Introduction
Modern assessment centers (ACs) are usually designed to evaluate, in individuals, a set of prespecified dimensions concerned with the knowledge, skills, abilities, or other characteristics (KSAOs) thought to be associated with performance in a particular job or job-set. In the context of ACs, these KSAOs are generally labelled “competencies” by practitioners and “dimensions” by academic researchers. In this article, following the academic tradition, I usually refer to the KSAOs which ACs are designed to measure as dimensions, but the reader may substitute the label competencies for dimensions.
Despite the widespread use of ACs globally (Lowry, Reference Lowry1997; Spychalski et al., Reference Spychalski, Quinones, Gaugler and Pohley1997; Thornton & Byham, Reference Thornton and Byham1982), the question of whether they are capable of reliably assessing the dimensions they are designed to measure has been a source of concern and debate in the academic literature for at least 70 years (Lance, Reference Lance2008; Sakoda, Reference Sakoda1952). Addressing this issue, Lance (Reference Lance2008) reviewed the research evidence accumulated to that point on the construct validity issue. He concluded that there was insufficient evidence that ACs measure dimensions, and recommended that attempts to do so should be abandoned. He further suggested that instead of seeking to measure dimensions, ACs should be redesigned to focus on assessing how well people perform in various situations of relevance to a job. Given the widespread use of ACs globally, almost all of which are designed to evaluate dimensions, the resources that have been and still are committed to them, and the high stakes decisions regularly taken in organizations worldwide on the basis of dimension assessment, his proposal has far-reaching consequences.
In response to Lance’s (Reference Lance2008) work, 10 academics and practitioners, all experts in the field of ACs, considered his argument and the evidence on which it was based (Arthur et al., Reference Arthur, Day and Woehr2008; Brannick, Reference Brannick2008; Connelly et al., Reference Connelly, Ones, Ramesh and Goff2008; Howard, Reference Howard2008; Jones & Klimoski, Reference Jones and Klimoski2008; Lievens, Reference Lievens2008; Melchers & Konig, Reference Melchers and Konig2008; Moses, Reference Moses2008; Rupp et al., Reference Rupp, Thornton and Gibbons2008; Schuler, Reference Schuler2008). Almost all disputed Lance’s conclusion that the weight of available evidence indicated that ACs do not measure dimensions, offered alternative explanations for research which Lance had cited in support of his argument, and made various suggestions for further improving the construct validity of ACs as measures of dimensions.
Since this debate, further research evidence of direct relevance to the construct validity of ACs has emerged. Of most significance here is research using Generalizability (G) theory which has assisted, for the first time, in isolating the variance in AC ratings which is uniquely associated with dimensions, and therefore statistically independent of other sources of variance in AC measurement (Jackson et al., Reference Jackson, Stillman and Atkins2005, Reference Jackson, Michaelides, Dewberry and Kim2016, Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022; Putka & Hoffman, Reference Putka and Hoffman2013). This novel approach to the construct validity issue offers a way of finally resolving the long-standing controversy about the construct validity of ACs, and in particular the extent to which they successfully evaluate dimensions. It also offers insight into the extent to which dimension scores can be used to validly predict relevant outcomes, including the future work performance of individuals, and their training and development needs.
In order to place a presentation and discussion of this new evidence concerning the construct validity of ACs in context, I begin with a brief history of ACs, discuss the history and nature of the KSAOs (i.e., competencies/dimensions) they are designed to measure, and consider the growth of competencies in organizations since the mid 20th century, including the increasing use of competency models and frameworks, and competency-based management. I then present and evaluate evidence on the construct validity of ACs, particularly that which has emerged since this issue was last debated in 2008, and draw conclusions about whether, in the light of this, the claim that well-designed ACs measure stable dimensions is justifiable. I consider also the validity and practical utility of other outputs which can be derived from ACs, including the measured performance of candidates on exercises, and their overall performance in an AC.
A short history of ACs
The personnel selection system which was eventually to evolve into the modern day AC was introduced in the late 1920s to select officers for the German army, navy, and air force (Fitts, Reference Fitts1946; Hayes, Reference Hayes1995). The most documented is that for the selection of army officers. From 1930, the psychologist Dr. Max Simoneit led this army officer selection system which by 1936 had expanded to involve 114 psychologists, operating in 16 separate selection stations across Germany (Ansbacher, Reference Ansbacher1941). Candidates were assessed over two days by three psychologists (each of whom had received three years training in selection), a colonel, and a medical officer. The tasks performed by the officer candidates were those in which their behavior could be observed and recorded. They included tests of sensory-motor coordination; a 45 minute obstacle course, a between-candidate discussion; an orders test in which candidates were required to instruct a group of soldiers in carrying out a mechanical task such as making a coat hanger out of a piece of wire; and one in which the candidate was required to pull on an electrified metal spring as hard as possible, with the expression on his face being photographed with a hidden camera, and later analyzed. These tasks were designed to examine prespecified personal characteristics. For example, the obstacle course was used to measure a candidate’s physical endurance and will-power. The assessors of each candidate collectively produced on him a comprehensive case report.
The selection procedures developed by German military psychologists in the 1930s included core elements of many modern ACs: the use of several trained assessors for each candidate, the observation and assessment of candidates over several different situations designed to replicate aspects of the role for which they are being considered, the avoidance of paper and pencil tests of underlying characteristics such as intelligence and personality, and the evaluation of candidates in relation to pre-designated characteristics (e.g., will-power). In the late 1930s, Simoneit described these military selection methods in great detail in a number of articles and a book (Ansbacher, Reference Ansbacher1941), and by the early 1940s, several favorable English language reviews of the methods had appeared (Fitts, Reference Fitts1946). With the onset of the Second World War, military selection systems based on the German model were introduced, first in Britain (the War Office Selection Board) and shortly after in America, Canada, and Australasia (Hayes, Reference Hayes1995; Highhouse & Nolan, Reference Highhouse, Nolan, Jackson, Lance and Hoffman2012).
Of these, the most documented is that set up in the United States by the Office of Strategic Services (OSS), the forerunner of the Central Intelligence Agency (CIA) (see Handler (Reference Handler2001) for a detailed description). The purpose of this process, which was designed by a team of psychologists led by Henry Murray, was to select spies and saboteurs to work behind enemy lines. Each candidate was assessed over three days by several psychologists. They completed an intensive battery of tests and tasks including paper-and-pencil tests of intelligence and personality; inventories assessing health, previous work conditions, and personal history; an in-depth personal history interview; and situational exercises such as an obstacle course, a task in which they were required to build a small structure while being undermined and criticized by stooges, and a mock interrogation. Ten characteristics, including energy and initiative, effective intelligence, leadership, and emotional stability, were assessed with the situational exercises, with an average of six such exercises used to assess each trait (Sakoda, Reference Sakoda1952).
After the assessment was completed, the ratings made independently by the three psychologists involved in assessing a candidate on each trait, on each situational exercise, were used to obtain, by consensus discussion, an average score on each trait. The psychologists then combined this rating data with the information about a candidate from the considerable number of other evaluations that had taken place, with the intention of deriving an overall and coherent picture of the individual’s personality and ability, and with it their suitability for the role as spy or saboteur. The OSS system had many of the features common to that developed previously in Germany. However, it added two features often found in modern ACs: the use of a trait by situation scoring matrix, and consensus meetings to pool scores derived from individual situational assessments. Just after the war ended, in an article on the OSS system, Murray and Mackinnon (Reference Murray and MacKinnon1946) described the OSS selection system as an “assessment center”. This is the first known use of this term (Highhouse & Nolan, Reference Highhouse, Nolan, Jackson, Lance and Hoffman2012).
In the aftermath of the war, non-military organizations began to use the AC approach developed in the German and allied military. One of the first was set up in Scotland in 1948 to select candidates applying for work in an engineering organization (Handyside & Duncan, Reference Handyside and Duncan1954). The first non-research AC in North America was introduced in 1958 at the Michigan Bell Telephone Company (Jaffee, Reference Jaffee1965), and soon afterward a revised version was introduced for operational selection decisions at AT&T (Bray et al., Reference Bray, Campbell and Grant1974; Bray & Grant, Reference Bray and Grant1966). At first, the non-military adoption of the AC approach to selection was slow, with only 12 organizations operating them in 1969 (Byham, Reference Byham1977). However, after a Harvard Business Review article was published on ACs (Byham, Reference Byham1970), the method quickly took off, and by the early 1970s they were operating in hundreds of North American organizations (Highhouse & Nolan, Reference Highhouse, Nolan, Jackson, Lance and Hoffman2012). This expansion continued throughout the 1980s and beyond in many countries, and today the AC plays a major and global role, particularly in the selection of graduates and candidates for supervisory and management roles (Lowry, Reference Lowry1997; Spychalski et al., Reference Spychalski, Quinones, Gaugler and Pohley1997; Thornton & Byham, Reference Thornton and Byham1982).
In contemporary AC designs, participants are assessed on several dimensions, across several exercises, by multiple trained assessors. Although there are several ways of carrying out these ratings (Lance, Reference Lance2008; Robie et al., Reference Robie, Osburn, Morris, Etchegaray and Adams2000), usually the assessor(s) evaluating the performance of a particular participant, on a particular exercise, rates them on each dimension after that exercise is completed (Woehr & Arthur, Reference Woehr and Arthur2003). The resulting dimension ratings are known as PEDRs (post-exercise dimension ratings). A participant’s overall assessment rating (OAR) may be obtained by summing her scores across exercises and dimensions, or through a consensus discussion between assessors. A participant’s score for a dimension may be obtained by summing her scores on that dimension across exercises. For further details of AC practices see Eurich et al. (Reference Eurich, Krause, Cigularov and Thornton2009), Krause and Thornton (Reference Krause and Thornton2009) and Spychalski et al. (Reference Spychalski, Quinones, Gaugler and Pohley1997).
ACs, dimensions, and individual differences
ACs were developed in the military for a particular purpose: to assess the suitability of candidates for war-related roles. The focus therefore was on the development of a robust and effective selection method. As a consequence, the psychologists involved in their development, notably Simoneit in Germany and Murray in North America, were more concerned with developing and improving their chosen method of selection than with abstract theoretical issues concerning the origin, structure, or nature of the dimensions on which they were assessing candidates. In contrast, mainstream academic psychologists concerned with personality developed complex theoretical perspectives (e.g., psychoanalytic theory, learning theory, cognitive theories, humanistic theories, trait-based theories, and biological theories), as did psychologists working on cognitive ability (e.g., general intelligence “G”, multifactor theories, multiple intelligences, and the componential subtheory).
As a consequence, a bifurcation, which exists to this day, developed between on the one hand the study of the nature of cognitive ability and personality, an area that now constitutes “individual differences”, an established constituent of mainstream psychology, and on the other hand the measurement of the specific characteristics and abilities thought to predict work performance, which is primarily undertaken by human resource professionals and studied by IO psychologists. Two other distinctions between these two perspectives are of note. First, the method generally adopted by academic psychologists to measure ability and personality, “paper-and-paper” tests and inventories, is quite unlike the AC method, based on the direct observation of behavior, developed by applied psychologists and HR practitioners. Second, the constructs assessed by those working in the individual differences tradition (e.g., general mental ability, and extraversion-introversion) are concerned with tendences of people to behave differently in any context, whereas those working with ACs (e.g., communication skills and leadership) concern the specific personal characteristics thought to predict the performance of people at work.
Competencies/dimensions and competency models
An important difference between the original military ACs and their modern-day counterparts is the intended purpose of the testing. Whereas the military assessment stations operated in Germany, the United States, and elsewhere sought to use observations and evaluations of candidates across a variety of measurement techniques, and measurement situations, to build up a whole integrated picture, or gestalt, of each candidate’s personality (Ansbacher, Reference Ansbacher1941; Highhouse, Reference Highhouse2002), in modern ACs the intention is to systemically and independently measure several work-related dimensions that are considered relevant to job performance, such as an individual’s leadership skills, and communication ability (Meriac et al., Reference Meriac, Hoffman and Woehr2014).
The importance of focussing on personal characteristics that appear relevant to performance in the work context was emphasized in an influential article by McClelland (Reference McClelland1973). McClelland argued that there was little reason to believe that intelligence or aptitude tests predict work performance, and little or no evidence that they did so. Based on the premise that intelligence tests do not predict work performance, McClelland proposed that such tests should be abandoned in personnel selection and replaced by an examination of a candidate’s “competencies”, giving, as examples, communication skills, patience, the tendency to set moderate and achievable goals, and ego development. The competency movement in personnel selection was given further impetus by a book, The Competent Manager (Boyatzis, Reference Boyatzis1982), in which competencies were defined and research on their nature in managers reported. Boyatzis defined a job competency as “an underlying characteristic of a person which results in effective or superior performance” that “may take the form of a motive, trait, skill, aspect of ones’ self-image or social role, or a body of knowledge which he or she uses” (p.21). He gave as examples: proactivity, concern with impact, self-confidence, use of oral presentations, logical thought, conceptualization, and use of socialized power, and managing group processes. Although McClelland’s (Reference McClelland1973) article and Boyatzis’ (Reference Boyatzis1982) monograph were flawed in several respects (Barrett & Depinet, Reference Barrett and Depinet1991), including McClelland’s false claim that cognitive ability tests do not predict job performance (Schmidt & Hunter, Reference Schmidt and Hunter1998), both had considerable impact (Shippmann et al., Reference Shippmann, Ash, Battista, Carr, Eyde, Hesketh, Kehoe, Pearlman and Prien2000; Stevens, Reference Stevens2013).
The notion that competencies should be defined very broadly and inclusively was widely adopted (Shippmann et al., Reference Shippmann, Ash, Battista, Carr, Eyde, Hesketh, Kehoe, Pearlman and Prien2000; Stevens, Reference Stevens2013). Consequently, as the AC movement gathered momentum from the 1970s onwards, a large number of competencies were proposed (Arthur et al., Reference Arthur, Day, McNelly and Edens2003; Woehr & Arthur, Reference Woehr and Arthur2003), often taking the form of “a muddled collection of learned skills, readily demonstrable behaviors, basic abilities, attitudes, motives, knowledge, and other attributes, including traits, that are often ambiguously defined and difficult to rate” (Howard, Reference Howard2008, p.100). Unlike the field of personality and cognitive ability, where multiple evidence-based models of the structure of relevant dimensions have been proposed, little theoretical work on competencies has taken place (Arthur et al., Reference Arthur, Day and Woehr2008; Arthur, Reference Arthur, Jackson, Lance and Hoffman2012), and little consideration has been given to developments in the study of leader skills and behaviors that may be relevant to them (Austin & Crespin, Reference Austin, Crespin, Bennett, Lance and Woehr2006; Meriac et al., Reference Meriac, Hoffman and Woehr2014).
In attempts to bring parsimony and order to the large and muddled competency landscape, Tett et al (Reference Tett, Guterman, Bleier and Murphy2000), Arthur et al. (Reference Arthur, Day, McNelly and Edens2003) and Meriac et al. (Reference Meriac, Hoffman and Woehr2014) created research-based taxonomies of competencies. Arthur et al., suggest that the measurement properties of ACs would be considerably enhanced by the adoption of a small number of such refined competencies.
As well as being central to ACs, competencies have been adopted more generally as part of HR strategy in the form of competency models or frameworks (Campion et al., Reference Campion, Fink, Ruggeberg, Carr, Phillips and Odman2011). This movement followed Prahalad and Hamels’s (1990) article proposing that an organization’s competitive strategy is underpinned by a set of “core competencies”. Core competencies are positioned at the organizational level of analysis, referring to “the collective learning in the organization, especially how to coordinate diverse production skills and integrate multiple streams of technologies” (Prahalad & Hamel, Reference Prahalad and Hamel1990, p. 82). However, they are also linked by the notion of “people-embodied skills” (Shippmann et al., Reference Shippmann, Ash, Battista, Carr, Eyde, Hesketh, Kehoe, Pearlman and Prien2000, p. 712), to individual level competencies of the type popularized by McClelland (Reference McClelland1973) and Boyatzis (Reference Boyatzis1982). Shippmann suggests that by providing a conceptual link that connects organizational success with the notion of individual competencies, the concept of core competencies had a great influence on human resource management, creating conditions for the perceived need to develop organized systems of individual-level competencies. These systems, referred to as competency models or frameworks, consist of “collections of knowledge, skills, abilities and other characteristics (KSAOs) that are needed for effective performance in the jobs in question” (Campion et al., Reference Campion, Fink, Ruggeberg, Carr, Phillips and Odman2011, p. 226). That is, they take the same form as the dimensions measured in ACs.
By developing and adopting these competency models it is thought that the goals, objectives, and strategy of an organization can be aligned with the job requirements of individuals (Campion et al., Reference Campion, Fink, Ruggeberg, Carr, Phillips and Odman2011). In this way, they can viewed as having the potential to create a common language of desirable characteristics for a given job or job type (Lievens et al., Reference Lievens, Sanchez and de Corte2004), and to provide the basis for the integration of disparate HR functions including employee selection, training, promotion, development, compensation, and retention. For example, if the organizational-level competency of “communication” is identified, HR professionals might include this in a competency model and focus on communication ability in the process of selecting and promoting staff (though it should be noted that empirical evidence that organizational-level competencies are maintained or enhanced by recruiting or selecting individuals considered to have those competencies is lacking (Stevens, Reference Stevens2013)). Competencies are further seen as having a role to play in the management of organizational change and the organization and storage of HR data on employees (Campion et al., Reference Campion, Fink, Ruggeberg, Carr, Phillips and Odman2011).
The success of competency models in integrating areas such as employee selection, promotion, and development necessarily depends on the existence and use of reliable and valid systems and techniques for measuring the competencies of individuals. Of the techniques currently available to do so, the most sophisticated and resource intensive is the assessment center. For this reason, the construct validity of ACs is of considerable interest in the context of competency modeling, and of competency-based management more generally.
The construct validity of ACs
The exercise effect
Most modern ACs are designed to assess participants on a set of pre-defined dimensions. As such, dimensions constitute the underling constructs that most ACs are designed to measure. It is therefore critically important that there is clear evidence that ACs have satisfactory construct validity. That is, that they reliably measure the extent to which participants, when compared to each other, have higher or lower scores on a measured dimension, and that differences in scores reflect, to an acceptable extent, differences in the underlying psychological or behavioral construct being measured.
Unfortunately, serious concerns about the construct validity of ACs have existed since the middle of the 20th century (Sakoda, Reference Sakoda1952). For example, Sacket and Dreher (Reference Sackett and Dreher1982) factor analyzed AC ratings and found that the factors identified mapped onto the exercises used, rather than the dimensions the AC was designed to measure. In other words, the performance of the candidates appeared to depend on the particular tasks they were required to carry out and not on the dimensions (competencies) that were supposed to be relevant across the tasks. The associated tendency for the correlations between the ratings given to candidates on different dimensions within the same exercise, to be greater than the correlations between the ratings obtained by candidates on each dimension across different exercises, is often referred to as the exercise effect. The exercise effect suggests that AC ratings may not reflect stable individual differences in dimensions, but rather that the rated performance of candidates differs across exercises irrespective of the dimensions on which they are measured.
Multitrait-multimethod analyses
Initial empirical analyses of the construct validity of ACs adopted the mutitrait multimethod (MTMM) approach (Campbell & Fiske, Reference Campbell and Fiske1959). The key assumption here is that the multiple dimensions measured in ACs are stable traits, and the multiple exercises ACs are methods for the measurement of these traits. Construed in this way, if dimensions are measured reliably across exercises, we would expect the rank-order of participants’ scores on a particular dimension to remain stable across exercises, indicating good convergent validity. We would also expect the rank order of participants’ scores on different dimensions within the same exercise to differ, indicating good discriminant validity. To investigate this, numerous studies of real-world AC rating data compared mean correlations for same-dimension, different-exercise ratings (high r values indicating good convergent validity) to mean correlations for same-exercise, different-dimension ratings (low r values indicating good discriminant validity). These studies, mostly carried out between 1980 and 2000, generally found that correlations between ratings of different dimensions on the same exercise were greater than those for the same dimension across different exercises (Lievens & Klimoski, Reference Lievens, Klimoski, Cooper and Robertson2001), thereby indicating that exercises had more impact than dimensions on AC ratings.
Beginning in the early 1990s (e.g., Schneider & Schmitt, Reference Schneider and Schmitt1992), a second phase of research on the exercise effect applied confirmatory factor analysis to examine the extent to which the latent factors underling AC ratings were concerned with exercises or with dimensions. Initially, these studies were undertaken on data derived from single ACs, but in later studies (Bowler & Woehr, Reference Bowler and Woehr2006; Lance et al., Reference Lance, Lambert, Gewin, Lievens and Conway2004; Lievens & Conway, Reference Lievens and Conway2001), CFA was used to analyze multiple data sets. Taken together, the CFA studies yielded two primary findings. First, almost all found that more AC rating variance was associated with exercises than with dimensions (Guenole et al., Reference Guenole, Chernyshenko, Stark, Cockerill and Drasgow2013). Second, the CFA models used were beset with problems associated with model convergence and admissibility (see Monahan et al., Reference Monahan, Hoffman, Lance, Jackson and Foster2013, pp. 1012–1015 for a summary). Specifically, unless biased models (Lievens & Conway, Reference Lievens and Conway2001) or arbitrary post hoc parameter constraints (Bowler & Woehr, Reference Bowler and Woehr2006) were applied, CFA modeling was rarely able to find evidence for dimension factors in ACs (Monahan et al., Reference Monahan, Hoffman, Lance, Jackson and Foster2013).
Post multitrait-multimethod analyses: CFA studies
The core design features, or architecture, of ACs are that trained assessors evaluate the performance of participants on several dimensions across multiple exercises. The method is therefore the evaluation of candidates on dimensions across exercises. However, when the MTMM approach to construct validity is applied to ACs, the measurement method is construed as the exercises only. An outcome of applying the MTMM approach to ACs is therefore an oversimplification and distortion of the essential elements of the AC measurement architecture. This has consequences for the way in which AC data are analyzed, with the MTMM framework suggesting that dimensions and exercises are the only variables relevant in the consideration of construct validity. It also has implications for the interpretation of the results of these analyses, with variation in a participant’s scores across exercises necessarily interpreted as methodology-related measurement error rather than meaningful variation in the performance of that participant when placed in different situations (Lance, Reference Lance2008).
Recent research (Buckett et al., Reference Buckett, Becker and Melchers2020, Reference Buckett, Becker and Roodt2021; Hoffman et al., Reference Hoffman, Melchers, Blair, Kleinmann and Ladd2011; Merkulova et al., Reference Merkulova, Melchers, Kleinmann, Annen and Tresch2016; Monahan et al., Reference Monahan, Hoffman, Lance, Jackson and Foster2013) on the construct validity of ACs is no longer underpinned by the MTMM approach. Rather than assume that ACs simply measure traits, the possibility that they may measure meaningful variation in the performance of participants across exercises, and meaningful variation in general performance irrespective of dimensions or exercises, is explored. This change, together with other steps, such as increasing the number of behavioral items used as indicators of each dimension (Buckett et al., Reference Buckett, Becker and Melchers2020, Reference Buckett, Becker and Roodt2021; Monahan et al., Reference Monahan, Hoffman, Lance, Jackson and Foster2013), and modeling two or three broad dimension factors instead of the larger number typically used in ACs (Hoffman et al., Reference Hoffman, Melchers, Blair, Kleinmann and Ladd2011; Merkulova et al., Reference Merkulova, Melchers, Kleinmann, Annen and Tresch2016) has largely been successful in resolving the admissibility and convergence problems associated with the MTMM approach to CFA analyses.
These post-MTMM CFA studies of AC construct validity, like almost all previous ones carried out across different organizations, different organizational levels, and diverse countries including the United States, China, Australia, and the United Kingdom report more variance associated with exercises than with dimensions (Lievens & Christiansen, Reference Lievens, Christiansen, Jackson, Lance and Hoffman2012). Importantly, this exercise effect persists even though a variety of “fixes” to reduce or eliminate it have been trialled (Lance, Reference Lance2008; Sackett & Lievens, Reference Sackett and Lievens2008), including reducing the cognitive load on assessors (Reilly et al., Reference Reilly, Henry and Smither1990), using frame of reference training (Monahan et al., Reference Monahan, Hoffman, Lance, Jackson and Foster2013), and adopting post-consensus dimension ratings (i.e. asking assessors to agree a rating for each competency after a participant had completed all exercises).
By finding evidence of a general performance factor in ACs (Buckett et al., Reference Buckett, Becker and Melchers2020, Reference Buckett, Becker and Roodt2021; Hoffman et al., Reference Hoffman, Melchers, Blair, Kleinmann and Ladd2011; Merkulova et al., Reference Merkulova, Melchers, Kleinmann, Annen and Tresch2016), recent CFA studies have contributed considerably to knowledge about the measurement structure of ACs. However, this contribution is limited, because, as Putka and Hoffman (Reference Putka and Hoffman2013) and Jackson et al. (Reference Jackson, Michaelides, Dewberry and Kim2016) point out, the measurement process used in ACs is complex, and the variance associated with exercise and dimension ratings, and with general AC performance, are confounded with many other sources of variance. Examples include the interaction between exercise ratings and competency ratings, and variance associated with the specific items on which candidates are rated. Putka and Hoffman and Jackson et al., point out that in order to provide an estimate of the variance in assessment ratings that is uniquely associated with dimensions (or with exercises, or any other component of the measurement process), it is necessary to adopt an analytic approach that permits every important source of variance in an AC to be separately isolated. Because of the complexity of AC designs, in which multiple assessors rate candidates on each competency (often using several items to do so) across several situations, a fully controlled analysis of the variance requires that a large number of sources of variance be estimated.
Post multitrait-multimethod analyses: G-theory studies
To date, most research designed to estimate the impact of various sources of variance (e.g., exercises, dimensions) on AC ratings have utilized confirmatory factor analysis (e.g., Guenole et al., Reference Guenole, Chernyshenko, Stark, Cockerill and Drasgow2013). Although CFA is helpful in this respect, it has several limitations. One limitation is that CFA cannot be used to distinguish between sources of variance that, in a particular measurement context, are considered reliable and unreliable. For example, if the purpose of an AC is to assess participants on dimensions, variance in the ratings given to participants on those dimensions is assumed to represent a reliable source of variance; and if each pair of assessors rating a given participant on the same dimension and the same exercise differ in their ratings, such variance would be considered unreliable. A second limitation of CFA in the context of research on AC construct validity is that assessors are typically allocated to participants nonsystematically, and this takes a heavy toll on acceptable participant-to-parameter ratios. A third limitation concerns the aggregation of ratings. In real-world ACs, it is usual for practitioners to sum (or average) the ratings given to each participant on a particular dimension in order to obtain that person’s overall dimension score. Such aggregation has a marked effect on variance estimates (Jackson et al., Reference Jackson, Michaelides, Dewberry and Kim2016; Putka & Hoffman, Reference Putka and Hoffman2013), but this cannot be addressed with CFA. A fourth limitation with the CFA approach is that it cannot be used to examine the three-way interaction associated with Participants x Exercises x Dimensions. Not only does this mean that in a CFA analysis this potentially important three-way source of AC rating variance is ignored, but also that variance associated with the interaction may be wrongly attributed to variance estimates relating only to dimensions, and/or only to exercises. A fifth limitation is that CFAs cannot be used to model the effects of assessors on AC rating variance, as in order to do so each assessor would need to be represented with a unique latent variable, resulting in extremely low participant-to-parameter ratios.
All of these limitations can be successfully addressed with the data analytic approach known as generalizability theory (G-theory) (Brennan, Reference Brennan2000, Reference Brennan2001a). G-theory, which was developed by Cronbach and others (Cronbach et al., Reference Cronbach, Rajaratnam and Gleser1963, Reference Cronbach, Gleser, Nanda and Rajaratnam1972), provides an extension, and an alternative, to classical test theory, and is primarily intended for multifaceted measurement of the type used in ACs. Designed to evaluate the reliability of behavioral measurements, it can be viewed conceptually as a combination of classical test theory and analysis of variance (Shavelson & Webb, Reference Shavelson and Webb1991).
The application of G-theory to isolate the many different sources of variance in ACs offers a potentially definitive way to resolve the long-running uncertainty about the construct validity of ACs. For example, if after partialling all sources of variance in ACs, designed and run according to best practice, it is found that the specific variance associated with exercises is at or close to zero, it would be possible to confidently conclude that exercises have no or almost no impact on AC ratings. Alternatively, if analysis of well-designed ACs showed that after partialling all other sources of variance in ratings, the specific variance associated with dimensions is at or close to zero, this would provide compelling evidence that ACs do not measure dimensions.
To date, nine studies using G-theory data analytic methods to analyze real-world AC data have been published (Arthur et al., Reference Arthur, Woehr and Maldegan2000; Bowler & Woehr, Reference Bowler and Woehr2009; Jackson et al., Reference Jackson, Stillman and Atkins2005, Reference Jackson, Michaelides, Dewberry and Kim2016, Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022; Lievens, Reference Lievens2001a, Reference Lievens2001b, Reference Lievens2002; Putka & Hoffman, Reference Putka and Hoffman2013), though for methodological reasons they differ considerably in the degree to which they clarify issues concerning AC construct validity. Specifically, five of the studies (Arthur et al., Reference Arthur, Woehr and Maldegan2000; Bowler & Woehr, Reference Bowler and Woehr2009; Lievens, Reference Lievens2001b, Reference Lievens2001b, Reference Lievens2002) contain methodological and technical issues of concern, including the suboptimal use of available data, and the misspecification of random effects models (Putka & Hoffman, Reference Putka and Hoffman2013). Reviewing these issues, Putka and Hoffman conclude that the five studies do not provide unconfounded estimates of the sources of variance in ACs. For this reason, only the four G-theory studies undertaken by Jackson et al., (Reference Jackson, Stillman and Atkins2005, Reference Jackson, Michaelides, Dewberry and Kim2016, Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022) and Putka and Hoffman (Reference Putka and Hoffman2013) are included in the current review.
Before setting out the results of the four Jackson et al., and Putka and Hoffman studies, the methodology used in each is described in some detail (see also Table 1). All four studies focused on ACs designed and implemented in ways that closely follow best practice, including recommendations set out in international guidelines produced by practitioners and academics who are experts in the AC field (International Taskforce on AC Guidelines, 2015), and were responsive to various suggestions made from research directly investigating the impact of various methodological factors that may influence the construct validity of ACs (Woehr & Arthur, Reference Woehr and Arthur2003).
Table 1. Details of Assessment Centers Studied, Data Analysis, and Presentation of Results, in Four G-Theory Studies
![](https://static.cambridge.org/binary/version/id/urn:cambridge.org:id:binary:20240604103418431-0073:S1754942624000051:S1754942624000051_tab1.png?pub-status=live)
Note. FC = Fully crossed design; PC = Partially crossed design; GE = Henderson’s method in urGENOVA; HP = HPMIXED procedure in SAS; BA = Stan and Rstan procedure in R (Bayesian); BB = Stan and brms procedure in R (Bayesian); MS = Managerial staff; Su = Supervisors; SuM = Supervisors and managers; SeM = Senior managers; * Personal communication.
Jackson et al. (Reference Jackson, Stillman and Atkins2005) obtained data from an AC used to select 199 applicants for posts as retail and customer service workers in New Zealand. The development of the AC followed best-practice principles (Jackson et al., Reference Jackson, Stillman and Atkins2005, pp. 217–220). It began with a competency analysis of the target position, including interviews with subject matter experts (SMEs), the selection of critical tasks, and the development of behavioral checklists for these tasks that included specific behavioral indicators for successful performance, for use in the AC. Job analysis for the identification of dimensions was guided by Threshold Traits Analysis (Lopez et al., Reference Lopez, Kesselman and Lopez1981). SME suggestions guided the linking of dimensions to task performance. Participants were rated on five dimensions across three exercises. A fully crossed design was used in which each dimension was assessed in every exercise. The assessors, 11 managers from the organization for which the customer service workers were being selected, were given behavioral observation training, frame of reference (FOR) training, and guidance on the use of the provided behavioral checklists.
Putka and Hoffman (Reference Putka and Hoffman2013) obtained data from three separate implementations (N = 153, 198, and 298) of a high stakes AC run in a large government organization in the United States. The purpose of this AC was to evaluate for promotion internal candidates for first-line supervisory positions. Exercises and rating scales were based on job analysis data over a series of 2-day workshops led by industrial and organizational (I-O) psychologists, pilot testing the exercise prior to operation, using rating scales with multiple behavioral anchors, and providing 32 hours of training to each assessor (see Putka & Hoffman, Reference Putka and Hoffman2013, p. 121 for details).
Jackson et al. (Reference Jackson, Michaelides, Dewberry and Kim2016) analyzed data obtained from three administrations of a high-stakes operational AC involving a total of 698 candidates. The purpose of the center, run in South-East Asia, was to inform decisions about promotion from line level to senior management positions. As with the Putka and Hoffman study, the design and implementation of the AC followed guidance on best practice. For example, the number of dimensions in any exercise were kept to a minimum in order to reduce the cognitive load on assessors (Chan, Reference Chan1996; Lievens, Reference Lievens1998): at least two assessors were used to rate each candidate on each exercise; assessors were randomly assigned to candidates; the choice of appropriate exercises was informed by job analysis and consultation with subject matter experts; assessors were provided with behavioral descriptions relating to the dimensions being used; rating items were directly traceable to tasks list and job analysis information; and assessors were trained for two days by experienced and academically qualified consultants (see Jackson et al., Reference Jackson, Michaelides, Dewberry and Kim2016, pp. 980–981 for details).
Jackson et al. (Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022) examined data derived from an AC designed to aid in promotion decisions in the UK police force. The roles for which promotion was being considered ranged from junior (constable) to senior (chief inspector). Ten samples were obtained involving a total of 2,917 participants. For each sample between two and four exercises were used (chosen from six). For five of the exercises the ratio of assessors to participants was 2:1, and in the remaining exercise it was 1:1. For each exercise the assessors were given one day’s training by experienced psychologists. This included familiarization with the assessment procedure, awareness of common rater errors, and frame of reference training (see Jackson et al., Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022 pp. 744–745 for details).
In analysing their data, Jackson et al. (Reference Jackson, Stillman and Atkins2005) computed 7 unique variance components using urGenova (Brennan, Reference Brennan2001b) for their fully crossed measurement design. Putka and Hoffman (Reference Putka and Hoffman2013), using REML to deal with the sparsely populated model, were able to decompose 15 sources of variance in the ACs they studied. Using a Bayesian approach to analysis, Jackson et al. (Reference Jackson, Michaelides, Dewberry and Kim2016) fitted a linear random effects model to the ratings given by assessors, decomposing all 29 sources of variance (including those associated with items and samples). Jackson et al. (Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022) also used a Bayesian approach, fitting a crossed linear random effects model to decompose the variance obtained from each sample into six sources.
In all four studies, after the unique variance associated with each source had been estimated, variance estimates were rescaled using formulae specified in the G-theory literature in order to estimate the contribution of each component to aggregated dimension-level scores, aggregated exercise-level scores, and, in the case of Putka and Hoffman (Reference Putka and Hoffman2013) and Jackson et al., (Reference Jackson, Michaelides, Dewberry and Kim2016, Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022), overall AC scores also. When aggregated to dimension-level scores (replicating the usual practice of averaging scores across exercises in order to obtain a set of competency scores for each candidate), the proportion of overall AC variance associated with all sources of reliable variance (i.e., between-candidate differences in overall AC performance, in dimension-related performance, in exercise-related performance; and in performance differences reflecting the interaction between exercises and dimensions and performance) are shown in Table 2. The results of the four studies in relation to estimates of general performance, exercise-related performance, dimension-related performance, and the Dimension × Exercise interaction are discussed below.
Table 2. Estimated Proportions of AC Variance Associated with Reliable and Unreliable Components When Scores are Aggregated to Dimensions
![](https://static.cambridge.org/binary/version/id/urn:cambridge.org:id:binary:20240604103418431-0073:S1754942624000051:S1754942624000051_tab2.png?pub-status=live)
Note. All sources refer to reliable variance unless indicated otherwise. Dashes indicate where results were unavailable. Percentages shown for Jackson et al. (Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022) are across-sample averages weighted by sample size.
General performance
Table 2 shows that a large source of variance in all four studies, accounting for between a third and a half of all rating variance, was associated with the general performance of participants. This general performance factor indicates that candidates were judged to be performing well, or less well, irrespective of the dimensions or the exercises on which their performance was being measured. It is consistent with Kuncel and Sackett’s (Reference Kuncel and Sackett2014) finding that by combining multiple dimension scores it is possible to obtain a reliable construct, but that this construct is dominated by general rather than a dimension-specific variance, and with the results of several post-MTMM CFA studies in which a general performance factor has been successfully modelled (Buckett et al., Reference Buckett, Becker and Melchers2020, Reference Buckett, Becker and Roodt2021; Hoffman et al., Reference Hoffman, Melchers, Blair, Kleinmann and Ladd2011; Merkulova et al., Reference Merkulova, Melchers, Kleinmann, Annen and Tresch2016). It is also consistent with the evidence that there is a large general component in job performance (Viswesvaran et al., Reference Viswesvaran, Schmidt and Ones2005). Given that the exercises used in ACs are designed to simulate real-world work situations, the replication of a general performance factor in AC performance might have been anticipated. That it wasn’t may be due in large part due to the long-standing MTMM focus in AC research on the degree to which variance in AC ratings is associated with only two influences: dimensions and exercises (Lance et al., Reference Lance, Foster, Nemeth, Gentry and Drollinger2007).
Exercise-related performance
Another component with a large influence on assessor ratings, as shown in Table 2, was exercise-related, with the three studies estimating that between a quarter and a third of the variance in AC ratings is uniquely associated with between-exercise differences in how well candidates were judged to have performed. This finding is consistent with the exercise effect, and with multiple meta-analytic investigations finding a large exercise-related factor in AC ratings (Bowler & Woehr, Reference Bowler and Woehr2006; Lance et al., Reference Lance, Lambert, Gewin, Lievens and Conway2004; Lievens & Conway, Reference Lievens and Conway2001; Woehr & Arthur, Reference Woehr and Arthur2003).
Dimension-related performance
The issue of central focus in this article concerns whether or not stable differences in competencies/dimensions are measured in ACs. It is this issue that lies at the heart of 70 years of concern, research, and debate about the construct validity of ACs. As shown in Table 2, all four G-theory studies found the unique variance associated with dimensions to be trivial (1%–2%). Whereas almost all CFA-based studies have found more exercise than dimension variance in ACs (Lievens & Christiansen, Reference Lievens, Christiansen, Jackson, Lance and Hoffman2012), the four G-theory studies show that after decomposing all important variance components, there was almost no rating variance associated with dimensions. All four studies conclude that the inconsequential variance measured indicates that, contrary to the fundamental design principle underlying dimension-based ACs, there is no evidence that stable dimensions (e.g., decisiveness, judgement and problem solving, relating to others, communication skills etc.) were measured by the assessors.
Interaction between exercises and dimensions
Table 2 shows that in the Putka and Hoffman and the Jackson et al. (Reference Jackson, Michaelides, Dewberry and Kim2016) studies there was a notable (9% and 15%, respectively) proportion of variance associated with an Exercise × Dimension interaction. This interaction suggests that the candidates assessed performance on the dimensions depended upon the particular exercises on which they were being measured. This Dimension x Exercise effect was not modeled in the Jackson et al. (Reference Jackson, Stillman and Atkins2005) study. When modeled in the Jackson et al. (Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022) study the variance associated with it was trivial (2%).
Putka and Hoffman suggest that the Dimension × Exercise effects is consistent with explanatory frameworks that have emerged in the personality literature, such as trait activation theory (Lievens et al., Reference Lievens, Chasteen, Day and Christiansen2006; Tett et al., Reference Tett, Toich, Ozkum and Morgeson2021; Tett & Guterman, Reference Tett and Guterman2000). The central idea here is that the extent to which a given trait (e.g. extraversion) influences an individual’s behavior depends on the situation they are in. To take a simple example, a strong extravert may be talkative at a party but quiet at a funeral. If the dimensions measured by ACs (e.g., communication) are construed as traits, trait activation theory would predict that if some exercises (i.e., situations) activate a given trait more than others this would result in the observed Dimension × Exercise interaction. However, some (e.g., Howard, Reference Howard2008) argue that ACs do not, or should not, measure traits, but rather “observable behaviors that are logically organized into categories related to job success” (p.99), and that can be trained and developed. Indeed dimensions (e.g., “resources management,” “organizational awareness,” and “policy planning”) measured in the Putka and Hoffman (Reference Putka and Hoffman2013) and Jackson et al. (Reference Jackson, Michaelides, Dewberry and Kim2016) studies in which a Participant × Trait × Dimension interaction was found, appear conceptually dissimilar from the American Psychological Association’s definition of a personality trait as “a relatively stable, consistent, and enduring internal characteristic that is inferred from a pattern of behaviors, attitudes, feelings, and habits in the individual” (VandenBos, Reference VandenBos2007).
An alternative, and possibly more parsimonious explanation for the Dimension x Exercise interaction effect is that it is an artefact of partially crossed designs often used in ACs, including those studied by Putka and Hoffman (Reference Putka and Hoffman2013) and Jackson et al. (Reference Jackson, Michaelides, Dewberry and Kim2016). To explain this, it is useful to refer to the example of a partially crossed AC design shown in Figure 1. Let us assume, consistent with the findings of Putka and Hoffman and Jackson et al. (Reference Jackson, Michaelides, Dewberry and Kim2016), that dimensions are not measured in ACs, but that general performance and exercise-related performance are measured. Referring to the upper portion of Figure 1, if, in an AC using this partially crossed design, a participant performed well in exercise 1, this would enhance their scores on dimensions 1, 2, and 3. If they performed well on exercise 2 this would enhance their measured performance on dimensions 1, 2 and 4. And if the performance was rated high in exercise 3 their measured performance on dimensions 2 and 4 would be enhanced. That is, between-participant differences in exercise performance, in the context of a partially crossed measurement design, will necessarily and inevitably result in a Dimension × Exercise interaction effect. Such an effect would not be due to real differences in the dimension-related performance of candidates, but rather would merely arise as an artefact of the partially overlapping relationship between exercises and dimensions. This explanation for the interaction between candidates, dimensions, and exercises predicts that it would not arise in a fully crossed AC design (see Figure 1) in which all dimensions are measured in every exercise, because here differential participant performance across exercises will have an equal impact on all dimensions.
![](https://static.cambridge.org/binary/version/id/urn:cambridge.org:id:binary:20240604103418431-0073:S1754942624000051:S1754942624000051_fig1.png?pub-status=live)
Figure 1. Partially and Fully Crossed Measurement Models.
The implications of the results of G-theory studies
The data analyzed by Putka and Hoffman (Reference Putka and Hoffman2013), and by Jackson et al., (Reference Jackson, Stillman and Atkins2005, Reference Jackson, Michaelides, Dewberry and Kim2016, 2022), were obtained from ACs of different designs, run in different countries, for different purposes. It is therefore particularly striking that the results obtained from them are very similar. This gives weight to the suggestion that these results are generalizable, particularly as the core design characteristics (number of exercises, number of dimensions, participant to assessor ratio, rating approach (i.e., within exercise), assessors occupation, length of assessor training, and AC purpose) in all four studies are generally consistent with the those used in the 48 AC studies examined in Woehr and Arthur’s (Reference Woehr and Arthur2003) meta-analysis.
The findings have other, equally important, implications. The results presented in Table 1 indicate that a candidate’s aggregated dimension scores have nothing, or extremely little, to do with her dimension-related performance. Instead, these scores are product of a participant’s (1) general performance, (2) performance on different exercises, (3) in the case of the Putka and Hoffman (Reference Putka and Hoffman2013) and Jackson et al. (Reference Jackson, Michaelides, Dewberry and Kim2016) studies only, performance on dimensions that are entirely a function of the particular exercise in which the participant is involved, and (4) various sources of random error. This in turn suggests that all previous research using aggregated ratings to obtain dimension scores, and then using these scores to assess the construct validity of dimensions, or the criterion-related validity of dimensions, is without value. Put simply, the common practice of aggregating the scores that a candidate has been given for a given dimension across exercises does not provide a measure of how well that candidate has performed on that dimension.
In addition, the findings provide an explanation for the puzzling question of how OARs predict job performance if ACs don’t measure the dimensions they are designed to measured. If rating scores are aggregated across dimensions and exercises to provide an OAR, the OAR will reflect the two stable sources of variance that do influence ratings: general performance, and averaged performance on exercises. ACs predict job performance because they measure a general performance factor and an exercise performance factor, not because they measure dimensions (competencies).
Revisiting the 2008 debate on the continued use of dimensions
In the 14 years since Lance’s (Reference Lance2008) proposal that due to evidence that ACs do not measure stable dimensions, researchers and practitioners should instead focus on the performance of candidates in different exercises, significant new evidence of direct relevance to the construct validity of ACs has emerged. Most of this new evidence has concerned studies of the relative contribution of exercises, dimensions, and other variables to candidate ratings. It is therefore timely to reconsider the question of whether, in the light of new evidence, the construct validity issue can now be resolved.
As explained earlier, almost all of the academics and practitioners invited to comment on Lance’s (Reference Lance2008) article disagreed with its conclusions, suggesting that Lance was incorrect to claim that dimensions were not reliably measured in ACs, and presenting several other objections to his conclusions. A review of these objections reveals that they fall into a limited set of categories. I next discuss each of these in turn in relation to the evidence obtained on the construct validity of ACs by the G-theory studies described above.
Objections to Lance’s (Reference Lance2008) conclusion that ACs do not measure dimensions
There is persuasive evidence that ACs do measure dimensions
Several scholars (Arthur et al., Reference Arthur, Day and Woehr2008; Connelly et al., Reference Connelly, Ones, Ramesh and Goff2008; Melchers & Konig, Reference Melchers and Konig2008) dispute Lance’s (Reference Lance2008) conclusion that ACs do not measure dimensions, citing evidence from Bowler and Woehr’s (Reference Bowler and Woehr2006) confirmatory factor analytic (CFA) study in which it is concluded that dimensions account for a sizeable proportion (22%) of rating variance. However, as explained earlier, CFA studies such as Bowler and Woehr’s, based on the MTMM approach to construct validity, are undermined by model convergence and admissibility issues (Monahan et al., Reference Monahan, Hoffman, Lance, Jackson and Foster2013) and only focus on two sources of variance: exercises and dimensions. Studies using G-theory to control for all, or most of the sources of variance in ratings (Jackson et al., Reference Jackson, Stillman and Atkins2005, Reference Jackson, Michaelides, Dewberry and Kim2016, 2022; Putka & Hoffman, Reference Putka and Hoffman2013), estimate the variance in dimensions associated with assessees to be trivial.
Dimensions have criterion-related validity
Connelly (2008) and Howard (Reference Howard2008) argue that the construct-related validity of AC dimensions is supported by research (Arthur et al., Reference Arthur, Day, McNelly and Edens2003; Donahue et al., Reference Donahue, Truxillo, Cornwell and Gerrity1997; Kudisch et al., Reference Kudisch, Ladd and Dobbins1997; Lievens & Conway, Reference Lievens and Conway2001) demonstrating that dimensions predict job performance. However, because these studies measured dimensions by aggregating scores across exercises, the dimension scores used in the analysis were confounded with non-dimension sources of variance, including the large exercise-related and general performance-related components identified by Putka and Hoffman (Reference Putka and Hoffman2013) and by Jackson et al., (Reference Jackson, Stillman and Atkins2005, Reference Jackson, Michaelides, Dewberry and Kim2016, Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022). As a consequence of a failure to measure unconfounded dimension-related performance, the claim that the studies cited by Connelly and Howard demonstrate the criterion-related validity of dimensions is untenable.
Dimension ratings are unreliable because they tend to be based on one-item measures, and these contain large amounts of specific and random error variance
Arthur et al. (Reference Arthur, Day and Woehr2008) and Howard (Reference Howard2008) suggest that studies of the construct validity of ACs have failed to identify substantial variance associated with dimensions because they have focused on ACs in which dimensions are only measured with single items. These single items are not broad enough to capture the behaviors underpinning the dimensions in question. However, the Putka and Hoffman (Reference Putka and Hoffman2013) and Jackson et al. (Reference Jackson, Michaelides, Dewberry and Kim2016) studies examined ACs in which multiple items were used to measure each dimension, yet both found trivial amounts of dimension-related variance.
Stable differences in across-exercise performance must be caused by dimensions
Connelly et al. (Reference Connelly, Ones, Ramesh and Goff2008) suggest that there is evidence of across-exercise stability in performance, and that this means there must be some stable factors which account for this. The Putka and Hoffman (Reference Putka and Hoffman2013) and Jackson et al., (Reference Jackson, Stillman and Atkins2005, Reference Jackson, Michaelides, Dewberry and Kim2016, 2022) studies confirm the existence of stable across-exercise performance. However, given that no evidence for stable dimensions was found in their studies, the factors responsible for stable differences in performance across exercises cannot have been the dimensions measured in the ACs under investigation. It therefore follows that this across-exercise stability must be due to other factors, possibly including individual differences in the cognitive ability and personality of participants.
Failure to measure dimensions may be due to poor interrater reliability
Connelly et al. (Reference Connelly, Ones, Ramesh and Goff2008) argue that poor interrater reliability may be responsible for a tendency for studies to identify dimension-related differences in performance. However, after unconfounding over half of the sources of variance in an AC, Putka and Hoffman found the average reliability for PEDRs to be satisfactory to very good: .76 for single raters and .87 when two raters were used for each participant. These estimates are consistent with more general work on the reliability of observer ratings (Hoyt & Kerns, Reference Hoyt and Kerns1999).
ACs fail to measure meaningful dimensions
Arthur et al. (Reference Arthur, Day and Woehr2008) and Connelly et al. (Reference Connelly, Ones, Ramesh and Goff2008) suggest that the exercise effect may be due to the inappropriate choice of dimensions. Both recommend the use of the six dimensions identified in Arthur et al.’s (Reference Arthur, Day, McNelly and Edens2003) study. However, Putka and Hoffman’s (Reference Putka and Hoffman2013) study was based on an AC in which the dimension used closely resembled the six listed in Arthur et al.’s (Reference Arthur, Day, McNelly and Edens2003) work, but they still found virtually no variance associated uniquely with dimensions.
Failure to identify dimensions is due to poor quality ACs
Melchers and Konig (Reference Melchers and Konig2008) and Moses (Reference Moses2008) suggest that ACs are often of poor quality. Yet the studies conducted by Putka and Hoffman (Reference Putka and Hoffman2013) and by Jackson et al., (Reference Jackson, Stillman and Atkins2005, 2014, Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022) were of ACs designed and operated according to best-practice principles. Melchers and Konig (Reference Melchers and Konig2008) suggest that dimensions are more likely to be measured reliably if assessors are given frame of reference training. The Putka and Hoffman and Jackson et al. studies focused on ACs in which assessors had been given frame of reference training. Both Melchers and Konig, and Schuler (Reference Schuler2008), suggest using psychologists as assessors. As with most ACs, the assessors in the ACs studied by Putka and Hoffman (Reference Putka and Hoffman2013) and by Jackson et al., (Reference Jackson, Stillman and Atkins2005, 2014, Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022) were not psychologists. However, they were experienced supervisors and managers and were given extensive training, and research (Gaugler et al., Reference Gaugler, Rosenthal, Thornton and Bentson1987) suggests that whether assessors are psychologists or managers has only a modest impact on the criterion-related validity of ACs.
The measurement of dimensions is a scientific and practical necessity
Arthur et al. (Reference Arthur, Day and Woehr2008) argue that abandoning dimensions in research on ACs is scientifically untenable because identifying constructs which lead to effective role and task performance is central to industrial organizational psychology. However, good science involves not just identifying explanatory constructs for phenomena, but also seeking evidence that these constructs can be observed and/or measured. If, after controlling for all other sources of variance in AC ratings, the variance in dimensions is trivial, it follows that these dimensions are not being reliably measured. Although it is possible to seek to improve the conditions for measurement by fixing various features of ACs (e.g., improving assessor training), many such fixes are apparent (see Table 1) in the ACs studied by Jackson et al., (Reference Jackson, Stillman and Atkins2005, Reference Jackson, Michaelides, Dewberry and Kim2016, Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022) and by Putka and Hoffman (Reference Putka and Hoffman2013), yet these have not led to evidence of the reliable measure of dimensions in these studies. From a practical perspective, Arthur et al. suggest that employees prefer dimensions based on human attributes to tasks, and Howard (Reference Howard2008) suggest that a focus on dimensions is essential because they allow practitioners to generalize about the human qualities needed for particular organizational positions and responsibilities. These arguments are based on the reasonable grounds that for practitioners it would be preferable if ACs were to reliably measure dimensions. However, if as the results of the Putka and Hoffman and Jackson et al., studies indicate, ACs do not reliably measure dimensions, it would be unscientific, practically damaging to participants and organizations, and unethical, to continue to use meaningless dimension scores on the basis that there is a practical preference for them.
Summary and conclusions
The prototype ACs pioneered for officer selection in the German armed forces in the 1930s, and later adapted for use in the UK, United States, and Australasian military, were designed to assess a candidate’s whole personality. Later, as the assessment process was increasingly used in non-military organizations, the purpose of the AC evolved, and by the 1970s and 1980s when the AC movement rapidly expanded, the idea of assessing the whole personality had given way to a focus on the measurement of dimensions (otherwise known as competencies). With the growth of the HR movement, competencies have taken on an increasingly central role in organizations, with competency models or frameworks now seeking to provide an integrated system for such areas as selection, promotion, development, and training (Campion et al., Reference Campion, Fink, Ruggeberg, Carr, Phillips and Odman2011; Shippmann et al., Reference Shippmann, Ash, Battista, Carr, Eyde, Hesketh, Kehoe, Pearlman and Prien2000).
Running parallel with the growth of ACs, and the competency movement more generally, has been the troubling and recurring finding that ACs, the most sophisticated and resource-intensive method for assessing competencies, may not work as intended. Specifically, when assessors observe candidates performing across several different job-relevant tasks or situations and rate their performance on these tasks across a series of job-relevant dimensions (e.g., communications skills, decision-making), these ratings, rather than being independent of the tasks, are highly influenced by them. This exercise effect has been widely replicated across different countries, different types of organization, and different levels of seniority (Lievens & Christiansen, Reference Lievens, Christiansen, Jackson, Lance and Hoffman2012).
The issue fundamental to the exercise effect controversy is the question of what it is that ACs measure. Because of the complex measurement design of ACs, in which multiple sources of variance may influence ratings, it has until recently not been possible to provide a conclusive answer to this question. However, with the adoption of G-theory approaches, it is now possible to isolate all relevant systematic sources of variance in ACs, and in doing so to finally provide a conclusive resolution to the measurement controversy.
To date, four methodologically acceptable G-theory studies using data analytic techniques capable of dealing with the complex measurement designs inherent in real-world assessment have been carried out. These studies (Jackson et al., Reference Jackson, Stillman and Atkins2005, Reference Jackson, Michaelides, Dewberry and Kim2016, Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022; Putka & Hoffman, Reference Putka and Hoffman2013), although focussing on ACs of different design, operating in different countries, have yielded strikingly similar results (see Table 2). All four studies indicate that two factors have substantial impacts on AC ratings: the general performance of assesses across all exercises/dimensions, and differences in their performance across exercises. Of particular relevance to the exercise effect debate, all four studies find that dimensions (competencies) have little or no impact on AC ratings. Put simply, the implication of the failure to identify significant and unique dimension-related variance, even in ACs that have been designed according to best practice principles, is that ACs do not measure dimensions/competencies.
When Lance (Reference Lance2008) drew this conclusion 14 years ago, most of the academics and practitioners with expertise in the field of ACs, invited to comment on his paper (Arthur et al., Reference Arthur, Day and Woehr2008; Brannick, Reference Brannick2008; Connelly et al., Reference Connelly, Ones, Ramesh and Goff2008; Howard, Reference Howard2008; Jones & Klimoski, Reference Jones and Klimoski2008; Lievens, Reference Lievens2008; Melchers & Konig, Reference Melchers and Konig2008; Moses, Reference Moses2008; Rupp et al., Reference Rupp, Thornton and Gibbons2008; Schuler, Reference Schuler2008) disagreed. The reasoned objections made to Lance’s conclusions were considered above in light of the four G-theory studies undertaken by Putka and Hoffman (Reference Putka and Hoffman2013) and by Jackson et al., (Jackson et al., Reference Jackson, Stillman and Atkins2005, Reference Jackson, Michaelides, Dewberry and Kim2016, Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022), and it was concluded that all are now difficult to uphold.
It might be argued that the “strong” claim that ACs do not measure dimensions/competencies is not substantiated by the findings of four G-theory studies (even though these involved a total of 19 separate AC implementations) because as Carl Sagan, the world-renowned astrophysicist, famously claimed, “extraordinary claims require extraordinary evidence”. However, by extraordinary claims, Sagan was referring not to claims that are contrary to current orthodoxy, but rather to those that are contrary to the weight of existing evidence (Deming, Reference Deming2016). The weight of existing evidence, from almost all of the many studies carried out across the world on the construct validity of ACs, including large scale meta-analytic studies based on CFA (Bowler & Woehr, Reference Bowler and Woehr2006; Lance et al., Reference Lance, Lambert, Gewin, Lievens and Conway2004; Lievens & Conway, Reference Lievens and Conway2001; Woehr & Arthur, Reference Woehr and Arthur2003), indicates that exercises, rather than dimensions, dominate AC rating variance. The G-theory studies of Putka and Hoffman (Reference Putka and Hoffman2013) and Jackson et al (Reference Jackson, Stillman and Atkins2005, Reference Jackson, Michaelides, Dewberry and Kim2016, Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022) build upon and refine this work by providing estimates of dimension-related variance which are largely unconfounded, and which refer to the aggregated dimension scores of the type used by practitioners. All four studies indicate that at least 98% of the variance in aggregated dimension scores has nothing to do with stable dimensions. Taking these findings together, the claim that ACs do not measure stable dimensions is now beyond reasonable doubt.
For industrial-organizational psychologists seeking to refute this claim, there appear to be only two options. The first would be to explain inadequacies in the design or data analysis used in the Putka and Hoffman and/or the Jackson et al., studies of sufficient importance that they can explain their failure to identify dimension effects, and to present an alternative study (or studies) in which these faults are rectified, and very substantial dimensions effects are identified. The second would be to demonstrate, through a large number of G-theory studies in which the dimension effects are isolated using appropriate designs and analytic techniques, including those able to deal with the sparsely populated designs typical of ACs, that the results obtained in Jackson et al., (Reference Jackson, Stillman and Atkins2005, Reference Jackson, Michaelides, Dewberry and Kim2016, Reference Jackson, Michaelides, Dewberry, Nelson and Stephens2022) and Putka and Hoffman (Reference Putka and Hoffman2013) are atypical and extreme outliers.
Practitioner implications
The compelling evidence that ACs measure the extent to which participants perform well on different exercises is consistent with the use of task-based ACs. Rather than seeking to identify stable competencies, task-based ACs are designed to examine how well participants perform in simulations of the tasks central to the performance of a particular job. For example, a task-based AC designed for a managerial job might include simulations of tasks such as business plan development and employee coaching. A turn toward the design and use of task-based ACs is therefore worthy of consideration (Goodge, Reference Goodge1988; Jackson et al., Reference Jackson, Stillman and Atkins2005; Jackson, Reference Jackson2007; Jackson et al., Reference Jackson, Ahmad, Grace, Yoon, Povah and Thornton2011; Jackson, Reference Jackson, Jackson, Lance and Hoffman2012). However, it should be noted that this approach does not permit the level of generalization across work situations which the competency approach promised. It also raises issues about the choice of tasks, and the extent to which performance on a particular task can be generalized to other situations (Melchers & Konig, Reference Melchers and Konig2008).
One interpretation of the finding that, at least in some circumstances, there is a Candidate × Exercise × Dimension interaction (Jackson et al., Reference Jackson, Michaelides, Dewberry and Kim2016; Putka & Hoffman, Reference Putka and Hoffman2013), is that ACs do measure competencies, but that these competencies are unstable across situations. Such a view would provide a justification for the continued measurement of competencies in ACs, but would necessitate any interpretation of them to be directly related to particular exercise or tasks. For example, it might be concluded that a candidate shows above average leadership ability in a leaderless group discussion, but not in a role play exercise. However, as Connelly et al. (Reference Connelly, Ones, Ramesh and Goff2008) point out, exercise-based feedback provides little information relevant to novel situations, and the amount of information given to participants about their performance may be overwhelming: a typical AC measuring 10 dimensions across 5 situations, would produce 50 distinct pieces of feedback per candidate. In addition, this approach would necessitate the assumption that the dimension-exercise interaction effect provides meaningful information about a participant’s performance on different exercises. As explained earlier, the Exercise x Dimension interaction effect may be no more than an artefact of partially crossed AC designs, from which it would follow that the measured performance of a participant on a particular dimension, in a particular exercise, would be spurious.
ACs are also used as a method for predicting a candidate’s overall job performance. This is achieved by obtaining an overall candidate score across exercises and dimensions, or by consensus discussion between assessors. The legitimacy of this practice is supported by meta-analytic research indicating that in comparison to many other selection techniques, ACs have relatively strong criterion-related validity (Sackett et al., Reference Sackett, Zhang, Berry and Lievens2021; Schmidt & Hunter, Reference Schmidt and Hunter1998), though it should be noted that they are also associated with above average Black versus White adverse impact (Sackett et al., Reference Sackett, Zhang, Berry and Lievens2021). Although meta-analytic research suggests that there are selection techniques with higher criterion-related validity and lower adverse impact than ACs (e.g., structured interviews) (Sackett et al., Reference Sackett, Zhang, Berry and Lievens2021), strong face validity and candidate acceptance may maintain their continued use in predicting overall job performance in many organizations.
Finally, and returning to the central focus of this article, all four methodologically sound G-theory studies carried out to date (Jackson et al., Reference Jackson, Michaelides, Dewberry and Kim2016; Putka & Hoffman, Reference Putka and Hoffman2013), building on and refining decades of previous research, indicate that even ACs designed and run according to best-practice principles do not measure stable dimensions (competencies). For practitioners, the implication of these findings is that organizational decisions relating to selection, promotion, training or any other matters, when based on AC competency scores, are unsound. Furthermore, when participants are given feedback on their AC performance on measured competencies, this feedback is spurious, misleading, and potentially harmful. For these reasons, and associated ethical considerations, practitioners should stop using ACs to measure competencies.