemic factors. At this point, the association of the intersystemic variations with the intra-systemic differences would be examined. For example, if in a group of systems political participation is positively related to education but the remaining differences in political participation cannot be explained by any other variable measured within systems, it would be necessary to identify the systemic factors associated with these differences. It should be emphasized that the systemic characteristics need not be dichotomous. For example, one may relate the within-system correlations between budgetary requests and budgetary appropriations to characteristics of American states, such as their per capita income or the degree of interparty competition.
Both of these strategies are based on some expectations about social reality. The most similar systems design is based on a belief that a number of theoretically significant differences will be found among similar systems and that these differences can be used in explanation. The alternative design, which seeks maximal heterogeneity in the sample of systems, is based on a belief that in spite of intersystemic differentiation, the populations will differ with regard to only a limited number of variables or relationships. On the one hand, if it turns out that Swedes, Finns, Norwegians, and Danes are alike in all of the examined aspects of their social behavior, then the study of these countries will not permit the identification of the systemic factors relevant for a particular kind of behavior. If, on the other hand, Americans, Indians, Chileans, and Japanese show no common patterns of behavior, a study of these countries will end up with four separate sets of statements contributing equally tittle to general theory.
Univariate Comparisons
Underlying the preceding discussion is a set of statements concerning the "sameness" of samples derived from different social systems. Systemic factors can be attributed to within-system variables if the systems are found to be "different" either with respect to a single variable, aggregated at the system level, or with regard to within-system relationships. By the same token, systemic factors can be eliminated from exp'^nation if within-system patterns are found to be the "same." Any formulation of a problem of inquiry as comparative is based on the assumption that factors subsumed

40 Research De'.igns
under the proper names of systems may potentially influence the phenomena that are being explained.
If systemic factors do indeed influence the within-system patterns, whether univariate or multivariate distributions, then identification of the system within which an observation is made raises our ability to predict a score on the dependent variable above the prediction based only on the mean score for the entire, or "total" population. The coefficient of regression of an individual's score on a variable representing his membership in a particular system must be larger than zero if the population is heterogeneous in terms of systems.
As an example, suppose we are examining individual propensity to vote for the parties of the right among Western European countries. If the proportion of the electorate voting for the parties of the right is the same in all countries, it becomes quite irrelevant whether an individual is a Frenchman or an Italian. Other factors are important, for example, social class or religion. If the members of the Western European elites share similar attitudes toward European integration, again it is not important whether a particular person is a member of the Dutch or the Italian elite. To the extent that identifying the social system does not help predict individual characteristics, systematic factors are not important. The total population is homogeneous, and further research is. not distinct from investigations customarily conducted within a single social system. The analysis can proceed at the level of. individual characteristics without resorting to any system-level variables.
If it can be assumed that the measurement of a given variable is relatively free of systematic error at the system level and if the scale of measurement is known, a simple test concerning differences among means (one-way analysis of variance) can be used to ascertain whether social systems differ with regard to this variable. The question we want to answer is whether the extent of variation of a given characteristic within each country is smaller than variation among countries. If all trains in England move at a speed of 50 miles per hour and all trains in France move at a speed of 60 miles per hour, then knowing the fact that someone is traveling in France rather than in England will be helpful in predicting the duration of a journey. But if the speed of trains in both England and France varies between 30 and 70 miles per hour, the difference of 10 miles per hour in average speed may not be sufficient to improve a prediction about the duration of a trip. The type of train or time of the year may be much more important than the country.
The nature and the extent of intersocietal differences have long been subjects of theoretical formulations in the social sciences. Anthropologists

Univaruile Comparisons -11
tend to perceive societies as highly different. Although individual personalities are "potentially" the same, culture, social organization, child-rearing practices, or some other factors result in the predominance of certain personality types in particular societies. These cultural configurations, or "patterns of cultures," were originally identified from folk themes, customs, and so forth. Patterns of culture were not based on the notion of frequency distribution of personality types within a culture but on an ideal-type personality model. Subsequently, however, the concept of modal personality replaced the concept of patterns. Modal personality, denned as the product of interaction between "fundamental physiologically and neurologicalty determined tendencies and experiences common to all human beings" and their cultural milieu, became a subject of statistical analysis of distributions of personality types. Furthermore, if Singer's conclusions are correct, projective techniques indicate that the distributions within societies are flat, and within-culture differences of personalities are therefore larger than the between-culture differences.1" It is not clear to what extent these findings can be generalized, but they are certainly surprising. Concepts of "cultural patterns," "modal personality," and "social character" and the problems of relating sociocultural settings to individual traits have an extensive theoretical tradition, but the empirical findings are scarce and thus inconclusive. As Inkeles and Levinson emphasize, "If national character refers to modes of a distribution of individual personality variants, then its study would seem to require the psychological investigation of adequately large and representative samples of persons, studied individually."19
One set of attitudes that has been extensively studied concerns evaluations of occupational prestige in different societies.20 Although the methodology of these studies has not been uniform and the samples have varied greatly, the general findings seem to indicate a high degree of intersocietal uniformity- These findings again run counter to our theoretical intuitions in light of which the prestige of occupations ought to be related to industrialization or social division of labor. But if the methodology of
"This discussion is based on Milton Singer. "A Survey of Culture anil Personality Theory and Research," in Bert Kaplan, ed.. Sliidyinf: Personality Cross Culturally, Row, Peterson, Evanston, 111., 1961.
1B Cited in Singer, op. c'u.. p. 55.
;a Alcx Inkeles and Peter Rossi, "National Comparisons of Occupational Prestige, American Journal of Sociologv, 61, 1956; Alex Inkeies. "Indiislrial Man: The Relation of Status lo Experience, Perception, and Value," American Journal of Sociology, 66, 1961; E. M. Thomas. "Reinspecting a Structural Position on Occupational Prestige." American Journal of Socinlofy. 67, I9fi2; A- 0. Haller, 0. 1. Lewis, and Iwao Tshino, "The Hypothesis of Intersocielal Simllaril.- in Occupational Prestige Hierarchies," Anii-rican Sum-mil of Swiflo.vy. 71, 1966. .lesearch reports on occupational status are available from at least 16 countries.

42 Research Designs
these studies is sound-if Americans and Japanese, Poles and Brazilians, Germans and Indonesians evaluate particular occupations alike-theories relating the socioeconomic structure to these attitudes will have to be revised. Social science theories may in general overstate intersocietal differences and the role of system-level factors, and in this era of empirical truth many myths might have to be revised. When Lipset and Bendix slated that "the overall pattern of social mobility appears to be much the same in the industrial societies of various Western countries," they felt it necessary to emphasize that this finding "runs counter to widely help impressions concerning the different social structures of American and Western European societies."21
IF no differences are found among systems, the population is homogeneous and systemic factors cannot be expected to be important as determinants. Thus the test of differences between or among national means-either a mean test or a variance test-provides a general estimate of the relevance of systemic factors and a guideline for the choice of the proper level of analysis. If the sample is differentiated in terms of systemic characteristics, generalizations beyond the examined sample of countries seem relatively safe. If the Indian, Polish, Yugoslav, and American local leaders do not differ in their orientation toward change, it can be expected that local leaders in other countries are not significantly different, and, in general, that systemic factors are not important in explaining this particular attitude.
These examples of inters ystemic similarities with regard to a single phenomenon, such as personality types, evaluation of occupations, social mobility, or values of local leaders, are by no means intended to support a thesis that social systems do not differ. Illustrations, both of an impressionistic and systematic nature, of inters ystemic differences are abundant. The examples discussed were merely intended to show that the assumption of intersystemic similarities, underlying the most different systems design, should not be discarded a priori as invalid. To our surprise and contrary to many theories, such similarities are indeed being discovered. The validity of this assumption, of course, will depend upon the nature of the social phenomena under consideration: one may expect that psychophysiological phenomena will be less dependent upon the social system than are political phenomena.
A limitation on comparing systems with regard to individual-level phenomena must be emphasized: the problems of measurement. Cross-system comparisons of single variables will be dependent upon the units and the scale of measurement within each social system. Very often such direct
*'S. M. I-ipsel anil Rcinhard Hendix. Swia! Mol'ilily in Industrial Society. University of California Pnfes, Berkeley. I960, pp. 11 and 13.

Comparing Relationships 43
comparisons will not be possible, either because the scales of measurement are unknown (e.g., is political participation in the Soviet Union higher than in the United States?) or because the investigator may choose to quantify the variables in a way that precludes this kind of comparisons (e.g., by dichotomizing at the national medians). This limitation will be discussed in greater detail in Part Two.
Comparing Relationships
Descriptive, univariale comparisons may often not only be difficult, they may also be less interesting than the multivariate patterns of determination. Since most theoretical propositions are formulated in terms oE predicting one variable by some other variables, the form and the fit of these predictions arc of central importance for a theoretically minded social scientist. Within-system predictions and the fit of these predictions, or "relationships," often constitute the focus of analysis. When leaders and citizens in several countries are studied, one can ask whether membership in India or Yugoslavia has more effect on the values of an individual than a position as a local leader. When perceived freedom to discuss politics is studied, one can ask whether education or the system better predicts individual perceptions. If achievement motivation is studied in Brazil and in the United States, one can ask whether social class or nationality is a better predictor.
The question is whether the relationship between the variable being explained and an independent variable is the same within all systems:
whether systemic characteristics are important in determining the form and the fit of theoretical predictions in different social systems. Again, if values in all countries are in the same way associated with political positions, or if freedom to discuss politics is related to education, or achievement motivation to social class, then systemic factors are not important in explaining the dependent variable. And again, as additional independent variables are considered, it may very likely transpire that at some point systemic characteristics do influence the observed relationships. But each finding of similarity of relationships across social systems reduces the number of potentially relevant systemic characteristics. The most different systems design implies an analytical strategy in which the overall influence of systemic factors is assessed step-by-step with the addition of each new variable,
Illustrations of similar relationships in various social systems are plentiful. Most recent comparative studies of political behavior seem to discover that relationships mnoni; individual attitudes are the ''ame regardless of political system. In his inventory of research on politiol participation, Milbrath found only two instances in which a relationship was not the same in all

44 Researcil Uesigns

political systems.'-'2 The study of civic culture consistently shows that education is the most powerful determinant of political attitudes in five countries. Indeed, Almond and Verba conclude the following:
"It is ... among the most important facts we discovered that most of the relationships between education and political orientation are of the first type: educational groups differ from one another substantially, and in a similar way, in each nation.23
Rokkan reports similar findings in the study of attitudes toward European integration:
". . . Galiup International, in its study of Public Opinion and the Europe of the Six, found that 62 percent of the Dutch sample was strongly in favor of unification, and only 36 percent of the Italians. This difference, however, tells us very little about the chances of strains between the two countries in the articulation of policies toward Europe. It turns out that the belter educated in the two national samples think practically alike'. 70 percent of them were strongly in favor of European unification. The difference between the two countries resulted almost entirely from a contrast in levels of education and information. . . . [emphasis added)"24
Converse and Dupeux report major differences in the frequency of party identification between France and the United States. Seventy-five percent of Americans identify themselves with a political party, while only 45 percent of the Frenchmen perceive themselves in partisan terms. This difference, however, can be attributed to the higher rates of political socialization through the family in the United States. The authors show that in both countries those persons who know their father's party preference are very likely to have a party preference themselves-79.4 percent in France and 81.6 in the United States. Converse and Dupeux conclude:
"Where the socialization processes have been the same in the two societies, the results in current behavior appear to be the same, in rates of formation of identification. The strong cross-national differences lie in the socialization processes. In other words, we have come full circle again: we
o L. W. Milbrath, Political ParSicipasion, Rand McNally, Chicago, 1965.
**G- A. Almond and Sidney Verba, op. cil., p, 317.
"Stein Rokkan, "Comparative Cross-National Research: The Context of Current Efforts." in R. L. Merritt and Slein Rokkan. eds.. Comparing Nations: the Use of Quaniiiaiive Data in Cross-National Research, Yale University Press, New Haven, Conn., 1966, p. 19.

have encountered large national differences but have once again succeeded in moving them to the marginals of the table."'25
One could expect that in all the cases cited above the social system does not increase the accuracy of prediction of the dependent variable. If an illiterate Italian were an illiterate Dutchman, his attitude toward integration would have been the same. If an American who does not know his father's party preference were a Frenchman who did not know his preference, it would still be unlikely that he would have party identification. As long as the independent variables remain the same, membership in a social system is not important in predicting the dependent variable. Education is a good predictor; social system is not. Class is a good predictor; social system is not- What matters is not whether an individual's name is John Smith or Giovanni Bianco, but whether he went to school or not, whether he knows his father's party preference or not, whether he has a high income or not. The countries differ with regard to their levels of education, class structure, and family socialization, but they do not differ as systems so long as their patterns of relationships are the same. System's differ not when the frequency of particular characteristics differ, but when the patterns of the relationships among variables differ.26
The fact that a single independent variable measured within systems yields a gain in prediction of the dependent phenomenon does not preclude the possibility that systems may al.-so contribute to the explanation. If a set of independent variables, measured wthin each system, predicts the dependent phenomenon independently of all systemic characteristics, the initial variation of the dependent variable will disappear when the means of the independent variables are adjusted. If the difference between Americans and Frenchmen disappears when the frequency of knowledge of father's party identification is adjusted, then systems cannot contribute to the
a1 A large number of examples of the structure of relationships among attitudes can be seen in several studies attempting lo develop allitude-measunng instruments. These findings are nios( impressive in that, in spite of the difference among the cultures and the differences in the mtensiey of parlicular altitudes, the structure of interrelationships among atliludes is highly invariant. See, for example, D. H- Smith and Atex Inkeles, "The OM Scale: A Comparative Socio-Psychological Measure of Indi-'o'idiial Modernity," Socionu'sry, 29, 1966; J. A. Kahl. "Some Measurements of Achievement Orientation." Ainrricw Journal of Sociology, 70. 1965; Howard Maclay ^d E. R. Ware, "Crnss-Ciillural Use of the SemEi,-<ic Differential." Behavioral o^CK'KCC, 6, I96I; Salomon Rellig and BenJ;imin Pasamanick. "Invariance in Factor Structure of Moral Value Judgments from American and Korean College Students," socio^nl'l^v, 29, 1966.

System Level Variables: Changing the Level of Analysis
Differing Relationships. Comparative Study and Levels of Analysis. System-Level Variables: Diffusion Patterns, Settings, and Contexts. Level of Analysis and Inference: Interpreting Ecological Correlations. Inferences when Within-System Relationships are Similar. Inferences when Within-System Relationships Differ Systematically. Conclusion.
Differing Relationships
It should be clear from the preceding chapter that whenever the within-system relationships are sufficiently different, identification of the social system will improve explanation. When systemic factors are introduced, the level of analysis is changed. Problems involved in this change constitute the topic of this chapter.
The most simple case requiring a change of level of analysis occurs when a bivariate relationship is different in two or more systems. Bendix and Lipset cite several such cases concerning political behavior of various occupational groups. For example, they report the following:
"Among workers in Germany and Sweden, the better pf'^ and more skilled are more likely to be class-conscious, and vote social democratic or communist, than those who are less well paid and less skilled. In Britain, the United States, and Australia, however, the lower paid and less skilled
47

48 Syslem Level Variables: Changing liu' Level of Analysis
proved better supporters of left parties than do (lie upper strata of the working class."1
We also know that teachers and physicians are further to the right than other professionals in Germany, while they are further to the left in France and Britain. Workers have a lower rate of political participation than people with higher incomes in the United States and Britain, but a higher rate in France. Upwardly mobile persons in the United States tend to be more conservative than those who retain the social position of their fathers, whereas in several European countries mobile persons are less conservative. Alker demonstrates that the correlation between achievement motivation and per capita income is positive in Latin America and negative in the European countries.2 Almond and Verba indicate that the correlation between church attendance and party identification is almost nonexistent in the United States, while it is highly negative in Italy and Mexico.3 Dogan shows that the relationship between land ownership and Communist voting among poor peasants is positive in France but negative in Italy.4 Janowitz et al. demonstrate that the intensity of class-based political cleavages is significantly higher in Great Britain than in the United States.5
In all of these examples "systems differ," But to say that systems differ is to say that some characteristic that distinguishes these systems influences the observed relationships. Whenever identification of particular social systems contributes to explanation, one must ask what it is about these systems that influences the phenomenon being explained. What is it about Germany and Sweden, on the one hand, and Great Britain, the United States, and Australia, on the other, that makes skilled workers behave differently in politics? What is it about the Latin American and European countries that determines the different relationships between achievement motivation and per capita income? What distinguishes the group of countries where the relationship between level of economic development and domestic violence is negative from another group where the relationship is positive?
' Reinhard Bendix and S. M. Lipset, "The Field of Political Sociology." in L- A, Coser, ed.. Potiiicnl Soriohgy. Harper & Row, New York. 1966, p. 32.
'H. R. Alker. "The Comparison of Aggregate Political and Social Data: Potentialities and Problems." Social Science information, 5. 1966.
'G. A. Almond and Sidney Verba, The Civic Culture, Princeton University Press, Princefon, N.I., 1963-
4 Mallei Dogan. "Political Cleavage and Social Stratification in France and Italy," in S. M, Lipsel and Stein Rokkan, eds.. Party Systems ami Voter Alignments: Cross-National Perspectives, Free Press, New York. 1967.
' Morris Janowilz, Klaus Liepcit, and D. R. Segal, "An Approach to the Comparative Analysis of Political Partisanship," an unpublished paper, no date.

Comparative Study and Levels of Analysis
The questions posed above concern the impact of systemic characteristics on the behavior of individuals within those systems. All the questions are based on an assumption that the dependent variable is measured within systems, and therefore systemic factors enter the theory only as independent variables. However, not all studies conducted in several systems are based on this kind of design. Some simple designs based on the number of levels of analysis and the nature of the dependent variable will be discussed.
1. AH variables, dependent and independent, are observed at the same level. Neither aggregation nor analysis of within-system relationships is possible. Survey research usually involves this kind of design: individuals are the only units of analysis- But studies conducted exclusively at the level of countries also share this design. This is true, for example, when democracy measured by judgmental classification is related to the level of economic development.
2. Some variables are observed within systems and some are observed at the level of systems, but analysis is confined exclusively to the cross-systemic level. Thus some of the variables constitute aggregates of individual characteristics, and other variables are observed directly at the level of systems. If the dependent variable is observed directly at the level of systems, research concerns the impact of individual behaviors on the behavior of systems. Huntington's theory of institutionalization requires this kind of design since the theory concerns the impact of political mobilization of individuals on the institutionalization of the political system.8 If the variables observed at the level of systems are the independent variables, research concerns the impact of systems on the behavior of individuals within them. For example, what is the effect of interparty competition on voting turnout7 or the effect of an "authoritarian culture" on the degree of authoritarianism of individuals?8 Even though individuals are observed within systems, their properties, such as age or literacy, are aggregated and treated as system level variables. The resulting findings relate "an average person" within a system or a "part of the population" of a system to some characteristics of the system. Levels of observation are multiple, but the analysis is conducted only at one level.
' S. P. Huntington, "Polilica! Development and Political Decay," World Politics, 17,1965.
'For example. Dawson reports a correlation of circa .60 (Gamma) between inter-party compelition and voting turnout in the Americ.-? states. R. E. Dawson, "Social Development, Party Competition and Policy." in W. N, Chambers and W. D. Bum-ham. cds., The American Parly Systrm.v. Oxford University Press, New York, 1967.
' E. T. Prolhro and I-enov H, Melikian, "The California Public Opinion Scale in an Authoritarian Culture," Public Opinion Quarterly, 17, 1953.

50 System Level Variables: Changing the Level of Analysis
3. Variables are observed at multiple levels, and the analysis is conducted at multiple levels. It is not necessary for any characteristics to be observed directly at the level of systems because system-level analysis may investigate the influence of within-system distributions on individual behavior. Ordinarily, however, some variables will be observed at the system level. If the dependent variable is a phenomenon observed at the level of systems, the research concerns the impact of the patterns of relationships within systems on the behavior of the system. For example, Hoselitz argues that if the income of individuals is related to their prestige, then the system will develop economically." Feierabend shows that if individual expectations exceed individual satisfaction, a system will be unstable.10 If the dependent variable is observed at a within-system level, then the question concerns the impact of the system on the pattern of relationships within it. For example, the presence of a poll tax will tend to increase the relationship between income and voting. This is the typical paradigm of explanation in the "functional" analysis. When a system is in a state that is not an "equilibrium state," some structures become activated to bring the system back to equilibrium. Thus a property of the system-being in a nonequilibrium state-is used to explain a behavior of the elements of the system.
Although all of the above examples were formulated at the level of individuals and countries, these designs can be applied regardless of what the levels of observation and analysis are. For example, the design of the International Studies of Values in Politics involved the study of the values of individual local leaders and of institutional behavior of local communities1 T Individual values were aggregated at the community level within each country. Questions were subsequently asked about the impact of country-level variables, such as degree of autonomy of local governments, Upon the relationship between values of leaders and the "activeness" of the community within each country- Thus the study employed triple levels of observation-individuals, communities, and countries-and double levels of analysis-community and country.
In comparative research, we are concerned with studies in which analysis proceeds at multiple levels. Even if the levels of observation are multiple but the levels of analysis are not, such studies will not be considered as "comparative." In other words, we are only concerned with studies in
' B. F. Hoselitz, SodnloKical ^specis of Economic Growth, Free Press, Glencoc, 111., 1960.
"I- K. Feierabend and R. L. Feierabend, "Aggressive Behavior within Politics, 1948-1962: A Cross-National Sludy." Journal of Con/Sirl Resolution, 10, 1966.
" P. E. Jacob, Henry Teune. and T. M. Watts. "Values, Leadership, and Development," Social Science information, 7, 1968.

51
Sysfem-[,evel Variables

which both the patterns of relationships within each system and the role of systemic factors are examined.
System-Level Variables: Diffusion Patterns, Settings, and Contexts
In order to discover what it is about systems that influences the behavior of individuals within them, we must first distinguish among types of characteristics of systems. Locating these types is particularly important since the number of systemic factors associated with differences in the patterns of relationships is always larger than the number of systems that can be observed. If some rules were available to determine at least what type of systemic characteristics operate on the dependent variable, the number of potentially explanatory system-level variables could be reduced. It should be emphasized that our attempt to distinguish among various types of systemic factors is not identical with the attempts to classify, conceptually or empirically, "group properties"'2 or "dimensions of nations."'3 We are concerned only with factors that may potentially influence or be influenced by within-system behaviors, not with properties of systems as potential variables in system- or group-level analyses.
To say that a group of social systems shares a certain characteristic that in turn distinguishes it from some other systems is to specify one of three types of systemic factors, which will be called "diffusion patterns," "settings," and "contexts."
1. Diffusion Pallerns.^ One interpretation of the similarity of political behavior of skilled workers in Anglo-Saxon countries may be that this behavior is a result of the diffusion of a cultural pattern. The relationship between occupation and political-attitudes is not based on independent events-neither the social nor the political system of the Anglo-Saxon countries influences the skilled workers to be conservative and the unskilled workers to be leftist. In this interpretation, rather, the relationship is a result of historical learning.
This problem, known in anthropology as "Gallon's problem," has been
"P. F. Lazersfeld and Morris Roscnberg. The Language of Social Research. Free l*ress. Glencoe. 111., 1955. Seclion IV.
"*R- J. Rumniel. "The Dimensionality of Nations Project," in R. L. Menilf and Stein Rokkan. eds.. Comparing Nafinns. Yale University Press, New Haven. Conn., 1963; P. M. Gregg and A. S. Ranks. "Dimensions of Political Systems: Factor Analysis Of a Cross-Polity Survey." American foUticdl Science Review. 59, 1965.
"We [hank Professor Raoul Narull for making us aware of this problem.

53
52 System Level Variable.1:: Changing the Level of Analysis
Svneifs-Level Variables

discussed several times by NarolJ.16 Statistically the question is how many independent events can we observe? If the similarity within a group of systems is a result of diffusion, there is only one independent observation, and the number of degrees of freedom is zero. Naroll cites some fascinating examples of such situations:
"Klimek shows that in aboriginal California, patrilinear totemic clans are to be found invariably and exclusively in tribes (of the southeast corner of the state) which also play tunes on flageolets, use carrying frames made of sticks and cords, make oval plate pottery, use a squared muller, and favor twins. . . . Debt slavery was practical only in the northwest corner of the present state . . . [and it is] found invariably and exclusively among the tribes whose women wear flat caps made of overlay twined basketry, whose men wear painted deerskin capes, who cook in low cooking baskets, who use pipes inlaid with haliotis and who levy a fine for adultery."18
Several solutions designed to determine whether a pattern of relationships is a result of diffusion or "functional" interdependence were proposed by Naroll.17 All of these solutions, however, are based on the assumption that geographical proximity determines communication among cultures. It is doubtful whether this assumption is as useful in the study of modern societies as it is with regard to primitive cultures.
Problems of determining whether what is observed is a diffusion pattern or a functional relationship are frequent and important to other social sciences as well as anthropology. A classical controversy of this nature concerns the meaning of the Weberian hypothesis relating Protestant values to capitalist orientations. Is it a "functional" proposition that a person who is a Protestant wilt be more likely than a Catholic to be an entrepreneur or that Protestant countries are more likely than Catholic countries to de-
" Raoul Naroll, "Gallon's Problem- The I.ogic of Cross-Cultural Analysis," Social Research, 32, 1965.
" Ibid.. pages 434-35.
"For the specific solutions see Raou! Naroll, "Two Solutions to Gallon's Problem." Philosophy of Sc-iciirf. 28. 1961; Raoul Naroll and R. G. D'Andrade, "Two Further Solutions lo Gallon's Problem." American Aislhropoiogist, 65, 1963; and Raoul Naroll, "A Fiflh SoliHion lo Gallon's Problem." American Anthropologist, 66, 1964.

velop economically?'" Or is the Weberian hypothesis reaUy a description of a unique historical event that took place once when an expanding system of religious values turned people to earthly preoccupations? Similar questions are raised with respect to evaluations of occupational prestige. Are they a function of the division of labor in a society or of exposure to a foreign system of values that is gradually adopted as societies go through the process of functional differentiation?19 Do Indonesian high school students evaluate occupations in the same way as Americans because their social structure resembles that of the United States or because they are learning a system of values broadcast by the United States Information Agency? Such questions become even more acute in studies of social change. Does change take place because transformations took place within a country or because the country was exposed to some values and behaviors originating from an alien source? Is economic development a function of internal changes or external exposure? How many times did economic development take place spontaneously-whenever a country passed some threshold of structural changes or only once, when the first pattern was formed?20
No general solutions to these problems are readily available. If we had a chance to observe some social systems that were not exposed to external communication, the impact of diffusion could be assessed. But, as Levi-Strauss convinces us, no primitive culture can resist exposure to the "modern" world.21 When contact is established, only one civilization survives. Precise statistical controls could compare amounts of foreign contacts with internal transformations. But whether we wilt be ever able to determine whether "a society changed" or "a society was absorbed" remains doubtful.
2. Settings. A second type of system properties consists of characteristics that are neither diffusional patterns nor aggregates of observations.
"1 H. H. Andersen and G. L. Anderson, "Cultural Reactions to Conflict: A Sludy of Adolescent Children in Seven Countries," in G. M. Gilbert, ed.. Psychological Approaches to fnter^roup and hilfrnaliofnii Under.-;! finding, University of Texas, Austin, 1956. The authors compared Protestant and Catholic children with regard to concern over money and showed that Protestants are indeed more concerned about rnoney than Catholics. But whether this is a test of Weber's (heory is questionable.
"Zygmunt Bauman, "Social Concomitanis of Economic Development," a paper presented at the UNESCO Conference on Social Prerequisite-^o Economic Growth, Kyrenia, Cyprus. April, 1963.
For a view arguing that development took place spontaneously and as a result of indigenous conditions only once and that subsequent developments in other countries were a result of adaptation of foreign patterns, see W. Kula, Probfemy i Melody rlisiorit Gospoddrcn'j, Panstwowe Wydawniciwo Naiikowe.Warsaw, 1964.
" C. Levi-Slrauss, Tristes Tropuim-s. Atheneum Publishers, New York, 1969.

54 System Level Variables: Chaii^in^ the Level of Analysis
These characteristics cannot be observed at the level of individuals. They correspond to what Lazersfeld and Rosenbcrg call "global" characteristics 22 although they are treated here in broader terms than Cattell's "syntality" variables.23 Settings constitute characteristics to which all individuals within a system are, at least potentially, exposed. Settings may be (1) historical, (2) institutional, (3) external, (4) behavioral, and (5) physical.
All individuals within a social system may be affected by the past history ot that system. For example, it is important for the present political behavior of individuals whether universal suffrage was extended before or after education became universal.24 We find that the number of years a country has been independent, the number of years the same constitution has been in operation, or the average length of time a chief executive or a party has been in office may influence the present behavior of individuals within a system.25 Such historical factors may not only directly affect the behavior of individuals, but may also influence properties of the system that in turn affect individual behavior. A particularly interesting example of a quantitative application of a historical-setting variable is presented by Allardt, who shows that the number of persons who were killed in a Finnish commune during the civil war of 1918-21 is related to votes for the Communist party in that commune today.28
Institutional-setting factors are used extensively In political science. An old controversy between political sociologists and political scientists concerned the question of the impact of institutional factors on individual political behavior. Sociologists, naturally, tend to seek socioeconomic determinants of political behavior, whereas political scientists tend to look at institutional characteristics such as constitutions. The focus of this controversy centered on the relative importance of socioeconomic and institutional factors in influencing party systems and voting behavior-27 Rae*s study of the political consequences of electoral laws is an empirical attempt
** Lazersfeld and Rosenberg, op. cH.. p. 287.
* R. B. Callell. "Types of Group Characteristics." in Lazersfeld and Rosenberg, op. cil.
"Reinhard Bendix and Stein Rokkan, "The Extension of National Citizenship to the Lower Classes: A Comparative Perspective," a paper presented to the Fifth World Congress of Sociology, Washington. 1962.
"For example, D. A. Rustow. A Wor\d of Nations. The Brookings Institution, Washington. D.C.. 1967.
"Erik Ailardt, "Patterns of Class Conflict and Working Class Consciousness in Finnish Politics." Publications of the Institute of Sociology, University of Helsinki, No. 30.1964.
" For example, S. M. Lipset. "Parly Systems and Representation of Social Groups," Institute of Industrial Relations. University of California, Berkeley, 1961.

Sysiem-LL'vet Variables 55
to determine the importance of institutional-setting factors.2" Rae finds that, although the effect of the electoral laws on distribution of seats is only marginal, these effects are nonetheless sufficient to produce important political consequences. Analyses of the influence of institutions on individual behavior are numerous, but not unambiguous. All Americans have only one president, at least at one point of time, and to this extent the properties of this office and of the person who occupies it may be important in explaining individual behavior. The degree of centralization of the educational system, the extent of economic planning, the degree of autonomy of local governments-all such factors may be important as determinants of the behavior of individuals within a system. Institutional-setting variables are deceptively easy to assess, and this often leads to misleading inferences from institutions to behaviors, either of systems or individuals within them. For example, the number of parties has been frequently used to indicate interparty competition, democracy, opposition, and participation. However, it is still not clear what effect, if any, the number of parties actually has on the behavior of individuals.
External relations of a system may influence the behavior of individuals within it and vice versa. Several studies attempt to relate internal and external conflict. It seems that internal sociopsychological conflict but not political conflict is related to the external conflict of a system.20
The behavior of a system or any of its subsystems may have an impact on or be influenced by the behavior of individuals within it. Upset argues that if the political system is effective, it will gain legitimacy in the eyes of its members.30 Interparty competition has a different effect on voting participation in different systems.31 Conversely, behavior of individuals may influence the behavior of the system. Huntington's theory perceives institutional ization as a function of relationships among individual behavior within a system. Marx's theories systematically formulate the influence of
" Douglas Rae. The Polilicai Consequences of Electoral I.aws. Yale University Press. New Haven. Conn.. 1967.
"R. J. Rummel, "Testing Some Possible Predictors of Conflict Within and Between Nations." Peace Rfsriiri-h Society. Papers. 1. 1964; Michael Haas, "Social Change and National Aggressiveness. 1900-1960." in J. D. Singer, ed.. QuanlUalive International Politics. Free Press. New York. 1967.
*° S. M. Lipset, Political Man, Doublcday. New Ybrk, 1960. ' American findings show lhat when the Agree of interparty competition is high, "lore people turn out to vole. Like many other findings from American research. this one seems lo be of a general nature. But Allardt indicates that in Finland more people vote in those communes in which one of the parties is snfcly dominant, unless 'I is the Social Dcmncr.itic Party. See F.rik Allardt and PcrUi Pesoncn. "Cleavages in Finnish Politics," in S. M. Lipset and Stein Rokkan, cds.. Puny Systemx anil Voter Alignment.';, op. cil.

56 Sysfem Level Vfiriahh's: Changing ifie Level of Analysis
interrelations of individuals on the change of a system. But in general such theories are scarce.
The last type of setting variables consists of properties of a physical nature. These variables may concern some characteristics of the physical or material environment or some physical characteristics of a society. Physical and material characteristics such as resources, accumulated capital, and the like, are used most frequently in economics and to some extent in anthropology when a culture is described in terms of the influence of physical environment on individuals.
3. Contexts. Within each social system, individuals hold certain attitudes and interact both with each other and with their physical environment. When the characteristics of individuals-whether predispositional, behavioral, or relational-are aggregated, the social system of which they are members acquires a parameter. Context factors constitute aggregates of individual characteristics. A useful distinction among context factors is that which Cattell calls "structural" and "population" variables.32 Structural contexts are aggregates of relational properties; population contexts are aggregates of individual properties.
Structural contexts are system-level variables generalized from individual characteristics in which "a reference is needed either to other members of the unit or the unit as a whole."^ Although the question of whether structural variables can be reduced to individual attributes is controversial, and in fact structural variables are rarely formed by simple aggregation when systems are large, one can view structural contexts such as division of labor, class structure, income inequality, and communication flow as aggregates of individual relations within a system. To say that the division of labor within a system is high is to aggregate observations about any two persons sharing an occupation. To characterize income inequality is to generalize the distance between incomes of pairs of persons. To state that a system has a dense communication network is to generalize a matrix describing individual interactions. Although these structural contexts can be observed and measured directly, if structure is viewed as a generalization of relations among individuals within a system, it is in principle observable at the within-system level. Therefore they constitute a context rather than a setting.
Population contexts are intuitively clear. They constitute aggregates of individual characteristics, whether they are predispositional or behavioral.
"Cattell, "Types of Group Characleristics," op. cit.
" P. L. Kendali and P. F. Lazersfeld. "The Relation between Individual and Group Charade rislics in Tin' American Sivltlier" in Lazersfeld and Roscnberg, op. cil., p. 293,

57
Level of Analysis and Inference

For example, Turks are on the average more achievement-oriented than Iranians; Frenchmen are less frequently identified with a political party than Americans; Syrian students are more authoritarian than American students. All of these observations are expressed either as means, based on some units of measurement assumed to be common across systems, or as proportions of populations in which an individual is the unit of measurement and the attributes are dichotomous. The measurement of contextual variables presents several serious problems that will be discussed in Part Two.
For system description or analytical studies conducted exclusively at the system level, the importance of contextual variables is evident- But since we are concerned here with system-level factors as determinants or consequences of within-system behavior, their role may be much more limited. In comparative studies it is not sufficient to characterize the average au-thoritarianism of Syrian students, the party identification of American voters, or the achievement motivation of Turks. Nor is it sufficient to identify the extent to which an average Indian is poorer or richer than other Indians or the extent to which an average Brazilian shares his occupation with other Brazilians. If a system attribute is to be treated as a system-level variable in comparative terms, it must be demonstrated that some characteristic of the distribution of the individual attributes influences individual behavior within the system. Does the fact that a system has a certain property influence the individuals within it? Although over 90 percent of the population of Ghana is black, at any one point of time this does not affect the color of the skin of individual Ghanians. Only 50 percent of the population of Brazil is literate. If the percentage of literates is to be treated as a system-level variable rather than just as an aggregate descriptor, however, it must be demonstrated empirically that it affects some attribute of individual Brazilians. It is not sufficient for comparative purposes to state the aggregate parameter of a system. It is necessary to treat it as a potential determinant of behaviors at a different level of analysis.
Level of Analysis and Inference:
Interpreting Ecological Correlations
Comparative studies involve a population derived from "natural" groupings of individuals such as societies, economies, polities, or cultures. This population is sampled in a two-step fashion: systems are selected first and individuals or other units within them next. Relationships among variables can be analyzed within each system. But when individual characteristics are aggregated, these relations are also observable at the level of systems. In other words, for any set of variables measured within each system, three

58 System Level Variables: Changing the Level of Analysis
types of predictions can be made: (1) Individual values of the dependent variable can be predicted from the individual values of llic independent variables on the basis of regressions within particular systems ("within-systems regression"). (2) System means of the dependent variable can be predicted from system means of (lie independent variables on the basis of regression across systems ("among-systcm regression"). (3) Individual values of the dependent variable can be predicted from the individual values of the independent variables on the basis of regression common lo the entire population of individuals, regardless of the social system involved ("total regression").
Although many regression analyses in the social sciences are not guided by a theory, any prediction of a functional dependence of one variable on other variables should be derived from a theory. It is not sufficient to hypothesize that "X is related to Y." A "relationship" can almost never be assessed in the absence of a prediction. A relationship, as measured by correlation coefficients, has meaning only in terms of the fit of a prediction that has a specific functional form. With regard to interval data, the most frequent assumption is that the function describing the dependence of two variables is linear, that is, that it generates a straight line. A linear prediction states that any change of the independent variable by a fixed interval is accompanied by a constant change in the dependent variable. If the independent variable changes by one unit, the dependent variable is different by b units. Coefficient b, usually called the "regression coefficient," describes, therefore, the magnitude of differences in the dependent variable corresponding to a difference of one unit in the values of the independent variable. "Regression intercept," usually symbolized by a, shows the value of the dependent variable when the independent variables equal zero. Thus in the case of two variables, the regression equation generating a straight line has the following general form:
v'i = a 4- bXf,
where i varies along units of observation,
For reasons discussed in the Introduction, however, we cannot expect that this prediction will be deterministic. Even a larger number of independent variables will rarely predict every single value of the dependent variable. For each value of the independent variable, X, we expect to find some (constant) amount of variation of the dependent variable, Y. The larger the amount of variance of the Y's about their least-square linear regression on X. the worse i-s the (it of prediction and the lower is the linear relationship between X and Y. Rut even if the variance of Y's around their linear regression on X is equal to the total variance of Vs about their mean,

Level of A naiysis and Inference

that is, even when & equals zero and the best-fitting linear regression coincides with the mean, one cannot conclude that "there is no relationship" between X and Y. A simple quadratic function may still provide an ideal fit, and thus there may be a perfect curvilinear relationship between the two variables. A form of prediction is logically primary to its fit, or relationship.34 The amount of variance about the linear least-square equation is measured by product-moment correlation coefficient. The square of this coefficient tells us what percentage of the total variance of the dependent variable is predicted by its linear regression on tile independent variables. Thus correlation is a measure of fit of prediction. Underlying the subsequent discussion will be the assumption that the variances of both the dependent and the independent variables do not vary greatly fromi system to system and, thus, that differences in the slopes of within-system regressions are proportional to differences of their fit. A relationship will mean a fit of a linear least-square equation relative to constant variances of both variables.
The total relationship between two variables, each observed within a number of systems, equals a sum of the relationships within systems and the relationship of system means. In other words, the fit of the total regression can be expressed as a sum, with appropriate weights, of the fits of regressions within particular systems and of the regression of means. More precisely, it is the total covariance that is equal to a weighted sum of wittnn-and between-systems covariances- Covariance is the average product of the simultaneous deviations of two variables from their respective means, or, in other terms, it is the product of the correlation between two variables and their respective standard deviations.
Since in comparative research all three types of regression can be analyzed, it is possible that three different relationships will be found. The relationship between race and illiteracy may be nonexistent within each American state, but it may be highly positive when the percentages of Negroes and of illiterates are examined in each state or region. The relationship between industrialization and Communist voting may be found to be negative at the level of countries but positive when administrative districts within countries are used as units of analysis. In general, what inferences can be made on the basis of comparing the within-system relationships, the relationship of system means, and the tolal relationship?
Much has been said about this problem, although in a somewhat different context. The problem was identified in Robinson's article as "ecological
** A reader who needs further explanation of repression should consult a basic statistics textbook, such as H. M- Bhilock. Sm-ial Slalisiics. McGraw-Hill. New York, I960.

fau ^ysiem Level ytinaon's: ^nangm^ 1111' i^evvi UJ ^inuiysts
correlation."-'15 He pointed out that inferences about individual-level relationships drawn from relationships between aggregated parameters may be fallacious. The force of this warning lias been weakened, both by Goodman's reformulation of the problem as one of comparing regressions30 and by the increasing interest in theoretical interpretations of ecological correlations. The problem is not to list the "fallacies" involved in making cross-level inferences, but to interpret theoretically the observed differences between individual and ecplogical correlations. As Duncan, Cuzzort, and Duncan argued, "correlations in which the units of observation are areal units are by no means always computed merely as an inferior substitute for the theoretically preferable individual correlations. . . .":!7 Blalock demonstrated that a change in the level of analysis involves a change in the system of variables operating on the dependent phenomenon. As he has stated, "In shifting from one unit of analysis to another we are very likely to affect the manner in which outside and possibly disturbing influences are operating on the dependent and independent variables under consideration-"311 Most of Blalock's subsequent discussion however, deals with artificially created groupings, whereas our focus is on the naturally formed units.
Inferences when Within-System Relationships are Similar
Decisions concerning the proper level of analysis and the type of system-level factors that should be considered depend upon the difference between the within-systems regressions, the regression of means, and the total regression for the pooled population. In the clearest case, within-system regressions would all have the same slope and would share it with the total regression.
In this situation the relationships are the same within particular systems, and the prediction based on the within-system regressions does not differ from the prediction based on the regression of the means. Systemic factors clearly do not have to be taken into consideration since the form and the fit of predictions is the same regardless of social system.
For example, bt us assume that in several countries the extent of inter-party competition in an electoral district is positively related to voting
^W. S. Robinson. "Ecological Corretalions and the Behavior of Individuals," American Sociological Review. 15, 1950.
" L. A. Goodman. "Ecological Regressions and Behavior of Individuals," American Sociological Review. 18, 1953-
"0, D. Duncan. R. P. Cuzzort. and B. Duncan. Sfalixlical Geography, Free Press, New York, 1961. p. 27.
" H. M. Blalock, Causal Inferrnres hi Noncxperiinentat Research, University of North Carolina, Chapel Hill, 1964. p. 98.

infcrenci'x w/icii wnism-^ysinn KeUiSionslnps are Similar 61
turnout. Wherever the competition is high, voting turnout tends to be high. If this relationship is similar in all countries under study and if a similar relationship is discovered when the extent of competition and turnout are aggregated for each system, all three regressions will have a similar slope. In this situation a single general statement can be formulated, according to which these two variables are positively related. There is no need to concern ourselves with the problems of ecological inferences.

Figure I

In other situations, however, either the within-system regressions-the same in particular systems-differ from the regression based on the means or the within-system regressions are not the same. An interpretation of these differences will be helpful in elucidating the nature of the systemic factors influencing the dependent variable. We will discuss later a few situations in which the difference between the within- and the between-system regressions is theoretically interpretable, rather than develop general rules for making inferences. The discussion will be restricted to two variable relationships since they are easier to understand intuitively. The argument is applicable, however, to multivariate relationships in which the dependent variable is measured as a deviation from its regression on a set of independent variables already considered.
The classical situation often discussed in this context is one in which the Within-system regressions are the same for all systems and are approximately equal to zero, but the slope of the regression of system means is different from zero. This is the example of the relationship between being black and being illiterate: there is no relationship within American states, but there *s a strong positive relationship a( the level of states and even a stronger

62 Syslem Level Variubfi-'s: Changing the Level of Analysis
one at the level of regions. For the sake of illustration we will assume interval measurement at all levels.
Thus, although a person who is black is not more likely to be less literate than a person who is white, states that have a high percentage of blacks also
have a high percentage of illiterates.

Figure 2
It can be generally shown that whenever within-systems relationships are low, the relationship observed at the level of systems will lend to be larger than the total relationship. This is particularly true in the extreme case when the relationship within systems is zero or nearly zero, as in the example discussed above. A simple mathematical formulation is useful in illustrating this difference. According to the covariance theorem discussed above,3"

w==l
where r is the correlation between two variables, o\r o-y are the standard deviations, t denotes observations made for the total population, a denotes observations made at the level of groups, w denotes observations made within groups, and k is the number of groups.
In other words, the total covariance (left hand of the equation) equals the sum of the among-group and within-group covariances. Thus the among-
"See also H. R. Alker, Mathematics anil Politics, Macininan, New York, 1965, for a fuller discussion of this Iheorem-

Inferemes when Within-System Relationships are Similar 63
group covariance equals the difference between the total covariance and the sum of within group covariances:
k
''xr.s O"J." o'r," == rrr,f Tr.* o'r.i - 2 rxr.w crx.w o^r,^">o w= 1
We are discussing here the situation when the within-group correlations are equal to zero (or nearly zero). In this case, the last term equals zero, and
rxvA o"J.o o'r,o Therefore,

's.t o'r,* ;--~-^r.i-
X.a CTr.a
""oo*= ^--~-r^
<^X,a 0-v.a

But whenever the within-group variances are different from zero, the total variance of a variable will be larger than the variance of system means. Thus the numerator above will be larger than the denominator, and consequently the observed among-group correlation, r^y. n, will be larger than the true correlation for the entire population, r.vr. i.
When this type of situation is observed, it seems clear that a third variable operates on both variables observed within systems-a variable such as the level of industrialization of a state. In such situations "setting" variables are more likely to provide explanation at the system level than "context" variables. Although this two-variable relationship could be controlled for another individual-level variable, such as family income, the difference between within-system and among-system regressions will remain- The among-system relationship is spurious since a system-level variable influences both of the means. In this sense the ecological correlation between the two variables is not "true."
A different situation is encountered when the among-sysfem correlation equals zero white the within-system correlations are different from zero, or when the within-system regressions (and according to our assumptions, correlations) have different sign than the among-syslem regression.
For example, imagine that we are studying the relationship between exposure to urban life (years in the city) and attitudes of modernity. In each system the relationship is positive, but the mean of modernity is the same in all countries regardless of the mean of urbanization.
In this situation, the within-systems regression coefficients, h, are the. same, ^d the means of the dependent variable are also the same. Thus for two systems the difference between the means of the independent variable, X,^ (where ". varies across systems) is a function of the difference between the intercepts of the regressions ii,r.

64 System Level Variables: Changing the Level of A naiysis
02- <h
X. - X; =
Since the intercept, Ow. of the regression line indicates the value of the dependent variable, when the independent variable equals zero it is clear that the difference between within-country and among-country regressions is a result of the differences in the value of (he dependent variable when the independent variable does not operate. This point may be more easily understood in stochastic terms. In one country the value of the dependent variable, modern attitudes, is higher than in another country before the independent variable, exposure to the city, begins to operate. In one country individuals have "further to go." Although once the independent variable operates, the countries move at the same pace-relationship is the same within all systems-system means of the dependent variable cannot be predicted from the system means of the independent variable, in our example this may mean that before the peasants in one country, say Chile, move to the city, they are more modernity-oriented than, for example,
peasants in India.

Inferences when Wilhin'Sysiem Relationships are Similar 65
of votes for the Communist party among fifteen Western countries. He concluded:
"Thus, there is a strong negative relation between the extent to which societies are industrialized and the strength of communism within the Western world. . . ."10
This finding is generally used as a refutation of Marx's theory. However, there is ample evidence, as Kornhauser himself shows, not only that workers in larger factories tend to vote Communist, but also, that within each country the extent of industrialization of a certain region or administrative unit is positively related to the Communist vote. Thus there are two seemingly contradictory findings: at the level of countries the relationship between industrialization and Communist vote is negative; within countries it is generally positive. How can these findings be reconciled?
Let us revert again to a diagram. We see that the difference among the means of the dependent variable of each system is a function

Figure 3 ' "l *a -.1
i In the example discussed above, the within-system relationships were
positive and the among-system relationship (slope and fit) equaled zero. But even more interesting are those situations in which the within- and the among-systcms regressions have a different sign. For example, Kornhauser found a highly negative (-.76) rank-order correlation between the proportion of male labor force in nonagricultural occupations and the proportion

Figure 4
of the regression intercepts. But since the means of the dependent variable are no longer equal from system to system, this difference is also a function °t the differences between means of the independent variables. Since, according to regression.
"William Kornhauser, The Politics of Mass Society. Free Press, Glcncoe, III., 1959, P. ISO.

the difference between means of the dependent variable of two systems equals

It would be outside the scope of this book to discuss formally the general conditions under which Ihe slopes of the within- and the among-systems regressions have different signs. Let us look at a specific case of two systems. Since in the general case we would be seeking the criterion that would permit us to determine when b^bn < 0, in this specific case one must establish the conditions under which the means of independent and dependent variables of two systems are related differently than the observations within these systems. When the within-systems regressions, assumed thoughout this discussion to be constant, are positive, we are interested in the case in which the difference between the means of the dependent variables has a different sign than the difference between the means of the independent variables. When the slopes of the within-system regressions are negative, it is necessary to find the conditions under which the differences of means have the same signs. In other words, we are interested in the following situation:
when ft, > 0, (X, - X,) (V. - V,) < 0 and - _ _ -^ _
when b^ < 0, (^ - X^ (V, - V,) > 0.
Let us assume that these two systems are ordered in such a way that (X, - X,) > 0.
We can then expect the among-system regression to have a slope signed differently than the within-system regressions if and only if the difference of the intercepts has a sign different from the coefficient of within-system regressions, and the difference between intercepts is larger (in absolute value) than the difference of means multiplied by this regression coefficient. Figure
4 presents the case when the among-system regressions are negative; Figure
5 presents the case when they arc positive. The effects of relaxing either of the criteria can be studied by manipulating the slope of regression and the difference between the means of the independent variables.
In the light of this general discussion, Kornhauser's finding must be attributed to the fact that he considered countries that differ relatively little with regard to the independent variable, but differ more with regard to the regression intercepts. As pointed out earlier, regression intercepts indicate the value of the independent variable, in this case Communist vote, when the independent variable, nonagricullural labor force, equals zero. In other

Inferences when Within-System Relationships are Similar 67

Figure 5

words the ecological correlation observed by Kornhauser is a function of the behavior of persons employed in agriculture, not outside of it. What Kornhauser found is that in the most-industrialized countries, peasants are less likely to vote Communist, while in the less-industrialized countries they are more likely to do so. Thus what we find is that the process of industrialization brings a progressive differentiation of Communist support between an agricultural and a nonagricultural population-a finding not contrary to Marx's theory. Allardt's finding of the two types of Communism in Finland confirms this prediction of the model analyzed above.41
In general, in this situation there is no need to change the level of analysis to system-level variables. The within-system regressions are the same, and the differences in contexts-average modernity of peasants before exposure to city life or their rate of Communist vote in nomndustrialized regions-can be easily adjusted. A difference of initial contexts constitutes the theoretical interpretation of the differences between the within- and among-systems regressions. The system-level or "ecological" interpretation does not merit an independent theoretical interpretation, and is therefore "spurious." But an additional theoretical statement relating the means of the independent variables to regression intercepts can be formulated on the basis of this analysis.
In both of the above situations we concluded that among-system or ecological" regression does not have a meaningful theoretical interpretation

independent of wilh in-system regressions. When regression coefficients within systems equal zero, then differences can be attributed to a system-level variable, most likely of a setting nature, operating at the level of systems. When regression coefficients within systems differed from zero, we concluded that the difference between the within-systems and ecological regressions stems from the differences of (he context. In general the ecological relationship is spurious whenever within-system regressions have the same slope, hence on the basis of (he assumption of similar variances, the same fit. There is no need to change the level of analysis.
Inferences when Within-System Relationships Differ Systematically
Particularly interesting from the point of view of comparative research are situations in which the slope of regression lines within each system is a function of the means of the independent variable in each system. Of several such situations, two will be discussed here.
In the first situation, the within-system observations are well predicted by linear regressions, but the slope of these regressions assumes a different sign within different ranges of the independent variable. An example can

Figure 6

be formulated with countries and regions as levels of analysis. The dependent variable is domestic violence, and the independent variable is the level of economic development. In Africa the level of economic development is low, and the level of violence increases with economic development-In Latin America the level of economic development is higher, and the level

of violence is not related to economic development. Finally, in Western Europe, where the level of economic development is high, violence decreases with the increase of the level of economic development.42 The relationship between means of economic development and violence is curvilinear. Within-system observations fit a linear model only because the range of variation of the independent variable is limited in any particular region. Thus within-system relationships are true only for a limited range of the independent variable, whereas the "true" regression is curvilinear.
A frequent speculation in the social sciences concerns the impact of the social "context" upon the behavior of individuals. Extensive evidence, derived mainly from social psychology, indicates that individuals behave differently when they act alone than when they are members of groups having some specific norms.43 Assume that a number of groups is available, and within each group a linear prediction fits observations. In other words, within each system the dependent variable linearity depends upon the independent variable. However, if the social context operates, the slope of these linear regression lines differs from system to system in such a way that it either systematically increases or decreases when the group mean for the independent variable increases. The within-systems regressions are linear;
among-system regression is again curvilinear and this time monotonic.
AIlardt and Pesonen report, for example, that the relationship between the proportion of Swedish-speaking people and the vote for the Swedish party among the Finnish communes increases as the proportion of Swedes increases.4'1 If the proportion of Swedes is low in a commune-the system in this analysis-then being a Swede is not associated with a greater likelihood of voting for the Swedish party. If, however, the proportion of Swedes is high, they are much more likely to behave as Swedes, that is, vote for this party. This is a context effect, whereby the context operates in an infer-active manner.^ The probability of behaving in a certain way depends upon the proportion of the individuals Af a given class within each system. Again, stochastic language may be helpful in elucidating this kind of an effect.
" Actually, whal we know is that (he correlation between a nonmonelary index "f economic development and domeslic violence is positive (-33) among the underdeveloped countries and negative (-.68) among the developed countries. See H. R. Alker in B. M. Russett et at.. World Handbook of Social and Political Indicators. 'ale University Press, New Haven, Conn., 1964.
13 For a summary of ihis evidence see E. L. Walker and R. W. Heyns, Anatomy of Conformity. Prentice-Hall, Englewood Cliffs, N.J,. 1962.
i* AHardt and Pesonen, op. cil.
"This is a generalization of a model constructed by professor Raymond Boudon to explain the Communist vote in France. We thank Professor [iniidon for several discussions about this class of models.

70 System Level Variables: Changing the Level of Analysis

Figure 7
Coleman, et al., discovered that the rate of diffusion of innovation among persons isolated from their social environment does not depend upon the number of persons who have already accepted this innovation. If, however, these persons are integrated into their social context, the rate of adoption does depend upon the level of acceptance at a previous time.16 In these terms, the stooges in Asch's conformity experiment can be viewed as a social context- The probability of the /cth person behaving in a certain way is a function of the (k - 1) number of persons who have behaved this way previously.
The difference of the form between the linear within-system regressions and the nonlinear among-system regression can be explained with a simple model. There are several groups, administrative units, regions, or countries. We are interested in a behavior of the individuals within these systems, for example, the proportion of the vote for a given party-Swedish party in Finland, Communist party in France, Democratic party in the United States. This proportion is explained in terms of the behavior of a particular subgroup within the population of these kystems-Swedes in the Finnish communes, workers in the French ciepartemenfs, or ethnic population in the American states. In the simplest form, a model without context would express the proportion of the vote for a given party as a sum of the vote of
" Reported in J. S. Coleman, Introduction 10 Mathematical Sociology. Free Press, New York, 1964.

groups other than those considered (assumed here to be constant only for the sake of simplicity) and of the size of the relevant group multiplied by its propensity to vote for a given party: '
V. ^ a + bX,,
where Y^ is the proportion of the vote for the party in system i, a is the proportion of this vote contributed by persons other than those under consideration, constant across systems, Xi is the proportion of the persons under consideration in system (", and b is the propensity of these persons to vote for the given party, constant across systems.
To use a specific example, let the party be the Swedish party in Finland. Then a is the proportion of the vote for the Swedish party contributed by non-Swedes, X^ is the proportion of Swedes in a commune, b is the propensity of the Swedes to vote for the Swedish party. At this moment, this propensity is considered to be constant throughout Finland: in all communes the probability of a Swede voting for the Swedish party is the same.
How can the context be introduced into this model? If the context within which the Swedes vote is an element of the explanation, the propensity to vote for the Swedish party cannot be constant but must depend in turn upon the proportion of Swedes in each community. Thus, instead of a constant b, we have
b^d-\-cX,.
where hi is the propensity of a Swede to vote for the Swedish party in system /, d is that part of their propensity that does not depend upon the presence of other Swedes, c is the "coefficient of context," and Xi is again the proportion of Swedes in system i. Substituting the new values of b/s into the original equation, we have
V, = cX^ + dX, + a,
where V, is the proportion of persons who behaved in a given way in system /", Xi is the proportion of persons under consideration in the population of system /', c is the coefficient of context, which indicates the importance of social context in bringing about behavior V, d is the propensity of the individuals of this group to behave in the specified way, regardless of the social context, and a is the contribution of persons other than those whose behavior is used to explain Y.
It is apparent that the function generated by the equation of the context model is curvilinear in the manner portrayed in Figure 7. Thus when the slopes of linear within-systems regressions change systematically with the mean of the independent variable and, hence, the among-system or "ecological" regression is curvilinear* one may expect that the social context

72 System Level Varitihlc's: Changing {he Level of Analysis
influences the behavior under consideration. In this situation, both the within-system and the among-system regressions require independent theoretical interpretation in terms of the coefficients d and c. denned above. Neither relationship is spurious, and no "ecological fallacies" are impending. Two theoretical statements are necessary to explain individual behavior; one that states that an individual behaves in the way Y with probability d when he is an isolated member of group X, and another that states that this probability is increased by c when the social context operates. A system-level variable-the context of individual behavior- must be introduced into the analysis.
Conclusion
We have distinguished above some simplified situations that can be used as analytical models in interpreting the operation of system-level variables on within-system relationships. What is necessary now, even if only in an introductory manner, is to (1) distinguish between "spurious" and "true" correlations when relationships are observed at different levels of analysis and (2) distinguish the effects of the variables observable only at the level of systems (diffusion patternsiand settings) from the variables aggregated from within-system observations (contexts).
The currently available theoretical and statistical techniques allow us to distinguish spurious correlations between two variables, X and Y, measured as deviations from their respective regressions on a third variable Z, only when Z is measured at the same level as X and Y. If an assumption can be made concerning the direction of influence, a prediction that the correlation between two variables measured as deviations from their regressions on a third variable is zero can be tested against a set of data. What we need in comparative research, however, are statistical techniques that would allow the control variable to be measured at a level different from the two variables that are tested. The situations discussed above cast only partial light on the problem of interlevel spuriousness. The following criteria can be introduced:
1. If the within-system regressions are the same in all systems and the total regression is also the same, then the relationship between the variables X and Y may at the most be spurious at the individual level. Since our goal is to develop general theories, it seems useful to identify as spurious only within-system relationships that change uniformly (most often disappear) in all systems when a third within-system variable is introduced. Thus when the introduction of a third variable changes the relationship between two

variables only in some systems, the gain of prediction does not justify the loss of generality.
2. If within-system regressions do not differ from zero in all systems, but the total regression does differ from zero, the ecological correlation is spurious. In other words the relationship between X and Y (race and illiteracy) is "true" at the individual level and "spurious" at the system level. Controlling a relationship between the system means for a system-level variable would reduce the among-system relationship to zero-the true value observed at the individual level.
3. If within-system regressions are the same and differ from zero in all systems but the total regression does not differ from zero, the ecological correlation is spurious. An adjustment of the values of the intercepts of within-system regressions would again adjust the among-system regression so that its slope would be the same as the within-system regressions.
4. If the regression coefficients differ from system to system, the among-system regression is equally true as the within-system regressions, since both regressions require theoretical interpretation. These interpretations depend upon the nature of the factors operating at the level of systems.
We have assumed in this discussion that within-system relationships are linear or, in other words, that Ihere are no interaction effects at the individual level. The relationship between education and achievement motivation, for example, was assumed to be independent of the level of education, that is, an increase in achievement motivation associated with an increase of education is constant. This assumption is often unwarranted. Interactions may occur in all or in some systems. But the discussion of individual-level interactions would exceed the limits of the present problem.
We argued in Ihe introduction to this chapter that whenever the within-system relationships are not the same. the analysis should be shifted to the level of systems. Subsequently an additional criterion was introduced consisting of the difference between the within-system regressions and the regression among means. In light of the discussion, however, the original criterion does not require modification because whenever relationships within systems are the same the relationship among means differing from the within-system relationships is considered spurious.

CHAPTER F|OUR
Formulating Theories Across Systems
Introduction: A Restatement. Formulating General Statements. Comparative Explanation in the Social Sciences: A Conclusion.
Introduction: A Restatement
The role of theory in the social sciences and some of the assumptions and implications of the accepted model of theory in the context of comparative research have been discussed in the preceding chapters. In this chapter we shall construct a procedure for formulating general statements. In order to set the context within which this procedure is justified some of the definitions and assumptions will be restated.
"Comparative" studies were defined as those in which the influence of larger systems upon the characteristics of units within them is examined at some stage of analysis. Consequently comparative studies involve at least two levels of analysis. Tn this sense not all of the studies conducted across systems or nations are comparative, but all studies that are comparative are cross-systemic. If national social, political, or economic systems constitute one of the levels of analysis, the study is a cross-national comparative study. If, however, the aniysis is conducted exclusively at the level of nations, then according to this definition it is not comparative.
A theory explains and predicts social phenomena. The explanations should be accurate, genera!, parsimonious, and causal. The implications of this role of theory for comparative studies are the following: (1) General theoretical statements, valid regardless of the social systems involved, should be sought. (2) Whenever they can be assessed validiy across systems, general concepts should be used. (3) Whenever necessary, the influence of

Introduction: A Restatement 75
system-level factors on a class of phenomena should be incorporated into the explanation. General statements can be formulated across systems if within-system relationships do not differ-if systems do not contribute to explanation. Whenever systems differ, some factor operating at the system level is influencing the within-systems relationships.
To illustrate the logic of this argument, let us return to M. Rouget, the French worker, age 24, employed in a large factory, He votes for a party of the Left, and we want to understand why. We explained his behavior in terms of some general statements found to be true of Frenchmen. But if several studies have confirmed that in all systems in which the option of voting for the Left is present, young workers employed in large factories are likely to vote for a leftist party with the probability of .60 to .70, the vote of M. Rouget can now be explained in exactly the same way as the behavior of Senor Martinez, a Chilean, or of a young worker in Norway. Regardless of the social system in which the behavior of individuals occur, the same theory is valid: young workers employed in large factories tend to vote Communist. But if an additional explanatory factor is considered, this theory is no longer equally true. When the sex of a French or a Chilean worker is considered, the explanation of the vote becomes more complete. Males in France and Chile are more likely to vote Communist than females. The introduction of this explanatory factor increases the probability of the Communist vote of young workers employed in large factories to .80. But in Norway voting for the Left is independent of sex. At this stage the explanation must include a statement of the relevant characteristics of France and Chile, on the one hand, and Norway, on the other. If this explanation is to be theoretical, then it is not sufficient to state that "in France males vote Left more often than females, while in Norway there are no differences between sexes in voting."
An explanatory statement must be logically open to extension to other case's. Instead of specifying the names of social systems, therefore, a variable operating at the level of systems must be added to the explanation. For example, it may be that wherever the role of established religious organizations is strong, there will be a difference in the voting behavior of men and women. The explanation of the vote of M. Rouget would then assume the following form:
1. M. Rouget is a young male worker employed in a large factory in a social system in which the church plays an important role, and
2. young workers employed in large factories tend to vote I.eft with the probability of .60 to .70, and in those systems in which the role of

/& formulating l heones Across Systems formulating Comparative theories: A I'roceuure if

church is strong, men vote Left more often than women; therefore, it is highly likely (probability of .80) that
3. M. Rouget votes for a party of the Left.
The premises of the explanation of the vote of Mr. Janson, a Norwegian, would be the same. but the second premise concerning the systemic conditions under which the behavior of men differs from that of women, would not provide any gain in prediction. Thus the behavior of both individuals, a Frenchman and a Norwegian, would be explained in terms of the same theory, but the explanation of the behavior of a Frenchman would be more complete-the probability that the conclusion is true would be higher.
An explanation of a specific property of an individual or a social unit calls for a confirmed general statement pertaining to this class of properties. In this sense accumulation of knowledge is a sequence of confirmations and/or modifications of theories. Theory-building and theory-testing are aspects of the same process. Theories not tested by any reference to empirical evidence have a zero degree of confirmation.1 As predictions drawn from a theory are tested against observable instances, the theoi^ gains confirmation. If the predictions do not hold, a theory may be modified or even rejected. Theories are not falsified by specific tests of the derivable hypotheses, however, but as a result of their overall usefulness for predicting a given class of events. The process of confirming and/or modifying theories will be referred to as the "formulation" of theories. The formulation of theories is conceived of as an interplay between constructing and modifying deductive systems and testing hypotheses through empirical research. This concept has an advantage over the more familiar concepts of theory-building and theory-testing since it avoids connotations of an inductive approach to theory-building and does not equate tests of specific hypotheses with evaluations of general theories.
Formulating Comparative Theories: A Procedure
Let us denote all factors analyzed within systems as either D (a dependent variable) or /, (independent variables), where ;' varies from 1 to k. Let us denote the factors operating at the level of systems as S,. where i varies from k 4- 1 to p. Finally, let us define the en-or of prediction, the
' For the concept of confirmation see C. G. Hempel. Aspects of Scientific Explanation and Other Emilys in the Philosophy ol Science, Free Press, New York, 1965, particularly the essay on "Confirmation. Induction, and Rational Belief," pp. 1-81 and pp. 381-94; and Rudolf Carnap, 'The Aim of Inductive Logic," in R. Nagel, P. Suppes, and A. Tarski, eds.. Logic Methodology, dint Philosophy of Science, Stanford University Press, Stanford, Calif., 1962.

unexplained part of variation of the dependent phenomenon, as N*, where / varies across the social systems, from 1 to n, and indicates the degree of completeness of explanation in each system. We shall assume that there are no systematic errors operating at the level of systems. The dependent variable, D, will be a function of the two sets of independent variables, defined above.
?>=/(/,,5i) +/V*.
The form of this function is defined by a particular theory.
If measurement is free of error, that part of the variation of the dependent variable unexplained by the theory can be treated as a function of the factors that influence the dependeni phenomenon but have not been included in the theory. Let us denote this set of factors relative to each social system as N, (without the asterisk). Then the error of prediction is a function of variables not included in the theory:
/v^=/(7v,).
The dependent variable can now be defined as a function of three sets of variables:2
D=/(/<.St,/V,).
Prior to any explanation, when the dependent variable is only measured within each system, the residual error, A'*J, is the same as the total variation. In other words, all of the independenTvariables^re at this moment subsumed under the names of sysTems. \

The procedure is, fti/sf, to find determinants of the dependent variable that account for its variance without reference to any systemic factors and, second, to develop explanation with system-level factors. This procedure can be thought of in', terms of a multivariate statistical technique that partials the variance of the dependent variable: stepwise multiple regression applied "simultaneously" to different systems; an interaction detector technique analyzing the relative importance of different predictors; or analysis of variance comparing the influence of within-system classifications to classifications based on systems. Which of these approaches is most suitable is a matter of the nature of the data under consideration.
The logic described above best conforms to a stepwise multiple regression conducted separately within each system and with intersystem comparisons
'This was suggested in a somewhat less detailed form by Johan Gallung in "Some Aspects of Comparative Research," folli. 2, 1965.

78 Formulating Theories Across Systems
at each step. 1. The first step in comparative research is the definition and measurement of a dependent variable, D. Let us assume we are formulating a theory of political participation and the dependent variable is "mobilization into politics," defined as psychological and behavioral "membership" in the political system.:1 Political mobilization is measured through various indicators of political awareness and participation, such as knowledge of the political system, interest in politics, voting, membership in a political organization, and attendance at political meetings. Let us assume that we have obtained sets of interval-scale scores of political mobilization describing random samples of individuals within several countries.
The first question is whether systems, in this case nations, are the "same" or "different" on the dependent variable. As was emphasized earlier, if nations do not differ on the dependent variable, the problem of explanation is not a comparative one. System-level factors do not operate on the dependent variable, and therefore explanation can remain at the individual level- All observations can be treated as if they were derived from the same population. In this situation the explanation of the dependent variable can proceed as if it were a one-system stui':y: the samples of respondents can be pooled, and political mobilization can be related to other variables common to and measurable within all countries. We may later discover that a multivariate relationship is curvilinear, and we may then examine whether this curvi linearity does not coalesce with the systems.
Various sets of data will indicate the extent of political mobilization in several countries. For example, if party identification is used as an indicator, we would find that mobilization is high in the United States and Germany (above 70 percent),1 somewhat lower in Japan (62 percent),5 and even lower in France (45 percent).8 Since the available information is highly scattered and derived from different sources, however, this information should be treated merely as an illustration of intersystemic differences. Further discussion will concern only such cases in which the existence of intersystemic differences is assumed.
"The concept of political mobilization used here was defined in Krzysztof Ostrowski and Adam Przeworski. "A Preliminary Inquiry into the Nature of Social Change:
the Case of the Polish Countryside." International Journal of Comparative Sociology. 8.1967.
' Data for the United States are from P. E. Converse and Georges Dupeux, "Po-litlcization of the Electorate in France and in the United States." in I.. A. Coscr, ed., Poliiicu! Sociology. Harper & Row. New York, 1966. and from A. j. Heidenheimer, The Governmcnis of Germany, Crowell. New York, 1966. Dala for Germany are from Heidenheimer, ibid.
'Polls, 1, i966.
'Converse and Dupeux, op. cil.

Fornisilaling Coiftparafive Theories'- ^ Procedure
2. If there is a cross-system difference, the question is what are the factors that influence the dependent variable-why did the difference occur? The presence of a difference, however, is treated, merely as an indication that the problem is comparative.
The question that must be answered is whether a within-system variable, /,, explains the dependent variable, D, in all of the systems. Formally the test is whether the fit of prediction is more or less the same for all systems. We are testing a hypothesis that the correlation between the dependent variable and the first-selected independent variable is the same in all systems. If the statistical technique was analysis of variance, we would ask whether the sum of squares accounted for in each social system is the same.
2.1. If the answer is positive-the two variables are positively, negatively, or not at all related in all countries-we can formulate our first theoretical statement. For example, Converse and Dupeux discovered that, although the rates of party identification differ between France and the United States, in both systems the crucial variable explaining this identification is political socialization through the family.7 To continue our example, let us say that the first statement that holds across systems is that political mobilization is a result of exposure to mass media; in all countries there is a positive correlation between exposure and political mobilization.
Since the independent variable is related in the same way to the dependent variable in all systems, we can apply the procedure that corrects the original scores on the dependent variable. We can ask what the scores on the dependent variable would have been if the dependent variable were the same for all countries. The larger the variations of the independent variable and the stronger the relationship, the more the values of the dependent variable will change as a result of this adjustrnent.
2.2. Let us suppose that the correlations do differ among systems. For example, in some countries persons who own radios are more likely to be politically mobilized than those who do not, while in another group of countries this relationship is negligible. An additional question can now be asked. It can be formulated in two ways-
2.2.1. If one finds that the relationship between political mobilization and the exposure to radio is positive in one country and negligible in another, it is possible to ask whether there is another variable in the latter country that is related to the same dependent variable and can be considered as belonging to the domain of the same general concept. Let us say, for example, that in the country where the original correlation
Converse and Dupeux, op. cil.

80 Formulating Theories Across Systems Formulating Comparative Theories: A Procedure 81

was negligible, exposure to television is related to political mobilization. The theory can now be modified: instead of talking about "exposure to radio" and "exposure to television," we can treat both of them as specific instances of a more general concept, "exposure to mass media." The new, more general statement will now say that "political mobilization is related to exposure to mass media, wherein ownership of radio and television sets indicate exposure to mass media." This kind of statement must be based on an explicit model of measurement that allows this kind of operation on uncorrelated indicators. In this case the level of generality of the concept will be shifted upward, and a statement analogous to one saying that relationships are the same for all systems (2.1.) can be made.
2.2.2. An alternative question can be posed in terms of controlling for some other variable rather than of increasing the generality of variables. Instead of asking whether there are indicators in the two systems that can be generalized into an index correlated with the dependent variable, one can ask whether there is a common variable that. if controlled, "modifies the original correlation so it becomes the same in all countries. Let us assume that urbanization is such a variable in two countries under study. In the first country political mobilization is related to farm size, and in the second country it is not. But in the first country, urbanization is correlated with both variables; in the second country urbanization influences the size of a farm, but not political mobilization, the dependent variable. A diagram will best portray this example:

The relationship between the dependent variable and the first considered independent variable is spurious in the first country since both of these phenomena in turn depend (share most of their covariance) upon urbanization. When urbanization is controlled, the relationship is negligible in both countries. In other cases, controlling for a third variable may bring out a relationship between two variables in the country in which it was not initially present.

2.3. We have found lhat the first independent variable is related in the same way to the dependent variable in all countries. In some cases, this relationship became the same only when either specific variables were generalized to a new concept or a common variable was controlled in all systems. The first theoretical statement can now be formulated. This statement will describe the dependent variable as a function of one independent variable and residual factors.
where
If a difference of relationships cannot be removed either by a reformulation of the independent variable or by controlling for a third variable, however, this independent variable, /i, will not enter the theory at this time. The search should continue until another variable is discovered that does provide a uniform explanation in all systems.
3. Having identified one common explanatory variable, we can now repeat the testing operations- This repetition will involve all steps described under 2., plus some new procedures. The question is whether another independent variable contributes to the explanation of the dependent phenomenon after the one variable has been accepted, that is, after the variance of the dependent variable is reduced in each system by the part that is accounted for by the first variable. For example, if exposure to mass media was found to explain political mobilization in all systems, the question could be asked whether the degree of education further contributes to the explanation of political mobilization. The hypothesis to be tested is that the relationship between the second independent variable, 1^, and the dependent variable, D. is the same in all systems ami that /^ contributes more to the explanation of political mobilization than exposure to mass media alone. First, we ask whether the new variable, /.., education, is related in the same way to the dependent variable, political mobilization, in all systems.
3.1. If the answer is negative, we must ask whether there are any more variables (or degrees of /freedom) available. If the answer is positive, we can retrace the steps described under 3. and continue testing until either the number of variables or degrees of freedom is exhausted. Whenever there are no more variables that provide a similar explanation regardless of the socw! syslcms, she level of analysis must be changed to S variables. This change is described under 5.
3.2. Now (he answer is positive: the new variable is related to D in ^e same way in all systems. The first part of the test-uniformity of

Formulating Theories Across Systems
relationships-has been satisfied. In order to establish that /^ contributes something new to ihc explanation of 0, it must be/shown that /r is not related in the same way to the first independent variable, /i, -whenever this relationship is different from zero. In oilier words, if 1-2 together with /i is to account for more variance of D lhan /i atone, then /^ must be correlated with D, but not with J^^Educalion can provide an independent explanation of political mobilization if and only if it explains a different part of political mobilization than does exposure to mass media-education and mass media must be unrelated.
3.3. If /^ and D are related positively or negatively in ail systems and /oJ is related to /i in all countries, then one of the correlations-r; p or ri.,n-can be considered as spurious. When education is related both to exposure to mass media and to political mobilization, the relationship between exposure to mass media and mobilization is "spurious": the degree of education constitutes an explanation of both the exposure and the mobilization. If an individual is educated, lie is both exposed to ma*^ media and politically mobilized. In this case our original statement, according to which
D=/(/i,/V/), must be replaced by the statement that
D=/(/.,/2,/v,"),
which can be read as "whenever /a is considered, it replaces /i as a predictor of D."
3.4. If the second independent variable contributes independently to the explanation of the dependent phenomenon, the inclusion of /.J will increase the proportion of the variance of D explained by the theory. Let us now say that /i and /:> have either an independent or an interactive effect on D-an effect that is larger than that of either /i or /a alone. For example, if an individual is exposed to mass media and is a member of nonpolilical organizations, he is very likely to be mobilized politically. The dependent variable is a function of two independent variables and the national residuum:
where
4. Regardless whether the first variable, /i, or the second one, /a, or both, contribute to the explanation of the dependent phenomenon, testing

Formulating Comparative Theories: A Procedure 83
can continue until there are either no more variables or no more degrees of freedom left. The same operation, gradually growing in complexity, must be repeated. If the addition of new independent variables continues to bring explanations that do not differ from country to country, the lest is positive and can be continued. When no more variables arc available, the level of analysis must be shifted. The two-step description of the test of independent contribution to the explanation of the dependent phenomenon is discussed here for the sake of logical rather than technical clarity. Most multivariate techniques provide both tests simultaneously by showing the increases in the variance of the dependent variable accounted for with the addition of each new variable, and the partial correlations between the ;th independent variable and the dependent variable when (/-I) variables are controlled for.
The statements resulting from these operations express the dependent variable as a function of the independent variables that are measurable within all systems and that provide a uniform explanation regardless of the systems. These statements will provide the most parsimonious and general explanation. In our example, we have formulated the following explanatory statements; political mobilization depends upon exposure to mass media and membership in nonpolitical organizations. This general form of explanation uniform across systems is the following:
D = /(/" /2.
where
5. Since all the variables that in all systems are related in the same way to the dependent variable have been considered, we now face situations in which the relationship between an additional independent variable and the dependent phenomenon varies from system to system. Let us use as an example the influence of involvement in the market on political mobilization. In one group of countries this relationship is positive and in another negligible. The question is what system-level variables or characteristics are associated with the differences of relationships between market involvement and political mobilization.
System-level variables most closely associated with the within-system correlations would enter into the explanation of the dependent phenomenon. For example, suppose that the relationship between market involvement and political mobilization is high wherever the market is free, that is, wherever the government does not regulate prices. The nature of the market will then become the system-level variable explaining political mobilization. The general statement will now have the following form:

This statement can be read as "the dependent phenomenon, D, depends upon /i and /;> regardless of systems and upon /.i, depending upon the systemic characteristic S," In our example, we can say that political mobilization depends upon exposure to mass media and urbanization regardless of the system involved, and in the systems that have a free market it also depends upon the extent of market involvement. A system-level variable need not be dichotomous: for example, the relationships between exposure to mass media and political mobilization may be a function of the level of economic development of a country.
The procedure of shifting the level of analysis to system variables and formulating explanatory statements relative to systemic characteristics can be repeated until either the number of variables or degrees of freedom is exhausted. We may find that among those countries that have a free market and, thus, where political mobilization depends upon market involvement, the relationship between political mobilization and income in turn depends upon the competitiveness of the party system. If a party system is competitive, persons with high incomes are more likely to be politically mobilized than persons with lower incomes. If, however, a party system is not competitive, personal income does not contribute to the explanation of political mobilization.
Since the number of systems is often limited, the number of degrees of freedom will be exhausted almost immediately. Thus most system-level explanations will be overdetcrmined: several system-level variables will equally well account for the same differences of within-system relationships, Which of these variables "causes" the dependent phenomenon will not be known. This is the point at which social science practice sharply departs from the model of theory. At some stage in the formation of theory, we will be forced to stale that "m all systems, political mobilisation depends upon /i and /-J; in systems with high 5i, it depends upon /.oi; in systems with low .S'i, it depends upon /4; /'" systems with low 5"i and high S-^ it depends upon /.-" but in Poland it also depends upon /i,." At this point the proper name of the social system will be introduced. The name still serves to identify the residual factors that have not been isolated. The incom-pleteness of explanation or the error, /v*,, depends upon the factors subsumed under the name of the social system. Thus the "unique" features of a nation, as they arc traditionally termed, should be considered broadly as including not only phenomena that do not have counterparts in other countries-liberum veto in Poland or the "melting pot" in the United States-but also all other characteristics of a nation that were not isolated

in the process of formulating a theory. An example of a theory concerning political mobilization is presented below.
Theory of Political Mobilization, All Systems
"Political mobilization depends upon exposure to mass media, membership in nonpolilical organizations, and political socialization through the family."
/ \
Systems with a Free Market "Residual political mobilization depends upon market involvement."

Systems with a Regulated Market "Residual political mobilization depends upon intergenerational mobility."
Systems without Competitive Parties
y Systems with
Competitive
Parties
"Residual political mobilization depend^ upon individual income."

7
India 'Residual political mobilization depends upon size of the community."
\ Chile
'Residual political mobilization depends upon
religiosity."
Theory: Political mobilization depends upon exposure to mass media, membership in nonpolitical organizations, and political socialization through the family; in those systems that have a regulated market, political mobilization depends upon individual mobility; in systems with free markets, political mobilization depends upon market involvement; if parlies compete, it depends upon individual income; in addition, in India it depends upon size of the community and in Chile, upon religiosity.
The theory in the illustration is certainly not very general and its deductive structure is almost nonexistent, but it is an example of a "middle-range" theory particularly frequent in political science. In a more general and deductively more powerful version, one might attempt to formulate a theory in terms of reinforcement learning and structural patterns of rewards and sanctions. Such a theory would state that in all systems individuals enter politics if they have learned that political participation is rewarded. In systems with a free market these rewards are mainly economic, whereas in systems in which the market is regulated the rewards are status and political power. The theory presented in the diagram would then be deductible from a general theory of learning, with systemic characteristics defining patterns of rewards.

Comparative Explanation in the Social Sciences: A Conclusion
This systematic exposition of the steps involved in the comparative confirmation and/or modification of theories closes the discussion of the role of comparative studies in the formulation of social science theories. The main role of a theory is to provide explanations of specific events. These explanations consist of inferring, with a high degree of probability, statements about particular events from general statements concerning classes of events- The procedure for formulating such statements places emphasis on finding explanations that are vdlid across all social systems and on replacing names of social systems with system-level variables. It is easier to formulate theories that are parsimonious than those that are causal and general. But as knowledge accumulates, the invariance of particular explanatory statements wilt become known and theories will become more "causal." The gradual increase in the generality of theories seems to be a concomitant of the development of a science.
This interpretation of the nature of theory and the role of comparative social research in the formulation of theories is intended to apply to all social science disciplines. Although the phenomena under consideration vary from discipline to discipline, the logic of scientific inquiry is the same for all social sciences. As the theories explaining social events become general, the explanations of particular events will cut across presently accepted borders of particular disciplines. Some political phenomena may be best explained in terms of learning and, in turn, the explanation of personality characteristics may require societal factors.8 The explanation of historical events, in spite of their alleged "uniqueness," is not an exception to this interpretation, General laws are necessary in order to explain properties of any event.
As has been stated, the most serious challenge to this interpretation arises from the alleged idiographic nature of historical events. All events are unique- But this does not imply that their explanation cannot be based on general theories. Unless uniqueness is seen in a highly literal sense in which every property of an event is in a class by itself, even unique events do not defy theoretical explanation, What is unique about Catherine It writing letters to Voltaire or Napoleon attacking Russia? Even though the combination of specific circumstances that accompanied the particular act of an individual might have been unrepeated, each of
B An interesting example showing the influence of nationality, religion, and socio-economic status on propensity toward conflict and interpersonal emph.'isis of children can be found in K. W. Tcrhiinc. "An Examination of Some Contributing Demographic Variables in a Cross-Nntkinal Study." The Journal of Social P.iycltology, 59, 1963.

the properties of these circumstances constitutes an clement of a class of such properties. Thus theories formulated in terms of classes of properties will provide an explanation of an event that was never repeated in its entirety. It would be nonsensical to believe that no other person behaved like Catherine II or Napoleon, or that there was only one cosmopolitan elite. Unique events manifest properties of general classes and can be explained by general statements, even if incompletely.
What is needed is systematic accumulation of knowledge about social reality-confirmation and/or modification of the same theories under broadly varying sets of social circumstances. What is needed is comparative research guided by explicit theories and replicating tests of the same general propositions.

• [ На главную ]
• [ Библиотека ]
• [ Reader ]
• [ Программы ]
• [ Абитуриенту ]
• [ Фотоальбом ]
• [ О ФП ]
• [ Гостевая книга ]
• [ Ссылки ]
• [ Обратная связь ]