Coh

Cohort Analysis

A cohort is a set of individuals entering a system at thesame time. Individuals in a cohort are presumed tohave similarities due to shared experiences that differ-entiate them from other cohorts. Cohort analysisseeks to explain an outcome through exploitation ofdifferences between cohorts, as well as differencesacross two other temporal dimensions: ‘age’ (timesince system entry) and ‘period’ (times when anoutcome is measured). This article exposits difficultiesinherent to cohort analysis, indicates promising direc-tions, and provides context.

1. Cohorts

Cohorts as analytic entities appear in the socialsciences, the life sciences, epidemiology, and else-where. A cohort can be a set of people, automobiles,trees, whales, buildings; the possibilities are endless.System entry can refer to birth—a person is born, or toany dated event—a machine is assembled on a par-ticular date. A set of individuals who begin serving aprison term at the same point in time might also definea cohort for certain purposes, in which case ‘birth’refers to initiation into a particular role system and‘age’ becomes duration (time since system entry). Thebreadth of the time interval that defines membershipin a particular cohort depends on analytic consider-ations and the nature of the phenomenon under study.

The main difficulties inherent to cohort analysis canbe illustrated with data from the NORC GeneralSocial Survey, a national sample survey conductedannually or biennially in the United States. Respon-dents in this repeated cross-sectional survey arequeried about their emotional wellbeing. Each cell inTable 1 shows the percentages, organized by age andsurvey year, of those who identified themselves asbeing ‘very happy.’ Each row shows how happinesslevels change across survey year (period) for peoplewithin a given age group. Each column reveals agevariation in any survey year. The diagonals permittracking of a single birth cohort over time. Those aged20–29 in 1973 were born between 1943 and 1953, aswere those aged 30–39 in 1983, and those aged 40–49in 1993. Ten-year age categories lead here to cohortsoperationalized as individuals born in contiguous 10-year intervals.

Nominal attempts to extract information fromTable 1 would summarize percent ‘very happy’ byrows, columns, and diagonals. The most immediatelyvisible pattern is the association between age andhappiness. In all three periods the percentages ofrespondents who identify themselves as ‘very happy’are smaller for those under age 40. This reading of thetable ignores the possibility that observed variation isdue to birth cohort, with younger cohorts less likely toreport being happy. Although visual inspection of thediagonals does not suggest a consistent associationbetween decade of birth and eventual happiness, theabsence of an observable relationship does not confirmthat no such relationship exists. A potential associ-ation between birth cohort and adult happiness mayhave been obscured by the effects of age and period.Finally, there appears to be little systematic variationbetween years in percent ‘very happy.’ In this instance,however, a potential association between period and

Table 1Percent ‘very happy’ by age and period

1973 1983 1993

20–29

30–39

40–49

50–59

60–69

70–79

N = 4454

Age

29% 30% 28%

36% 28% 29%

40% 31% 30%

41% 29% 38%

38% 37% 33%

38% 38% 33%

(347) (372) (278)

(294) (354) (381)

(247) (228) (329)

(253) (212) (205)

(192) (201) (166)

(117) (123) (155)

Source: General Social Survey, 1973–1993Note: Numbers in parentheses are base Ns for the percentages

Period

2189

Mason & Wolfinger: Cohort Analysis, IESBS, 2001 Page 17 of 17

C:\g2_e_on_042702\iesbs\cohort\cohort-public040901.doc 5/20/03 10:25 PM

(a) Single Cross-Section with

Retrospective Data (b) Panel Study

Period Period

Age

Age

Period of Cross-Sectional Survey

Figure 1 Retrospective and Prospective Data Structures that Engender Age, Period, and Cohort

current happiness may have been obscured by age andcohort. A further, linked difficulty is that the datastructure is imbalanced. Although periods are repre-sented over all ages, and ages are represented over allperiods, the observed age spans of cohorts necessarilydiffer.

This imbalance cannot be extirpated from analysesthat take into account age, period, and cohort sim-ultaneously. Comparable points hold for other datastructures supporting the analysis of multiple cohorts.

Table 1 shows that data structures that allow formeasurements on multiple cohorts necessarilymeasure age and period. Furthermore, knowledge ofplacement on any two of age, period, and cohortdetermines placement on the third. This dependencycan be expressed as:

Cohort¯Period®Age

which raises the questions of whether and how allthree of age, period, and cohort can be included incohort models. The linear dependency between age,period, and cohort, also known as the cohort analysisidentification problem (see Statistical Identificationand Estimability), is the point of departure for allmodern discussions of techniques of cohort analysis.The identification problem is present irrespective ofdata structure.

2. Data Structures

Three commonly seen data structures engender cohortanalysis (for a fuller discussion of data structures, seeFienberg and Mason 1985). Excluded from thisdiscussion, however, would be the single cross section.Reducing Table 1 to a single column indicates the datastructure of a single cross section. With this design,differences on an outcome variable by age can be

Figure 1Retrospective and prospective data structures thatengender age, period, and cohort

interpreted either as age or cohort differences. Thedata structure provides no basis for choosing becausethere is but a single period. Interpretation of agedifferences on an outcome as being attributable tofactors thought to vary with age or duration requiresthe assumption that there are no cohort or perioddifferences, or that they are known. Parallel assump-tions are necessary for interpreting the age differencesas attributable to cohort. Thus, the single cross sectiondoes not permit cohort analysis.

Some cross-sectional surveys elicit retrospectivedata on birth, marital, or other kinds of histories.Figure 1(a) structures this design, a single cross sectionwith retrospecti�e data, as an upper-triangular age byperiod array with cohorts defined by diagonals.

The retrospective data structure provides infor-mation on ages, cohorts and periods, and introduceslongitudinal information on individuals (see Longi-tudinal Data). When the longitudinal data in thisdesign are used without regard to age, period, andcohort, the analyst is implicitly, and possibly inad-vertently, assuming that at least one of age, period,and cohort is not essential. The same point holds forpanel study data structures (Fig. 1(b)). A panel studybegins with a single cross section, and is followed byone or more panels on the same individuals (units ofanalysis). This design is intended for the creation oflongitudinal data. Figure 1(b) structures the panelstudy as a lower-triangular age by period array withcohorts on the diagonal. It is triangular because, in thesimplest case, panel designs do not replenish the datastructure with the addition of new cohorts after theinitial cross section. A panel study could also bedesigned to include new cohorts at successive waves ofdata collection; the age, period, cohort dependencywould remain.

In the replicated cross-section design or time seriesof cross sections (see Longitudinal Data; E�ent His-tory Analysis in Continous Time) illustrated by Table 1,

2190

Cohort Analysis

typically the same individuals are not tracked fromone period to the next. However, each cross sectioncan have a retrospective component, and thus thisdesign can be longitudinal. Like the other multiplecohort designs, this structure permits cohort analysis,although the class of models it supports is less richwhen retrospective information is unavailable.

3. Cohort Models

Cohort models may be fixed or random effect (seeHierarchical Models: Random and Fixed Effects);terms for age, period, and cohort may enter the modelas discrete or continuous; one or more of the age,period, and cohort dimensions may be included in themodel via an explicit, substantive measure of thatdimension; interactions are possible. These are themost prominent possibilities in the literature on cohortanalysis.

3.1 Fixed Effect: Discrete Age, Period, and Cohort

Assume an I¬J age by period array (Table 1 is a 6¬3illustration), with age groups and period intervals ofidentical widths. The K¯ I­J®1 diagonals of thearray correspond to cohorts. The basic fixed-effectmodel treats a parameter (θ

ijk) associated with a

response variable as a linear function of discrete age,period, and cohort. Using dummy coding for age,period, and cohort, let:

θijk

¯ β!­3

I

i=#

βiA

i­3

J

j=#

γjP

j­3

K

k=#

δkC

k(1)

where the Ai, P

j, and C

k(k¯ i®j­J ) are dummies for

ages, periods, and cohorts, respectively. This is a fixed-effect model because inference is conditional on theages, periods, and cohorts represented by a particulardataset. Although Eqn. (1) manifests usual normaliz-ation restrictions (omission of one dummy from eachclassification), this is insufficient to break the lineardependency between age, period, and cohort. Omis-sion of all terms in one of the age, period, or cohortclassifications eliminates the dependency. This is asatisfactory strategy if prior theory and informationsuggest that age, or period, or cohort is superfluous.On the other hand, if all three dimensions are deemedindispensable to the analysis, the dependency must beeliminated by one or more further restrictions oncoefficients if the fixed-effect, discrete model is to beemployed. Problems with this approach include col-linearity among termson the right hand side ofEqn. (1)remaining after additional coefficient restrictions havebeen introduced, and coefficient bias. Moreover,restrictions that might appear to be innocuouslydifferent, such as equating the coefficients of different

subsets of adjacent categories, can lead to quitedifferent sets of age, period, and cohort effects, withthe various models all fitting the data equally well, ornearly so. One response to the problems of estimatingEqn. (1) is to massively overidentify one or moredimensions. For example, a priori knowledge maysuggest that period can be represented by several,rather than many, categories. The data cannot, how-ever, be relied on to contain the information on whichto base overidentifying restrictions (Fienberg andMason 1985, Glenn 1989, Heckman and Robb 1985,Kupper et al. 1983, Mason and Smith 1985, Wilmoth1990).

3.2 Fixed Effect: Continuous Time

Let age (A), period (P), and cohort (C ) be measured incontinuous time, with C¯P®A. Then the continuoustime equivalent to Eqn. (1) is:

θAPC

¯ fA(A)­f

P(P)­f

C(C ) (2)

where fA(A) is an (I®1)th-order polynomial in A, f

P(P)

is a (J®1)th-order polynomial in P, and fC(C ) is a

(K®1)th-order polynomial in C. Because of the C¯P®A linear dependency, the coefficients of the linearterms for A, P, and C are not estimable. As in thediscrete case, at least one linear restriction must still beimposed and in any event the discrete variable ap-proach is generally preferable.

3.3 Random Effect, Discrete Age, Period, Cohort

It is possible to view the modeling of age, period, andcohort effects from a Bayesian, hierarchical perspec-tive (Nakamura 1986; see Bayesian Statistics; Hi-erarchical Models: Random and Fixed Effects). In thisapproach, it is convenient to characterize age, period,and cohort effects through first differences, as in

θijk

¯ β*

!­3

I

i=#

β*

iA*

i­3

J

j=#

γ*

jP*

j­3

K

k=#

δ*

kC*

k(3)

where A*

i¯A

i®A

i−", P*

j¯P

j®P

j−", and C*

k¯

Ck®C

k−". The approach assumes that the β*

i,γ*

j, and δ*

k

are separately distributed, and that they are random orexchangeable (Nakamura 1986, Sasaki and Suzuki1987, 1989, Glenn 1989, Miller and Nakamura 1996).The exchangeability assumption requires that withinthe age dimension, within the period dimension, andwithin the cohort dimension, all permutations of thefirst-difference coefficients must be equally acceptable.This assumption makes possible the determination ofage, period, and cohort effects without resorting torestrictions on coefficients. To the extent that theassumption fails, and it can when there are shocks

2191

Cohort Analysis

(e.g., war, famine, plague, a stock market crash) in theprocess being modeled, the random effects approach isnot a panacea.

3.4 Substanti�e Measurement of Age, Period, orCohort

When one or more of the age, period, and cohortdimensions, whether represented as discrete or con-tinuous, is replaced by a variable chosen to measurethe underlying process thought to be captured by age,or period, or cohort, the linear dependency is almostalways broken. Attention is then appropriately fo-cused on the theoretical and substantive merits of thespecification. Models that include age, period, andcohort should be thought of as starting points, giventheir inferiority relative to models that are able to testideas of how and why a cohort, or period, or agemechanism affects an outcome. Concretely, supposethe cohort dimension is held to reflect the impact ofmeasured variable X, then Eqn. (1) might change to:

θijk

¯ β!­3

I

i=#

βiA

i­3

J

j=#

γjP

j­δX (4)

where X is either constant over ages within a cohort, orvaries by age within a cohort. Relative cohort size is anexample of a variable that could be defined at the birthof each cohort, or allowed to vary as cohorts ‘age’through the life cycle. Measured variables can also beemployed in extensions and revisions of Eqns. (2)–(3).For example, in Eqn. (2) the polynomial in C can beomitted upon inclusion of some function of X (X itself;a polynomial in X; interactions between X and age orperiod).

In the extension of the random-effects approachthat includes measured variables, cohort analysisbecomes a specific case of random effects multilevelanalysis (see Statistical Analysis: Multile�el Methods).This development can be expected in the course of thecontinued deployment of Bayesian statistical solu-tions, because the use of measured variables canenhance the validity of the exchangeability assum-ption. In the random-effects approach, substantivevariables can be written into the model in the followingway, for one or all of age, period, and cohort:

β*

i¯ λ

!A­λ

"AX

A­τ

iA

γ*

j¯ λ

!P­λ

"PX

P­τ

jP

δ*

k¯ λ

!C­λ

"CX

C­τ

kC

(5)

where, for example, XC

is a measured variable forcohort, and for simplicity only one measured variableper dimension has been included, and only as a linearterm. In this extension it is the τ

iA, τ

jP, and τ

kCthat are

assumed to be exchangeable within the age, period,

and cohort dimensions, and this is likely to be moredefensible—because substantively plausible underly-ing, measured variables should covary with thephenomena that produce shocks. Moreover, theexchangeability assumption becomes comparable tothe assumption of a random-error term in a fixedeffects approach with measured variables.

4. Interactions

The need for interactions often arises, either forsubstantive reasons (Converse 1976), or for technical,adjustment purposes (Mason and Smith 1985). It ispossible to include interactions in both fixed- andrandom-effect models, regardless of the presence ofmeasured variables. Fienberg and Mason (1985) elab-orate readily implemented strategies for doing so inthe fixed-effect framework, using either discrete orcontinuous age, period, and cohort. Not all inter-actions are estimable, and hence they cannot be addedinto the model at will.

5. Conclusions

Panel studies, cross-sectional studies with retrospec-tive information, and replicated cross sections (in-cluding age by period arrays created from process-generated data) engender the analysis of a responsevariable as a function of age, period, and cohort aswell as other factors. Such analyses must contend withthe linear dependency between age, period, and cohortmembership. The use of one or more measuredvariables held to underlie at least one of age, period, orcohort can break the linear dependency. So too canapplication of credible prior information, whetherexpressed as constraints on coefficients in fixed-effectmodels, or as exchangeability assumptions in random-effects or hierarchical models. Measured variablescan, of course, be incorporated into both fixed-effectand random-effect models. This strategy is to bepreferred, since it makes it possible to test ideas aboutsubstantive processes in the most direct way. Modelsthat include age, period, and cohort can also includeinteractions between these dimensions, though not allsuch terms have estimable coefficients. Models that donot explicitly consider all three of age, period, andcohort, and yet are based on data structures thatpermit their inclusion, rest on the implicit assumptionthat age, or period, or cohort is irrelevant. Thisassumption should and can be assessed.

6. Further Reading

Fienberg and Mason (1985) discuss the formalizationof the identification problem; identifiability of non-linear components; identifiability of certain inter-

2192

Cohort Analysis

actions beyond those implicit in the simultaneousinclusion of age, period, and cohort; polynomialmodels; and other topics. Kupper et al. (1983) andKupper et al. (1985) explore in depth several issuesraised by the identification problem for the fixed effectdiscrete case, and take the stance (Kupper et al. 1985)that age-period-cohort models are no more informa-tive than exploratory graphical displays. Ploch andHastings (1994) illustrate the use of smoothed per-spective plots. Robertson and Boyle (1998b) providean overview of different strategies for graphical displayof age-period-cohort data.

In Nakamura’s (1986) Bayesian formulation, identi-fying linear restrictions are replaced by the assumptionof exchangeability of first differences in age, period,and cohort effects (although Nakamura chooses toemphasize the closely related assumption of ‘smooth-ness’). Hodges’ (1998, Sect. 2) development of hi-erarchical models as linear models contributes to anunderstanding of how Nakamura’s specification over-comes the singularity of the fixed effects model matrix.Berzuini and Clayton’s (1994) Bayesian formulation,which is focused on second differences of effects, doesnot solve the identification problem. The socialsciences literature in which Nakamura’s specificationis used or discussed (Sasaki and Suzuki 1987, 1989,Glenn 1989, Miller and Nakamura 1996) fails to makeclear that it is not the Bayesian approach per se thatprovides an alternative to linear restrictions on coeffic-ients, but rather Nakamura’s particular formulation.Nakamura’s (1986) computational approach has beensupplanted by the use of Markov Chain Monte Carlomethods (The BUGS Project 2000; see Monte CarloMethods and Bayesian Computation: O�er�iew;Marko� Chain Monte Carlo Methods).

Robertson and Boyle (1998a), and Robertson et al.(1999) review the largely disciplinary-specific epi-demiological literature on the methodology of cohortanalysis, which has focused primarily on additivemodels, and conclude that only the nonlinear com-ponents of age, period, and cohort can be used reliably.Holford et al. (1994) employ substantive reasoningabout cell malformation in carcinogenesis, developspecifications in which age is inherently nonlinear(e.g., logarithmic) and thus eliminate the identificationproblem through choice of functional form. Masonand Smith’s (1985) extended study of tuberculosismortality combines substantive reasoning based onprior information and expectations, uses one body ofdata to guide modeling of another, includes aninteraction term, employs a substantive, measuredvariable, and concludes that potential interactionsrequire at least as much attention as the identificationproblem itself.

Hobcraft et al. (1985) focus on theoretical reasonsfor the use of age, period, and cohort in different areaswithin demography. Nı! Brolcha! in (1992) takes aim atthe relevance of the cohort dimension for under-standing temporal variation in human fertility. Dis-

cussions of this kind can help cohort analysts becomemore substantively grounded. Research using measur-ed variables in place of accounting categories (e.g.,cohort size instead of a cohort classification) is not,however, in need of such assistance (Easterlin 1980,Ahlburg and Shapiro 1984, Welch 1979). Blossfeld(1986) provides a persuasive example of the use ofmassive overidentification within a single dimension.

See also: Bayesian Statistics; Hierarchical Models:Random and Fixed Effects; Longitudinal Data;Longitudinal Data: Event–History Analysis in Dis-crete Time; Markov Chain Monte Carlo Methods;Statistical Analysis: Multilevel Methods; StatisticalIdentification and Estimability

Bibliography

Ahlburg D A, Shapiro M O 1984 Socioeconomic ramificationsof changing cohort size: An analysis of U.S. postwar suiciderates by age and sex. Demography 21: 97–108

Becker H A (ed.) 1992 Dynamics of Cohort and GenerationsResearch. Thesis Publishers, Amsterdam

Berzuini C, Clayton D 1994 Bayesian analysis of survival onmultiple time scales. Statistics in Medicine 13: 823–38

Blossfeld H-P 1986 Career opportunities in the Federal Republicof Germany: A dynamic approach to the study of life-course,cohort, and period effects. Eur. Sociol. Re�. 2: 208–25

Converse P E 1976 The Dynamics of Party Support: CohortAnalyzing Party Identification. Sage, Beverly Hills, CA

Easterlin R A 1980 Birth and Fortune: The Impact of Numberson Personal Welfare. Basic Books, New York

Elder G H Jr 1999 [1974] Children of the Great Depression: SocialChange in Life Experience, 25th Anniversary edn. Universityof Chicago Press, Chicago

Fienberg S E, Mason W M 1985 Specification and implemen-tation of age, period, and cohort models. In: Mason W M,Fienberg S E (eds.) Cohort Analysis in Social Research: Beyondthe Identification Problem. Springer-Verlag, New York, pp.44–88

Glenn N D 1989 A caution about mechanical solutions to theidentification problem in cohort analysis: Comment on Sasakiand Suzuki. American Journal of Sociology 95: 754–61

Heckman J, Robb R 1985 Using longitudinal data to estimateage, period and cohort effects in earnings equations. In:Mason W M, Fienberg S E (eds.) Cohort Analysis in SocialResearch: Beyond the Identification Problem. Springer Verlag,New York, pp. 137–50

Hobcraft J, Mencken J, Preston S 1985 Age, period, and cohorteffects in demography: A review. In: Mason W M, FienbergS E (eds.) Cohort Analysis in Social Research: Beyond theIdentification Problem. Springer Verlag, New York, pp. 89–135

Hodges J S 1998 Some algebra and geometry for hierarchicalmodels, applied to diagnostics (with discussion). Journal of theRoyal Statistical Society: Series B 60: 497–536

Holford T R, Zhang Z, McKay L A 1994 Estimating age, periodand cohort effects using the multistage model for cancer.Statistics in Medicine 13: 23–41

Kupper L L, Janis J M, Karmous A, Greenberg B G 1985Statistical age-period-cohort analysis: A review and critique.Journal of Chronic Disease 38: 811–30

2193

Cohort Analysis

Kupper L L, Janis J M, Salama I A, Yoshizawa C N, GreenbergE 1983 Age-period-cohort analysis: An illustration of theproblems in assessing interaction in one observation per celldata. Commun. Statist.–Theor. Meth. 12: 2779–807

Mason W M, Fienberg S E (eds.) 1985 Cohort Analysis in SocialResearch: Beyond the Identification Problem. Springer Verlag,New York

Mason W M, Smith H L 1985 Age-period-cohort analysis andthe study of deaths from pulmonary tuberculosis. In: MasonW M, Fienberg S E (eds.) Cohort Analysis in Social Research:Beyond the Identification Problem. Springer Verlag, NewYork, pp. 151–227

Miller A S, Nakamura T 1996 On the stability of churchattendance patterns during a time of demographic change:1965–1988. Journal for the Scientific Study of Religion 35:275–84

Nakamura T 1986 Bayesian cohort models for general cohorttable analyses. Ann. Inst. Statist. Math. 38: 353–70

Nı! Bhrolcha! in M 1992 Period paramount? A critique of thecohort approach in fertility. Pop. De�. Re�. 18: 599–629

Ploch D R, Hastings D W 1994 Graphic presentations of churchattendance using general social survey data. Journal for theScientific Study of Religion 33: 16–33

Robertson C, Boyle P 1998a Age-period-cohort models ofchronic disease rates. I: Modeling approach. Statistics inMedicine 17: 1305–23

Robertson C, Boyle P 1998b Age-period-cohort models ofchronic disease rates. II: Graphical approaches. Statistics inMedicine 17: 1325–39

Robertson C, Gandini S, Boyle P 1999 Age-period-cohortmodels: A comparative study of available methodologies.Journal of Clinical Epidemiology 52: 569–83

Ryder N B 1965 The cohort as a concept in the study of socialchange. Am. Sociol. Re�. 30: 843–61

Sasaki M, Suzuki T 1987 Changes in religious commitment inthe United States, Holland, and Japan. American Journal ofSociology 92: 1055–76

Sasaki M, Suzuki T 1989 A caution about the data to be used forcohort analysis:Reply toGlenn.AmericanJournal ofSociology95: 761–5

The BUGS Project 2000 http:}}www.mrc-bsu.cam.ac.uk}bugs}Welch F 1979 The effects of cohort size on earnings: The baby

boom babies financial bust. Journal of Politics and Economics87: 565–97

Wilmoth J R 1990 Variation in vital rates by age, period, andcohort. Sociological Methodology 20: 295–335

W. M. Mason and N. H. Wolfinger

Cold War, The

The term is widely accepted in historical writings. Itrefers to the epoch between 1947–8 and 1989–90. Thenature of the Cold War, its causes and effects, and thereasons for its long duration are, however, contro-versial. Its meaning would exclude Asia because of theoutbreak of wars there; the wars in Korea, Vietnam,and Afghanistan made a difference to how the UnitedStates (US) and the Soviet Union (USSR), the twosuperpowers, waged their global contest in Europe.The situation in Europe resembled neither peace nor

war; East–West Conflict: Confrontation and DeU tenteexpresses the ambiguity, but this label is less eye-catching.

1. The Concept and its Salience

The persistent central features of the Cold War are theglobal contest between the US and the USSR, thedependence of the allies on the security guarantee oftheir respective superpower, and bipolarity. The latterwas reinforced in the core area, the arms race, due tothe widening gap between the extensive and highlydiversified weapons systems of the superpowers andthe arsenals of Britain, France, and the Peoples’Republic of China (PRCh), the other three establishednuclear powers. This explains that American andRussian writings hold to Cold War as a concept fittingthe entire epoch. It allows for distinguishing betweenmoments of imminent war—Cuba, October 1962, andYom Kippur war, October 1973—times of hightensions and war outside Europe—Korea, 1950–53;Vietnam, 1964–75; Afghanistan, 1979–88—long inter-vals of deU tente, and even moments of concerted crisismanagement (June 1967 Near East war) or collab-oration (during the Laos crisis and Vietnam war in the1960s and ending wars in Africa in the late 1980s).

This version pays less attention to the fact thatthroughout the four decades not all parties to the ColdWar were involved on the same issue-at-stake and atthe same time. France joined Britain (UK) and the USin 1948 in founding the West German state as theWestern allies’ response to Stalin’s anchoring ofPoland, Bulgaria, Romania, Hungary, and Czecho-slovakia firmly into the Soviet security-zone, andmaintained staunch opposition to proposals for aneutralized, but nationally rearmed united Germany(Hitchcock 1998). Since themid-1960s, France becamethe spokesman for a Europe less dependent on the USand discussing terms of settlement with ‘Russia.’ TheFederal Republic of Germany (FRG) converted ColdWar into cold peace with the ratification of Ost- undDeutschland�ertraX ge (1970–3). Both governments co-operated in defending deU tente in and for Europe; byimplication, this aspiration for a European peace-zonerejected the American concept of the indivisibility ofthe global contest between East and West. The PRCh,the only principal ally the USSR ever had (1950–58}9),made its peace with the US in 1972 on Pejing’s term of‘One China.’ While the PRCh had provoked the USduring the 1950s to continuous nuclear sabre-rattling,it turned its antihegemonic posture throughout the1970s and 1980s against the USSR.

1.1 The Components of the Term

The definition refers to four aspects of the conflict: (a)the ideological antagonism between the Western con-cept of freedomof choice for the people of the domestic

2194

Cohort Analysis

Copyright # 2001 Elsevier Science Ltd. All rights reserved.International Encyclopedia of the Social & Behavioral Sciences ISBN: 0-08-043076-7