Você está na página 1de 5

Reply to Glenn Author(s): David Knoke and Michael Hout Source: American Sociological Review, Vol. 41, No.

5 (Oct., 1976), pp. 905-908 Published by: American Sociological Association Stable URL: http://www.jstor.org/stable/2094740 . Accessed: 25/07/2013 14:15
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

American Sociological Association is collaborating with JSTOR to digitize, preserve and extend access to American Sociological Review.

http://www.jstor.org

This content downloaded from 186.18.32.91 on Thu, 25 Jul 2013 14:15:08 PM All use subject to JSTOR Terms and Conditions

COMMENTS
to be conceptually inadequate for the broad range of "'nondemographic" variables. But in so doing, he makes no comparisons with alternative models and he does not establish his conclusion in the context of specific substantive problems. Moreover, in rejecting additive models, Glenn totally ignores the criterion of empirical adequacy. Thus, in failing to relate additive formulations to other formulations, either conceptually or empirically, Glenn ignores the purpose of models. Glenn's third point is that the results of those cohort analyses affected by the problem of estimability are necessarily tentative. We agree, but the point is hardly unique to cohort analyses. The results of all modeling efforts are tentative in the sense that they rest on theory and assumptions and run the risk of being discarded in favor of results based on a competitive modeling effort using different theory and assumptions. What deserves special emphasis here is that in models of the sort considered in our article, age, period and cohort are proxies for unmeasured mechanisms or variables. If these mechanisms or variables were measured and available for analysis, the estimability problem dealt with in our paper either would not occur or would be different. For example, if cohort size is held to be the variable which causes cohort differentiation in the context of a particular substantive problem, then, if size measurements can be constructed, it is unnecessary to include cohorts as such in the specification because the preferred variable is available. Under circumstances like this the results of cohort analysis become less tentative, since the estimability problem as we described it is eliminated. The replacement of proxies by the variables they index is a universal goal of research. In conclusion, we reiterate a point made in our article. Specifications of the kind we dealt with are not universally appropriate or desirable. Cohort analysis can take many forms. Whatever the approach the analyst thinks promising, care should be taken to develop a model which is sufficiently formal to enable determination of the necessity of estimability restrictions. Such restrictions require theoretical and substantive justification. William M. Mason Karen Oppenheim Mason University of Michigan H. H. Winsborough University of Wisconsin, Madison In our original article, "Some Methodological Issues in Cohort Analysis of Archival Data"

905

(ASR April, 1973), equation (2), page 246, should read: YJ =,+ +Pryj +Sk+ E;J. ,81 REFERENCES Greenberg, B. G., J. J. Wright, C. G. Sheps 1950 "A technique for analyzing some factors affecting the incidence of syphilis." Journal of the American Statistical Association 251:373-99. Hall, R. E. 1971 "The measurement of quality change from vintage price data." Pp. 240-71 in Z. Griliches (ed.), Price Indexes and Quality Changes. Cambridge: Harvard University Press. Winsborough, H. H. 1975 "Age, period, cohort, and education effects on earnings by race-an experiment with a sequence of cross-sectional surveys." Pp. 201-17 in K. Land and S. Spilerman (eds.), Social Indicator Models. New York: Russell Sage.

REPLY TO GLENN

We welcome this opportunity to clarify explicitly our application of the cohort-ageperiod model to our studies of party identification and voting turnout in the United States (Knoke and Hout, 1974; Hout and Knoke, 1975). First, we reanalyze previously published cohort data to demonstrate the reasonableness of our assumption of constant cohort effects on party identification. Next, we clarify some of the semantic confusion which seems to attend the use of the terms "additive" and "interactive" statistical models. Finally, we raise some questions concerning the ultimate implications of Glenn's suggestions for developing a sociological analysis of cohort behavior. Glenn questions whether our assumption of constant cohort effects on party identification within age and period categories is reasonable. Here we offer evidence supporting our assumption, based on survey data reported in Glenn and Hefner (1972), Glenn (1972) and Knoke (1976:ch. 7). Each article reports three-way cross-classifications of survey respondents by cohort, year of survey and party identification (three major party groups). We subjected these data to log-linear analyses, fitting to each table several models corresponding to different assumptions about the relationships among the three variables. * We appreciate Norval D. Glenn's comments on several earlier drafts of this reply, which have much improved the quality and value of this exchange of views.

This content downloaded from 186.18.32.91 on Thu, 25 Jul 2013 14:15:08 PM All use subject to JSTOR Terms and Conditions

906

AMERICAN SOCIOLOGICAL REVIEW

Table 1. x2 Values for Some Log-Linear Models Fitted to Three Cohort-Period-PartyIdentification Cross-Tabulations Glenn-Hefnera Model Marginals x2 df Glenn b x2 df
e

Knoke

x2

df

1 (C)(T)(P) 2 (CT)(P) 3 (CT) (TP) 4 (CT) (CP) 5 (CT)(TP)(CP) Sample Size

18,869.9 (132) 1,491.5 ( 86) 990.2 ( 74) 384.6 ( 74) 122.5 ( 62) (56,235)

1,145.3 (44) 922.9 -(34) 597.3 (30) 348.6 (24) 33.9 (20) (29,852)

2,941.0 (61) 523.1 (44) 371.0 (32) 79.2 (36) 33.3 (26) (10,927)

C: Cohort T: Period P: Party Identification a 7 cohorts, 7 periods, 3 parties. b 6 cohorts, 3 periods, 3 parties. e 4 cohorts, 6 periods, 3 parties. (Where data were missing due to the absence of a cohort from a time period, we constrained the entries to fixed zeros, and fitted "quasi-independence" models to the remaining cells; see Goodman, 1968.) The results are summarized in Table 1. Model 1 fits only the three one-way marginal distributions of each variable, under the assumption that none of the variables are associated. Of course, this "baseline" model provides a poor fit to each data set, but the resulting x2 value may be taken as a measure of the "unexplained variance" to be accounted for by other models which allow relationships among the variables (see Goodman, 1972: 1054). The next three models all show some improvement in fit by permitting party identification to be associated separately with cohort or year. Finally, model 5 permits all two-way marginals to be fitted, testing the hypothesis that only constant relationships occur between cohort, period and party identification. Model 5 asserts that the odds on being Republican, Democrat or Independent in a given cohort are unaffected by the period of observation. As the x2 values for model 5 show, only the Knoke data attain a significant fit by the usual criterion of a probability greater than .05. But, given the enormous size of the samples used by Glenn and Hefner and by Glenn, to expect a statistically significant fit is unreasonable, since x2 values for a given relationship will increase with sample size. A better criterion against which to judge the adequacy of model 5 in each case is the proportion of "unexplained variance" in the baseline model 1 which is reduced after fitting the additive model 5. In each case, model 5 accounts for more than 97 percent of the x2 values for model 1. Even judged against model 2, which permits cohort and period measures to associate, model 5 still accounts for more than 90 percent of the baseline x2 Thus, we conclude that model 5, which depicts only two-way relationships between variables in the cohort-period-party tables, provides a highly satisfactory fit to the observed data. These results indicate that, contrary to Glenn, the relationship between cohort and party preference does not vary over periods. Period effects on party identification do occur, as demonstrated by the necessity to include the (TP) marginal, but these period effects operate identically across cohorts. Hence, when shifts in party identification occur over time, they tend to be as large at older as at younger age levels. Insofar as conclusions can be drawn from empirical data from the post-World War II era, our assumption of constant effects among cohort, period and age on party identification does not seem unreasonable. Glenn calls the Mason et al. (1973) cohort model, which we used with minor modification, "additive" in the sense that it assumes age effects are the same at each level of period and cohort, and so forth. Then he acknowledges in a footnote that our models are "interactive," in the sense that cohort effects are an interaction between age and period. We suspect that this use of terminology is confusing and we want to set the record straight on the nature of the statistical models we employed. Our models do contain interaction terms, since each age, period and cohort variable is a linear combination of the other two. Indeed, it is interaction that leads to the identification problem which motivated the development by Mason et al. of procedures to deal with statistical indeterminacy by imposing certain constraints on the variables. Persons familiar

This content downloaded from 186.18.32.91 on Thu, 25 Jul 2013 14:15:08 PM All use subject to JSTOR Terms and Conditions

COMMENTS
with the long controversy on the existence of status inconsistency and social mobility effects will recognize that our models are one of a family of such interaction-effects models and not an "additive" effect model in the usual meaning of. the term (see Duncan, 1966; Knoke, 1973). We feel that the term "additive" should be applied to models which employ only two sets of demographic dummy variables (for example, age and period) in a regression equation to avoid unnecessary confusion. Perhaps Glenn had in mind, as an alternative to our interaction model, a statistical model specifying one dummy variable for every cell in the age-by-period cross-classification (see Blau and Duncan, 1967:ch. 11, for an example with social mobility and fertility). We could have fit such a "total" interaction model to the data, but to do so would have produced uninterpretable results since no parameter could be assigned meaningfully to age, period or cohort effects.' The interaction model we use specifies only one type of cohort effect, which we believe to be most consistent with the theoretical definition of that concept. As explicitly stated in Hout and Knoke (1975:61), our equation assumes that the effect of a given cohort on the dependent variable remains constant. This specification comes from our theoretical interest in the cohort succession hypothesis: "Aggregate changes in societal characteristics, whether in fertility rates or political attitudes, need not involve changes in the attributes of individuals. The succession of birth cohorts through the population may be sufficient to account for aggregate shifts over time if the entering cohorts differ enough from cohorts exiting through death" (Knoke and Hout, 1974:701). If the attributes of individuals change during their lifetimes, "demographic metabolism" is not the only mechanism op-

907

rating to alter aggregate attributes of the population. By specifying constant cohort effects, we explicitly limit the causal mechanism to the cohort succession process. Our assumption of constant cohort effects is the apparent reason for Glenn's labeling of our model "additive." However, as he acknowledges in a footnote, ours is an interaction model and, clearly, only one of a large variety of such interaction models which might have been hypothesized. The particular form of the interaction was chosen on theoretical, not on ad hoc or arbitrary grounds. Clearly, alternative interaction specifications are possible, although not all will be theoretically plausible or result in meaningful empirical findings. For example, Glenn appears to argue that cohort effects vary by age and/or period; but since the cohort variable is already an interaction term (cohort-periodage), conceptual clarity among the three demographic dimensions becomes seriously blurred when one starts positing second- and third-order interactions among age, period and cohort. However, we see a reasonable possibility that researchers could specify meaningful interactions other than the type we chose. Such alternative models could be fit to data and the gains in both statistical fit and substantive understanding could be assessed. Since alternative specifications are likely to involve more complex interactions than ours, we rightfully must insist that they provide a superior explanation to compensate for the loss of parsimony involved. Until such revised model(s) are forthcoming, we stand by our results as the most substantively and statistically plausible explanation of cohort succession effects on party identification and voter turnout which has appeared to date. We appreciate Glenn's exemption of our efforts from the accusation of "mechanical, atheoretical cohort analysis," and we join him in condemnation of research based on shoddy 1 We did fit our interaction model to the cell theoretical work. We are sorry he found our means of the age-by-period cross-tabulation for conclusions stated too "dogmatically," and we both the party identification and voting turnout data. The proportion of variance in cell means herewith acknowledge that our (as anyone's) explained by our interaction model were 56 and results will be valid only to the extent that 66 percent, respectively (df=40). The "total" our assumptions hold. All empirical results interaction model, of course, accounts for 100 are ultimately tentative, since there is no of the cell mean variance, but requires an addi- proof positive, only disproof. However, we tional 55 degrees of freedom. On the individual believe we have further strengthened the case level, the age, period and cohort variables added for the tenability of our interaction model of only about two and four percent to the R2 the age, cohort, period effects on party idenabove that contributed by the social covariates. tification and voting turnout. We would welinterThus, the replacement of our constrained evidence to the conaction model by the "total" interaction model come further empirical only marginally would improve the statistical fit trary. Finally, we feel uneasy about the implicaof the individual-level data, while resulting in tions we see in Glenn's comment for the desubstantivelychaotic effect parameters.

This content downloaded from 186.18.32.91 on Thu, 25 Jul 2013 14:15:08 PM All use subject to JSTOR Terms and Conditions

908

AMERICAN SOCIOLOGICAL REVIEW


Glenn, Norval D. 1972 "Sources of the shift to political independence: some evidence from a cohort analysis." Social Science Quarterly 53:494-519. 1976 "Cohort analysts' futile quest: statistical attempts to separate age, period and cohort effects." American Sociological Review 41:900-4. Glenn, Norval D. and Ted Hefner 1972 "Further evidence on aging and party identification." Public Opinion Quarterly 36:31-47. Goodman, Leo A. 1968 "The analysis of cross-classified data: independence, quasi-independence, and interactions in contingency tables with of without missing entries." Journal of American Statistical Association 63: 1091-131. 1972 "A general model for the analysis of surveys." American Journal of Sociology 77:1035-86. Hout, Michael and David Knoke 1975 "Changein voting turnout, 1952-1972." Public Opinion Quarterly 39:52-68. Knoke, David 1973 "Intergenerational occupational mobility and the political party preferencesof American men." American Journal of Sociology 78:1448-68. 1976 Change and Continuity in American Politics: The Social Bases of Political Parties. Baltimore: Johns Hopkins University Press. Knoke, David and Michael Hout 1974 "Social and demographic factors in American political party affiliations, 1952-1972." American Sociological Review 39:700-13. Mason, Karen Oppenheim, William M. Mason, H. H. Winsboroughand W. Kenneth Poole 1973 "Some methodological issues in cohort analysis of archival data." American Sociological Review 38:242-58.

velopment of a sociological analysis of cohort behavior. His main objection to age-periodcohort analysis as we have employed it seems to stem from a belief that every cohort experiences age and/or period effects in a fashion unique from all other cohorts. Thus, he asserts, information provided by the behavior of other cohorts offers no clue to determine the relative impact of age or period on any cohort. But if age and/or period effects experienced by the hypotheticalcohort in his Table 1 are "completely unique," as he insists, in what meaningful sense can they be called effects? (In fact, why stop there; since every individual experiences his or her life uniquely, should we not insist that no generalizations across individualsare permissible?)To argue, as Glenn does, that cohort effects ("additive effects" in his terms) cannot be found is to destroy the original meaning of the concept. "Cohort effects" are restricted to those influenceswhich persistthroughthe life cycle of membersof the group; thus a "cohort effect" which changes at every new age and period is a contradiction.Glenn's analysis would leave us with a juxtaposition of biographical and historical situations and, while a valid theoretical scheme using only these explanatory concepts might be possible, such a scheme is neither our objective nor, we infer, Glenn's. For if every cohort is unique in its behavior at every point in time, then no generalizations can be made and we have no science, even by the weak standardsof contemporary social science. We fear that, if Glenn's suggestions are taken literally, sociological research must forego the aim of discovering generallyvalid relationshipsand settle instead for historical description.We personally find that conclusion untenableand we hope Glenn does too, but that is where his arguments seem inescapablyto lead. David Knoke
Indiana University

ON THE USE OF ORDINAL DATA IN CORRELATION ANALYSIS * (COMMENT ON LABOVITZ, ASR JUNE, 1970)

Michael Hout
University of Arizona REFERENCES

The practice of treating ordinal data as interval has been the subject of considerable Blau, Peter M. and Otis Dudley Duncan debate. A major concern has been the pos1967 The American Occupational Structure. sible sensitivity of statistical results to the New York: Wiley. choice of the particular scale used. The Duncan, Otis Dudley primary purpose of this paper is to present a 1966 "Methodological issues in the analysis * I wish to thank my colleagues C. R. Plott, of social mobility." Pp. 51-.97 in N. Smelser and S. M. Lipset (eds.), Social John Ferejohn, Morris Fiorina, Robert Bates and Structure and Mobility in Economic Morgan Kousser for their helpful comments and Development. Chicago: Aldine. suggestions.

This content downloaded from 186.18.32.91 on Thu, 25 Jul 2013 14:15:08 PM All use subject to JSTOR Terms and Conditions

Você também pode gostar