"To Carve Nature at Its Joints": On the Existence of Discrete Classes

Psychological Review1985, Vol. 92. No-3. 317-349

Copyright 1985 by the American Psychological Association, Inc.0033.295X/85/S00.75

"To Carve Nature at Its Joints": On the Existenceof Discrete Classes in Personality

Steve Gangestad and Mark SnyderUniversity of Minnesota

In principle, units of personality may be of two varieties: dimensional variables,which involve continuously distributed differences in degree, and class variables,which involve discretely distributed differences in kind. There exists, however, aprevailing and rarely questioned assumption that the units of personality arecontinuous dimensions and an accompanying prejudice against class variables.We examine this prejudice, the arguments that generated it, and those that upholdit. We conclude that these arguments are applicable to class variables as theyoften have been explicated, in phenetic terms; by contrast, genetically explicatedclass variables are not vulnerable to these arguments. We propose criteria forconjecturing and present methods for corroborating the existence of class variablesin personality. Specifically, we test a class model of a construct whose conceptualstatus makes it reasonable to evaluate whether or not the differences betweenindividuals represented by this construct constitute discrete classes. Finally, weexamine the implications for conceptualizing and investigating the nature andorigins of personality.

As a psychological concept, personality re-fers to regularities and consistencies in thebehavior of individuals and to structures andprocesses that underlie these regularities andconsistencies. Such phenomena, to the extentthat they exist, ought to distinguish individualsfrom other individuals and to render their

actions predictable. Typically, in personalitytheories, these distinguishing features havebeen treated as comparative individual differ-ences on the assumption that one can mean-

The research and preparation of this manuscript weresupported in part by National Science Foundation GrantBNS 82-07632 to Mark Snyder, in part by a NationalScience Foundation Graduate Fellowship to Steve Gan-gestad, and in part by a grant from the University ofMinnesota Computer Center.

We are grateful to Auke Tellegen for recommendingthe Monte Carlo simulation and valuable additionalexternal validation analyses; in addition, he and Allan R.Harkness provided helpful advice on conducting the

simulation. Also, we thank Jeffry A. Simpson for assistingin the empirical phase of the real variable control.

Finally, we appreciate the constructive comments of PaulE. Meehl, Auke Tellegen, Jack Block, William G. Gra-ziano, Allan R. Harkness. Jeffry A. Simpson, DaveSmith, and Jason Young.

Requests for reprints should be sent to Steve Gangestador Mark Snyder, Department of Psychology, Universityof Minnesota, 75 East River Road, Minneapolis, Min-nesota 55455.

ingfully compare any two individuals in termsof the same variables.

In principle, these personality variablesmay be of at least two varieties. On the onehand, there are dimensions or characteristicsthought to be possessed in some quantitativedegree by all individuals. As a result, thedistributions of such characteristics are con-tinuous ones. Dimensions hardly need furtherintroduction; they are well-known in the tra-ditional vocabulary of personologists as traits.Indeed, most existing comprehensive person-ality inventories purport to measure the basicdimensions of personality. For instance, dom-inance is a putative trait, assumed to bedistributed continuously and measured byone of the scales of the Personality ResearchForm (Jackson, 1974).

On the other hand, certain units of person-ality may not be dimensions, but rather, classvariables. Class variables are expressed notcontinuously, but as differences distributedinto discrete categories, and thus are differ-ences that can be introduced by the collo-quialism "There are two (or, for that matter,any finite number of) types of people in theworld." That is, when it comes to personality,class variables seek, as Plato put it, "to carvenature at its joints" by identifying true dis-cretenesses in personality.

317

318 STEVE GANGESTAD AND MARK SNYDER

If comparative individual differences canbe distributed either as continuous dimensionsor as discrete classes, then we may ask whetherany specific difference between individuals isproperly conceptualized as a dimension or aclass variable. Is, for instance, extraversion acontinuous dimension, possessed in some de-gree by all individuals (e.g., Eysenck, 1953),or a class variable, with individuals belongingeither to the discrete class of extraverts or tothat of introverts (e.g., Myers, 1962)? A littlereflection, however, reveals that the appropri-ate roles of continuous dimensions and dis-crete classes in personality rarely ever surfacesas an issue for theoretical or empirical inquiry.The reason, it seems, is that this fundamentalissue is almost universally a prejudged one.With few exceptions, the prejudgment is onecaptured by a statement borrowed from Mil-ton: "Differing but in degree, of kind thesame." Overwhelmingly, the basic units ofpersonality are presumed to be dimensions.

In this article, we first examine this pre-judgment, the arguments that generated itand those that uphold it. We conclude thatthere are cogent arguments against class vari-ables as they have commonly been explicated.We next discuss a different explication ofclass variables, one that is not vulnerable tothese same arguments. We then propose cri-teria for conjecturing and methods forcorroborating the existence of class variablesin personality. Specifically, we test a classmodel of a construct whose conceptual evo-lution makes it reasonable to evaluate whetherit identifies discrete classes of individuals.Finally, we examine implications of classmodels for understanding the nature and theorigins of personality.

Presumption of Dimensionalism and thePrejudice Against Class Variables

Most personologists immediately recognizethat the presumption of dimensionalism isno straw person. One need not look far tofind explicit statements of the preferred statusof continuous dimensions, for example, theassertion that

The method of choice in personality scale constructionis one which focuses on a linear relation between itemsand a single underlying latent continuum. Other ap-proaches, embodying more complicated models such as

. . . class models have the status of curiosities (Jackson,1971, p. 239).

Neither is it difficult to find explicit statementsof prejudice against class variables, for in-stance:

In a course at the New School for Social Research,Muhammed Ali was reputed to offer this typology:People come in four types, the pomegranate (hard on theoutside, hard on the inside), the walnut (hard-soft), the

prune (soft-hard), and the grape (soft-soft). As typologiesgo, it's not bad—certainly there is no empirical reason

to think it any worse than those we may be tempted totake more seriously (Mendelsohn, Weiss, & Feimer, 1982,p. 1169).

Of course, some typologies do tempt oneto take them seriously. Oddly enough, how-ever, some of their proponents endorse adimensional view of personality, not proposingthe existence of true discrete categories, butclaiming only that their proposed types reflectextreme ends of continuous distributions. Forexample, in Eysenck's (1953) typology ofextraversion and introversion, each type existsonly as an idealization defined by a clusterof attributes. Few, if any, individuals arethought to perfectly fit either idealization.Instead, Eysenck assumed that individualscan be assessed, for degree of fit to each ofthe ideals, along a continuous dimension.

Several classic typologies—such as those ofJung (1921/1923) and Kretschmer (1948)—also may be construed to postulate dimen-sional variables rather than true classes (cf.Eysenck, 1953). Thus, Jung (1921/1923)wrote, "Hence, there can never occur a puretype in the sense that he is entirely possessedof the mechanisms with complete atrophy ofthe other. A typical attitude signifies the mererelative predominance of one mechanism"(p. 9). And, contemporary theorists haveregarded typologies as idealized concepts ortheoretical fictions, whose usefulness is derivedfrom "heuristic value rather than . . . relationto real world examples" (Hogan, 1983, p.61). Seldom is any real discreteness claimed.It is thus difficult to argue with the claimthat, "The widespread notion that typologiesimply discontinuities, bimodal distributions,and the like, does not accurately reflect thewritings of modern typologists" (Eysenck,1969, p. 17).

It appears, then, that even typological^oriented personologists have joined with their

DISCRETE CLASSES IN PERSONALITY 319

nontypologist counterparts in the presump-tion that personality units come in quantita-tive degrees, rather than in true, discreteclasses. Of course, this presumption is notnecessarily unjustified. It is substantively jus-tified to the extent that theoretical or empir-ical considerations yield the conclusion thatreal class variables simply are not to be foundin personality. Accordingly, we ask: Whatmight constitute such considerations?

Arguments Against Class Variables

Some of these arguments can be traced totraditions in the measurement of psycholog-ical properties (cf. Tyler, 1965). Sophisticatedmeasurement methods were first developedfor cognitive abilities assumed to be quanti-tative variables and, thus, are heavily steepedin dimensional analysis. Development ofmethods for measuring class variables, bycontrast, proceeded more slowly (cf. Loevin-ger, 1957). As a result, methods for testconstruction and evaluation generally assumeunderlying latent continua, and personalityresearchers who adopt these methods mayalso adopt, without question, the underlyingassumption of dimensionality.

Factor analysis presents a similar situation.Although not restricted by an assumptionthat factors are continuously distributed di-mensional variables, the method has becomeassociated with uncovering basic dimensions,not class variables, and thus factors are typ-ically interpreted as dimensions. Thurstone(1947) himself encouraged this association:"The factors in psychological investigationsare not ordinarily to be thought of as ele-mental things which are present or absent,like heads or tails in the tossing of coins" (p.56). Hence, many factor analysts who believethey are uncovering primary units of person-ality also believe these primary units aredimensions.

These statistical traditions may be respon-sible for a methodological inertia that createsan unfavorable environment for class modelsof personality. There may, however, exist amore fundamental and pervasive reason whythe search for continuous dimensions hasbecome the "method of choice" in personality.This reason, we suspect, is a fundamentaldistrust of class variables ingrained in the

traditions of North American personality the-ory, a distrust rooted in the following (orsome similar) argument. The minor premiseof this argument states that generally we canmeasure basic, low-level, phenotypic charac-teristics of personality on continuously dis-tributed dimensions. If so, the major premisestates, to propose the existence of discreteclasses in personality is bound to be anoversimplification or, worse yet, an entirelyarbitrary fiction. Thus, one personality theo-rist has asserted that "simplicity limits their(class variables') value. Generally, an individ-ual's behaviors are so complex, diverse, andvariable that he cannot be sorted into asimplistic category or slot" (Mischel, 1976;p. 16). And, in a related vein, others haveargued that "Because personality variablesare distributed continuously and are inevita-bly subject to consensual definitions, person-ality types are bound to be arbitrary con-structions rather than discoveries of how to'carve nature at its joints' " (Mendelsohn etal., 1982, pp. 1168-1169).

But just how cogent are these arguments?Are class variables really unable to deliver ontheir promise "to carve nature at its joints"?The answer to this question depends criticallyon what form of classification system is underscrutiny. We may distinguish between twogeneral approaches to classification, only oneof which is vulnerable to criticisms of over-simplification and arbitrariness. The otherapproach—which is not so readily dismissedas leading to, at best, oversimplifications and,at worst, arbitrary fictions—may make itpossible to assert the genuine existence ofdiscrete classes in personality.

Approaches to Classification: Phenetic andGenetic Explications

One approach has been explicated phenet-ically (e.g., Gingerich, 1979; McNeil!, 1979;Ruse, 1973; Sneath & Sokal, 1973), withclasses denned by covarying phenotypic (ob-servable and first-order dispositional) char-acteristics that, when considered in a multi-dimensional phenotypic hyperspace, demon-strate densification or multimodality. Thequestion the pheneticist asks, in effect, is"How can I reduce the information in thephenotypic hyperspace to a few categories


such that similarity between personalitieswithin categories is maximized and differencesbetween personalities between categories ismaximized?" This classification process isundertaken primarily for descriptive conve-nience, with the classification scheme servinga summarizing function. As such, no under-lying meaning is directly ascribed to theconstructed classes. Thus, one who asks whatit means for individuals to belong to thesame class simply will be told that they sharerelatively many common features in the clus-tered hyperspace. Examples of phenetic ex-plication include attempts to delineate bio-logical species, phyla, and kingdoms; and,within personality psychology, Cattell's types,denned as modes or central locations ofdensification in uni- or multidimensionalspaces.

Probably, some notion of phenetic classi-fication was what personologists (at least thosewe have quoted) had in mind when arguingagainst class variables. For, the case againstphenetic explications of class variables inpersonality is a rather persuasive one. Phe-notypic characteristics in personality may notcluster very tightly. Thus, classifications basedon phenotypic characteristics may well beoversimplifications of the "real" state of af-fairs. And, given the multitude of currentlyavailable clustering methods based on differ-ent similarity indices that can yield quitedifferent results, phenetic classifications inpersonality may be, if not arbitrary, at leastquasi-arbitrary constructions (cf. Everitt,1974).

We are, thus, sympathetic to the argumentsagainst phenetically explicated class variablesin personality, arguments to which the phe-netic approach, by its very nature, may beintrinsically vulnerable (cf. Farris, 1979). But,there is a second approach to classification,one explicated in genotypic (rather than phe-notypic) terms. In this approach, to say thata true class exists is to say that all membersshare some single genotypic source of influ-ence upon their phenotypic characteristics.That is, a class exists when all members sharesome latent underlying entity, structure, orevent that affects their current behavior orother outward characteristics. The phenotypiccharacteristics of the class members will, notincidentally, tend to be similar in domains

where the common genotypic source is acausative influence, and this phenotypic co-variation may be the basis for inferring theexistence of the genotypic source. Neverthe-less, in contrast to the phenetic approach,phenotypic similarity does not constitute theexplicit definition of the classes. Rather, thelatent genotypic class variable's nature is onlyimplicitly defined by its causal relationshipto the phenotypic characteristics. The latentclass variable is instead defined as that some-thing that must exist to explain the patternof phenotypic covariation.

Viewed from the perspective of geneticexplication (not to be strictly identified withbiological genetics: Ashlock, 1979; Bock,1973; Mayr, 1969; Meehl & Golden, 1982),a class variable diners from a dimensionalone in that the latent variable is thought tobe distributed into discrete classes rather thanalong a continuous dimension. And, whenexplicated in the causal-theoretical terms ofthe genetic approach to classification, classvariables simply are not subject to the criti-cisms that they represent oversimplificationsof reality or that they are arbitrary construc-tions. Conjectured class variables explicatedas latent constructs either really exist or theydo not, and class variables explicated in theseterms either really exercise influence uponphenotypic characteristics or they do not.That is, these class variables purport to iden-tify real entities; truly, they purport "to carvenature at its joints." And we, at least, fail tosee how entities that really exist and thatexert real influences can represent oversim-plifications or arbitrary constructions.

Genetically Explicated Class Variablesin Personality

The genetic explication of class variables,we hasten to point out, should not be totallyalien to personality psychologists. For, it sim-ply identifies the class variable as a causal-dispositional construct, one that refers to anunderlying attribute or structure within theperson conjectured to influence some domainof observable behavior (Cronbach & Meehl,1955; Loevinger, 1957). Moreover, this geneticexplication of classification (which identifiesclass variables as causal-dispositional con-structs) is highly compatible with one of the


primary goals of personality (to identify un-derlying dispositional attributes that organizeand structure behavior; cf. Snyder & Ickes,1985). Yet, to our knowledge, class variablesin personality, with few exceptions, have beenexplicated not in terms of the causal-dispo-sitional genetic approach, but instead in termsof phenotypic characteristics or as extremevalues on dimensions. We suggest that, tounderstand why genetically explicated classvariables have not emerged in personalitytheorizing, one must examine the etiology ofpersonality variables.

Consider first the etiology of a continuousvariable. The ideal conception of the contin-uous variable as normally distributed is jus-tified if one believes numerous independentantecedent events have all contributed insmall amounts to its development, that is, ifwe believe a many-trial Bernoulli process hasgenerated the distribution of the variable (cf.Hays, 1973). Of course, this etiological modelis an idealization. Nevertheless, when oneinfers the existence of a dimensional con-struct, one infers a latent variable that itselfhas been influenced by numerous quasi-in-dependent causal events.

This diffuse process of etiology is to beexpected if personality development conformsto many of the dominant models of devel-opmental psychology. Of these, the most in-fluential has been the environmental learningaccount (see, for instance, Mussen, Conger,& Kagan, 1963). It proposes that generalizedbehavioral orientations develop across expo-sure to numerous relatively independentstimulus-response reinforcements, whose ef-fects are continuously distributed dimensionalcharacteristics. Other traditional accounts ofdevelopment (those stressing imitation, iden-tification, or observational learning; e.g.,Bandura & Walters, 1963) entail this samediffuse process of etiology. And, the predom-inant model within behavior genetics (theadditive influence of alleles at multiple genesites) also entails a diffuse etiology.

Class variables, however, are discretely dis-tributed and, thus, the diffuse Bernoulli pro-cess is not a possible etiological model. Re-cently, Meehl (1977) outlined several modelsof specific etiology (which exists when a spe-cific causal agent was present in the originsof an entity) for class variables. In its strongest

sense, specific etiology refers to operation ofa necessary and sufficient factor, or a neces-sary but not sufficient factor, which is itselfa discrete entity. A somewhat weaker formof specific etiology is the threshold effect,which occurs when the class variable is causedby an underlying continuous variable dichot-omized by some threshold, below which thevariable exerts one effect, and above which itexerts another effect. These forms of specificetiology are largely incompatible with theviews of personality development (emphasiz-ing diffuse etiological processes) that have, atleast until recently, dominated North Amer-ican psychology. Nevertheless, at least twoincreasingly significant traditions, each ofconsiderable relevance to developmental pro-cesses, could predict class variables in person-ality.

First, accumulating evidence suggests thatgenetic variation plays a significant, if notsubstantial, role in determining personalitydifferences (e.g., Henderson, 1982; Plomin,DeFries, & McClearn, 1980). Although itseems unlikely that any single Mendelizinggene has large effects on psychological param-eters, more complex epistatic and polygenicthreshold models can account for discretelydistributed physiological parameters (e.g.,Carter, 1969; Falconer, 1967). Quite possibly,these same models may account for discretelydistributed psychobiological parameters hav-ing wide effects on individual and socialbehavior. Second, theoretical accounts of sys-tematic and discontinuous structural changesduring critical periods (e.g., Hall, 1907;Kohlberg, 1969; Piaget, 1965; Werner, 1957)are also compatible with the development ofclass variables. These discontinuous structuralchanges may occur differently for differentindividuals, dependent on thresholds of ge-netically or environmentally determined vari-ables.

We do not deny the possibility that specificetiologies in personality may be rare events.It may well be that diffuse etiologies are thegenerally appropriate models of personalitydevelopment (after all, they have served thepurposes of developmental psychology for solong) and that specific etiologies are exceptionsto the general case. Moreover, we recognizethat, because of their possible rarity, somemay choose to refer to class variables as


curiosities. We, nonetheless, believe that, byvirtue of their origins through specific ratherthan diffuse etiologies, class variables mayrepresent especially intriguing and importantpsychological phenomena.

Admittedly, it still remains to be seenwhat, if any, particular class variables exertpervasive influences on human social behav-ior. Biological sex, a clear class variable, mayhave substantial impact on development;however, the precise domains of influence arestill open to question (Block, 1976; Maccoby& Jacklin, 1974). Several of the psychoses(e.g., schizophrenia; Meehl, 1962) have beenconjectured to have underlying class variablesas causes, but because their base rates areprobably low, these class variables may notaccount for much variation in behavior inthe general population. Let us now, then,address the critical issues of just how onegoes about conjecturing and corroboratingthe existence of a class variable in personality.

Conjecturing a Class Variable

One may, of course, conjecture a classmodel of any personality construct onechooses. Nevertheless, one should have atleast some minimal theoretical or empiricalreasons to conjecture a class model beforeproceeding to test one. At least two kinds ofreasons may suffice. First, one may havereason to believe that a particular specificetiology produces consistent behavioral man-ifestations. Such a reason provides an etiolog-ical springboard to a class model. Second,one may be aware of a theoretical networkof contemporaneous causal relationships thatspecifies differences between individuals inkind rather than in degree, or that proposesthat individuals possess discretely differentinternal structures that influence behavior.Such a reason constitutes a contemporaneous-theoretic springboard to a class model.

On the basis of these considerations, wewere able to identify a personality constructthat could be conjectured to exist as a classvariable, and to demonstrate that differencesbetween individuals in this conjectured classvariable are distributed into discrete catego-ries. The variable that we identified is onethat is known to have wide-ranging influenceson individual and social behavior. Not acci-

dentally, it is one with which we are highlyfamiliar. Specifically, we conjectured that amajor source of variance underlying the re-sponses to the Self-Monitoring Scale (Snyder,1974) and accounting for a major portion ofthe extensive network of observed relationsbetween responses to this measure and pre-dicted criterion behaviors (see below andSnyder, 1979a, 1979b) is a dichotomous classvariable.

According to theoretical analyses of self-monitoring (the first propositions of whichappeared in the mid 1970s; Snyder, 1974),individuals differ in the extent to which theymonitor, through self-observation and self-control, their expressive behavior and self-presentation. Some individuals, those high inself-monitoring, are thought to regulate theirexpressive self-presentation for the sake ofdesired public appearances, and thus arethought to act in ways that are highly sensitiveto social and interpersonal cues of situation-ally appropriate performances. Other individ-uals, those low in self-monitoring, are thoughtto lack either the ability or the motivation toso regulate their expressive self-presentation.They, instead, are thought to display expres-sive behavior that truly reflects their ownattitudes, traits, feelings, and other currentinner states.

A number of hypotheses follow from thesebasic initial propositions. Thus, for example,according to the self-monitoring construct,high self-monitoring individuals should bebetter able than low self-monitoring individ-uals to convincingly display a wide variety ofemotions without actually experiencing theemotions. Furthermore, the behavior of highself-monitoring individuals should be moresensitive than that of low self-monitoringindividuals to shifts in what constitutes"good" performance in a social situation.And, in behavioral domains where variationin what is socially appropriate across situa-tions is great, high self-monitoring individualsshould report greater behavioral variabilityacross situations than should low self-moni-toring individuals. By contrast, low self-mon-itoring individuals should exhibit greater con-sistency of behavioral expressions of feelingsand thoughts expected to be stable, and thusthey should show greater correspondence be-tween self-report measures of attitudes and


preferences and nontest behavioral indicatorsof these attitudes and preferences.

Research on self-monitoring, designed totest these and other hypotheses, typically hasemployed a 25-item true-false self-reportmeasure to identify individuals high and lowin self-monitoring. Sample items are "I wouldprobably make a good actor," "I guess I puton a show to impress or entertain people,""I have trouble changing my behavior to suitdifferent people and different situations," and"I would not change my opinions (or the wayI do things) in order to please someone elseor win their favor"; see Snyder, 1974, p. 531,for a complete list of the 25 items. In researchusing this measure, all of the hypothesesstated above, and many others, have receivedempirical support (e.g., Ajzen, Timko, &White, 1982; Becherer & Richard, 1978;Caldwell & O'Reilly, 1982; Danheiser & Gra-ziano, 1982; Ickes, Layden, & Barnes, 1978;Krauss, Geller, & Olson, 1976; Kulik &Taylor, 1981; Lippa, 1976, 1978a, 1978b;Lippa & Mash, 1979; Lippa, Valdez, & Jolly,1979; Lutsky, Woodworth, & Clayton, 1980;McCann & Hancock, 1983; Paulhus, 1982;Rarick, Soldow, & Geiser, 1976; Ross, Mc-Farland, & Fletcher, 1981; Shaffer, Smith, &Tomarelli, 1982; Snyder, 1974; Snyder, Ber-scheid, & Glick, 1985; Snyder & Cantor,1980; Snyder & Gangestad, 1982; Snyder,Gangestad, & Simpson, 1983; Snyder &Kendzierski, 1982a, 1982b; Snyder & Mon-son, 1975; Snyder & Swann, 1976; Snyder &Tanke, 1976; Tunnell, 1980; Tybout & Scott,1983; Zanna, Olson, & Fazio, 1980; Zucker-man & Reis, 1978). Further validation andextensions of the self-monitoring constructhave been reviewed by Snyder (1979a, 1979b)and Shaw and Costanzo (1982).

From the perspective of construct validation(e.g., Cronbach & Meehl, 1955), then, thereappears to be strong evidence for the existenceof a latent causal factor—as yet not directlyobserved—that partially accounts for a spec-ifiable, yet extensive and extensible, domainof observable behavior and that we call theself-monitoring variable. More recently, at-tempts have been made to explicate the natureof this causal factor and the larger causalnetwork within which it is embedded. Thus,for instance, Snyder and Campbell (1982)argued that individuals high and low in self-

monitoring are guided by different meta-controls of behavior in social contexts. Ac-cording to this view, the interactions of highself-monitoring individuals are guided bypragmatic organizing structures that lead oneto consider the self's performances and thestrategic appearances they generate. The in-teractions of low self-monitoring individuals,on the other hand, are purported to be guidedby principled organizing structures that leadone to consider the consistency of the self'sdeeds with one's private world of beliefs andfeelings.

These theoretical treatments implicitlysuggest, then, that the organizing structuresthat guide the expressive behavior of individ-uals high and low in self-monitoring differnot in degree, but in kind. That is, withinthese treatments are hints that underlyingself-monitoring propensities may be a causalfactor that is discrete, not continuous, innature. As such, these hints provide us witha contemporaneous-theoretic springboard forconjecturing a class variable. We thus havebeen led to ask the questions: Do there existtwo classes of individuals that differ in self-monitoring propensities, each homogeneousin the sense that the behavior of all individualswithin the class is affected by a shared etio-logical entity? Does there exist a specificdichotomous etiological factor or thresholdeffect operating in the development and thusthe eventual expression of these differing self-monitoring propensities that organize expres-sive behavior?

Evaluating the Class Model: Structural andExternal Components

How does one distinguish variables thatentail classes that are discrete, nonarbitrary,and real in nature from continuous quanti-tative variables? Generally, if a class variableexerts strong influence on some domain ofobservable events, then these events are dis-continuously distributed in the multidimen-sional hyperspace that they define. To theextent that unusual densification (or, in ex-treme cases, multimodality) can be detected,there exists empirical evidence consistent witha class model.

Unfortunately, however, precise mathemat-ical criteria for determining whether a class


variable exists cannot be specified withoutarbitrariness. Class variables simply do notmake themselves known in phenotypic hy-perspaces in distinct mathematical forms.Rather, they are manifested in varying degreesof discontinuity. Thus, setting a minimumdegree of densification to be regarded asevidence of a class variable is necessarilyarbitrary. Moreover, continuous, quantitativevariables may manifest the same mathemat-ical forms as class variables under certainconditions of sampling and measurement (seeMeehl & Golden, 1982, for examples).'

In light of these considerations, empiricalevidence beyond observed discontinuity isrequired for a variable to be construed as anatural class variable. Hempel (1965) assertedthat class concepts must be given systematicimport—that is, demonstrated to fulfill anexplanatory or predictive function—to beconstrued as natural classifications that "carvenature at its joints" (pp. 146-147). For ourpurposes, we suggest that a construct conjec-tured to be a class variable be embedded ina network of empirically supported relationsthat indicate that the construct does representa class variable. The empirical support shouldinclude a complete program designed to es-tablish construct validity of a classificationscheme purported to measure the class vari-able through the same means that a contin-uous quantitative variable is validated—interms of the consistency between theoreticallyderived and empirically observed relations oftest behavior with nontest behavior (Cronbach& Meehl, 1955).

Thus, our efforts to test a class model ofself-monitoring propensities involved a two-stage process. First, we examined a set ofevents within the domain of events relevantto self-monitoring—verbal responses to itemsappearing on the Self-Monitoring Scale—fordiscontinuities consistent with a class model.Second, we developed a classification schemebased on the observed discontinuities andexamined the relation between the classifi-cation scheme and nontest behavior relevantto self-monitoring. We refer to these twofeatures of our test of the class model as,respectively, the structural component andthe external component of the class model(cf. Loevinger, 1957).

The Structural Component of the ClassModel: The Strategy of Taxometric Analysis

Over the past two decades, Paul Meehl andhis colleagues (e.g., Golden, 1982; Golden &Meehl, 1979; Meehl, 1965, 1968, 1973; Meehl& Golden, 1982) have developed a numberof taxometric models designed to detect latentclass structures in sets of relevant empiricalobservations. These methods can be appliedwhen the state of knowledge allows one toconjecture the existence of a dichotomousclass variable and to supply a list of indicatorsbelieved to discriminate (admittedly only im-perfectly) between the two classes. Thesemodels thus were well suited to our presenttask. We had conjectured the existence of alatent class variable, and there clearly doesexist a set of indicators believed to discrimi-nate between the two classes, namely theitems of the Self-Monitoring Scale (Snyder,1974).

Maximum Covariance Analysis

These methods are based on a maximumcovariance model (Meehl, 1965, 1968, 1973;Meehl & Golden, 1982). One starts with asubstantive theory involving a conjectureddichotomous taxonomy and a set of indicatorspurported to discriminate between the classes.It is tentatively assumed that the indicatorsare more or less independent of one anotherwithin each class of individuals. That is, theindicators ideally intercorrelate in a sampleincluding individuals from both classes onlybecause the indicators discriminate betweenclasses. (This assumption is only tentativebecause it can be verified later in the proce-

1 Given that we have conjectured that self-monitoringis a dichotomous class variable, should we not expect the

distribution of self-monitoring scores to be bimodal?Although some theorists have denned class variablesexplicitly in terms of multimodality (e.g., Cattell, 1979),class variables viewed as latent entities do not necessarily,and perhaps only rarely, make themselves known throughbimodality of an indicator scale. In fact, it is a simpleexercise to demonstrate that bimodality occurs only whenthere exists very little overlap between latent classes onan indicator scale (e.g.t Murphy, 1964), a situation onecannot readily expect when self-report items are used asindicators (Meehl & Golden, 1982). Clearly, then, bimo-dality is too stiff a criterion for purposes of evaluatingthe class model.


dure.) If these conjectures and assumptionsare correct, then the identified indicatorsbehave in relation to one another in predict-able ways that are not predictable from adimensional model. The series of procedureswe applied were designed to assess the con-sistency between conjectured and observedbehavior of the indicators. We make no claimthat any single procedure is sufficient, initself, to identify a class structure underlyingthe set of indicators. Rather, it is observedconsistency across the entire series of proce-dures that provides evidence for the latentclass structure.

In extensive Monte Carlo runs, as well asin a real empirical trial, the maximum co-variance model has been shown to effectivelydetect real latent class structures, and not tospuriously detect class structures when noneexists (Golden & Meehl, 1973, 1979; Meehl,1978). Nevertheless, as a precaution againstspuriously inferring the existence of real em-pirical classes in the self-monitoring domain,we instituted two series of control analyses—first, a real variable control series and, second,a Monte Carlo control series. Each series ofcontrol analyses (to be discussed in fullerdetail) was designed to demonstrate that, inaddition to being able to detect a real latentclass structure when they should detect a reallatent class structure, the taxometric methodsfail to detect a latent class structure whenthey should fail to detect such a structure.

We first selected, on both empirical andintuitive grounds, eight self-monitoring items(Table 1) that would meet the assumptionsand requirements of the taxometric proce-dures. On empirical grounds, we consideredonly items that correlated relatively highlywith the rest of the measure, reasoning thatif the items are influenced by a latent self-monitoring class variable, then any individualitem that discriminates between the classesshould correlate with the entire measure. Onintuitive grounds, we attempted to not includetwo items that would correlate substantiallywithin the two conjectured classes. The gen-eral rule of thumb here was to not nominatetwo items of extremely similar manifest con-tent. Thus, the items "I have never been goodat games like charades or improvisationalacting" and "I find it hard to imitate the

Table 1Eight Self-Monitoring Items Selected forTaxometric Analyses

Itemno. Keying* Item

4 F I can only argue for ideas

which I already believe.6 T I guess I put on a show to

impress or entertainpeople.

8 T I would probably make agood actor.

12 F In a group of people 1 amrarely the center ofattention.

13 T In different situations and

with different people, Ioften act like very

different persons.18 T I have considered being an

entertainer.21 F I have trouble changing

my behavior to suitdifferent people anddifferent situations.

22 F At a party I let otherskeep the jokes andstories going.

• Keying indicates high self-monitoring direction of scoring.

behavior of other people" were not included,despite relatively high item-total correlations,because a similar item with a higher corre-lation, "I would probably make a good actor,"was included. Seven of the items included onthese grounds fell into the 9 best items (outof the original 25) in terms of high item-total correlation. (The two items above werethe exceptions.) The eighth item included, "Ican only argue for ideas which I alreadybelieve," had content quite different from theother items included. As well, it had a fairlyhigh item-total correlation.2

2 The reader may ask why we chose eight items. Asour discussion of the method reveals, either seven oreight items is probably the minimum number of itemsthat the method allows. And, given redundancies incontent within the set of items on the measure highlycorrelating with the total scale score, we would haverisked including too many item pairs with substantialwithin-class correlations had we included more than eightitems. Thus, although the choice of eight items wasnecessarily somewhat arbitrary, the choice was also rea-soned.

326 STEVE GANGESTAD AND MARK. SNYDER

In the first phase of the taxometric proce-dures, we selected two items, which we call /and j, from the pool of the eight conjectureditems. We constructed a 7-point scale fromthe remaining six items (all keyed in the highself-monitoring direction). Out of a sampleof 1,918 University of Minnesota studentswho had completed the Self-Monitoring Scaleas part of larger questionnaire sessions be-tween winter quarter, 1979, and fall quarter,1981, we formed seven subsamples, eachcomprising the set of individuals who ob-tained a given score on the 7-point scale.3

(Thus, one subsample consisted of all thoseindividuals who obtained a zero; a secondsample, all who obtained a one, etc.) Foreach sample, we computed the covariancebetween i and j (with i and j also keyed inthe high self-monitoring direction). If a classvariable underlies the responses to the itemsas conjectured, if the classes are of not toounequal size, and if auxiliary assumptions tobe verified later hold, then the seven samplecovariances between / and j plotted as afunction of the values on the 7-point scaleassociated with the samples should be peaked,maximal toward the middle values and nearerto zero toward the extremes. If no classvariable exists, no reason exists to expect apeaked covariance curve. See the Appendixfor a complete explication of this prediction.

We need not restrict ourselves to the resultsof a single covariance curve. We could—andwe did—repeat this procedure using differentpairs of indicators associated with their ap-propriate six-item scales. Thus, given all pos-sible pairs of the eight items, using with eachpair the scale composed of the remaining sixitems, we made 28 sets of observations thatwere largely independent and, to reduce sam-pling variation and the effects of small as-sumption departures, we aggregated acrossthese 28 sets of observations. To furtherreduce sampling variation and the effects ofassumption departures, we also (as recom-mended by Meehl, 1973) smoothed the av-erage of trie covariances as a function of the7-point scale by means of methods describedby Tukey (1977; in particular, Tukey's "3RH,twice" method of smoothing). All furtheranalyses, however, were performed on boththe smoothed and raw, unsmoothed covari-ance curves. The examination of these 28

covariance curves considered as a whole,then, constituted the first step of the taxo-metric analyses.

Figure 1 shows the average covariance be-tween two items (across the 28 pairs) as afunction of the score on the corresponding7-point scale that emerged from our analyses.This covariance curve averaged across all 28item pairs appears consistent with the oneexpected if a class model is the appropriatemodel for self-monitoring propensities. Thefunction is peaked toward the middle (values3 & 4) and relatively drooped toward theextremes.4

Estimates of Base Rates

Observation of covariance curves acrossthe 28 pairs consistent with a dichotomousclass model can only be taken as provisionalsupport for the class model. For us to justi-fiably conclude that a class model is appro-priate, we must observe additional evidencethat exhibits consistency with the model. If alatent dichotomous class variable does un-derlie the responses to the indicators, then aspecific parameter—the proportion of thepopulation belonging to one class and theproportion belonging to the other class—defines the distribution of the class variable.

Naturally, any set of mathematical for-mulae derived from the class model and used

3 Unless otherwise noted, all analyses reported in thisarticle are based on this same sample of 1,9IS individuals.We examined the distribution of self-monitoring scoresfor this sample. Although the distribution was not bi-modal, it was not normal either. Compared to a normalcurve, it bulged on its sides and was short near its mode,consistent with what we would expect if there existedtwo partly overlapping latent distributions. More formally,when we assessed the fit of the kurtosis of the distributionwith that of a normal curve, it was clearly different;Fisher's (1946) & = -.36, z = -3.21, p < .005.

4 Because a different 7-point scale was constructed foreach pair of items, the 28 covariances averaged to calculateeach data point in Figure 1 are not calculated fromprecisely the same groups of individuals. Any given six-item scale, however, shares at least four items with anyother scale, and thus, if two classes exist, the groups ofindividuals upon which the 28 covariances are basedshould be similar in their expected relative compositionof individuals from the two latent classes. Thus, theshapes of the covariance curves should be similar foreach of the 28 item pairs. For ease of presentation, wehave averaged the covariances as if they were calculatedfrom the same groups.


Figure 1. Covariance between self-monitoring items av-

eraged across all 28 item pairs as a function of corre-sponding 7-point scales. (Solid line represents smoothedfunction; broken line represents raw function.)

to estimate base rates in our sample (that is,proportions of high and low self-monitoringindividuals) should give approximately thesame estimates if two real classes do exist,given that these base rates are fixed. If realclasses do not exist, making the class modelinappropriate, and if different sets of mathe-matical formulae used to estimate the "baserates" (which themselves, in such an instance,would not exist) are not derivable from oneanother or from an interpretive text otherthan the class model, then consistency of theoutput of the different mathematical formulaemust be considered coincidence. Stated sim-ply, if different methods to estimate the baserates fail to provide consistent estimates, eitherthe class model is inappropriate or someauxiliary assumptions of the methods havenot been satisfied. Conversely, it is unlikelythat different, nonredundant methods wouldprovide consistent estimates if actual baserates corresponding to real empirical classesdid not exist. Thus, consistent base rate es-timates ought to increase faith in the classmodel.

One method (Method 1) of estimating thebase rates operates on the covariance curveyielded by the first stage of the taxometricanalysis (see Meehl, 1973; Meehl & Golden,1982). We applied this method to both theunsmoothed and smoothed covariance curvesto obtain two base rate estimates. We alsoapplied the method to the covariance valuesaveraged across all possible pairs involving agiven item for each item to obtain eightadditional estimates. These base rate esti-mates, although not fully redundant, are alsonot totally independent. Therefore, we appliedtwo additional methods to our sample toobtain two estimates that are largely indepen-dent of those yielded by the maximum co-variance method. The mathematical proce-dures of these two additional methods (Meth-ods 2 and 3) are described by Golden &Meehl (1979) and Golden (1982), respectively.

The three methods did yield remarkablyconsistent estimates of the base rates in oursample. Specifically, Method 1, when appliedto both the smoothed and the unsmoothedcovariance curves, yielded an estimated baserate of .42 for the high self-monitoring class.In addition, the eight estimates correspondingto each of the indicators were consistent andaveraged .40. Furthermore, and most impor-tant, Methods 2 and 3 provided estimateshighly consistent with those provided byMethod 1: estimates of .43 and .42, respec-tively. Table 2 provides a tabulation of allbase rate estimates.

A Real Variable Control: The Caseof Impulsivity

The peaked covariance curves and theconsistency of the base rate estimates areboth consistent with a class variable under-lying the self-monitoring indicators. We mustadmit that with a methodology examiningcomplex multivariate relations, some unex-amined formal aspects of the model maymake these results not fully unexpected evenunder a dimensional model.

Could it be, for instance, that despite thefact that the base rate estimate methods arenot derivable from one another, their resultsare actually highly constrained to be similarunder a dimensional model? And, is it alsopossible that despite the fact that peaked


Table 2Estimates of Class Base Rates?

Item no.

468

1213182122

AverageAveraged curveAveraged curve smoothed

1

.40

.42

.42

.39

.36

.44

.41

.38

.40

.42

.42

Method

2

.43

.43

.40

.39

.40

.42

.49

.45

.43

——

3

.43,43.39.40.39.41.49.44

.42

——

" Values given are estimates of base rate for high self-mon-itoring class. Estimates for low self-monitoring class canbe calculated by subtracting values from 1.

covariance curves are not expected undergeneral dimensional models, they are, in ourparticular instance, the result of the specificconfiguration of difficulty levels of, and inter-correlations between, our indicators? Moregenerally, is it possible that when the methodsshould fail to detect a latent class structure,they do not fail to detect a latent classstructure?

These possibilities are legitimate concernsand should be addressed. Fortunately, theyneed not be addressed solely on formalgrounds. They may be addressed empirically.Specifically, we performed two series of con-trol analyses. The first series applied thetaxometric methodology on real data tappinganother variable presumed to be dimensional.The question we asked here was, Can thetaxometric analyses fail to detect a classstructure underlying a set of indicators whenthe latent variable is presumably dimensionaland, thus, when the methods should fail todetect a class structure?

For these analyses, we chose the variableof impulsivity, a variable we had no reasonto suspect to be taxonic. As specific indicators,we chose items that load on the impulsivityfactor of the Eysenck Personality Inventory(Eysenck & Eysenck, 1968). We selected theseitems to match, as closely as possible, theself-monitoring items in terms of (a) averageintercorrelation between items; for self-mon-

itoring, r — .16; for impulsivity, r = .16; <b)range of intercorrelations between items; forself-monitoring, rs range from .01 to .41; forimpulsivity, rs range from .03 to .47; and (c)range of item difficulties; for self-monitoring,difficulties range from .29 to .72; for impul-sivity, difficulties range from .32 to .75.5

We administered the items to 934 Univer-sity of Minnesota undergraduates in psychol-ogy between spring quarter, 1979, and winterquarter, 1984, and applied the taxometricmaximum covariance procedures. Figure 2shows the covariance curve averaged acrossall possible 28 ij pairs. Clearly, in contrast tothe smoothed self-monitoring covariancecurve, this smoothed function demonstratesno peakedness.

Is it the case that peakedness would emergeif we examined only the covariances betweenthose items deemed, by an independentmethod, to best discriminate between classes?Method 2 of estimating the latent base ratesalso yields estimates of item validities (seeGolden & Meehl, 1979). On the basis ofthese estimates, we identified the half (4) ofthe items that best discriminates betweenpossible latent classes. We also went back tothe self-monitoring data and similarly iden-tified the four items that best discriminatebetween conjectured self-monitoring classes.We then calculated and averaged the covari-ance curves of ij pairs involving only theidentified items. As Figure 3 clearly illustrates,the peak of the covariance curve involvingthe best self-monitoring items is, as would bepredicted under a class model, markedly ac-centuated. In contrast, the covariance curveinvolving impulsivity items is, once again,not peaked.

Furthermore, the base rate estimates ofpossible latent classes underlying impulsivitywere not consistent. Thus, for instance, theestimates yielded by Method 1 applied to the

5 These eight items were "Are you usually carefree?"(keyed yes, Y), "Do you generally do and say thingsquickly without stopping to think?" (Y), "Do you oftendo things on the spur of the moment?" (Y), "Can youusually let yourself go and enjoy yourself a lot at a livelyparty?" (Y), "Do you hate being with a crowd who playjokes on one another?" (N), "Do you like doing thingsin which you have to act quickly?" (Y), "Can you easilyget some life into a rather dull party?" (Y), and "Do youlike playing pranks on others?" (Y).


COftRC9MNnNG 7-POINT SCALf

Figure 2. Covariance between impulsivity items averagedacross all 28 item pairs as a function of corresponding7-point scales. (Solid line represents smoothed function;broken line represents raw function.)

unsmoothed and smoothed covariance curveswere .36 and .34, respectively, whereasMethod 1 applied to covariance curves asso-ciated with individual items gave estimatesaveraging .63. Methods 2 and 3 yielded esti-mates of .53 and .49, respectively, and werethus inconsistent with each of the estimatesgiven by Method 1.

In sum, the taxometric methods appliedto indicators of impulsivity clearly did notindicate the existence of a latent class struc-ture. Thus, when the methods should havefailed to detect a latent class structure, theydid fail to detect a latent class structure.These findings thereby increase our faith thatthe taxometric methods appropriately de-tected a latent class structure underlying self-monitoring indicators.

A Monte Carlo Simulation

We also implemented a second series ofcontrol analyses. Perhaps support for thelatent class structure is attributable to theprecise configuration of self-monitoring itemdifficulties and interitem correlations such

that the response tendencies underlying theitems possess a multivariate normal distri-bution. Assuming a multivariate normal dis-tribution of the latent response tendencies tothe eight items (as is conventional in formallatent trait approaches; e.g., Lord & Novick,1968), we estimated the tetrachoric correla-tions between all possible pairs of responsetendencies. When factor analyzed, with unitiesin the diagonals, the correlations yielded eightprincipal components. To create a MonteCarlo sample that, if the actual distributionof response tendencies is multivariate normal,should be a facsimile of the actual sample,we generated 10,000 cases, each assignedeight normally distributed orthogonal prin-cipal component values. We then calculatedvalues of response tendencies for a given itemfor each case by summing the products ofprincipal component values and factor load-ings for that item. By then dichotomizingeach item response tendency at its difficultylevel within the real data sample, we created

SElF-MONiraflING

CORftESPONCMNG 7-POINT SCM-E

Figure 3. Covariance between best self-monitoring itemsand best impulsivity items as a function of corresponding7-point scales. (Solid lines represent smoothed functions.Upright triangles represent raw self-monitoring datapoints. Inverted triangles represent raw impulsivity datapoints.)


COflREtMMMNO 7-POINT SCALE

Figure 4. Covariance between simulation items averagedacross all 28 item pairs as a function of corresponding7-point scales. {Solid line represents smoothed function;broken line represents raw function.)

a sample with item characteristics (intercor-relations, difficulty levels) nearly identical tothose of our real sample, but generated by alatent dimensional model.

If the alternative explanation is appropriate,then the taxometric methods should yield thesame results when applied to the MonteCarlo sample and to the real data. In fact,the results were markedly different. The co-variance curve averaged across the 28 itempairs of the Monte Carlo sample was flattened,relative to the covariance curve averagedacross the 28 item pairs of the real data set.Figure 4 illustrates this covariance curve.Furthermore, base rate estimates were nolonger consistent. Method 1 yielded estimatesfor the base rate of the high self-monitoringclass when applied to the unsmoothed andsmoothed covariance curves of .33 and .29,respectively. The estimates provided by Meth-ods 2 and 3 were .42 and .45, respectively.

The results of the series of analyses appliedto the Monte Carlo data, then, indicate thatthe support for the existence of a latent classvariable underlying self-monitoring phenom-ena is not attributable to the specific item

difficulties and intercorrelations of the self-monitoring indicators.6 Once again, when thetaxometric methods should have failed todetect a latent class structure, they actuallydid fail to detect a latent class structure. Letus now return to an examination of furtherevidence relevant to evaluating the class modelconjectured to appropriately account for thelatent structure of self-monitoring.

Additional Considerations: Factor Analysisand the Class Model

The taxometric analyses have providedconsiderable corroboration for the conjecturethat the latent variable underlying responsesto the self-monitoring measure is a classvariable. We observed systematic variancebetween pairs of items consistent with theclass model, and we observed a consistent setof base rate estimates, as entailed by the classmodel. Moreover, two series of control anal-yses did not support alternative explanationsfor these findings.

If a class variable is, as we have claimed,a major source of covariation in the responsesto items on the self-monitoring measure, thenwe should be able to find this self-monitoringclass variable as a factor emerging from afactor analysis of the Self-Monitoring Scale.Furthermore, if the assumption of local in-

6 In constructing the Monte Carlo simulation sample,

we assumed that the latent response tendencies possesseda multivariate normal distribution. Technically, then, this

set of control analyses only addresses those alternativedimensional models for which response tendencies are

distributed, if not normally, at least quasi-normally.Therefore, the Monte Carlo analyses have unknown

relevance for other alternative models in which latentresponse tendencies are not distributed normally, forinstance, those for which response tendencies have mark-edly skewed distributions. Although this may appear to

be a possible limitation of the Monte Carlo controlanalyses, we should note that our other set of controlanalyses, the real variable control analyses involvingimpulsivity, is not hampered by such a limitation. Forthat set of control analyses, the distribution of responsetendencies possessed their naturally occurring form,whether normal, quasi-normal, or not normal at all.Furthermore, there is no a priori basis for believing that,whereas alternative models employing normally distrib-uted response tendencies underlying our item character-istics produced inconsistent base rate estimates, otheralternative models employing response tendencies notnormally distributed would produce consistent baserate estimates, as the real self-monitoring responses ac-tually do.


dependence holds reasonably, such that withinlatent classes items do not highly inlercorre-late relative to their correlation in the mixedsample, then the class variable should emergeas a general factor accounting for a relativelylarge amount of the common variance sharedacross items. In other words, the class variableshould emerge as the first unrelated factorextracted in a principal components analysis(Rummel, 1970).

The unrelated factor solution. To identifythis first unrelated factor, we performed aprincipal axes factor analysis with iterationto determine communalities (Statistical Pack-age for the Social Sciences, SPSS, subroutinePA-2) on the self-monitoring items respondedto by the 1,918 participants in the samplewe used for the taxometric analyses. Basedon the results of a scree test (Cattell, 1966),we extracted three factors. If the first unre-lated factor does, as conjeclured, closely cor-respond to Ihe self-monitoring class variable,Ihen Ihe faclor loading for a given ilemshould be a good eslimale of lhal particularitem's correlation wilh Ihe class variable (i.e.,ils abilily to predicl individuals' membershipin Ihe class of high self-moniloring individualsor in Ihe class of low self-moniloring individ-uals).

Using Bayes's theorem and informationobtained in the taxomelric analyses of Iheself-moniloring items (specifically Ihe esti-mated crude validities of Ihe items) we couldcalculate, for each individual, on Ihe basis ofhis or her responses to Ihe eighl selected self-moniloring items, the probability of belongingto the class of high self-monitoring individualsand, alternatively, the probability of belongingto the class of low self-monitoring individuals(see Golden & Meehl, 1979). Individuals witha probability of greater lhan .5 of belongingto Ihe class of high self-moniloring individualswere assigned to Ihe high self-monitoringclass. Those wilh a probabilily of greaterlhan .5 of belonging to Ihe low self-moniloringclass were assigned to the low self-monitoringclass. For 63% of our sample, we could be alleast 90% confident of their classification.Averaging across all individuals, we estimatethat 89% of classifications were correct ones.Given this taxomelric classification scheme,we could estimate each item's correlationwith the class variable by calculating the phi

Table 3Correlations Between Self-Monitoring Items andSelf-Monitoring Class Variable Estimated byTaxometric Procedures and Factor Analysis

Estimates

Itemno.

123456789

10111213141516171819202122232425

Taxometricprocedures

.25

.02

.14

.24

.28

.53

.06

.63-.01

.05

.06

.48

.25

.14

.04

.13

.13

.48

.07

.34

.30

.41

.17

.20

.11

Factoranalytic

.39

.04

.20

.24

.39

.48

.08

.59-.02

.02

.11

.45

.25

.28

.06

.22

.17

.41

.09

.49

.34

.45

.31

.30

.18

d'

.14

.02

.06

.00

.11

.05

.02

.04

.01

.03

.05

.03

.00• .14

.02

.09

.04

.07

.02

.15

.04

.04

.14

.10

.07

• Absolute value.

coefficient between the item and Ihe dichot-omous categorization of individuals intohigh self-moniloring and low self-moniloringclasses.

In Table 3 we presenl Ihe faclor loadingsof Ihe first unrelated faclor and the phicoefficienls derived from Ihe laxomelric anal-yses. As can be seen, Ihe orderings of Ihecoefficienls within Ihe two sels were remark-ably similar to one another. Indeed, a rank-order correlation between Ihe Iwo sels ofcoefficienls approached unity, r = .97.7

1 The taxometric classification scheme allowed us toassess the extent to which the auxiliary assumption of

independence of indicators within classes was violated.The items were largely independent within the subsamplesof high and low self-monitoring individuals, mean inter-item correlations being -.03 and -.02, respectively.Within the subsample of probable high self-monitoringindividuals, correlations ranged from -.16 to .21, with75% between -.1 and .1. Within the subsample ofprobable low self-monitoring individuals, correlationsranged from -.14 to .20, with 86% between -.1 and .1.


Thus, although we cannot claim that thefirst unrelated factor corresponds preciselyto the self-monitoring class variable, it doesappear that the self-monitoring class variableis expressed as the first unrotated factor. Ofcourse, given the means by which we selectedthe eight items for the taxometric analyses(on the basis of high item-total correlations),one cannot be too surprised that these twosets of coefficients do correlate strongly. Wethus do not suggest that this analysis consti-tutes independent evidence for the classmodel.8 Nevertheless, an understanding thatthe latent class variable, should it exist, isreflected as the first unrotated factor will leadus to nonredundant assessments of the classmodel on the basis of analyses involving therotated factors.

The rotated factor solution. Previous re-ports of factor analyses of the Self-MonitoringScale (e.g., Briggs, Cheek, & Buss, 1980;Furnham & Capon, 1983; Gabrenya & Arkin,1980; Sparacino, Ronchi, Bagley, Flesch, &Kuhn, 1983; Tobey & Tunnell, 1981) havenot discussed the possibility that a "true"axis corresponding to a "real" influence isthe general first unrotated factor. Instead,these reports have only offered discussions ofthe factor structure resulting from rotationin accordance with principles of simple struc-ture. Simple structure yields a location of theaxes (factors) that makes for easy descriptionof the factors, and thus lends interpretabilityto the factor structure (see Thurstone, 1947),but rotation to simple structure categoricallydoes not ensure identification of the "real"variables underlying the responses. That is,there is no indisputable justification for be-lieving that a rotated factor structure identifiesthe "real" underlying sources of variation(e.g., Eysenck, 1950; Guilford & Zimmerman,1963; Overall, 1964).

Nevertheless, the rotated factor structureis not at all uninformative. An oblique rota-tion (SPSS direct oblique criterion of simplestructure) of the factors extracted in oursample yielded three factors similar to thethree factors reported by Briggs et al. (1980).Because Briggs et al.'s analyses and reportingsof factor loadings are already available in theliterature and because their factor structurehas been employed by other investigators(e.g., Cheek, 1982; Riggio& Friedman, 1982;

Siegman & Reynolds, 1983; Sypher & Sypher,1983), we employed their recommended fac-tor scales for all analyses involving the factors.

These three factors are all positively cor-related with the first unrotated factor. Fromour perspective, then, these factors definethree content areas that discriminate the highself-monitoring class from the low self-mon-itoring class. We describe the three contentdomains captured by these factors in orderof their importance to the self-monitoringclass variable, as reflected by their correlationswith the first unrotated factor:

1. An expressive self-control content do-main. Items that load highly on this factorconcern ability to act, ability to play gamesthat require expressive control, and ability toconceal motives through expressive control(e.g., "I can look anyone in the eye and tella lie with a straight face"). Briggs et al. (1980)

8 This situation notwithstanding, it may be possible toprovide some such independent evidence. Maxwell (1971)noted that scores on a factor discriminating between twolatent classes may possess a bimodal distribution, provid-ing that several conditions exist, including the conditionthat the class means on the variables loading on thisfactor are substantially disparate. As Grove (1983) dis-cussed, this assessment is probably not very powerful forpurposes of testing a latent class model. Nevertheless, weexamined the distribution of factor scores on the firstunrotated factor. This distribution did appear to possesstwo distinct modes: a major mode at .3 standard deviationsbelow the mean and a second mode at .9 standarddeviations above the mean. If we assume that the midpointbetween these two modes is near the optimal cuttingscore for dissecting two latent distributions, we estimatebase rates of .38 and .62 for the high and low ends,respectively, base rate estimates very similar to the baserate estimates yielded by the taxometric analyses. As weshould also expect, the kurtosis of the distribution wassubstantially plalykurtic, Fisher's g2 = -.71, and signifi-cantly different from the kurtosis of a normal distribution,z = 6.31, p < .001. As a precaution against the possibilitythat the nature of this distribution is merely a functionof the fact that all variables factored are dichotomous,we examined the distribution of the first unrotated factoremerging from an analysis of the responses to the EysenckPersonality Inventory provided by our sample of 934individuals. This distribution did not possess kurtosissignificantly different from a normal distribution, ft =-.21, z = 1.34, ns. And, the distribution of the firstunrotated factor of the Self-Monitoring Scale was signif-icantly more platykurtic than was the distribution of thefirst unrotated factor of the Eysenck Personality Inventory,z = 2.53, p < .02. Thus, these analyses of the distributionof the first unrotated factor of the Self-Monitoring Scalemay provide some independent evidence for the claimthat this factor taps a latent class variable.


have interpreted this factor as acting ability.The estimated correlation between this factorand the class variable is .61.9

2. A social stage presence content domain.Items that load highly on this factor concernnot feeling awkward in public situations,being the center of attention in groups, andbeing a storyteller and jokester (e.g., "In agroup of people I am rarely the center ofattention"). Briggs et al. (1980) have inter-preted this factor as extroversion. The esti-mated correlation between this factor and theclass variable is .50.

3. An other-directed self-presentation con-tent domain. Items that load highly on thisfactor concern attempts to act in social situ-ations by displaying what others would likeor expect one to display (e.g., "I may deceivepeople by being friendly when I really dislikethem"). Briggs et al. (1980) have interpretedthis factor as other-directedness. The estimatedcorrelation between this factor and the classvariable is .32.

We should point out that in each case, webelieve that Briggs et al.'s interpretation ofthe content areas corresponding to the factorsis too broad. Thus, for instance, the itemsthat define the social stage presence contentarea, we believe, do not correspond to socia-bility or extraversion in general. Rather, asour interpretation connotes, they appear tobe concerned with feeling comfortable per-forming in typical group social situations.

But why do these three factors emerge? Ifall of the common variance were attributableto the self-monitoring class variable (i.e., ifthe self-monitoring measure were a "puretest," Lazarsfeld, 1950, with any pair of itemscorrelating only because they both discrimi-nate between the classes), then a factor anal-ysis of the items of the Self-Monitoring Scalewould, of course, yield a single factor, thatcorresponding to the self-monitoring classvariable. The items on the Self-MonitoringScale, however, do not correlate only becausethey discriminate between self-monitoringclasses. Some items having to do with similarcontent do correlate within classes. Theseitems should load on the same rotated factor.Items that load on different factors, we sug-gest, should not correlate within classes.

In fact, the average interitem correlationsfor items loading on separate factors are —.03

and -.01 within the classes of high and lowself-monitoring individuals. However, the av-erage interitem correlations for items loadingon the same factor are .11 and .12 within theclasses of high and low self-monitoring indi-viduals. It follows that each of the threefactors yields relatively internally consistentscales within each of the two samples: Withinthe sample of probable high self-monitoringindividuals, as = .45, .58, and .43 for thethree factors of expressive self-control, socialstage presence, and other-directed self-presen-tation, respectively; within the sample ofprobable low self-monitoring individuals,a's = .35, .64, and .51, respectively. (Giventhe small number of items on each subscale,these alphas are, in general, not unre-spectable.)

As revealed by their correlations with theself-monitoring class variable, these factorsthus serve as fallible indicators of the classvariable. And, as revealed by the relativelyhigh internal consistency of the factors withinthe samples of probable high and low self-monitoring individuals, these factors corre-spond to sources of variation in the itemsthat are partly orthogonal to the self-moni-toring class variable. Thus, we may speculatethat the three rotated factors correspond tothree dimensional variables that discriminatebetween the two classes and yet are, in part,independent of the class variable.

Of course, these speculations are all inferredfrom the taxometric analyses performed onthe eight items. Could we assess the conjecturethat the three factors are fallible indicatorsof a latent class variable quasi-independentlyof these analyses? We could nonredundantlyassess this conjecture by applying the taxo-metric method of maximum covariance, usingthe three factors as indicator variables. Onceagain, this method requires indicator variablesconjectured to discriminate between the twoclasses, but not correlate within classes. Toensure more orthogonal sources of variation,we removed the items that loaded on morethan one factor from the factor scales onwhich they loaded more weakly. In addition,

9 To estimate the correlations between the factors andthe class variable, we correlated each factor with thedichotomous variable constructed from the eight-itemclassification scheme.


OTHiR-DffiECTED SELF PRESENTATION

figure 5. Covariance between expressive self-control andsocial stage presence factors as a function of other-directed self-presentation factor. (Solid line representssmoothed function; squares represent raw data points.)

at least one of the variables must be able tobe sliced into a number of small intervals.We nominated the other-directed self-presen-tation factor to be this variable because,containing the largest number of items, itcan be sliced into the largest number ofintervals.

We separately calculated, for the 11 groupsof individuals obtaining each of the possiblescores on this factor, the covariance betweenthe other two factors. As predicted, and as isillustrated in Figure 5, a plot of these covari-ances revealed a curve peaked toward themiddle value of the other-directed self-pre-sentation factor scale. The estimated baserate of the high self-monitoring class yieldedby a method analogous to Method 1 was .43,remarkably close to the estimates given bythe three quasi-independent methods dis-cussed. The base rate yielded by the samemethod applied to the Tukey smoothed Co-variance curve was .44. Finally, as Figure 6shows, we estimated the latent distributions

of the two classes on the other-directed self-presentation factor to be very similar to thelatent distributions on this factor estimatedusing the taxometric classification schemederived from the eight-item analysis. Thesefindings thus provide yet more support forthe class model.10

The External Component of the ClassModel: Attributing Systematic Import

to the Class Variable

Clearly, the taxometric analyses indicatethat a dichotomous class model is highlyconsistent with the internal structure of themultinomial relations observed between self-monitoring items. The plotted covariancecurves have the form expected of a dichoto-mous class model. The estimated relativeproportions of the two classes based on threelargely nonredundant methods all convergeon approximately a 40/60 split of the classesof high and low self-monitoring individuals.Moreover, covariance analyses of factors pos-ited to be independent fallible indicators pro-vided yet additional convergent evidence.Considered together, these findings providesupport for the class model that is, indeed,strong.

Nevertheless, these findings are not yetsufficient to conclusively establish that wehave identified a class variable possessing the

10 Readers may wonder about the application of cluster

analyses to our data. We also applied a classical clusteranalysis to individuals distributed in the three-dimensionalspace formed by the three obliquely rotated factors.Specifically, we applied Ward's (1963) hierarchical clus-tering algorithm employing a criterion of minimumincrease in within-group sums of squares to 150 individ-uals on the basis of their factor scores (implemented bya source program derived from a routine of Anderberg,1973; central memory requirements of this programlimited the size of the sample we could cluster to about150). A number of evaluations indicate that Ward's(1963) clustering algorithm outperforms most, if not all,of the popular algorithms available in recovering knownmixtures (e.g., Blashneld, 1976; Kuiper & Fisher, 1975;Mojena, 1977). This analysis produced two classes with

base rates very similar to the base rates our othermethods yielded, .39 and .61. More impressively, theanalysis produced classifications of individuals that agreedwith 87% of the classifications produced by the taxometricmethods based on eight items. Classical cluster analysis,then too, corroborates our evidence for the appropriatenessof a class model.


200

150

100

50

\

10

OTHER-DIRECTED SELF PRESENTATION

figure 6. Latent class distributions on other-directed self-presentation factor estimated by maximumcovariance method applied to factors (solid lines) and maximum covariance method applied to eight items(broken lines).

sort of systematic import that is psychologi-cally meaningful and conceptually powerful.All of the analyses described thus far haveinvolved only verbal responses to self-reportitems observed in single test-taking sessions.To establish greater systematic import of theclass variable identified by the taxometricanalyses, and to further and even more strin-gently corroborate the class model, we needto examine empirical relations, derived fromself-monitoring theory, between responses tothe Self-Monitoring Scale and criterion vari-ables observed outside of the test-taking sit-uation. That is, we need to examine empiricalrelations to criterion variables with which theSelf-Monitoring Scale has been linked inempirical investigations. And, we must ask,are these relations consistent with the conjec-ture that a dichotomous latent class variableaccounts not only for relations between re-sponses to different self-monitoring items,but also for relations between responses toself-monitoring items and behavior observedoutside of the test-taking situation?

To address this question of the validity ofthe external component of the class model,

we first selected a number of empirical inves-tigations that involved phenomena clearlygrounded in theoretical considerations of self-monitoring processes. These studies cover arelatively wide range of self-monitoring phe-nomena and processes, involving distinctlydifferent manipulated independent variablesand measured dependent measures: (a) Snyderand Cantor (1980), Investigation 2; (b) Snyderand Gangestad (1982), Investigation 1; (c)Snyder and Gangestad (1982), Investigation2; (d) Snyder and Kendzierski (1982b), In-vestigation 1; (e) Snyder, Gangestad, andSimpson (1983), Investigation 1; and (f) Sny-der, Gangestad, and Simpson (1983), Inves-tigation 2.

If the class model is indeed appropriate,then, at a minimum, the classification schemeemerging from the taxometric analyses shouldbe able to predict criterion variables in thesestudies with success comparable to the full25-item self-monitoring measure. Thus, foreach participant in each investigation, wecalculated the probability of belonging to theclass of high self-monitoring individuals andthe probability of belonging to the class of


low self-monitoring individuals, using theclassification scheme based on the eight se-lected self-monitoring items that emergedfrom the taxometric analyses. Those with aprobability greater than .5 of belonging tothe high self-monitoring class were classifiedas high self-monitoring individuals, and thosewith a probability greater than .5 of belongingto the low self-monitoring class were classifiedas low self-monitoring individuals. We thenrecalculated the statistics (Fs or ft) that mostdirectly test the major hypotheses of thestudies using, as the measure of self-monitor-ing, the classification scheme in place of thefull 25-item Self-Monitoring Scale.

Table 4 displays the results of these reanal-yses. Clearly, the taxometrically derived eight-item classification scheme produced relationsalmost as strong, as strong, or even strongerthan the relations produced by the full Self-Monitoring Scale. It appears, then, that ameasure of the class variable detected by thetaxometric procedures effectively can stand

in for the entire Self-Monitoring Scale andcan account successfully for relations betweenthe full self-monitoring measure and behav-ioral criteria predicted by the theory of self-monitoring.

But there exist much stronger tests of theclass model. Indeed, the results of these testswere some of the most informative withrespect to appropriateness of the class model.The analyses discussed thus far demonstratethat the dichotomy produced by the taxo-metrically derived classification procedure isas related to the external criterion measuresas is the dichotomy produced on the Self-Monitoring Scale splits (either at the medianor sampling from the ends of the distribution).The most informative question regarding ex-ternal validation, however, is not addressedby these analyses. Specifically, what happensto the strength of the relations with theexternal criteria when we take into consider-ation the information thrown away by di-chotomizing?

Table 4Validation of the Self-Monitoring Class Model: Reanalyzes of Self-Monitoring Investigations

Effect size or % of varianceaccounted for by

classification scheme relativeDichotomized25-item self-monitoring

Study scale

8-item to full scalelaxumctric

classificationscheme Actual

Class modelpredictions*

Snyder & Cantor (1980)Investigation 2

mF(4, 53)F(l, 56)

Snyder & Gangestad (1982)Investigation 1

F(l, 117)Investigation 2

F(l, oo)

Snyder & Kenzierski (1982)Investigation 1

Snyder, Gangestad, & Simpson(1983)

Investigation 1'(28)

Investigation 21(58)

Average

2.838.29

11.49

5.894.24

11.90

3.93

2.53

2.424.69

11.23

6.693.83

9.43

4.21

2.48

1.49

2.70

1.75

1.14

0.88

1.51

1.58

1.32

1.62

2.44

2.20

1.11

1.35

1.67

" These values reflect corrections for the fallibility of the classification scheme; uncorrected values are not highlydissimilar.


It is the answer to this question for whichthe dimensional model and the class modelprovide clearly different predictions. Considerfirst the prediction derived from the dimen-sional model. If the latent variable underlyingresponses to the self-monitoring indicators isdimensional, and if the relations between theself-monitoring measure and external criteriaare essentially linear (as is generally assumed;Jackson, 1971), then the full self-monitoringmeasure should be more highly correlatedwith relevant external criteria variables thaneither the dichotomization of the full self-monitoring measure or the taxometric clas-sification scheme. On formal grounds, thisprediction is straightforward: A biserial cor-relation is as a rule greater than the pointbiserial correlation from which it is estimated.

Consider now the prediction derived fromthe class model. If there is relatively littlesystematic covariation between the self-mon-itoring measure and external criteria otherthan that attributable to a dichotomous classvariable, then to best predict external criteria,one should simply want to know to which ofthe two latent classes individuals belong. Thus,once one knows that an individual belongsto the high self-monitoring class, one shouldnot care whether the person scored 15 or 22.Similarly, once one knows that an individualbelongs to the low self-monitoring class, oneshould not care whether the person scored10 or 3. Indeed, saturating one's predictorvariable with this information only introduceserror and thus reduces the predictability ofcriterion variables. Using the taxometric clas-sification scheme based on the responses tothe eight items, we can classify individualswith considerable confidence (in fact, with anaccuracy rate estimated to be 89%). Thus, itfollows from the class model that the fullself-monitoring measure should be less highlycorrelated with relevant external criterionvariables than is the dichotomization yieldedby the taxometric classification scheme.

Each model, then, makes distinctly differentpredictions. Accordingly, we returned to oursix external validation studies and performednew sets of analyses. For each study, wecalculated the effect size achieved or percent-age of variance accounted for by the taxo-metric classification scheme, expressed as aproportion of the effect size achieved or per-

centage of variance accounted for by the fullscale." If a dimensional model is appropriate,these relative percentages of variance ac-counted for should be less than 1.00. Incontrast, if a class model is appropriate, theserelative percentages of variance accounted forshould be greater than 1.00.

The actual relative values are given inTable 4. Generally, they are well above 1.00,averaging 1.58. In all but one case, the taxo-metric classification scheme well outper-formed the full scale. These results, then,provide clear additional support for the classmodel.

The extent of this additional support canbe appreciated all the more when one com-pares the predicted values of these relativeproportions with the actually achieved values.If the class variable is the sole variable re-sponsible for the relations between the exter-nal critera and the self-monitoring measure,then these relative proportions should beapproximately equal to the reciprocal of theproportion of variance in the full scale sharedwith the class variable. A comparison of thesevalues predicted under the class model andthe actual values (see Table 4) reveals thatthey are very similar. Thus, these analysesnot only provide further evidence for theexistence of a latent class variable, but alsodemonstrate that this class variable can, byitself, largely account for the relations betweenthe entire self-monitoring measure and exter-nal criterion variables.12

" In three studies (Snyder & Cantor, 1980; Snyder &Gangestad, 1982, Investigation 1; Snyder & Kendzierski,1982), the predicted result involved differences betweencorrelations. In these cases, we calculated the effect sizesof these differences (i.e., z,, - za) for both the full scaleand the taxometric classification scheme. In the remainingthree studies (Snyder & Gangestad, 1982, Investigation2; Snyder, Gangestad, & Simpson, 1983, Investigations 1

and 2), the predicted result involved a single correlation.Here, we calculated the percentages of variance of thecriterion variable accounted for (r2) by both the full scaleand the taxometric classification scheme.

12 The performance of the rotated factors in predictingthe external criterion variables was also assessed. Thefactors were more weakly related to the nontest behaviorthan were either the full self-monitoring measure or theeight-item classification scheme. In no study did any ofthe factors outperform both the 25-item measure andthe taxometric classification scheme, despite the fact thatthe factors have reliabilities comparable to that of thefull scale.


Implications of Class Models in Personality

As we have seen, it is in principle possiblethat differences between individuals may beexpressed either as continuously distributeddifferences in degree or discretely distributeddifferences in kind. As we have also seen, this"in principle" situation notwithstanding, it iswidely presumed that such differences aremost appropriately thought of as continuousdimensions. This pervasive presumption isbolstered by an enduring prejudice againstthe viability of class models in personality, aprejudice rooted in part in claims that classvariables constitute oversimplifications ofreality or arbitrary fictions. We have arguedthat, although these claims may be well takenwith respect to phenetically denned classvariables, they do not apply to class variablespurported to represent genotypic sources ofcausative influence. Yet, even though geneti-cally explicated class variables ought to beparticularly appealing to personality re-searchers concerned with causal-dispositionalconstructs that underlie, organize, and struc-ture behavior, we were well aware that theexistence of no such class variable (with theexception of biological sex) exerting substan-tial influence on human social behavior inthe general population had been established.

We assert that the existence of one geneti-cally explicated, discrete class variable hasbeen demonstrated and its influence on hu-man social behavior corroborated. The struc-tural relations among individuals' responsesto items on the Self-Monitoring Scale arehighly consistent with those expected if alatent class variable were a major influenceon the responses. Moreover, the class modelderived from the self-report items served suc-cessfully as a predictor of nontest behaviorin diverse studies of self-monitoring. Finally,and equally important, the observations wemade are highly inconsistent with a purelydimensional model. Thus, we feel confidentthat a taxonic structure does underlie self-monitoring phenomena.

What, then, are the implications of thesefindings? We propose to begin by examiningthe empirical findings considered narrowlyand ask: What do they tell us about theconstruct and how it should be measured?Then, we proceed to issues of etiology and

development, and inquire: What new per-spectives does the class model provide whenit comes to conceptualizing development ofthe class variable and its manifestations? Fi-nally, we plan to address more general issuesin personality and pose the question: Whatare the implications of our inquiry into classmodels for defining strategies for understand-ing the nature of personality and the psy-chology of the individual?

Psychometric Implications of theClass Variable

Although all but one of the self-monitoringitems do correlate in the expected directionwith the latent class variable, there does existsubstantial variation in the size of these cor-relations. Among the best discriminatingitems are "I would probably make a goodactor," "I have never been good at games likecharades or improvisational acting," "In agroup of people I am rarely the center ofattention," "At a party 1 let others keep thejokes and stories going," and "I guess I puton a show to impress or entertain others."Items that perform moderately well concernacting like different persons in different situ-ations and attempting to be pleasant to others.

There do exist several items that discrimi-nate rather poorly, if at all, between theclasses. Several of these items concern insin-cere (e.g., "I sometimes appear to others tobe experiencing deeper emotions than I ac-tually am") or negative (e.g., "In order to getalong and be liked, I tend to be what peopleexpect me to be rather than anything else")behaviors. Whatever the reason may be, thefact that they do not perform well has impli-cations for measuring self-monitoring. Be-cause several items are not good discrimina-tors, we suggest an alternative measure of theself-monitoring class variable. We do notsuggest that the measuring instrument be theeight-item classification scheme that emergedfrom the taxometric analyses, largely becauseit is not a pragmatic research instrument. Touse it, one must know the probabilities ofclass membership for all 256 eight-item com-binations, and the probabilities may not gen-eralize across populations. Instead, we rec-ommend the 18-item measure presented inTable 5. It consists of only those items cor-


relating at least .15 with the latent classvariable, as estimated by first unrotated factorloadings.

This new measure has an internal consis-tency (a) of .70, higher than the internalconsistency of the original 25-item measure(.66 in our sample). Moreover, it is morefactorially pure than the original measure. Itsfirst unrotated factor accounts for 62% of thecommon variance, compared to 51% ac-counted for by the first unrotated factor ofthe 25-item measure. More important, totalscores on the new measure are uncorrelatedwith its second, relatively minor, unrotatedfactor, r = .03. By contrast, total scores on

Table 5The I8-Ilem Measure of Self-Monitoring

Item and Key

1. 1 find it hard to imitate the behavior of other

people. (F)2. At parties and social gatherings, [ do not attempt

to do or say things that others will like. (F)

3. 1 can only argue for ideas which I already believe.

(F)4. I can make impromptu speeches even on topics

about which I have almost no information. (T)5. I guess I put on a show to impress or entertain

others. (T)6. I would probably make a good actor. (T)7. In a group of people 1 am rarely the center of

attention. (F)8. In different situations and with different people, I

often act like very different persons. (T)9. I am not particularly good at making other people

like me. (F)10. I'm not always the person I appear to be. (T)11. I would not change my opinions (or the way I do

things) in order to please someone or win theirfavor. (F)

12. I have considered being an entertainer. (T)13. I have never been good at ga'mes like charades or

improvisational acting. (F)14. I have trouble changing my behavior to suit

different people and different situations. (F)15. At a party I let others keep the jokes and stories

going. (F)16. I feel a bit awkward in public and do not show up

quite as well as I should. (F)17. I can look anyone in the eye and tell a lie with a

straight face (if for a right end). (T)18. I may deceive people by being friendly when I

really dislike them. (T)

Note. Keying is given by either T (true) or F (false) inparentheses following item. High self-monitoring individ-uals tend to answer in the keyed direction; low self-mon-itoring individuals tend to answer in the alternative di-rection.

the original measure are mildly correlatedwith its second unrotated factor, r= .15. Fi-nally, the correlation between the new 18-item measure and the original 25-item mea-sure is .93.

Because the major underlying latent vari-able measured by this instrument is a classvariable, increasing scores represent higherprobabilities of belonging to the high self-monitoring class rather than progressivelygreater amounts of this latent variable. Thus,appropriate splits should be used to classifypeople. For our sample, we estimated thatthe probability of belonging to the high self-monitoring class was greater than .5 for thosewith scores of 11 or greater.

Developmental Origins of the Class Variable

Although these considerations of the rela-tions between self-monitoring items and theunderlying classes provide sketches of thebehavioral features associated with the twoclasses, they cannot fully explicate the latentclass variable. For, at least as far as we haveseen, such phenotypic manifestations are onlyfallible indicators of the class variable. Theexplicit nature is likely to be revealed only atsome latent underlying level.

Let us consider more closely at what latentunderlying level the true explicit nature ofthe class variable may exist. We need notassume that any underlying homogeneitywithin the classes necessarily exists at a leveldescribable in existing psychological terms.The homogeneity of the classes of biologicalsex, for instance, is identifiable with chro-mosome structure. The differing chromosomestructures play a necessary role in a complexcausal network, involving extensive sociallearning processes, that produces the behav-ioral characteristics associated with the sexes.Because the causal connections are not in-variant across individuals, however, the ho-mogeneity at the chromosomal level is notretained through the causal chain to the levelof phenotypic behavioral characteristics. Thepure homogeneity extends only as far as thechromosomal differences.

It is possible that the homogeneity of theself-monitoring class variable is similarlyembedded several levels deep in a causalnetwork producing phenotypic differences.


As with biological sex, perhaps there exists adichotomous structure, event, or thresholdeffect in the biological or psychological pastof persons that determines class membership.The resultant class variable then may play arole in a complicated causal network thatinfluences several domains of phenotypiccharacteristics (e.g., expressive self-control,social stage presence, other-directed self-pre-sentation). As with biological sex, however,the level of the specific etiological factor maybe the only level at which homogeneity ofthe classes exists.

Etiology of the Class Variable

Aside from brief speculations in an earlytreatment (Snyder, 1974) and in a recenttreatment (Snyder, 1983b) of self-monitoring,the question of development of differences inself-monitoring propensities remains virtuallyunanswered. In speculating that a specificetiological factor may be responsible for theclass structure detectable in adult behavioralmanifestations of self-monitoring, we do notsuggest that this specific factor directly pro-duces, in full, all of these same behavioralmanifestations in childhood. More likely, aspecific factor (assuming one exists) producesdifferences in a circumscribed domain ofbehavior that then, through a process ofdivergent causality, produces increasingly largedifferences over time. Divergent causality oc-curs when initially small differences betweenindividuals become amplified or extendedover time, producing larger differences be-tween individuals in similar domains or inadditional domains (Langmuir, 1943; London,1946; Meehl, 1978; cf. Waddington's, 1957,discussion of the epigenetic landscape). Withrespect to self-monitoring, perhaps initiallysmall differences in restricted domains existin children, and then snowball into the ex-tensive differences in wide-ranging domainsobserved in adults.

First, consider candidates for the specificetiology. One such candidate is a biologicalgenetic structure. Although class determina-tion cannot be so simple as to be traceableto a single Mendelian gene, it is possible thatit corresponds to a threshold character or anepistatic configuration of alleles at indepen-dently segregating sites. A second possibility

is that the class variable corresponds to dis-tinct behavioral strategies acquired by chil-dren (perhaps at some critical age) in responseto parenting styles, peer pressures, siblingrelationships, or in a combination of theseand other environmental events. Adoption ofone of two strategies may be sensitive to athreshold on some specific environmentalparameter, such as amount of attention re-ceived from caretakers. More specifically, per-haps children who receive little attentionfrom their caretakers adopt a strategy designedto gain the attention and regard of others—a high self-monitoring strategy. Finally, thespecific etiology may involve a combinationof environmental and genetic factors. Perhapsadoption of one of two strategies is sensitiveto a threshold on a specific environmentalparameter, but the threshold value variesacross individuals and is largely geneticallydetermined.

We do know of one twin study (Dworkin,1977) that examined genetic influences onself-monitoring. It is interesting that themonozygotic within-pair variance on the Self-Monitoring Scale was similar to the varianceon the Self-Monitoring Scale we estimated toexist within each of the two classes. Thedizygotic within-pair variance was, by con-trast, more than twice as great as monozygoticwithin-pair variance and similar to the vari-ance we estimated to exist in mixed collegeor adult populations. These results, then, areconsistent with the possibility that the classstructure can be largely traced to geneticorigins of an epistatic nature.

Domains of Initial Manifestations

Assuming some specific etiology or thresh-old effect does determine class membership,across what parameters of psychologicalfunctioning are initial differences produced,and what are their behavioral manifestations?Recent research on individual differences inlanguage acquisition and characteristic sym-bolic styles suggests one intriguing possibility.Nelson (1973, 1981) and, subsequently, others(e.g., Horgan, 1981; Ramer, 1976; Starr, 1974)observed two relatively distinct patterns inthe acquisition of language by children atages 2-3. Some children, referred to as ref-erential, tend to first learn to use language as


a referential system that allows one to com-municate actual happenings in the world. Forinstance, they develop a large vocabulary ofreferential nouns. Other children, referred toas expressive, tend to learn early on the socialuses of language. They learn that the pattern,structure, and intonation of linguistic expres-sions are dependent on social context, andthey learn that language is a means to obtainattention from others.

Wolf and Gardner (1979) suggested thatthese differences in language acquisition areassociated with differences in general symbolicactivities. Thus, expressive children (whomWolf and Gardner call dramatists) focus onthe social structure of events, learning at anearly age that different persons on differentoccasions can play the same role in the samesocial situation, and that persons can exchangeroles in play. Dramatists also like to telldramatic stories, and they are able to displayelaborate forms of imitation.

The behavioral overlap between the refer-ential-expressive distinction in children andthe adult manifestations of the latent self-monitoring class variable is striking. Bothexpressive children and high self-monitoringadults are relatively sensitive to considerationsof social context and are dramatists in theirsocial performances, being sensitive to theattention that dramatic performances gainand skilled at imitating others. In contrast,both referential children and low self-moni-toring adults tend to be relatively insensitiveto social context, and nondramatic. On thebasis of content validity, we have been led toentertain the possibility that the two sets ofindividual differences are manifestations ofthe same latent variable expressed at differentages.

Processes of Divergent Causality

Initial manifestations of the class variablemay involve propensities to learn varyingfeatures of the social world, later expressedas actual differences in social knowledge.Similarly, initial differences in potential toacquire role-playing skills eventually may beexpressed as actual differences in these skills.Moreover, these skill differences eventuallymay be reflected in behavioral repertoires,which then may be reflected in actual behav-

ioral enactments. Through a process of self-perception (Bern, 1972) these differences inbehavior may then be reflected in differencesin beliefs about one's self and the guidingprinciples behind one's actions (cf. Snyder &Campbell, 1982). Thus, substantial divergentcausality may occur through normal processesof acquiring knowledge and skills.

There does exist another scenario by whichdivergent causality could occur. Consider thecase of a dichotomous etiological factor thatinfluences individuals' choices of situationssuch that individuals gravitate toward envi-ronments conducive to the behavioral mani-festations of the etiological factor. By virtueof spending time in situations that are partic-ularly conducive to their initial behavioraltendencies, individuals develop behavioralrepertoires and skills increasingly consistentwith their initial tendencies. Over time, then,individuals differing with respect to the etio-logical factor increasingly differ with respectto behavioral repertoires and skills (for elab-oration of this analysis of choosing situations,see Snyder, 1981, 1983a; cf. Plomin, DeFries,& Loehlin, 1977).

It has, in fact, been demonstrated thatindividuals high and low in self-monitoringtend to enter into or create situations and tochoose friends in ways conducive to theexpression of their own behavioral orienta-tions (e.g., Snyder & Gangestad, 1982; Snyder,Gangestad, & Simpson, 1983; Snyder &Simpson, 1984). Certainly children of grade-school age exercise considerable control overtheir choices of friends as activity partners.Perhaps children belonging to the high self-monitoring class are more sensitive to choos-ing activity partners who are skilled at theparticular activity. And, perhaps children be-longing to the low self-monitoring class aremore sensitive to choosing activity partnerswho are well liked and similar to the self. Asa result of spending time with their friends,then, high self-monitoring childrens' behav-ioral repertoires may become increasinglydiverse, whereas low self-monitoring childrens'behavioral repertoires, in contrast, may be-come increasingly uniform, thus creating fur-ther divergence in the types of interpersonalsituations in which high and low self-moni-toring children find themselves. In like fash-ion, other differentiating choices could create


additional amplified manifestations of theclass variable.

We have devoted extended discussion todivergent causality because identification ofa class variable with a specific etiology invitesan explicit, systematic investigation of diver-gent causality. Once a specific etiology hasbeen determined, one can classify individualsand, in principle, determine manifestationsof the class variable across time by recurrentlysampling behaviors and abilities from widedomains over a long course of development.Clearly, then, an understanding of self-moni-toring phenomena can substantially benefitfrom identification of an underlying latentclass variable. Understanding the explicit in-ner nature of self-monitoring will benefit tothe extent that the homogeneity of the classvariable is identified. Understanding the rootsof self-monitoring will benefit to the extentthat a specific etiology or threshold effectdetermining class membership is identified.And, understanding the development of self-monitoring will benefit to the extent thatincreased differentiation between the classesover time is identified. Moreover, identifica-tion of the class variable not only raisesquestions whose answers will be informative,but also raises questions that are, in principle,highly researchable. Thus, identification ofthe class variable can produce a metamor-phosis of understanding, resulting in a con-fluence of assets rarely seen in personology.At the same time that an understanding ofself-monitoring gains the elegance and sim-plicity entailed by a dichotomous class vari-able, it gains content and precision.

Implications for Understanding the Natureof Personality

The issues we have raised about self-mon-itoring as a class variable can be raised aboutany class variable in personality. Given iden-tification of a class structure, we may ask:Where does the true homogeneity in the classvariable lie? Do the classes have a specificetiology? Do they result from a thresholdeffect? Does the etiological factor initiallyaffect some circumscribed domain of psycho-logical functioning that, through a process ofdivergent causality, becomes amplified or ex-tended over time? These questions gain special

significance—or only become applicable—when they concern a class variable. Theygenerally are not applicable or are not feasiblyresearched when the variable of concern isnot a class variable.

Two additional questions, which we didnot explicitly address with respect to self-monitoring, may be addressed with respectto other class variables in personality. First,we may ask, if the variable is the result of athreshold function, is the function a "pure"or an "impure" one? If it is not pure or"tight," a few individuals will end up in anintermediate range between the two classes.Thus, the classes will have fuzzy edges. Wesuspect that the etiological and developmentalconsiderations that are appropriate when oneis dealing with fuzzy classes are more similarto those appropriate for pure classes than tothose appropriate for normally distributeddimensions. Thus, fuzzy classes have importsimilar to that of pure classes, provided theclasses reveal true divergences in psychologicalfunctioning.

There does exist a type of fuzzy classesthat does not reveal true divergences. Recog-nition of this type of classes brings us to asecond question: Are the classes environmentalmold types (Meehl, 1972; cf. Cattell, 1979)?Environmental mold types exist not becausean important psychological parameter dis-cretely differentiates between classes. Rather,they exist because the structure of the envi-ronment dictates divergence. Thus, there arethe classes of college students and noncollegestudents, not because of some discrete diver-gence in internal structure that distinguishesthe two classes, but because of the structureof an educational system. Clearly, many ofthe intriguing issues concerning etiology anddevelopment that one can address with respectto class variables reflecting true divergencesare not applicable to environmental moldtypes.

Consideration of Methods of Discoveryin Personology

How does one go about discovering real,causal latent entities and dispositions in per-sonality? This methodological question hasbeen around about as long as has theoreticalinquiry into personality, and is one for which


we do not pretend to have a failsafe answer.The present effort to discover a real causalentity, however, does have implications foranswers to this question.

It is probably the case that the most fre-quently used method to discover real causalinfluences in personality—whether that goalis explicitly or implicitly stated—is factoranalysis (e.g., Cattell, 1978). As a data-reduc-tion tool, of course, factor analysis has manyuses, each of which may involve one or moreassumptions justifying its use. For use as ameans to discover real causal entities ordispositions underlying a collection of overtbehaviors, Cattell (1978) offered the followingcosmological assumption as justification forinterpreting factors rotated to minimize somecriterion of simple structure as real causativeinfluences: Given a set of overt behavioralvariables having multiple causative influences,any one underlying influence substantiallyaffects only a portion of the overt variables.It is also often assumed that any one overtvariable is likely to be influenced substantiallyby only one real underlying variable.

But how does one know that these as-sumptions are correct? In particular, howdoes one know that these assumptions arecorrect for the specific variables and thespecific sample one has selected? Two prom-inent factor analysts have aptly answeredthese questions in stating "As a general rule,the simple structure model may serve as thebest objective guide to the location of thebasic dimensions of human nature. There islittle in the way of proof, however, eitherlogical or empirical, that this is true" [italicsadded] (Guilford & Zimmerman, 1963, p.290; see also Eysenck, 1950).

Thus, identification of the real underlyinginfluences through factor analysis and rotationto simple structure depends in part on gettinglucky. If—luckily—the particular set of vari-ables of concern are consistent with the as-sumptions basic to simple structure rotation,then one discovers the real underlying sourcesof influence through factor analysis and ro-tation to simple structure. If these circum-stances do not hold, however, the real under-lying influences are not discovered in thesimple structure.

We should note that Cattell (1978) himselfhas recognized limitations to the assumptions

underlying the use of simple structure toidentify real sources of influence:

Simple structure requires a foresighted choice of thesample of variables. It is obviously absurd to expect toget any determinate rotation of a certain factor X if wehave chosen the variables in the study such that probablyall of them will have some significant loading on X.

(p. 112)

In this regard, do the present taxometricprocedures differ from the use of simplestructure factor analysis? To find a real sourceof influence through the application of thetaxometric methods, there must be, underly-ing a set of phenotypic variables, a realsource of influence that is a dichotomousclass variable. There also must be identifiablephenotypic variables that generally discrimi-nate between the two classes. Finally, thephenotypic variables identified must nothighly correlate within the two classes. Clearly,finding a real source of influence underlyinga set of phenotypic variables using the taxo-metric methods requires a modicum of for-tuitousness no less than does a successfulapplication of factor analysis.

But, we quickly add, there is one significantdifference. When this modicum of fortui-tousness is not forthcoming, the factor-ana-lytic simple structure solution itself does nottell us that the real underlying sources ofinfluence are not to be found in the rotatedsolution. The method itself does not distin-guish between successful and unsuccessfulfinds. Thus, as a result, one may be led toincorrectly believe that real sources of influ-ence have been found in the rotated solution(or that no real influences, outside of thoserepresented in the rotated solution, exist).

In contrast, the taxometric methods informtheir users that they have not found a realsource of influence when they have not re-ceived that modicum of fortuitousness—whenno real class variable exists, when the phe-notypic variables they have selected are notsufficiently influenced by the class variable,or when the phenotypic variables intercorre-late substantially within the classes. In thesecircumstances, the taxometric procedures willnot yield the predicted peaked covariancecurves, the consistent base rate estimates, anda classification scheme that will increase pre-dictability of external criterion variables (seeMeehl, 1978, 1979; Meehl & Golden, 1982).


Thus, the taxometric methods do allow theirusers to distinguish unsuccessful from suc-cessful finds.

Of course, we do not mean to suggest thattaxometric methods be applied indiscrimi-nately. Clearly, they are most appropriatelyapplied when there exists some reason tobelieve that a specific class variable exists.Also, we do not mean to suggest that factoranalysis and rotation to simple structureshould never be used. Even if and whenrotation to simple structure does not revealreal underlying sources of influence, rotationto simple structure generally does facilitatedescriptive interpretability of the factors. Forpursuing purely instrumental ends (e.g.,atheoretical prediction), researchers may re-quire no more than the convenience of easeof interpretability that simple structure offers.At the same time, however, researchers shouldbe aware that rotation to simple structure(or, in fact, rotation to any structure selectedon the basis of an a priori assumption) cannotbe counted on to locate nature's true joints.Indeed, the following statement made byCronbach and Meehl (1955) is as true todayas it was the day it was written: "Factors mayor may not be weighted with surplus meaning.Certainly when they are regarded as 'realdimensions' a great deal of surplus meaningis implied, and the interpreter must shouldera substantial burden of proof" (p. 63).

Prevalence of Class Variables in Personality

At the beginning of this article, we notedthat there exists in personology a widespread,implicit if not explicit, presumption thatdifferences between people are differences indegree, not differences in kind. We arguedthat many of the arguments against the ex-istence of discrete class variables in personalityare not applicable when class variables areexplicated genetically—as class variables that"carve nature at its joints"—rather than phe-netically. Still, we admitted, there had yet tobe demonstrated the existence of a classvariable (with the exception of biological sexand psychopathologies of relatively low inci-dence) that exerted important influences onsocial behavior. We believe that we are nowin a position to claim that one exception hasbeen demonstrated—self-monitoring does give

every indication of existing as a discretelydistributed class variable.

Discreteness in personality, then, is possibleand can be demonstrated to exist. However,by taking seriously the existence of classvariables in personality, a new question arises.How widespread are true class variables inpersonality? How common are personalityetiologies involving specific dichotomous fac-tors or threshold effects? It is, of course,possible that true class variables are abundant,that discreteness in personality developmentand organization are as much the rule as theexception. But we are well aware that theidentification of even a single class variablewith important influences on social behavior(excepting biological sex) may come as some-thing of a surprise to many personologists.Surely, much past observation, from whichdiscrete variables have been quite absent,speaks otherwise. But then, it may be thatwe have too long used yardsticks that havenot allowed us to detect the discreteness thatexists.

Of course, it is also possible that discretelydistributed class variables are rare phenomenain personality. In this respect, self-monitoringpropensities may reside in select company.This possible rarity, however, surely shouldnot blind us to seeing the value of identifyinga class variable where a class variable trulyexists. In the case of self-monitoring, identi-fication of a class variable has led us to askand pursue fundamentally different questionsand research strategies, as well as to developradically different conceptual frameworks ofetiology and development. So too with thecase of any other class variable in personalitywould one be compelled by its identificationto formulate and pursue different questionsand research strategies. Only by recognizingthe class variables that exist will these ques-tions be asked and answered. Only by askingand answering these questions will it be pos-sible to discover discreteness where it trulyexists in personality. Only by discovering thedomains of personality where true discretenessexists in personality will it be possible "tocarve nature at its joints." And, not inciden-tally, taking seriously the existence of discreteclasses in personality may even make it pos-sible to take seriously claims of "There aretwo types of people in the world."


References

Ajzen, I., Timko, C, & White, J. B. (1982). Self-monitoring and the attitude-behavior relation. Journal

of Personality and Social Psychology, 43, 426-435.Anderberg, M. R. (1973). Cluster analysis for applications.

New York: Academic.Ashlock, P. D. (1979). An evolutionary systematist's view

of classification. Systematic Zoology, 28, 441-450.Bandura, A., & Walters, R. H. (1963). Social learning

and personality development. New York: Holt, Rinehart& Winston.

Becherer, R. C., & Richard, L. M. (1978). Self-monitoringas a moderator of consumer behavior. Journal of

Consumer Research, 5, 159-162.Bern, D. J. (1972). Self-perception theory. In L. Berkowitz

(Ed.), Advances in experimental social psychology (Vol.6, pp. 1-62). New \brk: Academic Press.

Blashfield, R. K. (1976). Mixture model tests of clusteranalysis: Accuracy of four agglomerative hierarchicalmodels. Psychological Bulletin, 83, 377-388.

Block, J. H. (1976). Issues, problems, and pitfalls inassessing sex differences: A critical review of the psy-chology of sex differences. Merrill-Palmer Quarterly,

22, 283-308.

Bock, W. J. (1973). Philosophical foundations of classicalevolutionary classification. Systematic Zoology, 22,

375-392.Briggs, S. R., Cheek, J. M., & Buss, A. H. (1980). An

analysis of the Self-Monitoring Scale. Journal of Per-sonality and Social Psychology, 38, 679-686.

Caldwell, D. E, & O'Reilly, C. A. (1982). Responses tofailure: The effects of choice and responsibility onimpression management. Academy of Management

Journal, 25, 121-136.

Carter, C. O. (1969). Genetics of common disorders.British Medical Bulletin, 25, 52-57.

Cattell, R. B. (1966). The scree test for the number offactors. Multivariate Behavioral Research, 1, 140-160.

Cattell, R. B. (1978). The scientific use of factor analysis

in behavioral and life sciences. New York: Plenum.Cattell, R. B. (1979). The structure of personality in its

environment. New York: Springer.Cheek, J. M. (1982). Aggregation, moderator variables,

and the validity of personality tests: A peer-rating

study. Journal of Personality and Social Psychology,43, 1254-1269.

Cronbach, L. J., & Meehl, P. E. (1955). Construct validity

in psychological tests. Psychological Bulletin, 52, 281-302.

Danheiser, P. R., & Graziano, W. G. (1982). Self-moni-

toring and cooperation as a self-presentational strategy.Journal of Personality and Social Psychology, 42, 497-

505.Dworkin, R. H. (1977, August). Genetic influences on

cross-situational consistency. Paper presented at theSecond International Congress on Twin Studies, Wash-ington, DC.

Everitt, B. S. (1974). Cluster analysis. London: Heine-mann.

Eysenck, H. J. (1950). Criterion analysis—an applicationof the hypothetico-deductive method of factor analysis.Psychological Review, 57, 38-53.

Eysenck, H. J. (1953). The structure of human personality.London: Methuen.

Eysenck, H. J. (1969). Nature and history of humantypology. In H. J. Eysenck & S. B. G. Eysenck (Eds.),

Personality structure and measurement (pp. 3-137).London: Routledge & Kegan Paul.

Eysenck, H. J., & Eysenck, S. B. G. (1968). Manual for

the Eysenck Personality Inventory. San Diego: Educa-tional and Industrial Testing Service.

Falconer, D. S. (1967). The inheritance of liability todiseases with variable ages of onset, with particularreference to diabetes mellitus. Annals of Human Ge-

netics, 31, 1-20.Farris, J. S. (1979). On the naturalness of phylogenetic

classification. Systematic Zoology, 28, 200-214.Fisher, R. A. (1946). Statistical methods for research

workers (10th ed.). Edinburgh, Scotland: Oliver &

Boyd.Furnham, A. & Capon, M. (1983). Social skills and self-

monitoring processes. Personality and Individual Dif-

ferences, 4, 171-178.Gabrenya, W. K... Jr., & Arkin, R. M. (1980). Factor

structure and factor correlates of the Self-Monitoring

Scale. Personality and Social Psychology Bulletin, 6.

13-22.Gingerich, P. D. (1979). Paleontology, phytogeny, and

classification: An example from the mammalian fossil

record. Systematic Zoology, 28, 451-464.Golden, R. R. (1982). A taxometric model for the

detection of a conjectured latent taxon. MultivariateBehavioral Research, 17, 389-416.

Golden, R. R., & Meehl, P. E. (1973). Detecting clinicaltaxa, V: A Monte Carlo study of the maximum covari-

ance method and associated consistency tests (ReportNo. PR-73-4). Minneapolis: University of Minnesota,

Reports from the Research Laboratories of the De-partment of Psychiatry.

Golden, R. R., & Meehl, P. E. (1979). Detection of theschizoid taxon with MMPI indicators. Journal of Ab-

normal Psychology, 88, 217-233.Grove, W. M. (1983). The numerical taxonomy of endog-

enous depression. Unpublished doctoral dissertation,University of Minnesota, Minneapolis.

Guilford, J. P., & Zimmerman, W. S. (1963). Somevariable-sampling problems in the rotation of axes infactor analysis. Psychological Bulletin, 60, 289-301.

Hall, G. S. (1907). Aspects of child's life and education.

New York: Appleton.

Hays, W. L. (1973). Statistics for the social sciences (2nded.). New York: Holt, Rinehart & Winston.

Hempel, C. G. (1965). Fundamentals of taxonomy. InC. G. Hempel, Aspects of scientific explanation andother essays in the philosophy of science (pp. 137-

154). New York: Free Press.

Henderson, N. D. (1982). Human behavior genetics.Annual Review of Psychology, 33, 403-440.

Hogan, R. (1983). A socioanalytic theory of personality.

In M. M. Page (Ed.), Nebraska Symposium on Moti-vation: Vol. 30. Personality—current theory and research

(pp. 55-89). Lincoln: University of Nebraska Press.

Morgan. D. (1981). Rate of language acquisition andnoun emphasis. Journal of Psycholinguistic Research,

10, 629-640.

Ickes, W. J., Layden, M. A., & Barnes, R. D. (1978).Objective self-awareness and individuation: An empu-Tical link. Journal of Personality, 46, 146-161.


Jackson, D. N. (1971). Structured personality tests: 1971.Psychological Review, 78, 229-248.

Jackson, D. N. (1974). Personality Research Form manual.

New York: Research Psychologists Press.Jung, C. G. (1923). Psychological types (H. G. Baynes,

Trans.). New York: Harcourt, Brace. (Original workpublished 1921)

Kohlberg, L. (1969). Stage and sequence: The cognitive-developmental approach to socialization. In D. A.

Goslin (Ed.), Handbook of socialization theory and

research (pp. 347^(80). New York: Rand-McNally.Krauss, R. M., Geller, V., & Olson, C. (1976, September).

Modalities and cues in perceiving deception. Paperpresented at the meeting of the American PsychologicalAssociation, Washington, DC.

Kretschraer, E. (1948). Korperban undcharakter [Physiqueand character]. Berlin: Springer.

Kuiper, F. K., & Fisher, L. (1975). A Monte Carlocomparison of six clustering procedures. Biometrics,31, 777-783.

Kulik, J. A., & Taylor, S. E. (1981). Self-monitoring andthe use of consensus information. Journal of Personality,49, 75-84.

Langmuir, I. (1943). Science, common sense, and decency.Science, 97, 1-7.

Lazarsfeld, P. F. (1950). The logical and mathematicalfoundation of latent structure analysis. In S. A. StoufFerel al., Measurement and prediction (pp. 362-412).Princeton, NJ: Princeton University Press.

Lippa, R. (1976). Expressive control and the leakage ofdispositonal introversion-extraversion during role-played teaching. Journal of Personality, 44, 541-559.

Lippa, R. (1978a). Expressive control, expressive consis-tency, and the correspondence between behavior andpersonality. Journal of Personality, 46, 438-461.

Lippa, R. (1978b, September). Self-presentation and ex-

pressive display of personality. Paper presented at theannual meeting of the American Psychological Asso-

ciation, Toronto.Lippa, R., & Mash, M. (1979) The effects of self-

monitoring and self-reported consistency of personality

statements made by strangers and intimates. Unpub-

lished manuscript, California State University, Fullerton.Lippa, R., Valdez, E., & Jolly, A. (1979, September).

Self-monitoring and the consistency of masculinity-femininity cues. Paper presented at the annual meeting

of the American Psychological Association, New York.Loevinger, J. (1957). Objective tests as instruments of

psychological theory. Psychological Reports, 3. 635-694.

London, I. D. (1946). Some consequences for history andpsychology of Langmuir's concepts of convergence anddivergence of phenomena. Psychological Review, 53,170-188.

Lord, F. M., & Novick, M. R. (1968). Statistical theoriesof mental test scores. Reading, MA: Addison-Wesley.

Lutsky, N., Woodworth, W., & Clayton, S. (1980, May).Aaions-attitudes-actions: A multivariale, longitudinalstudy of attitude-behavior consistency. Paper presentedat the annual meeting of the Midwestern PsychologicalAssociation, St. Louis, MO.

Maccoby, E. E., & Jacklin, C. N. (1974). The psychology

of sex differences. Stanford, CA: Stanford UniversityPress.

Maxwell, A. E. (1971). Multivariate statistical methodsand classification problems. British Journal of Psychia-

try, 119. 121-127.Mayr, E. (1969). Principles of systematic zoology. New

York: McGraw-Hill.McCann, D., & Hancock, R. D. (1983). Self-monitoring

in communicative interactions: Social cognitive con-sequences of goal-directed message modification. Jour-

nal of Experimental Social Psychology, 19, 109-121.McNeill, J. (1979). Purposeful phonetics. Systematic

Zoology, 28, 465-482.

Meehl, P. E. (1962) Schizotaxia, schizotvpy, schizophrenia.American Psychologist, 17, 827-838.

Meehl, P. E. (1965). Detecting latent clinical taxa by

fallible quantitative indicators lacking an accepted cri-terion (Report No. PR-65-2). Minneapolis: University

of Minnesota, Reports from the Research Laboratoriesof the Department of Psychiatry.

Meehl, P. E. (1968). Detecting latent clinical taxa, II: A

simplified procedure, some additional hilmax locators,

a single-indicator method, and miscellaneous theorems

(Report No. PR-68-4). Minneapolis: University ofMinnesota, Reports from the Research Laboratoriesof the Department of Psychiatry.

Meehl, P. E. (1972). Reactions, reflections, projections.In J. Butcher (Ed.), Objective personality assessment:

Changing perspectives (pp. 131-189). New York: Aca-demic Press.

Meehl, P. E. (1973). MAXCOV-HITMAX: A taxonomicsearch method for loose genetic syndromes. In P. E.

Meehl, Psychodiagnosis: Selected papers (pp. 200-224). Minneapolis: University of Minnesota Press.

Meehl, P. E. (1977). Specific etiology and other forms ofstrong influence: Some quantitative meanings. Journal

of Medicine and Philosophy, 2, 33-53.Meehl, P. E. (1978). Theoretical risks and tabular asterisks:

Sir Karl, Sir Ronald, and the slow progress of softpsychology. Journal of Consulting and Clinical Psy-

chology, 46, 806-834.Meehl, P. E. (1979). A funny thing happened on the way

to the latent entities. Journal of Personality Assessment,

43, 563-581.

Meehl, P. E., & Golden, R. R. (1982). Taxometricmethods. In J. N. Butcher & P. C. Kendall (Eds.), The

handbook of research methods in clinical psychology

(pp. 127-181). New York: Wiley.Mendelsohn, G. A., Weiss, D. S., & Feimer, N. R. (1982).

Conceptual and empirical analysis of the typologicalimplications of patterns of socialization and femininity.Journal of Personality and Social Psychology, 42, 1157-1170.

Mischel, W. (1976). Introduction to personality (2nd ed.).New York: Holt, Rinehart & Winston.

Mojena, R. (1977). Hierarchical grouping methods andstopping rules: An evaluation. Computer Journal, 20,

359-363.Murphy, E. A. (1964). One cause? Many causes? The

argument from a bimodal distribution. Journal of

Chronic Diseases, 17. 301-324.Mussen, P. H., Conger, J. J., & Kagan, J. (1963). Child

development and personality (2nd ed.). New York:Harper & Row.

Myers, I. B. (1962). The Myers-Briggs type indicator

manual. Princeton, NJ: Educational Testing Service.


Nelson, K. (1973). Structure and strategy in learninghow to talk. Monographs of the Society for Research

in Child Development. 38 (1-2, Serial No. 149).Nelson, K. (1981). Individual differences in language

development: Implications for development and lan-guage. Developmental Psychology, 17, 170-187.

Overall, J. (1964). Note on the scientific status of factors.Psychological Bulletin, 61, 270-276.

Paulhus, D. (1982). Individual differences, self-perception,and cognitive dissonance: Their concurrent operationin forced compliance. Journal of Personality and Social

Psychology, 43, 838-852.

Piaget, J. (1965). The child's conception of the world.Totowa, NJ: Littlefield, Adams.

Plomin, R., DeFries, J. C, & Loehlin, J. C. (1977).Genotype-environment interaction and correlation inthe analysis of human behavior. Psychological Bulletin,84, 309-322.

Plomin, R., DeFries, J. C., & McClearn, G. E. (1980).Behavioral genetics: A primer. San Francisco: W. H.Freeman.

Ramer, A. (1976). Syntactic styles in emerging language.Journal of Child Language, 3. 49-62.

Rarick, D. L., Soldow, G. F., & Geiser, R. S. (1976).Self-monitoring as a mediator of conformity. Central

States Speech Journal, 27, 267-27'1.Riggio, R. E., & Friedman, H. S. (1982). The interrela-

tionships of self-monitoring factors, personality traits,and nonverbal social skills. Journal of Nonverbal Be-havior, 7, 33-45.

Ross, M., McFarland, C., & Fletcher, G. J. O. (1981).The effect of attitude on the recall of personal histories.Journal of Personality and Social Psychology, 40. 627-634.

Rummel, R. J. (1970). Applied factor analysis. Evanston,IL: Northwestern University Press.

Ruse, M. (1973). The philosophy of biology. London:Hutchinson.

Shaffer, D. R., Smith, J. E., & Tomarelli, M. (1982). Self-

monitoring as a determinant of self-disclosure reci-procity during the acquaintance process. Journal of

Personality and Social Psychology. 43. 163-175.

Shaw, M. E., & Costanzo, P. R. (1982). Theories of social

psychology (2nd ed.). New York: McGraw-Hill.

Siegman, A. W., & Reynolds, M. A. (1983). Self-moni-toring and speech in feigned and unfeigned lying.Journal of Personality and Social Psychology, 45, 1325-1333.

Sneath, P. H. A., & Sokal, R. R. (1973). Numerical

taxonomy. San Francisco: W. H. Freeman.

Snyder, M. (1974). Self-monitoring of expressive behavior.Journal of Personality and Social Psychology, 30, 526-537.

Snyder, M. (1979a). Cognitive, behavioral, and interper-sonal consequences of self-monitoring. In P. Pliner,K. R. Blankstein, & I. M. Spigel (Eds.), Advances in

the study of communication and affect: Vol. 5. Perceptionof emotion in self and others (pp. 181 -201). New York:Plenum.

Snyder, M. (1979b). Self-monitoring processes. In L.Berkowitz (Ed.), Advances in experimental social psy-

chology (Vol. 12, pp. 85-128). New York: AcademicPress.

Snyder, M. (1981). On the influence of individuals onsituations. In N. Cantor & J. F. Kihlstrom (Eds.),Personality, cognition, and social interaction (pp. 309-

329). Hillsdale, NJ: Erlbaum.Snyder, M. (1983a). The influence of individuals on

situations: Implications for understanding the linksbetween personality and social behavior. Journal of

Personality. 51. 497-516.Snyder, M. (1983b, May). The self in action. Paper

presented at the annual meeting of the MidwesternPsychological Association, Chicago.

Snyder, M., Berscheid, E., & Click, P. (1985). Focusingon the exterior and the interior: Two investigations ofthe initiation of personal relationships. Journal of

Personality and Social Psychology, 48, 427-439.Snyder, M., & Campbell, B. (1982). Self-monitoring: The

self in action. In J. Suls (Eds.), Psychological perspectives

on the self'(Vol. 1, pp. 185-207). Hillsdale, NJ: Erlbaum.Snyder, M., & Cantor, N. (1980). Thinking about ourselves

and others: Self-monitoring and social knowledge.Journal of Personality and Social Psychology, 39, 222-234.

Snyder, M., & Gangestad, S. (1982). Choosing social

situations: Two investigations of self-monitoring pro-cesses. Journal of Personality and Social Psychology,

43. 123-135.Snyder, M., Gangestad, S., & Simpson, J. A. (1983).

Choosing friends as activity partners: The role of self-monitoring. Journal of Personality and Social Psychol-

ogy. 45. 1061-1072.

Snyder, M., & Ickes, W. (1985). Personality and socialbehavior. In G. Lindzey & E. Aronson (Eds.), Handbook

of social psychology (3rd ed., pp. 883-948). New York,

NY: Random House.Snyder, M., & Kendzierski, D. (1982a). Acting on one's

attitudes: Procedures for linking attitude and behavior.

Journal of Experimental Social Psychology, 18, 165-183.

Snyder, M., & Kendzierski, D. (1982b). Choosing socialsituations: A strategy for generating correspondence

between attitudes and behavior. Journal of Personality,

50. 280-295.

Snyder, M., & Monson, T. C. (1975). Persons, situations,and the control of social behavior. Journal of Personalityand Social Psychology, 32, 637-644.

Snyder, M., & Simpson, J. A. (1984). Self-monitoring

and dating relationships. Journal of Personality andSocial Psychology, 47. 1281-1291.

Snyder, M., & Swann, W. B. (1976) When actions reflectattitudes: The politics of impression management.

Journal of Personality and Social Psychology, 34, 1034-1042.

Snyder, M., & Tanke, E. D. (1976). Behavior and attitude:Some people are more consistent than others. Journal

of Personality. 44. 510-517.

Sparacino, J., Ronchi, D., Bagley, T. K., Flesch, A. L.,& Kuhn, J. W. (1983). Self-monitoring and bloodpressure. Journal of Personality and Social Psychology,44, 365-375.

Starr, S. (1974). The relationship of single words to two-word sentences. Child Development, 46, 701-708.

Sypher, B. D., & Sypher, H. H. (1983). Self-monitoringand perceptions of communication ability in an orga-


nizalional setting. Personality and Social PsychologyBulletin, 9. 297-304.

Thurstone, L. L. (1947). Multiple factor analysis. Chicago:University of Chicago Press.

Tobey, E. L., & Tunnell, G. (1981). Predicting ourimpressions on others: Effects of public self-conscious-ness and acting, a self-monitoring subscale. Personalityand Social Psychology Bulletin, 7, 661-669.

Tukey, J. W. (1977). Exploratory data analysis. Reading,MA: Addison-Wesley.

Tunnell, G. (1980). Intraindividual consistency in per-sonality assessment: The effect of self-monitoring.Journal of Personality, 48, 220-232.

Tybout, A. M., & Scott, C. A. (1983). Availability ofwell-defined internal knowledge and the attitude for-mation process: Information aggregation versus self-perception. Journal of Personality and Social Psychol-ogy, 44, 474-491.

Tyler, L. E. (1965). The psychology of individual differences(3rd ed.). Englewood Cliffs, NJ: Prentice-Hall.

Waddington, C. H. (1957). The strategy of the genes.New York: Macmillan.

Ward, J. H. (1963). Hierarchical grouping to optimizean objective function. Journal of the American Statis-tical Association, 58. 236-244.

Werner, H. (1957). The concept of development from acomparative and organismic point of view. In D.Harris (Ed.), The concept of development (pp. 125-161). Minneapolis: University of Minnesota Press.

Wolf, D., & Gardner, H. (1979). Style and sequence inearly symbolic play. In N. Smith & M. Franklin (Eds.),Symbolic functioning in childhood (pp. 117-138).Hillsdale, NJ: Erlbaum.

Zanna, M. P., Olson, J. M., & Fazio, R. H. (1980).Attitude-behavior consistency: An individual differenceperspective. Journal of Personality and Social Psychol-ogy, 38, 432-440.

Zuckerman, M., & Reis, H. T. (1978). A comparison ofthree models for predicting altruistic behavior. Journalof Personality and Social Psychology, 36, 498-510.

Appendix

Predicting a Peaked Covariance Curve

Consider the two items, / and j, selected fromthe set of the eight conjectured items. If two classesexist within any sample (for convenience, let us

call them the class of highs and the class of lows),it is an algebraic truth that the sample covariance

between the two indicators is equal to the sum of

three terms:

cov(y) = p covH(y) + q covL(y) + pq&ibj,

where

p = the proportion of highs in the sample;

q = the proportion of lows in the sample;

covH(i/) = the covariance between the indicatorswithin the subsample of highs;

covL(i/) = the covariance between the indicatorswithin the subsample of lows;

A; = the difference between the mean i" scores

within the subsample of highs andwithin the subsample of lows; and

A/= the difference between the mean./ scoreswithin the subsample of highs and

within the subsample of lows.

We have ideally assumed that the two indicatorsare independent within the classes and thus that

the within-class covariances are equal to zero. Ifthis assumption holds, then the only source ofcovariance within the total sample will be thethird term in the expression above. Thus,

cov(y)

Of course, before we have started we do notknow what p and q are for any given sample, nordo we have estimates of A< or A '̂ for any given

population, nor, in fact, do we know whether twoclasses do actually exist. As the above formula

reveals, however, if two classes do exist (and when

Ai and Aj are held relatively constant), we expectthe covariance between /' and j in a sample to be

some function of the relative proportions of thetwo classes, p and q. Thus, for instance, if wecould somehow select a pure sample of high self-

monitoring individuals, we would expect cov(y')to be near zero because cov(;j') = (1.00) (.00) AfAj = 0. Similarly, if we could select a pure sample

of low self-monitoring individuals, we would alsoexpect cov ( i j ) to be near zero. Suppose now weselect a sample of 1 /4 of one class and 3/4 of theother. Then we would expect cov (ij) to be other

than zero because (.25) (.75) A/A/ = (.1875)Ai A/'. Moreover if j and j are keyed in the conjec-tured direction, as we assume here, then we wouldexpect this value to be positive. And if we select asample of 40% of one class and 60% of the otherclass, we would expect some larger value stillbecause (.40) (.60) AfA/> .1875 AiA/ Finally, it

is a simple mathematical truth that because theproduct pq is maximal when there exist equalnumbers from each class in the sample (i.e., p =q = 1/2), as long as Ai and A/ are held constant,cov (ij) is also expected to be maximal when p =

q = 1/2.


Given this fact, we can create a powerful boot-straps effect (Cronbach & Meehl, 1955). For ouritem pair, i and j, we take the remaining six itemsof our conjectured eight-item pool and constructa 7-point scale (with values ranging from 0-6). If,as we have already assumed, these six items dis-criminate between the classes, then this small scalealso discriminates between the classes. And, if ouritems i and j do not highly correlate with any ofthe six items within the classes, as we have alsoalready assumed, then ;' and j will not correlatevery highly with the small scale within the classes.Let us now use this 7-point scale to select differentsubsamples, each corresponding to the set of in-dividuals who obtained a given score on the scale.If the above conditions hold (once again, testablefor fit afterwords) and if two classes really do exist,then the seven different subsamples we have createdshould have a different p and q. The seven subsam-ples should, however, have similar Ai and Aj.(These latter values, in fact, should be similar toAf and A/for the entire sample.)

If two classes exist and if the smaller of the twoclasses is large enough so that the latent frequencydistributions on the 7-point scale cross, then therewill exist a scale value within which p ~ q ~ I/2. Moreover, if the latent frequency distributionsare monomodal and are not too unequal in size(so that the smaller of p and q equals at least .2),this value will be located somewhere toward themiddle of the scale. Samples associated with valuestoward the extremes are expected to be composedof more disparate p and q. Given our previousresults, this expectation yields the following pre-diction: If a class variable underlies responses tothe items as conjectured, the seven sample covari-ances between ;' and j plotted as a function of thevalues on the 7-point scale should be peaked;maximal toward the middle and nearer to zerotoward the extremes.

Received March 28, 1984Revision received November 28, 1984

Special Call for Papers on Psychology and Aging

The Psychology and Aging journal is gearing up for its first year of publication in 1986.Manuscripts have been received in the editorial office for a number of months, but more than50% of the original submissions have been experimental. The Editor, M. Powell Lawton, andthe Associate Editor, Donald H. Kausler, wish to emphasize that Psychology and Aging willbe a broad-ranging publication, and manuscripts from all areas of psychology are desired.

Papers on all aspects of issues related to psychology and aging are encouraged. As theproposed editorial policy statement outlines:

Psychology and Aging publishes original articles on adult development and aging. Such originalarticles include reports of research, which may be applied, biobehavioral, clinical, educational,experimental (laboratory, field, or naturalistic studies), methodological, or psychosocial. Although theemphasis is on original research investigations, occasional theoretical analyses of research issues,practical clinical problems, or policy may appear, as well as critical reviews of a content area in adultdevelopment and aging. Clinical case studies that have theoretical significance are also appropriate.Brief reports are acceptable with the author's agreement not to submit a full report to anotherjournal; a 75-100 word abstract plus 48-space lines of text and references constitute absolutelimitations on space for such brief reports.

Manuscripts should be directed to:

M. Powell Lawton

Philadelphia Geriatric Center

5301 Old York Road

Philadelphia, Pennsylvania 19141

Documents

"To Carve Nature at Its Joints": On the Existence of Discrete Classes