summary

503019 research-articleXXXX JOMXXX10.1177/0149206313503019Journal of ManagementMarcus et al. / Structure of Counterproductive Work Behavior Journal of Management Vol. 42 No. 1, January 2016 203­–233 DOI: 10.1177/0149206313503019 © The Author(s) 2013 Reprints and permissions: sagepub.com/journalsPermissions.nav The Structure of Counterproductive Work Behavior: A Review, a Structural Meta-Analysis, and a Primary Study Bernd Marcus University of Hagen O. Anita Taylor University of Western Ontario Stephanie E. Hastings Alberta Health Services Alexandra Sturm University of Hagen Oliver Weigelt University of Hagen Although counterproductive work behavior (CWB) has long been established as a broad domain of job behaviors, little agreement exists about its internal structure. The present research addressed alternative models of broadly defined CWB according to which specific behaviors can be grouped into (a) one general factor, or into (b) two, (c) five, or (d) eleven narrower facets, and a number of possible integrations of these models. First, conceptual differences between these models (including the nature of overall CWB as implying a reflective or formative model, boundaries of the domain, and relations among specific facets) are reviewed with regard to theoretical and practical implications. In Study 1, structural meta-analysis was then used to test whether a reflective higher-order factor underlies meta-analytically constructed correlation matrices of five CWB facets. Analyses supported a general factor model. For Study 2, a primary data set (N = 1,237 employees) was collected in order to test alternative structural models and Acknowledgments: This article was accepted under the editorship of Deborah E. Rupp. This research was supported by a grant from the Social Sciences and Humanities Research Council of Canada (SSHRC 160555) provided to the first author. We also thank Michael Eid and Gerhard Blickle for many helpful comments on an earlier version of this article. Further information on samples, analyses, and raw data may be obtained from the first author upon request. Corresponding author: Bernd Marcus, Department of Psychology, University of Hagen, D-58084 Hagen, Universitätsstr. 33, Germany. E-mail: [email protected] 203 204   Journal of Management / January 2016 possible integrations of these models. Confirmatory factor analyses revealed that the best fit was for a bimodal (nonhierarchical) model in which individual CWBs simultaneously load on one of the eleven facets describing their content (e.g., theft, absenteeism) and on one of three factors describing the target primarily harmed (organization, other persons, self). Less support was found for hierarchical models and for models involving fewer content factors. These findings suggest that CWB is best described by a reflective higher-order factor at the general level and by a complex set of bimodal facets at the more specific level. Keywords: deviant/counterproductive behavior; structural equation modeling; meta-analysis; performance assessment/management Counterproductive work behaviors (CWBs) can be defined as any volitional acts by employees that potentially violate the legitimate interests of, or do harm to, an organization or its stakeholders (e.g., Sackett & DeVore, 2001). This definition covers a broad range of specific acts (e.g., theft, substance use, sabotage, interpersonal violence, absenteeism) and partially overlaps with a number of related constructs such as workplace retaliation, aggression, or incivility, to name only a few (e.g., Spector, et al. 2006). In the present paper, we focus on CWB, as this encompasses the latter behavioral domains but is less restrictive in its construct definition. For example, unlike definitions of antisocial behaviors (cf. O’Leary-Kelly, Duffy, & Griffin, 2000), the definition of CWB does not assume that harm-doing is intentional (cf. Spector & Fox, 2005). The common defining element among CWBs is an effect (harm) that could be observed, rather than nonobservable antecedents such as motives (e.g., intention to harm) or other explanatory constructs (e.g., social norms, as in deviance). Thus, the definition of CWB avoids restricting theoretical approaches a priori. As a broad domain of employee behavior, CWB now seems well established in organizational research and practice. For example, there is evidence that managers’ perceptions of employees’ overall performance are strongly affected by CWBs (Rotundo & Sackett, 2002). Despite the attention general CWB has gained, there is still relatively little agreement on how the various acts belonging to that domain are related to each other and how they should be categorized. Specifically, although research has begun to address the issue of CWB’s internal structure and various models have been proposed in the literature, a systematic integration of this research and theorizing is currently lacking. Clarifying these issues seems worthwhile for a number of reasons. From a theoretical point of view, better understanding of relationships and overlaps among the many subsets of behavior covered by the term “CWB” (cf. Rotundo & Spector, 2010) would foster conceptual clarity and parsimony. From a practitioner perspective, structuring the domain of CWB would help to tailor interventions to the appropriate content and level of specificity (cf. Spector et al., 2006). In the present research, we are attempting to further our understanding of CWB on the basis of existing knowledge. That is, we are not presenting yet another “new” model of CWB but rather are reviewing some existing models theoretically and empirically in an attempt to clarify differences and overlaps between those models and identify ways in which various Marcus et al. / Structure of Counterproductive Work Behavior   205 models of CWB may be integrated. After a brief introduction to three of the most prominent models of CWB currently in the literature, we first discuss two conflicting conceptions of the nature of general CWB as either a coherent latent construct or an umbrella term denoting a composite of distinct behavioral domains. Next, we review issues of the scope (i.e., breadth and specificity) of the overall domain and its constituent facets. We continue with a discussion of the number of modes or organizing principles simultaneously needed to categorize each act of CWB, an issue that has garnered little attention in previous research. We close our narrative review with an outline of ways in which existing models of CWB may be integrated, given the differences and similarities between the models. In the empirical part of the paper, we first try to clarify the issue of whether a latent construct underlies different acts of CWB by means of a structural meta-analysis of one widely used measure of CWB. We then report data from a large-scale primary study to address the issues of scope, modality, and possible integrations of CWB models, which cannot be assessed empirically by means of meta-analysis. Models of CWB About 30 years ago, Hollinger and Clark’s (1983) now classical study on employee theft led to a broadened concept of theft labeled property deviance, which was distinguished from production deviance (e.g., lateness, substance use). Notably, Hollinger and Clark used the term “counterproductive behavior” interchangeably with “deviance.” We follow this tradition, as content overlap between measures of workplace deviance and CWB seems almost perfect. Since then, a number of researchers have proposed different models of how the overall domain of CWB may be internally structured. In the present article, we focus on three more recent models, which currently feature most prominently in the CWB literature: (1) Robinson and Bennett’s (1995; Bennett & Robinson, 2000) also dichotomous, yet different, distinction between organizational deviance (OD) and interpersonal deviance (ID); (2) Spector et al.’s (2006) five-dimensional model, which includes abuse, production deviance, theft, sabotage, and withdrawal; and (3) the even more fine-grained eleven-facet model proposed by Gruys and Sackett (2003). In the following, we first review a number of differences between these models and then derive a number of ways in which these models may be integrated and perhaps modified. Development of Three Models of CWB From among the models listed earlier, Bennett and Robinson’s (2000) two subcategories are probably the most well-established and often-researched dimensions of CWB. Numerous studies and at least two meta-analyses (Berry, Ones, & Sackett, 2007; Dalal, 2005) have examined this model directly. The distinction between OD and ID stemmed from an earlier multidimensional scaling study (Robinson & Bennett, 1995) in which behavioral incidents were rated for similarity. That study revealed that raters judge CWBs as more similar if they are directed at the same target (OD: organization as a whole; ID: particular persons within the organization) or if they yield comparably serious consequences (severe vs. less severe). In later research based on frequencies of self-reported CWB, Bennett and Robinson (2000) abandoned the severity dimension but were able to support the ID-OD distinction by means 206   Journal of Management / January 2016 of factor-analytic methods. Whereas the earlier study showed OD and ID as endpoints of the same dimension, these targets of CWB are now conceptualized as different constructs. Metaanalyses (Berry et al., 2007; Dalal, 2005) found some support for the discriminant validity of the latter conceptualization by confirming that correlation patterns of the two factors with outside variables are partially different. There are still reasons to be cautious about the comprehensiveness of the resulting two-factor solution. While sampling from a large set of original items developed to cover broadly defined deviance, Bennett and Robinson (2000) dropped items with low variance, which likely eliminated more severe forms of deviance because serious acts tend to be rare. Moreover, the authors imposed the theoretically expected two-factor solution on the data, which may have led to the exclusion of different or narrower facets of CWB. A model of CWB including five facets was proposed by Spector et al. (2006). One of these facets, abuse, largely overlaps with ID, whereas the other four (production deviance, sabotage, theft, and withdrawal) represent narrower facets of CWB mainly targeted at the organization. Unlike Robinson and Bennett (1995), Spector et al. developed each facet in their model independently based on differing explanatory theories rather than trying to structure the overall domain of CWB by empirical means. Spector and colleagues did not explicitly address the questions of why these facets in particular (as opposed to possible others), and why this particular number of dimensions, were chosen to describe the domain of CWB. For reasons reviewed later, the authors did also not report an attempt to validate the internal structure of their model empirically, yet they did test hypothesized relations to outside variables. In a still different approach, Gruys and Sackett (2003) examined the dimensionality of CWB. Similar to Bennett and Robinson (2000), Gruys and Sackett started by collecting a comprehensive set of CWB items, but the latter authors removed only redundancies and retained the more severe forms of CWB. Possibly due to this difference, they kept a much larger set of items (66) for empirical analyses than did Bennett and Robinson (28, of which 9 were later dropped). The 66 items were sorted into 11 categories based on expert judgment (similar to Spector et al., 2006). Gruys and Sackett stated that the sorting was done “based on the similarity of content” (2003, p. 33), but the exact rationales for forming these categories are not clear from the paper. The 11 categories include (1) theft and related behaviors, (2) destruction of property, (3) misuse of information, (4) misuse of time and resources, (5) unsafe behavior, (6) poor attendance, (7) poor-quality work, (8) alcohol use, (9) drug use, (10) inappropriate verbal action, and (11) inappropriate physical action. Although both models were developed independently, some of Gruys and Sackett’s dimensions obviously overlap with those proposed by Spector et al. (2006) (e.g., theft, poor attendance with withdrawal). Other dimensions seem to describe still narrower facets than those in Spector et al.’s model (e.g., Spector et al.’s abuse dimension seems to split into a verbal and a physical component). Furthermore, some of Gruys and Sackett’s facets did not appear in either of the models reviewed earlier (e.g., unsafe behavior; alcohol and drug use, except for one substance use item in Bennett and Robinson’s scale). Hence, Gruys and Sackett’s model seems to be both more fine-grained and more comprehensive than the other two models of CWB. To summarize, of the three models reviewed, those of Bennett and Robinson (2000) and of Gruys and Sackett (2003) were developed using similar research strategies and methods but resulted in quite different solutions in terms of breadth of domain coverage and of number Marcus et al. / Structure of Counterproductive Work Behavior   207 Table 1 Conceptual Differences Between Models of Counterproductive Work Behavior (CWB) Bennett and Robinson (2000) Spector, Fox, Penney, Bruursema, Goh, and Kessler (2006) CWB facet labels (1) organizational deviance (2) interpersonal deviance       (1) production deviance (2) sabotage (3) theft (4) withdrawal (5) abuse toward others Rationale for building facets Conceptualization of overall CWB structuring of overall domain not specified, but use of factor analysis implies reflective construct exclusion of severe and self-destructive acts facet-specific theories Coverage of domain Breadth of individual facets Modality Type of mode used broad unimodal, though based on an originally bimodal model target explicitly formative Gruys and Sackett (2003) (1) theft and related behaviors (2) destruction of property (3) misuse of information (4) misuse of time and resources (5) unsafe behavior (6) poor attendance (7) poor-quality work (8) alcohol use (9) drug use (10) inappropriate verbal action (11) inappropriate physical action structuring of overall domain not specified, but use of factor analysis implies reflective construct comprehensive not specified, but based on Bennett and Robinson (2000) narrow to broad narrow unimodal unimodal content, partially confounded with target content, partially confounded with target and specificity of single facets of CWB. In contrast, Spector et al.’s (2006) model, while apparently falling in between the other two models in terms of specificity of facets, was based on an approach explicitly abandoning empirical means for structuring overall CWB. Conceptual differences between all three models are summarized in Table 1 and reviewed in the following sections. At this point, it seems also noteworthy that the three reviewed models do not differ substantially in approaches to the measurement of CWB. Each of the models was operationalized in a self-report questionnaire containing brief statements of counterproductive acts. As far as content overlaps, items across CWB inventories tend to be very similar (cf. Rotundo & Spector, 2010). In fact, Spector et al. (2006) actually developed their fivefactor Counterproductive Work Behavior Checklist (CWB-C) with items adopted from an older measure of the two Robinson and Bennett factors. Like Bennett and Robinson’s (2000) original questionnaire, Spector et al.’s CWB-C is now widely used in empirical studies of CWB, whereas Gruys and Sackett’s inventory has not been used in published research since the original paper appeared. 208   Journal of Management / January 2016 Conception of Overall CWB: Formative Versus Reflective Construct All of the models described above specify subdimensions of the overall domain of CWB. However, there is an important difference between the models regarding the assumed nature of this overall construct. Unlike the other theorists, Spector et al. (2006) explicitly state that they assumed CWB to fit a causal indicator (or formative) model. As opposed to effect indicator (or reflective) models, formative models do not imply or require that relationships among indicators of a construct reflect underlying latent common causes or constructs (Bollen & Lennox, 1991; Edwards & Bagozzi, 2000). Justification for forming such composites is typically based on common consequences rather than causes. For example, we can speak of “work motivation,” assuming that motivation generally affects performance, but without assuming that different motives (or motivators) underlying motivation need to form a common latent construct. Those motives may substitute but may also complement each other, which could lead to all kinds of empirical relations, from negative, to zero, to positive. In contrast, narrow traits forming reflective higher-order personality dimensions such as extraversion are assumed to be affected by a common cause and thus positively correlated. The distinction of formative and reflective constructs has a number of important implications (cf. Bollen & Lennox, 1991; Edwards & Bagozzi, 2000). First, one could expect all elements of overall CWB to be positively correlated only if CWB were a reflective construct. In that case, assessment of the structure of CWB becomes subject to factor-analytic methods, whereas the boundaries of a formative construct are to some extent a matter of arbitrary decisions. In other words, if researchers refer to CWB (workplace deviance, aggression, retaliation, etc.) as a formative construct, there would be no empirical base for deciding whether or not an act belongs to that domain. Without empirical standards, it becomes difficult to identify a parsimonious set of nonredundant constructs among the many CWB-related concepts discussed in the literature. From a practical point of view, general CWB as a reflective construct implies that interventions that work against one type of CWB would also tend to be effective against any other type of CWB because at least one common cause underlies all CWBs. For example, selection devices successfully identifying applicants who refrain from stealing would also identify individuals who come to work on time. If overall CWB were a formative construct, there would be no reason to expect similar side effects across different types of CWB. In addition, the lack of a common cause among indicators in formative composite scales has a number of undesirable implications for measurement (Edwards, 2011). However, this refers to assumed relationships between theoretical constructs and measured indicators (i.e., items). In fact, it is for this type of relationships that the terms formative and reflective are typically used. It is noteworthy that none of the models reviewed appears to assume indicators within each facet of CWB fit a formative model. Hence, we restrict our discussion and later analyses to conceptual differences at the level of specific and general constructs of CWB and adopt the terms formative and reflective in an analogous fashion to describe these conceptual relations. The question then arises as to whether CWB actually corresponds to a formative or reflective model. Although others (e.g., Tepper, Henle, Lambert, Giacalone, & Duffy, 2008) have made the same assumption, Spector et al. (2006) did not provide an explicit rationale for their assumption of a formative model. In contrast, both Bennett and Robinson (2000) and Gruys and Sackett (2003) used factor-analytic methods, which imply the assumption of some kind of reflective model. Their results are less than clear-cut, however. Using confirmatory factor Marcus et al. / Structure of Counterproductive Work Behavior   209 analysis (CFA), both groups of researchers reported relatively poor fit for their expected (facet-level) models. However, although Gruys and Sackett found that a model of CWB with just a general factor fit the data even worse, a strong first component emerged with all facets loading highly on it when they subjected their composite facet scales to an exploratory principal components analysis1 (for similar results with different scales at the item level, see Ashton, 1998, who employed principal components analysis, and Marcus & Wagner, 2007, who conducted a CFA). Use of factor analysis per se does not indicate whether the underlying model is presumed to have the form of a first-order (direct reflective) or of a hierarchically organized higherorder (indirect reflective) model. Some additional evidence for the existence of such a general factor is provided by the substantial correlations between the two factors of OD and ID found in meta-analyses (ρ = .62, Berry et al., 2007, to .70, Dalal, 2005). This would correspond to a theoretical proposition by Sackett and DeVore (2001), who suggested that CWB is a hierarchically organized construct with a general factor at the top of the hierarchy and with narrower facets, each of which also carries unique variance, at lower levels. Some empirical support for this form of the CWB construct was provided by Marcus, Schuler, Quell, and Hümpfner (2002), who confirmed in a CFA that a general factor lies above a set of narrower facets of CWB (which differed from the facets in the CWB models reviewed here). Beyond the practical and psychometric implications already mentioned, why should researchers interested in the theoretical understanding of CWB care about the formativereflective distinction? Although this question has rarely been addressed in previous theories of CWB, some links between theoretical perspectives and positions on the overall nature of CWB seem plausible. If CWB is primarily seen as an outcome of external factors (i.e., situational variance), as is typical for approaches rooted in social psychological traditions, there is no need to assume that particular patterns of (between-person) covariation exist among different acts of CWB. For example, Spector and Fox (2005) proposed a stressor-emotion model of CWB, in which external stressors provoke emotional reactions that in turn may lead to CWB. In contrast, theoretical approaches that seek to explain CWB primarily by individual differences (i.e., person variance) may require explicit assumptions about the latent structure underlying CWB. A particularly illustrative example is the general theory of crime (Gottfredson & Hirschi, 1990), which posits that all acts of crime and deviance are driven by the same common cause of low self-control (see Marcus & Schuler, 2004, for supportive evidence of this theory in CWB research). Proponents of this theory argue that all acts of crime and deviance share certain features (in particular, immediate satisfaction of desires at the expense of long-term negative consequences) that are theoretically linked to the trait of self-control. Because those features make people low in self-control more likely to commit any kind of deviant act (or CWB), these acts should all be positively correlated in the form of a positive manifold and thus conceptually and empirically fit a reflective model. In more situation-centered approaches (that is, in the stressor-emotion model and related theories), the theoretical focus is on common motives or antecedents (e.g., some kind of triggering event) rather than on consequences of behavior. With this type of approach, there is no theoretical reason to assume that a common latent cause drives relationships among deviant acts. In this type of theory, personality is rather modeled as a moderator that helps to explain 210   Journal of Management / January 2016 why different people react to the same external triggers in different ways (e.g., either with withdrawal or aggression). This, in turn, would imply that researchers focus less on commonalities but on the unique features of different types of CWB (such as withdrawal vs. aggressive behavior, in the present example). In fact, the authors of the stressor-emotion model emphasized such differences in their later paper on the construct of CWB (Spector et al., 2006), although the original model did not distinguish between facets of CWB. In a nutshell, evidence of the reflective nature of CWB would support person-centered approaches that assume a common CWB factor while not necessarily refuting situational approaches that are relatively assumption free in that respect. Of course, it is possible that both types of approaches complement each other in contributing to our understanding of CWB, as the construct may be described by both common and unique factors. To summarize, the nature of the relation of overall CWB to its constituent facets remained an unresolved issue in previous research. Some researchers, including Spector et al. (2006), explicitly assume that CWBs correspond to a formative model but have provided few supportive arguments for that statement. Others, including both Bennett and Robinson (2000) and Gruys and Sackett (2003), were less explicit in their assumptions, but the methods used by these authors correspond to reflective rather than formative models. In addition, there is a theoretical rationale for expecting a reflective higher-order factor of CWB, yet the current state of empirical and theoretical evidence on these issues seems far from being conclusive. Content of CWB: Domain Coverage, Specificity, and Modality of Facets Apart from the assumed nature of overall CWB, there are a number of more obvious differences between the models reviewed. First, Gruys and Sackett’s (2003) model includes a number of behaviors barely covered by the other two models (namely, unsafe behavior, alcohol use, and drug use, though Bennett & Robinson’s, 2000, scale contains one drug use item and Spector & Fox, 2005, explicitly discussed some types of safety violations as acts of CWB). The question of domain breadth is of both theoretical and practical relevance, particularly if overall CWB were found to be a latent reflective construct underlying all acts of CWB. For acts of substance use and unsafe behavior, the immediate target potentially harmed is the actor. As mentioned in the introduction, the definition of CWB covers acts done both with and without the intention to harm others, yet some theorists of CWB-related domains (e.g., Skarlicki & Folger, 1997, in their concept of retaliation) assumed a motive of harmdoing that seems unlikely if actors first and foremost harm themselves. This difference would correspond to the theoretical proposition that cognitive and emotional antecedents differ between self-destructive and retaliatory forms of CWB (Martinko, Gundlach, & Douglas, 2002). If overall CWB were a reflective construct, it would thus be worthwhile to investigate whether this domain includes self-destructive acts. If so, this would imply at least some commonality in causes, whereas it would make little theoretical sense to lump together two entirely distinct domains. Moreover, practitioners may want to know which acts are potentially covered by interventions aimed at CWB. Furthermore, the three models obviously differ in the number and specificity of assumed facets of CWB. If CWB were a formative construct, but also if facets of CWB carry some unique variance in addition to a common latent factor (as proposed by Sackett & DeVore, 2001, among others), it would be important to know exactly which facets of CWB need to be Marcus et al. / Structure of Counterproductive Work Behavior   211 distinguished to fully understand the entire domain theoretically and tailor practical interventions accordingly. Related, though probably less obvious, is the issue of modality. We use this term to refer to the number of distinct organizing principles employed to assign each individual act of CWB to dimensions within a model of CWB. For example, in Robinson and Bennett’s (1995) original scheme based on similarity ratings, individual CWBs were simultaneously described by the two modes of target (organization vs. people) and severity (minor vs. severe). Using similar methodology, Gruys and Sackett (2003) also arrived at a bimodal description of CWBs in a second study published in the same article as their self-report study, yet they interpreted their second mode as task relevance rather than severity. Although Gruys and Sackett describe a number of additional modes or principles that could potentially be used to organize CWBs (private vs. public, planned vs. unplanned; see also Bowling & Gruys, 2010), none of the models based on frequencies of actual behavior explicitly distinguished between different modes. The distinction of OD versus ID in Bennett and Robinson’s (2000) revised model refers to just one mode, that of the target of CWB. The other two models reviewed herein are less clear with respect to their organizing principle, although each act is assigned to just one category, which would imply a unimodal underlying scheme. Whereas some dimensions, such as Spector et al.’s (2006) abuse factor, seem to parallel the target mode, most categories in both Spector et al.’s and in Gruys and Sackett’s (2003) models seem to refer to the nature or content of the deviant acts rather than their targets. For example, Gruys and Sackett’s facets of verbal and physical inappropriate behavior are distinguished by the form in which aggressive tendencies toward other people are expressed, not by different targets of aggression. The issue of modality becomes theoretically relevant especially if two modal categorizations are confounded within the same facet. Take Gruys and Sackett’s (2003) facet of misuse of information as an example. This facet is defined by its content; that is, all acts categorized as misuse of information share similar features in terms of the nature of the behavior itself. However, two of the five items grouped by Gruys and Sackett into that category (fail to give others necessary information, lie to others to cover up mistakes) refer to interpersonal targets, whereas the other three (destroy company records, discuss confidential matters with outsiders, provide organization with false information) appear to have the organization as the target. Confounding of the target mode within a category based on the content mode (or any other confounding) effectively precludes integrations of different models of CWB that are based on differing modal schemes. This would hamper development of comprehensive and parsimonious models of CWB, as one and the same act could be grouped into very different categories depending on the model used to organize behaviors. Moreover, if each act of CWB is best described by two or more modes, practitioners may need to consider both modes when tailoring interventions (e.g., prevention of organization-directed theft may not work equally well for coworker-directed theft). Possible Integrations of the Models Although each of the three models reviewed has its own merits and has already inspired CWB research, parsimonious theory development would benefit from an integration and clarification of issues on which alternative models disagree. Authors of the two more recent 212   Journal of Management / January 2016 models (Gruys and Sackett, 2003; Spector et al., 2006) explicitly refer to the distinction of OD and ID (Bennett & Robinson, 2000) and state that they regard the two dimensions as too broad to account for the domain of CWB with the optimal level of specificity. Spector et al. (2006) further mapped their five dimensions onto their own earlier operationalizations of OD and ID. ID showed a very strong association with abuse (r = .94), whereas OD correlated about as strongly with abuse (r = .65) as with Spector et al.’s remaining four dimensions (r = .65 to .74) that purportedly represent facets of OD. Other than that, to our knowledge, only one recent theoretical paper has attempted to integrate at least two of the CWB models reviewed. Wu and LeBreton (2011) assigned Gruys and Sackett’s (2003) 11 facets to Robinson and Bennett’s (1995) original four categories, which had distinguished between the modes of target and severity. As discussed earlier, categorizing Gruys and Sackett’s content facets to target categories is not without problems. As a starting point for more comprehensive integrations, we propose a number of general forms such integrative models may take. Figure 1 illustrates schematically three general ways in which facets of CWB may be related. Each part of Figure 1 hereby exemplifies one general form but stands for a set of similar specific models. For illustrative purposes, we use five CWB facets as a baseline in parts of the figure, but this is not meant to imply any preference for the Spector et al. (2006) model over other models. As per convention for path diagrams, boxes stand for observed indicators (limited to two per facet for the sake of simplicity), whereas ellipses represent latent constructs. Although our earlier review did not support such a conclusion, it is still possible that different models of CWB do not need to be integrated, because one model describes the true structure of CWB exhaustively. If that were the case, the most liberal form such a structure could take in terms of a formal path model is that of a measurement model in which all types of CWB are correlated freely. Solid lines in Figure 1(a) show this form for the example of the five-facet model of CWB (replace that number with eleven or two for the other CWB models). Whereas each facet of CWB is related to its indicators in the form of a direct reflective model, relationships between the facets in the structural part of the model would also correspond to a formative model of overall CWB. This is illustrated by the dotted lines in Figure 1(a), which point from CWB facets to general CWB, thus not assuming any particular pattern of facet intercorrelations. Evidence for a structure like that in Figure 1(a) thus would confirm the specific model tested and be consistent with, although not provide a sufficient test of, the assumption of overall CWB as a formative construct (cf. Edwards & Bagozzi, 2000). A second possible form integrates different models of CWB in a hierarchical fashion, which corresponds to the form of a higher-order factor-analytic model illustrated in Figure 1(b). The model depicted shows the most complex integration of all models plus one general factor at the top of the hierarchy, but its general form would be retained if one or two levels in the hierarchy were found to be redundant. A model of that form is based on the assumption of an indirect reflective model across all levels of CWB. Whereas a higher-order construct parsimoniously describes a common element across its constituent facets, lower-order facets may carry additional unique variance. Substantively, this implies that both higher- and lowerorder constructs are theoretically meaningful in their own right. Furthermore, Figure 1(b) represents a unimodal model of CWB within each level of the hierarchy. If constructs at different levels are described by different modes, the mode of the lower-order construct should ideally be nested in the higher-order mode. For example, if we knew that a lower-order facet Marcus et al. / Structure of Counterproductive Work Behavior   213 Figure 1 Schematic Representations of General Forms of Counterproductive Work Behavior (CWB) Models (a) (b) (c) Note: (a) measurement model for one particular dimensional structure of CWB (solid lines) with possible implications for the formative nature of general CWB (dotted lines) added; (b) multilevel hierarchical reflective model (solid lines) with possible cross-loading (dotted line) added; (c) bimodal model with five content and two target factors. 214   Journal of Management / January 2016 of theft belongs to a higher-order dimension of OD, we would know that each act that fits the content factor of theft also fits the target factor of organization. Any deviation from nestedness (e.g., theft from the employer or from coworkers, as exemplified by the cross-loading indicated by the dotted line in Figure 1[b]) indicates confounding of modes, which implies that both modes need to be considered simultaneously for the given facet. Finally, Figure 1(c) shows a fully bimodal model in an idealized form. In this form, each act of CWB is simultaneously described by two different modes. Unlike the hierarchical model, the two modes are modeled at the same hierarchical level here. Formally, tests of such bimodal models would correspond to those of multitrait-multimethod models (with one mode modeled as “traits” and one as “methods”). Bimodal modeling thus allows for disentangling possible confounds within hierarchical models. Methodologically, confirmation of a bimodal model would imply that explicit measurement of the second mode is not redundant, as opposed to higher-order factors in a purely hierarchical model. Substantively, a bimodal model implies that acts that fall into the same content (or target) category may still need to be treated differently if they fall into different target (content) categories, as this may imply partially different meanings. Further, it is possible (although not illustrated in the present figure) to integrate more than two modes at the same level, similar to the more than two hierarchical levels in Figure 1(b). To summarize, the various models of CWB may be tested separately in a form corresponding to formative conceptions of overall CWB, or they may be integrated in either a hierarchical or a multimodal fashion or by combinations of the latter two kinds of models. Each of these variations has different implications for theory, research, and practical interventions aimed at CWB. Hence, clarification of these issues seems worthwhile. Present Studies In the empirical part of our research, we address the above reviewed topics of (1) formative or reflective nature of CWB, (2) scope of the overall domain and its facets, (3) modality, and (4) possible integrations of alternative models of CWB, in two consecutive studies. In Study 1, we employ structural meta-analysis (Viswesvaran & Ones, 1995) to utilize the existing evidence on empirical relationships between facets of CWB in order to clarify the first issue listed above. In structural meta-analysis, matrices of (CWB facet) intercorrelations are first constructed by meta-analyzing each bivariate relationship in the matrix. Then, these matrices are analyzed by means of CFA. Specifically, we test whether or not a general factor underlies relations between facets of CWB, which would correspond to a reflective model of overall CWB. Ideally, one would utilize existing evidence on each of the three models for that purpose. However, two complications prevented us from following this strategy. First, it is technically not possible to test a general factor model with just two constituent facets as in Bennett and Robinson’s (2000) OD-ID distinction, because two indicators are insufficient to identify the structural part of that model in CFA (see Marcus et al., 2002, for a more detailed explanation in the present context). Second, although there is no technical limitation of modeling general CWB with Gruys and Sackett’s (2003) eleven facets, data on their original scale were reported for only one study. This problem could theoretically be overcome by assigning measures used in previous research to the eleven categories post hoc. However, a multitude of other problems may arise from that strategy, including unequal sample sizes across cells Marcus et al. / Structure of Counterproductive Work Behavior   215 in the matrix or even empty cells, combinations across conceptually and psychometrically different measures in each cell, or equivocal assignments of effect sizes to cells. These problems may accumulate to create hardly interpretable results or ill-defined matrices (cf. Viswesvaran & Ones, 1995). We therefore decided to restrict the meta-analyses presented here to Spector et al.’s (2006) facets as a baseline for testing a general factor model of CWB.2 These authors’ CWB-C scale has been used sufficiently often to avoid problems associated with post hoc categorization in structural meta-analysis. Moreover, this model provides for a conceptual extension of the OD-ID distinction and thus can be said to entail two of the models reviewed. Finally, Spector et al. explicitly reject the concept of overall CWB as a reflective construct. Hence, evidence supporting a reflective model based on Spector et al.’s facets can be said to provide for a conservative test of general CWB. The remaining issues listed above require item-level data to be addressed, which prevents use of structural meta-analysis. We therefore conducted a large-scale primary study to address these issues. Attempting to integrate all models reviewed requires inclusion of the most comprehensive and fine-grained model because more parsimonious models can be modeled as special cases of more complex models, whereas the reverse is not possible. For Study 2, we therefore collected data with Gruys and Sackett’s (2003) inventory as a baseline and model facets from other CWB models using this measure’s item content. Study 1 Method Literature search. Several sources were utilized for identifying published and unpublished studies of Spector et al.’s (2006) model of CWB. These included various search engines (ABI/Inform, Google Scholar, ProQuest Psychology, ProQuest Dissertation and Theses, PsycINFO, Sociofile, Sociological Abstracts, and Social Sciences Citation Index), manual searches in a number of academic journals (the Academy of Management Journal, Applied Psychology: An International Review, the European Journal of Work and Organizational Psychology, Human Performance, the International Journal of Selection and Assessment, the Journal of Applied Psychology, the Journal of Business and Psychology, the Journal of Management, the Journal of Occupational and Organizational Psychology, the Journal of Organizational Behavior, the Journal of Vocational Behavior, Organizational Behavior and Human Decision Processes, and Personnel Psychology), and scientific programs of the Society for Industrial and Organizational Psychology’s annual conference. However, conference articles that were eventually published and included in our previous searches were excluded to avoid duplications. Finally, we contacted a number of researchers who had previously published in the CWB field and requested unpublished data. Inclusion criteria, sample, and coding procedure. We retained studies that contained empirical data on the association between facets from Spector et al.’s (2006) model. As discussed earlier, it was important to retain relatively pure measures of these constructs to avoid any confounds or content overlap that may bias results. To meet this criterion, we only included studies that used the CWB-C (Spector et al., 2006). This resulted in a database of eleven independent samples with a total N of 3,931. Due to the restriction to one particular 216   Journal of Management / January 2016 measure of CWB, sample size as well as item content in each cell was virtually identical across cells (except that some studies used translations of the CWB-C), thereby avoiding most methodological problems typically associated with the combination of meta-analysis and structural equation modeling. Relevant data were extracted from each study including sample (e.g., demographic information), study (e.g., laboratory vs. field study), and statistical (e.g., correlations, reliabilities of measures) information. Data analysis. Structural meta-analysis, a procedure that combines the principles of psychometric meta-analysis and structural equation modeling, was used for the analyses (Viswesvaran & Ones, 1995). First, psychometric meta-analysis, as specified by Hunter and Schmidt (2004), was used to create matrices of bivariate correlations between the CWB constructs. This procedure involved corrections for unreliability and also accounted for the effect of sampling error on the variance of the correlations. As reliability estimates were not provided for all measures in the primary studies, corrections were conducted using the artifact distributions method. The reliability distribution generated from the primary samples was used for this purpose. CFA using maximum likelihood estimation was then applied to the correlation matrices to test whether a general factor model underlies correlations. Viswesvaran and Ones (1995) recommended using attenuation-corrected correlation coefficients as input for structural meta-analyses. However, if disattenuation has to be based on artifact distributions, resulting estimates may be biased (cf. Hunter & Schmidt, 2004; Wothke, 1993). Previously published structural meta-analyses reported results based on either observed or disattenuated correlations. Michel, Viswesvaran, and Thomas (2011) recently compared both practices and found little disagreement between them in terms of substantive conclusions, although matrices of observed correlations tended to produce slightly better fit. In the present study, we report findings obtained with both types of input matrices to allow for direct comparisons. Results Correlation matrices used as input in CFAs of the Spector et al. (2006) model are shown in Table 2. Uncorrected values are shown below the main diagonals, whereas correlations corrected for unreliability are presented above the diagonals. Table 2 also contains the reliability estimates used for corrections. Based on correlations in Table 2, we tested whether one general factor underlies the CWB facets in Spector et al.’s (2006) model. Confirmation of higher-order factors, while not technically falsifying a (basically assumption-free) formative model, corresponds conceptually to reflective models. It is also notable that Spector and colleagues had originally mapped their five facets onto two, rather than one, formative behavioral constructs. However, one of these constructs, ID, is represented only by the abuse facet in Spector at al.’s model. Incorporating single indicators in a CFA model requires fixing the amount of indicator variance not accounted for by the latent factor to some predetermined value, which in our case implies that the two-factor model is structurally equivalent to the general factor model (i.e., model fit is the same, and the standardized correlation between the two higher-order factors corresponds to the standardized loading of the single indicator in the general factor model; cf. Brown, 2006). As indicated by fit statistics presented in Table 3 (selection of fit indices Marcus et al. / Structure of Counterproductive Work Behavior   217 Table 2 Meta-Analytic Correlations of Spector, Fox, Penney, Bruursema, Goh, and Kessler’s (2006) Facets of Counterproductive Work Behavior 1. Abuse 2. Sabotage 3. Production deviance 4. Theft 5. Withdrawal 1 2 3 4 5 (.89) .59 .59 .56 .51 .79 (.64) .60 .59 .50 .73 .93 (.68) .56 .53 .71 .89 .82 (.72) .53 .63 .75 .76 .75 (.72) Note: Meta-analytic correlations not corrected for unreliability (bare-bones meta-analyses) appear below the main diagonal. Correlations corrected for unreliability (true score correlations) appear above the diagonal. Meta-analytic estimates of mean internal consistencies used for corrections are given in the diagonal. K is 11 for all coefficients; N varies from 3,930 to 3,932. All 98% confidence intervals exclude zero. Table 3 Fit Statistics of Structural Models Based on Meta-Analyses of Spector, Fox, Penney, Bruursema, Goh, and Kessler’s (2006) Counterproductive Work Behavior Facets Model 1. General factor model based on disattenuated correlations 2. General factor model based on observed correlations χ2 df TLI CFI SRMR RMSEA 527.43 30.53 5 5 .95 1.00 .97 1.00 .020 .010 .163 .036 Note: TLI = Tucker-Lewis Index; CFI = Comparative Fit Index; SRMR = Standardized Root Mean Square Residual; RMSEA = Root Mean Square Error of Approximation. was based on recommendations by Hu & Bentler, 1999, and Brown, 2006), evidence for this model is mixed. Whereas most indices meet conventional standards (Hu & Bentler, 1999) for acceptable or even good fit (Comparative Fit Index [CFI], Tucker-Lewis Index [TLI] ≥ .95; Standardized Root Mean Square Residual [SRMR] ≤ .08), the Root Mean Square Error of Approximation (RMSEA) of .16 clearly exceeds the recommended value of .06. In contrast, model fit was good to excellent according to all indices employed with observed correlations as input. Parameter estimates for factor loadings were uniformly strong with both the corrected (.77 for withdrawal to .99 for sabotage) and the uncorrected (.68 for withdrawal to .77 for sabotage) matrices used as input. Discussion Although a K of 11 may not seem large for a meta-analysis, the fact that all primary studies used the same instrument implies that the present meta-analysis methodologically approximates a primary study with a sample size close to 4,000. Given this situation, the evidence we found for one factor underlying the uncorrected correlations between the five facets seems strong indeed. Even though the evidence is not quite as clear if corrected values are used as input, uncorrected covariance matrices are typically used as input for CFAs based on 218   Journal of Management / January 2016 primary studies as well. Nevertheless, use of uncorrected values may have obscured differences in reliabilities between subscales of the CWB-C. The abuse (ID) subscale is longer and more reliable than the other scales, and it is also conceptually different with respect to the target mode. Hence, after disattenuation one would expect the abuse scale to yield a smaller loading on the general CWB factor than the four OD scales if two, instead of one, factors would describe the higher-order structure most adequately. That was not the case, though, as the loading of abuse was substantial with corrected values (.80) and even higher than the loading of the OD factor withdrawal (.77). Together with the uniformly strong loadings on all other facets, this suggests that a general factor corresponding to a reflective model of CWB in fact seems to underlie relations among facets proposed by Spector et al. (2006), who explicitly dismissed a latent construct of general CWB. Whereas this finding thus indirectly supports some theories of CWB that assume a reflective nature of that construct (as outlined in the introduction), it by no means refutes (nor corroborates) theories that lack such an assumption but conceptualize CWB as formative. Conceptually adequate identification of formative models requires different procedures with item-level data (see Treiblmaier, Bentler, & Mair, 2011). Perhaps even more importantly for our present purpose, our findings are restricted to just one model of CWB and cannot be integrated or compared with other models by means of meta-analysis. We therefore collected an additional primary data set. Study 2 Study Objectives The objectives of the second study were to replicate findings from the meta-analyses and to clarify the issues listed in the present research that we were not able to address on the basis of existing data. Toward that end, it was useful to first establish a measurement model for the lowest order of CWB facets as a baseline standard against which to compare alternative structural models. The form of such a measurement model corresponds to the model represented by solid lines in Figure 1(a). In addition to the five-factor model shown there, we also chose the eleven-facet model as baseline because it represents the most fine-grained structure of CWB proposed in the literature. These two alternative baseline models enabled us to compare the two theoretical models directly (i.e., side by side in a nonintegrated manner). Based on the two measurement models, we next tested a number of hierarchical higherorder models that correspond to the form illustrated in Figure 1(b). Variants of second-order structures tested include models with one, two, three, and five higher-order factors added to the eleven facets, which corresponds to integrations of Gruys and Sackett’s (2003) model with the models of Spector et al. (2006), Bennett and Robinson (2000), and a single factor of general CWB, respectively. In a similar fashion, one or two higher-order factors are also added above the five-factor baseline model. Finally, we test the possibility that a bimodal structure describes item covariation better than a hierarchical structure. A bimodal structure corresponds to the form of the model depicted in Figure 1(c). As outlined in the introduction, the modes of target and content of CWB are particularly salient in the CWB models reviewed but are also partially confounded. We therefore tested models in which each CWB item loads simultaneously on one target and one content factor. Content factors correspond to the original eleven or five facets of the measurement models, whereas target factors correspond to the OD-ID distinction, partially augmented by the self as a third target. Marcus et al. / Structure of Counterproductive Work Behavior   219 Method Sample and procedure. Data were collected online with employees enrolled in the psychology program of a German distance university. Notably, students enrolled in distance learning programs differ fundamentally from typical undergraduates; the former tend to be considerably older, and most hold regular employment and have diverse backgrounds often including previous academic degrees (e.g., Dabbagh, 2007). Unlike undergraduates who work part-time to earn their living, our participants were regular employees who study after work. The present study was explicitly advertised to employed students only. Participants who indicated in a control question that they were self-employed or held an employment of less than six months’ tenure were excluded from analyses, as employees in Germany tend to work on probation for the first six months of their employment, which could create bias with regard to CWB. A total of N = 1,321 participants met these inclusion criteria. Of these participants, 24 percent were men. Their mean age was 35 years (SD = 8.9, range = 20 to 69). Participants worked in a wide range of industries, with many in medical and social services (22%), followed by other services (17%); education (14 %); manufacturing, retail, and public administration (8 to 9% each); financial and security services (5% each); and miscellaneous (11%). The mean of hours worked per week was 32.7 (SD = 11.6). About 32% held a management position. We used listwise deletion of missing data for our analyses because (1) loss of power seemed acceptable with the present sample size, (2) there was no indication of bias due to listwise deletion (number of missing items was uncorrelated with mean CWB reported: r = –.03, n.s.), and (3) CWBs tend to have modal values of zero, which implies little additional information to be expected from imputing missing values (cf. Switzer & Roth, 2002). Thus, sample size for all analyses reported is N = 1,237. Students received course credit in response (upon completion of the survey, participants were led to a different website for obtaining course credit in order to keep the survey anonymous). Measure and item coding. We used a German translation of Gruys and Sackett’s (2003) original questionnaire, which to our knowledge is the only available measure that taps the eleven-facet model of CWB. Translations were done using back-translation procedures, followed by discussions among three German/English bilinguals familiar with the topic of CWB until consensus was reached. The instrument has sixty-six items measuring the eleven categories with three to thirteen items each. Responses are scored on seven-point scales using a format adopted from the original scale. The format refers to hypothetical behavior (“would engage in the behavior,” with options ranging from “[would not] . . . no matter what circumstances” to “[would] . . . under a wide variety of circumstances”). The original authors explained use of this format by being able to eliminate possible individual differences in the opportunity to actually perform the behavior in question, and they justify it with the established link between intentions and actual behavior. Whereas assignment of items to the eleven categories was done by the original source (Gruys & Sackett, 2003), an additional coding was needed to categorize items to target factors and to Spector et al.’s (2006) facets. Three of the present authors independently coded each item on the target and the Spector et al. categories. Cases of disagreement were discussed until consensus was reached. Five items were dropped from all analyses because either the target was unspecified (e.g., “fail to read the manual outlining safety procedures”) or the item specified a target seriously restricting the opportunity to perform the act (e.g., 220   Journal of Management / January 2016 “make unwanted sexual advances toward a subordinate”). In analyses involving the Spector et al. facets as the baseline model, four additional items were dropped because they could not be assigned unequivocally to any of the categories (all originally from the misuse of information scale, for example, “provide organization with false information to obtain a job”). Data analysis. Data were analyzed using LISREL 8.5 (Jöreskog & Sörbom, 2001) with observed item scores as indicators. As items had an ordinal format, the matrix of polychoric correlations was used as input. A problem typical of CWB studies is that distributions of observed variables tend to be seriously skewed, which violates assumptions of normal-theorybased maximum likelihood estimation. The asymptotic covariance matrix is therefore used as input in all analyses to yield robust χ² estimates. Furthermore, even robust maximum likelihood is not recommended with severely skewed ordinal indicators for a variety of reasons (cf. Brown, 2006, for a summary). Moreover, we intended to test a number of partially contradictory models, which implies that some of the models might turn out to be misspecified. Recent simulation data suggest that various robust estimation techniques are superior to the asymptotically correct weighted least squares procedure under such conditions (Yang-Wallentin, Jöreskog, & Luo, 2010). Furthermore, in another simulation study (Forero, MaydeuOlivares, & Gallaro-Pujol, 2009), unweighted least squares (ULS) estimation outperformed the diagonally weighted least squares (DWLS, which is similar to robust weighted least squares; cf. Flora & Curran, 2004) method in terms of accuracy of parameter estimates and fit statistics if various undesirable properties (e.g., skewness, small factor loadings) combine. The major pitfall associated with ULS, its lack of scale invariance, is not relevant if ordinal indicators with the same format are used as input, as was the case in the present study (Forero et al., 2009). We therefore report ULS estimates for the following analyses (in line with Forero et al.’s findings, however, results for DWLS estimation tended to be very similar and are available upon request). An additional methodological difficulty occurs for tests of the bimodal models involving target and content factors simultaneously (see Figure 1[c]). These models technically correspond to the form of a multitrait-multimethod model with correlated traits and correlated methods (referred to as CT-CM). While theoretically able to separate trait from method (or content from target) variance, such models are unfortunately often afflicted by serious problems such as nonidentification or improper solutions (cf. Eid, Lischetzke, Nussbeck, & Trierweiler, 2003). As a possible solution, Eid and colleagues (Eid, 2000; Eid et al., 2003; Eid, Nussbeck, Geiser, Cole, Gollwitzer, & Lischetzke, 2008) proposed a variant of the CT-CM model in which one method factor is defined as a standard procedure and therefore not explicitly modeled. This method is referred to as “correlated traits-correlated methods minus one” (CT-C[M-1]) model. The CT-C(M-1) model defines a method factor as a residual common to that method that cannot be predicted by the true scores on the traits as measured by the standard method. Hence, there is a substantive rationale for expecting traits to be uncorrelated with methods, whereas this is an arbitrary expectation in CT-CM models. Further, the CT-C(M-1) is relatively assumption free and thus avoids most identification problems associated with CT-CM models. However, a major drawback of the CT-C(M-1) model is that estimates and interpretation of the parameters depend upon the method chosen as the standard. More precisely, the (unmodeled) “method” chosen as standard defines a reference against which variance components Marcus et al. / Structure of Counterproductive Work Behavior   221 specific to modeled method factors are compared. Method factors can be interpreted as deviations from the reference (consequently, Geiser, Eid, & Nussbeck, 2008, replace the term method with nonreference). Choice of the reference method thus should be based on theoretical rationales (Geiser et al., 2008). We chose organization (OD) as the reference target (method) and model other persons (ID) and the self as nonreference target (method) factors in the present study. The rationale for this decision is that all CWBs can be conceptualized as eventually affecting the organization as a social system. Whereas CWBs categorized as OD tend to harm the organization directly, CWBs directed at other individuals or the self may also (indirectly) harm the organization but be affected by unique factors like personal conflicts or risk-taking propensity that bear little relationship to the specific organization. Confirming a model that includes ID and self as target factors would thus provide evidence for such uniqueness, whereas parameters associated with these factors would help to quantify unique effects beyond organizational factors common to all CWBs. Results Table 4 shows fit statistics for a series of CFA models, all of which are based on the 11-facet model of CWB as a baseline. As only some of our models are conceptually nested in each other, we focus in our comparative interpretations mainly on so-called information criteria of fit, which, unlike the scaled χ² difference test available for robust estimation (cf. Brown, 2006), are designed for selecting among nonnested models. Specifically, we report Akaike’s information criterion (AIC) and the Bayesian information criterion (BIC). AIC and BIC are based on different mathematical models and were found to have different properties and largely opposing pros and cons under a variety of conditions (Vrieze, 2012). We therefore report both AIC and BIC. Lower values indicate better fit. The measurement model with eleven intercorrelated facets fits the data appropriately according to most indices (RMSEA = .041; SRMR = .077; TLI/CFI = .98). We could therefore proceed by adding structural components to that baseline. Although a model with a general second-order factor above the 11 facets still had an acceptable fit (RMSEA = .047; SRMR = .089; TLI/CFI = .97), both AIC (6,849 vs. 5,704) and BIC (7,530 vs. 6,611) indicated considerable increase in misfit compared to the measurement model. We next tested a model in which the second-order structure was described by the target factors of ID and OD. Cross-loadings of these two factors were permitted if the primary factors contained items for CWB directed at both the organization and other people (which was the case for theft, property destruction, and misuse of information). This specification led to some improvement over the general factor model (AIC = 6,618; BIC = 7,315). A still better fit (AIC = 6,337; BIC = 7,044) was found for a model in which self-directed CWB was added as a third target factor (with that factor loading on unsafe behavior, drug use, and alcohol use). We also tried to model a second-order structure corresponding to Spector et al.’s (2006) five factors in a number of theoretically meaningful variants (e.g., with or without the facet of misuse of information, which is difficult to assign to any of the Spector et al. factors), yet all specifications led to improper solutions (i.e., Heywood cases, namely, standardized parameters larger than one) and are therefore not reported in Table 4. To summarize, hierarchical CFA models varied in model fit with three target factors providing for the best fit, yet none of those models fitted the data as well as the measurement model. 222   Journal of Management / January 2016 Table 4 Fit Statistics of Structural Models Based on Primary Study Measures of Gruys and Sackett’s (2003) Counterproductive Work Behavior (CWB) Facets Model 1. Measurement model 2. General factor model 3. Two-factor model with interpersonal deviance (ID) and organizational deviance (OD) as higher-order factors 4. Three-factor model with self-directed CWB added to Model 2 5. Five-factor model based on Spector, Fox, Penney, Bruursema, Goh, and Kessler’s (2006) specification 6. Bimodal correlated traits-correlated methods minus one (CT-C[M-1]) model with ID as method factor 7. Bimodal CT-C(M-1) model with ID and self-directed CWB as method factors 8. Mixed bimodal/hierarchical model with general factor added to content factors in Model 6 9. Mixed bimodal/hierarchical model with general factor added to content factors in Model 7 10. Mixed bimodal/hierarchical model with three principal axes factor analysis–based higher-order factorsa added to content factors in Model 7 χ2 df TLI CFI SRMR RMSEA AIC BIC 5,350.45 1,714 6,582.54 1,758 6,346.60 1,755 .98 .97 .97 .98 .97 .97 .077 .089 .087 .041 .047 .046 5,704.45 6,610.77 6,848.54 7,529.56 6,618.60 7,314.98 6,061.45 1,753 .97 .97 .084 .045 6,337.45 7,044.07 inadmissible parameter estimates 5,218.99 1,713 .98 .98 .076 .041 5,574.99 6,486.43 5,210.78 1,711 .98 .98 .076 .041 5,570.78 6,462.46 6,251.21 1,757 .97 .97 .088 .045 6,519.21 7,205.35 6,218.12 1,755 .97 .97 .087 .045 6,490.12 7,186.50 5,948.54 1,752 .97 .97 .085 .044 6,226.54 6,938.28 Note: N = 1,237; TLI = Tucker-Lewis Index; CFI = Comparative Fit Index; SRMR = Standardized Root Mean Square Residual; RMSEA = Root Mean Square Error of Approximation; AIC = Akaike’s information criterion; BIC = Bayesian information criterion. In Model 3, Facets 1, 2, 3, 5, 10, and 11 from the right-hand column in Table 1 load on ID, and facets 1, 2, 3, 4, 6, 7, 8, and 9 load on OD. In Model 4, specification is the same as in Model 3, except that Facet 5 loads on both ID and self-directed CWB, whereas Facets 8 and 9 load exclusively on selfdirected CWB. In Model 5, various alternative specifications all led to improper solutions. a. See text for descriptions of these factors. In contrast, bimodal CT-C(M-1) models with targets specified as “method” factors fitted the data even better than the measurement model. If ID was specified as the sole addition to the OD reference target, AIC dropped to 5,575 and BIC dropped to 6,486. Adding selfdirected CWB as a second nonreference target led to slightly better fit statistics (AIC = 5,571; BIC = 6,462). In the latter model, item loadings on each target factor were constrained to equality, which is typically necessary for CT-C(M-1) models if method factors are not correlated substantially (r = .34 in this case; see Eid et al., 2003). Whereas loadings of items on the eleven content (“trait”) factors were uniformly substantial (all 61 standardized loadings ≥ .53), coefficients were much smaller for ID (.24; .23 if just ID is specified) and the self (.28) as primary targets. Finally, we tested the possibility that a general factor underlies relationships between the content factors as directed at the organization (i.e., mixed Marcus et al. / Structure of Counterproductive Work Behavior   223 hierarchical-bimodal models in which content facets are allowed to load on a general factor). These models performed better than the purely hierarchical general factor model but not as well as the purely bimodal models (two targets added: AIC = 6,490; BIC = 7,187; only ID added: AIC = 6,519; BIC = 7,205). As Spector et al.’s (2006) factors did not appear to describe a higher-order structure that could be integrated well with Gruys and Sackett’s (2003) 11 facets, we also tested the possibility that Spector et al.’s facets are more adequately modeled as first-order factors. Fit statistics for a series of models using the Spector et al. structure as baseline are presented in Table 5. A measurement model with the original five factors could be fitted to the data, but indices partially indicated barely acceptable fit (RMSEA = .067; SRMR = .099; TLI/CFI = .96). If the original abuse factor was split into a verbal and physical component, statistics for absolute fit improved slightly (RMSEA = .063; SRMR = .097; TLI/CFI = .96), and both AIC (10,279 vs. 9,250) and BIC (10,914 vs. 9,911) indicated that the latter model was superior. We therefore modeled higher-order structures above six instead of the original five factors. Again, specification of one general factor led to a drop in fit (AIC = 9,829; BIC = 10,443) compared to the measurement model, whereas two higher-order targets of ID and OD with cross-loadings allowed where appropriate fell in between the general and the measurement model (AIC = 9,743; BIC = 10,363). Attempting to add a third higher-order target factor of self-directed CWB led to inadmissible parameter estimates. Similar to findings reported above for 11 facets as baseline, bimodal CT-C(M-1) models fitted the data considerably better than hierarchical or even the baseline models. Unlike in the previous analyses, however, this time a model with both ID and self specified as targets (AIC = 7,718; BIC = 8,393) showed considerable improvement over a model with just ID as the nonreference factor (AIC = 8,602; BIC = 9,268). Factor loadings were again more substantial for content (.47 to .90) than target factors (.37 for ID; .48 for self; equalized within each factor), although the difference was not as pronounced as with eleven content factors and the targets were more substantially correlated (r =. 54). A model with a general factor above the six content facets plus one independent ID facet produced a fit between that of the purely hierarchical and bimodal versions (AIC = 8,947; BIC = 9,567; a general factor model with two additional targets did not converge with ULS; results with DWLS estimation indicated a similar trend of intermediate fit and are available from the first author). Overall, the bimodal model including self-directed CWB yielded the best fit with Spector et al.’s (2006) facets as baseline, yet none of these models came close to the fit of corresponding models based on 11 facets. Finally, we took some steps to try to resolve an apparent inconsistency in our findings on the higher-order structure of CWB. Whereas Study 1 had supported the assumption of a general factor of CWB, the present study indicated significant drops in model fit when we added higher-order factors. Notably, this drop occurred despite the uniformly large correlations among first-order content factors (.69 to .81) and loadings on the general factor (.72 to .97) observed for our best-fitting CT-C(M-1) Model 7 in Table 4 (similar coefficients were observed with other higher-order models). First, we subjected the matrix of latent correlations among content factors in that model to an exploratory principal axes factor analysis (PAF) to find out whether a different latent structure than specified may underlie these relations. This analysis, however, unambiguously indicated a one-factor solution (first three PAF 224   Journal of Management / January 2016 Table 5 Fit Statistics of Structural Models Based on Primary Study Measures of Spector, Fox, Penney, Bruursema, Goh, and Kessler’s (2006) Counterproductive Work Behavior (CWB) Facets Model 1. Measurement model with five factors 2. Measurement model with abuse split into verbal and physical component 3. General factor added to Model 2 4. Two-factor model with interpersonal deviance (ID) and organizational deviance (OD) as higher-order factors 5. Three-factor model with selfdirected CWB added to Model 2 6. Bimodal correlated traitscorrelated methods minus one (CT-C[M-1]) model with ID as method factor 7. Bimodal CT-C(M-1) model with ID and self-directed CWB as method factors 8. Mixed bimodal/hierarchical model with general factor added to content factors in Model 6 χ2 df TLI CFI SRMR RMSEA AIC BIC 10,031.16 1,529 .96 .96 .099 .067 10,279.16 10,914.10 8,992.37 1,524 .96 .96 .097 .063 9,250.37 9,910.90 9,589.45 9,501.46 1,533 1,532 .96 .96 .96 .96 .100 .100 .065 .065 9,829.45 9,743.46 10,443.91 10,363.03 inadmissible parameter estimates 8,341.93 1,523 .96 .96 .096 .060 8,601.93 9,267.59 7,453.82 1,521 .96 .97 .093 .056 7,717.82 8,393.72 8,705.14 1,532 .96 .96 .099 .062 8,947.14 9,566.71 Note: N = 1,237; TLI = Tucker-Lewis Index; CFI = Comparative Fit Index; SRMR = Standardized Root Mean Square Residual; RMSEA = Root Mean Square Error of Approximation; AIC = Akaike’s information criterion; BIC = Bayesian information criterion. In Model 4, theft, sabotage, verbal abuse, and physical abuse load on ID, and theft, sabotage, withdrawal, and production deviance load on OD. In Model 5, production deviance loads on self-directed CWB in addition to OD. eigenvalues: 8.172; 0.027; 0.024). Next, we computed composite scores for Gruys and Sackett’s (2003) facet scales and performed another PAF on these scores, in the same manner as reported in Note 1 for these authors’ original data. The sequence of PAF eigenvalues (5.350; 0.942; 0.316) closely paralleled those found with Gruys and Sackett’s data. A parallel analysis performed using O’Connor’s (2000) software indicated a three-factor solution (note that PAF eigenvalues do not need to exceed 1.00 to be meaningful). After oblique promax rotation, Factor 2 (loaded by destruction of property and inappropriate physical action) could be meaningfully interpreted as “physical aggression,” whereas Factor 3 (loaded by alcohol use and drug use) clearly referred to “substance use.” All remaining facets loaded on the first factor, which could be seen as general, mainly performance-related, CWB. Based on these findings, we next specified an alternative higher-order structure above the content factors of Model 7 in Table 4, which is referred to as Model 10 in that table. This model did not fit the data as well as the original Model 7, but it did fit better than its general factor variant (Model 9) (AIC = 6,227; BIC = 6,938). Correlations among the three higher-order factors ranged from r = .86 to .90. Marcus et al. / Structure of Counterproductive Work Behavior   225 Discussion Based on a sizable sample and, perhaps, on the most comprehensive set of CWB items available, we tested a number of possible variations and integrations across the various models of CWB proposed. First, if compared in a nonintegrative fashion, it became clear that the eleven-facet structure proposed by Gruys and Sackett (2003) described the primary factor structure of the present data set considerably better than Spector et al.’s (2006) five factors. Although this comparison may be perceived as unfair, as the present instrument was originally developed to measure the 11 facets, this finding held after restricting content to items that could be unequivocally assigned to Spector et al.’s facets. Our results also show that splitting Spector et al.’s more general facet of abuse into a verbal and a physical component may help overcome some deficiencies of this model (note that only two of eighteen abuse items in Spector et al.’s CWB-C refer to physical aggression). Yet the two models differ not only in the number and specificity of facets but also in the breadth of the overall domain covered. Whereas Spector et al.’s five factors may adequately describe a narrower specification of CWB, they appear insufficient to comprehensively cover a more broadly defined domain. Second, the present results help clarify how factors broader than those specified by both Gruys and Sackett (2003) and Spector et al. (2006) relate to the primary facets. Whereas the meta-analysis reported in Study 1 indicated that just one general factor seems to underlie correlations between the narrower set of facets measured in the CWB-C, the present data were more in line with a hierarchical structure that differentiates between targets at the secondorder level. Moreover, whereas Bennett and Robinson’s (2000) popular model distinguishes between just ID and OD as targets, the present analyses point to the conclusion that primarily self-directed CWB may be a useful addition to those two target factors. Specifically, we successfully modeled drug and alcohol use as directed at the self, whereas violations of safety procedures seem to be best modeled as immediately harmful for both others and the self. Further, the best fit was consistently found if targets were not modeled as higher-order factors above the level of content-driven facets but as primary factors that load simultaneously on the same items as the content factors. These findings appear to indicate that, across the three original models of CWB we attempted to test in the present research, at least two organizing principles or modes are confounded, which would be better kept separate. On one hand, there are meaningful differences between different types of CWB that could be categorized on the basis of similarities between behaviors themselves (e.g., stealing, failing to care for safety, verbally attacking others). On the other hand, at least some of these types of acts could be directed at different targets (e.g., stealing from the company or a coworker, endangering others, and/or endangering the self), which could lead to important distinctions within content categories. However, the different parameter values observed also indicate that differences in content of counterproductive acts tend to be more significant than differences between targets. Results of Study 2 were not as conclusive with respect to the question of whether (one or more) higher-order factors underlie facets of CWB. On one hand, adding such factors to either the measurement or bimodal models consistently impaired model fit, although the absolute fit still remained acceptable. On the other hand, latent facet correlations on which these analyses were performed consistently showed very high values, and exploratory factor analyses corroborated some form of positive manifold. 226   Journal of Management / January 2016 General Discussion In the present studies, we tried to address four research questions: (1) Is there one (or more) latent construct(s) underlying overall CWB? (2) Which types of behaviors should be subsumed under the domain of CWB? (3) Which, and how many, modes are needed to describe each act of CWB? (4) How can and should different models of CWB be integrated? Ad (1): Proponents of the view of “general CWB” as a formative construct (i.e., a collection of CWB facets; cf. Edwards, 2011) proposed that its constituent facets should be studied independently (e.g., Spector et al., 2006), whereas the position that all acts of CWB are, at least partially, driven by a common cause (as, e.g., posited in Gottfredson and Hirschi’s, 1990, general theory of crime) implies we should focus primarily on the common element in both research and practice aimed at CWB. To our knowledge, this was the first systematic attempt to resolve this (implicit) theoretical controversy empirically. We tried to approach the issue from different angles, using meta-analytic scale–level and primary item–level data. Unfortunately, results from both studies did not perfectly converge. Whereas the scale-level study supported a general-factor model, the item-level analyses led to impaired fit with any kind of reflective structure added above the original facet level. Ironically, there was more support for a reflective general factor based on Spector et al.’s (2006) explicitly formative CWB model than based on the model by Gruys and Sackett (2003), who seem more sympathetic to a reflective structure. The question then arises of how this apparent inconsistency can be explained. One potential explanation may lie in content differences between the two underlying models. Although based on quite similar definitions of CWB (with harm doing as the core defining element), the two resulting measures differ considerably in breadth of domain coverage. Besides the issues discussed below as related to our second research question, the narrower scope of Spector et al.’s (2006) measure may have happened to restrict the content to a set of CWBs that carry relatively little unique variance. Upon inspection, it is apparent that selfdirected and more severe acts of CWB are excluded from the CWB-C, which might explain why the remaining acts appear to fit a general factor better. In fact, the two additional factors we extracted in our PAF in Study 2 were related to more serious forms (physical aggression) and self-directed CWB (substance use), respectively, and adding these factors improved model fit to some extent. Still, all 11 facets are consistently and substantially correlated in both latent and observed matrices, as reported in the Results section. Overall, the present studies suggest that some general latent factor underlies all acts of CWB, but that this factor emerges less clearly as more complexity is added to the model, in terms of both content (e.g., self-directed CWB) and methodological factors (e.g., acts with differing severity and thus prevalence). If this were true, a number of implications for the theory and measurement of CWB would follow. As outlined in the introduction, some theoretical approaches to CWB correspond to reflective models, whereas others imply some focus on unique facets that do not need to form a general factor. Theorists focusing on the former kind of approach should make clear exactly which kinds of CWB fit their reflective definition and establish empirically that their measure fits this structure. This may not so much be a question of the intended structure of the measure used because, as shown in Study 1, CWB scales may fit a reflective model even if not designed to do so. If the focus is on situational triggers, and on personality more as a moderator of reactions to those triggers, Marcus et al. / Structure of Counterproductive Work Behavior   227 establishing a higher-order structure of CWB may not be as relevant. Yet in the latter case, expected uniqueness of CWB facets should be linked explicitly to theoretical antecedents, which implies that establishing the distinctiveness of CWB facet measures becomes more relevant. With few exceptions (e.g., Marcus & Wagner, 2007; Spector et al., 2006), CWB studies corresponding to both types have rarely been very explicit in that respect. If both kinds of approaches are integrated, it seems plausible to focus in person-centered approaches on what is simple and shared among CWBs, whereas situation-centered approaches appear particularly useful for explaining what is complex and unique to certain CWBs. Again, the issue of CWB as a formative versus reflective construct primarily is a theoretical one. There may not be an ultimate empirical test of either position, yet our findings are in line with a position that the two approaches are both legitimate but focused on different levels. Ad (2): Beyond the conceptual nature of overall CWB, Study 2 also allowed us to address the substantive structure of CWB in terms of the adequate breadth and content of the overall domain and its facets. Data from that study were based on a broad set of CWBs. With this broad scope of the domain, our results suggest that a relatively large number of factors are necessary to describe the facet level of CWB. On the content side, we found that Spector et al.’s (2006) five factors do not adequately cover the domain as defined by Gruys and Sackett’s (2003) set of 11 facets. Spector et al.’s model thus appears insufficient to describe the structure of CWB comprehensively. With regard to targets, we found that the dichotomous ID-OD distinction should be augmented to include self-directed CWB. Hence, the set of items in Gruys and Sackett’s measure provides for substantive additions beyond both other models and also in terms of both modes that seem to underlie those models of CWB. Evidence of a third target, which has been largely overlooked in empirical research on CWB, is a unique contribution of the present research. In line with Martinko et al. (2002), we propose that the intention to do harm, as associated with the idea of retaliation, is of only minor relevance for self-directed CWB. In addition to this theoretical implication, we propose that this finding also has implications for the measurement of CWB, as discussed in more detail in the following section. Ad (3): Study 2 also allowed us to compare hierarchical with bimodal integrations. In these comparisons, bimodal models tended to outperform hierarchical variants. Hence, the evidence seems relatively clear that acts of CWB can vary, in meaningful ways, in target and content simultaneously rather than being described by just one mode. Although for some content there may be little within-factor variation in targets (e.g., poor attendance overlaps with OD, aggression with ID, substance use with self-directed CWB), other types of CWB (e.g., theft, property destruction, misuse of information) may be directed at different targets. This may look like an intuitively plausible, if not obvious, feature of CWB, but, to our knowledge, it has not been formally tested in any previous research. It implies that, for research and practice aimed at the facet level of CWB, measures successfully taken to combat one specific content-target combination (say, stealing from the organization) do not necessarily work for the same content factor combined with a different target (stealing from coworkers). To some extent, different causes may drive CWBs of similar content directed at different targets, although effect sizes of parameter estimates (i.e., factor loadings) indicate that variation in content tends to be more important than variation due to different targets. Ad (4): Finally, we tested a number of possible ways in which the models reviewed could be integrated. Best fit was observed for bimodal integrations of Gruys and Sackett’s (2003) 228   Journal of Management / January 2016 content facets with Bennett and Robinson’s (2000) OD-ID targets augmented by the self. Although Spector et al.’s (2006) facets were generally not well supported, a bimodal integration with targets here also fit the data better than other variants of that model. Hence, it seems adequate to integrate different models of CWB content with target models, but the two content models may provide for alternative specifications that are hard to integrate. This interpretation differs from conclusions implied by a review of the original authors’ intentions. As discussed in the introduction, Spector et al. used Bennett and Robinson’s (2000) ID-OD distinction as a starting point for developing a more fine-grained model. Similarly, Gruys and Sackett refer to the same distinction as being overly broad to structure the domain of CWB. One might thus expect the two more recent models to map onto the ID-OD distinction in a hierarchical fashion (i.e., similar to narrow facets of big five personality factors). In contrast, the present results imply that either of the two more recent models is best integrated with the ID-OD distinction at the same level in the hierarchy. Theoretically, this implies that both Spector et al.’s and Guys and Sackett’s facets are not just more specific than those of Bennett and Robinson but have a conceptually different meaning. If, for example, a study looked at patterns of correlations between CWB facets according to one content model with outside variables, it would be difficult to draw conclusions from that study on relations of the same outside variables with CWB target factors. It is thus important, in terms of parsimony of research, to report findings at the facet level for both the target and content mode in primary research, as this would be difficult post hoc (e.g., in meta-analyses) in lieu of item-level data. Implications for Research and Practice The present research has a number of implications for the practice of research and human resources management apart from those already mentioned in the previous section. First, the two currently most widely used measures of CWB (Bennett & Robinson, 2000; Spector et al., 2006) may both not cover the full range of counterproductive acts. Bennett and Robinson (2000) decided to drop acts of CWB that produced low base rates or did not fit with their ID-OD distinction, and Spector et al. based their CWB-C on a previous measure originally tapping the same distinction. Such omissions may have restricted the scope of measurement in a way that leaves severe acts of CWB and particular facets (e.g., self-directed CWB) largely unmeasured. However, practitioners may be particularly interested in serious forms of CWB, as these acts are likely to have the most detrimental effects. As these acts are rare, including them in measures of CWB may impair technical psychometric properties of the measure and at times even create zero variance (note, however, that none of the items in Study 2 had zero variance) or spurious factors (cf. Gruys & Sackett, 2003). It is therefore tempting for researchers to drop those items. Yet we need to be aware that this practice may also change the meaning of measures in unknown ways. Furthermore, practitioners may see reason in treating primarily self-directed CWB differently than CWB directed immediately at the organization or its members. Whereas the latter types of CWBs call for sanctions aimed at protecting others, self-directed CWBs such as substance use may lead organizations to look for interventions aimed primarily at helping individual employees. In fact, substance use, safety violations, and related behaviors are topics often treated separately from other CWBs in fields such as occupational health, although there is evidence that the causes and correlates of these acts overlap with those of general CWB (e.g., Lehman, Farabee, Holcom, Marcus et al. / Structure of Counterproductive Work Behavior   229 & Simpson, 1995; Lehman & Simpson, 1992). Measures including severe and self-directed acts of CWB would allow for more direct examination of differences and parallels between these and other acts of CWB. Of course, the downside of comprehensiveness is impaired cost-effectiveness. For example, Gruys and Sackett’s (2003) scale that we used for Study 2 may be considered overly long for many practical applications. However, we believe that this instrument could be cut to a similar size as Spector et al.’s widely used CWB-C (which has 33 items) without losing much reliability. Some of the eleven facets are measured by 10 or more items, and some behaviors are repeated for several interpersonal targets (e.g., supervisor, coworker, customer). However, the potential nature of CWB as a construct that is multidimensional at the facet level but also partially driven by a common reflective factor may only be captured by using relatively lengthy scales. Beyond the already mentioned necessity to link facets of CWB to antecedents theoretically, another implication of our findings refers to study design. Situation-centered approaches typically assume that CWB follows external triggers, which implies that triggers and CWB occur in a certain chronological sequence. Person-centered approaches such as self-control theory do not make such assumptions (they basically require just some sort of opportunity on the situation side). An agenda for future research on CWB thus not only should include a more precise examination of the extent to which specific facets of CWB require specific explanations but should also rely more often on longitudinal designs than is current practice in CWB research. A more practical implication of our findings is closely related to the focus on CWB facets. As mentioned earlier, previous evidence suggests that there may be both common and specific antecedents across CWB facets. This is in line with the relatively complex structure we observed at lower levels in conjunction with substantial covariance among facets. The practically relevant question then becomes whether interventions tailored to specific facets are worth the additional effort, in terms of incremental validity, beyond interventions of empirically established utility at more general levels (e.g., integrity or personality tests used in personnel selection and organizational policies for improving justice perceptions; cf. Berry et al., 2007, among others). This is a question that needs to be addressed more systematically in future research. Again, the present results may be used to guide such efforts. Strengths and Limitations A particular strength of the present research is that it combined the evidence available from a coherent set of previous studies with a newly collected large data set that allowed us to examine alternative models of CWB in a comprehensive fashion. We were thus able to utilize strengths of both meta-analytic and primary studies while avoiding many of each method’s major shortcomings. Another unique contribution of this research is that, to our knowledge, no previous study of the structure of CWB has been based on a similarly systematic review of the various aspects of this topic highlighted (or at times neglected) in the currently most influential individual models of CWB. This theoretical synopsis allowed us to delineate ways in which these models may or may not be integrated, whereas the size and scope of the data sets collected allowed us to examine many of these possibilities empirically. Furthermore, application of CT-C(M-1) models allowed us to address the issue of modality for the first time in a methodologically adequate fashion. 230   Journal of Management / January 2016 Nevertheless, some limitations need to be mentioned as well. Most obviously, both our meta-analytic and primary data sets relied exclusively on self-report data, which could imply issues with common method variance. As argued recently by Conway and Lance (2010), this concern is not automatically justified but can be refuted on the basis of theoretical arguments and empirical evidence. Theoretically, our focus on the internal structure of CWB almost required to hold constant the source of ratings, as otherwise any observed structure could only be attributed equivocally to substantive or to method factors. Empirically, research in the area of CWB has shown that self-report and non-self-report data tend to produce similar relations to outside variables as well as similar estimates of internal consistency reliability (Berry, Carpenter, & Barratt, 2012; Berry et al., 2007). In the present research, it was not possible to address this issue directly because collection of non-self-reports of sufficient scope and sample size was impractical. Unfortunately, every kind of CWB measure tends to create its own set of problems, and those due to range restriction in objective measures may be even worse than those due to possible desirability in self-reports (Fox, Spector, Goh, & Bruursema, 2007). To the extent that correlations among CWB items may have been inflated, these inflations should have led to overrating commonality in CWB at the expense of specificity. As our major findings point to the potential significance of specific facets in spite of this potential bias, we hold that our conclusions were mostly conservative rather than liberal. Another potential drawback of the present research was that our set of CWBs, although relatively comprehensive, still lacked a number of potentially relevant additional modes, as well as CWBs specific to particular industries or job groups (cf. Bowling & Gruys, 2010). In a similar manner as we have established for the distinction of target and content, it is possible that the structure of CWB is even more complex if those additional factors are added. However, it seems wise to start with a relatively assumption-free (or nontautological) general definition of the domain of CWB (such as the definition based on harm doing we employed) and then proceed by examining necessary distinctions within that domain step by step, as implied by the specific theoretical approach. Toward that end, we hope the present results provide future researchers with some reasonable guidance as to what point to start from and where to go from there. Appendix References to Studies Included in the Meta-Analysis Bayram, N., Gursakal, N., & Bilgel, N. 2009. Counterproductive work behavior among white-collar employees: A study from Turkey. International Journal of Selection and Assessment, 17: 180-188. Bolton, L. R., Becker, L. K., & Barber, L. K. 2010. Big five trait predictors of differential counterproductive work behavior dimensions. Personality and Individual Differences, 49: 537-541. Goh, A. 2007. An attributional analysis of counterproductive work behavior (CWB) in response to occupational stress. Unpublished doctoral dissertation, University of South Florida, Tampa. Hunter, E. M., & Penney, L. M. 2007. The waiter spit in my soup! Counterproductive behavior toward customers. Poster session presented at the annual meeting of the Society for Industrial and Organizational Psychology, New York. Kessler, S., O’Brien, K., Spector, P., Bandelli, A., Borman, W., Nelson, C., & Penney, L. 2008. Is Machiavellianism inherently bad? A reexamination of previously held views. Poster session presented at the annual meeting of the Society for Industrial and Organizational Psychology, San Francisco, CA. Krischer, M. M., Penney, L. M., & Hunter, E. M. 2010. Can counterproductive work behaviors be productive? CWB as emotion-focused coping. Journal of Occupational Health Psychology, 15: 154-166. Marcus et al. / Structure of Counterproductive Work Behavior   231 Spector, P. E., Fox, S., Penney, L. M., Bruursema, K., Goh, A., & Kessler, S. 2006. The dimensionality of counterproductivity: Are all counterproductive behaviors created equal? Journal of Vocational Behavior, 68: 446-460. Villanueva, L. S. 2006. An examination of the role of self-control in the prediction of counterproductive work behaviors: Does cognition matter? Unpublished doctoral dissertation, University of Houston, TX. Vincent, R. C., Shoptaugh, C. F., & Miller, A. 2008. Personality, motivational, and behavioral antecedents to…

Collepals.com Plagiarism Free Papers

Are you looking for custom essay writing service or even dissertation writing services? Just request for our write my paper service, and we'll match you with the best essay writer in your subject! With an exceptional team of professional academic experts in a wide range of subjects, we can guarantee you an unrivaled quality of custom-written papers.

Get ZERO PLAGIARISM, HUMAN WRITTEN ESSAYS

Why Hire Collepals.com writers to do your paper?

Quality- We are experienced and have access to ample research materials.

We write plagiarism Free Content

Confidential- We never share or sell your personal information to third parties.

Support-Chat with us today! We are always waiting to answer all your questions.