Results suggest significant relationships between ELM and Job performance, satisfaction with prevision, overall satisfaction, commitment, role conflict, role clarity, member competence, and turnover intentions. The relationship between ELM and actual turnover was not significant. Leader and member ELM perceptions were only moderately related. Partial support was found for measurement instrument and perspective (I. E. , leader vs.. Member) as moderators of the relationships between ELM and its correlates.
Meta-analysis showed that the LAMB (7-item ELM) measure has the soundest psychometric properties of all instruments and that ELM is congruent with numerous empirical relationships associated with transformational leadership. Nice first proposed, this basic unit of analysis has remained unchanged. Within the broad area of organizational leadership, leader-member exchange (ELM) theory has evolved into one of the more interesting and useful approaches for studying hypothesized linkages between leadership processes and outcomes.
First proposed by Grater and colleagues (Danseuses, Cushman, & Green, 1973; Danseuses, Green, & Hag, 1975; Green, 1976; Green & Cushman, 1975), ELM is distinguished from other leadership theories by its focus on the dyadic relationship between a leader and a member. Unlike traditional theories that eek to explain leadership as a function of personal characteristics of the leader, features of the situation, or an interaction between the two, ELM is unique in its adoption of the dyadic relationship as the level of analysis.
Although the theory has been modified and expanded According to ELM, the quality of the relationship that develops between a leader and a follower is predictive of outcomes at the individual, group, and organizational levels of analysis. Dyadic relationship development is grounded in role and exchange theories (see Green & Lull-Been, 1995; Laden, Sparrows, & Wayne, 1997; Lifeline, Green, & Cassandra, 1997, or more detailed discussions of the theoretical background and development of ELM).
After more than 25 years of empirical research and theoretical development, ELM continues to provide an operable alternative to the traditional leadership approaches focused on leader traits and behaviors (e. G. , Bass, 1990; Integer, 1973; Stodgily, 1948). Although researchers remain enthusiastic about ELM, there is unresolved ambiguity about the nature of the construct, its measurement, and its relationships with other organizational variables. This ambiguity may be due in part to the evolving nature of ELM theory.
In a recent view, Green and Lull-Been (1995) classified the evolution of ELM theory into four stages: (a) work solicitation and vertical dyad linkage where the focus was on the discovery of differentiated dyads (I. E. , in-groups and outgrows), (b) ELM where the focus was on the relationship quality and its outcomes, (c) a prescriptive approach to dyadic partnership building, and (d) ELM as a systems-level perspective (I. E. , moving beyond the dyad to group and network levels). The latter two stages are fairly recent developments, and most of the work associated with them is theoretical. The majority of empirical research
Charlotte R. Serener and David V. Day, Department of Psychology, The Pennsylvania State University. We gratefully acknowledge the contributions made by Dan Brass, Jim Afar, John Mathieu, Janet Swim, and Francis Yammering on earlier versions of this article. We also thank Stephanie Klein for coding all the studies used in these analyses and Chuck Pierce for assisting us with data analysis. This article is based on Charlotte R. Greeter’s master’s thesis. An earlier version was presented at the 10th annual conference of the Society for Industrial and Organizational Psychology, May 1995, Orlando, Florida.
Correspondence concerning this article should be addressed to Charlotte R. Serener, Department of Psychology, The Pennsylvania State University, 429 Bruce V. Moore Building, University Park, Pennsylvania, 16802-3104. 827 828 SERENER AND DAY evaluating factors thought to contribute to high-quality exchanges and analyzing the connection between ELM and work-related outcomes occurred during the second stage of ELM theory development. It could be argued that the last two stages show the most potential for describing leadership behaviors in complex organizations. Clearly, more research is needed to understand these extensions of
ELM theory. The goal of this article is to quantitatively summarize and evaluate the existing ELM research base to build a foundation for future empirical work. More specifically, we consider issues related to the ELM construct, including measurement scale properties and leader-member agreement, as well as the correlates of ELM. We use the results of these quantitative summaries to draw conclusions about ELM and identify opportunities for future theoretical development and empirical research. An Analysis of ELM Issues Measurement and Dimensionality Instruments. Despite claims of an apparently robust phenomenon (e. G.
Green & Lull-Been, 1995), there is surprisingly little agreement on what ELM is or how it should best be measured. The theoretical progression of ELM described above is illustrated by the changes in ELM measurement instruments over the years. The construct of ELM has evolved from the two-item measure of negotiating latitude (ML) to more elaborate, multidimensional scales (e. G. , Laden & Measly, in press; Cherished, Needier, Cassandra, & Tipper, 1992). Because different studies use different ELM scales, it is unclear whether conflicting results are due to deficiencies in the theory or in the personalization of the core construct.
Through the use of meta-analytic techniques, it was possible to examine the type of measurement instrument as a potential between-study moderator of outcomes. On the basis of the amount of research that has gone into its development and refinement, and the recent recommendation provided by Green and Lull-Been (1995) that it be adopted as the standard measure of ELM, we expected that the seventies ELM measure (LAMB; Green, Novak, & Shoemakers, 1982) would demonstrate the highest reliability (I. E. , internal consistency) and largest correlations with other variables, as compared with other ELM measures.
Internal consistency and unintentionally. An issue that has been raised by numerous researchers (e. G. , Tidiness & Laden, 1986; Laden & Measly, in press; Laden et al. , 1997; Cherished et al. , 1992) concerns the potential multidimensionality of ELM. For example, Tidiness and Laden proposed that ELM is comprised of the dimensions of perceived contribution, loyalty, and affect. Green and Lull-Been (1995) argued that ELM is comprised of the interrelated dimensions of respect, trust, and mutual obligation.
We addressed the possibility of multiple dimensions underpinning the ELM scales by examining aggregated internal consistency (I. . , coefficient alpha) estimates of ELM. I Green and Lull-Been noted that Coronary alphas for multidimensional ELM measures were consistently in the . 80- . 90 range. Therefore, they concluded that ELM may comprise several dimensions, but they are al] highly related and can be adequately measured with a unidirectional measure of ELM. Thus, we expected that averaged alphas would be generally large (I. . , above . 80), indicating the likelihood of one underlying ELM dimension summarized by the centered item of the LAMB scale (I. E. , “How effective is your working relationship with your leader [follower]? ; Green & Lull-Been, p. 236). Perspective. ELM can be measured from both leader and member perspectives, but is it the same construct when measured from different perspectives? Empirical support for the relationship between leader ELM and member ELM has been equivocal. Green and Cushman (1975) found a correlation of . 0 between leader ELM and member ELM; however, others have reported much lower correlations (e. G. , Cassandra, Green, & Novak, 1986, reported a correlation of only . 24 between perspectives). The low correlation between leader and member exchange is consistent with meta-analytic research on elf-supervisor agreement on performance ratings, which has demonstrated relatively low agreement (r = . 35, corrected; Harris & Checkbooks, 1988). Green and Lifeline (1995) suggested that the degree of leader-member agreement can be used as an index of the quality of data.
They implied that an aggregate, sample-weighted correlation assessing the level of agreement between leader and member reports would be positive and strong. We tested this proposition by estimating the sample-weighted correlation between corresponding leader and member ELM ratings. We also examined measurement perspective as potential between-study moderator of the relationship between ELM and its correlates. Correlates ELM is generally found to be associated with positive performance-related and attitudinal variables, especially Although a large alpha (e. G. , > . 0) cannot always be interpreted as evidence of scale unintentionally (Ocarina, 1993), it does provide information regarding the average interim correlation. Thus, if the number of items is reasonably small (e. G. , < 10), a large coefficient alpha estimate can be construed as meaning that the average item intercorrelation is also relatively large. It is possible that a scale with a relatively high alpha may represent a multidimensional construct; however, the dimensions are probably highly intercorrelated. LEADER-MEMBER EXCHANGE META-ANALYSIS for members.
These include (a) higher performance ratings (e. G. , Laden, Wayne, & Stilwell, 1993), (b) better objective performance (e. G. , Green, Novak, & Shoemakers, 1982; Vehicle & Gobbled, 1984), (c) higher overall satisfaction (e. G. , Green, Novak, & Shoemakers, 1982), (d) greater satisfaction with supervisor (e. G. , Douche, Green, & Table, 1986), (e) stronger organizational commitment (e. G. , Nostrum, 1990), ND (f) more positive role perceptions (e. G. , Snyder & Binning, 1985). Support for other relationships, such as member competence (CB. Kim & Organ, 1982; Laden et al. 1993) and turnover (CB. Green, Laden, & Hole, 1982; Vehicle, 1985), has been equivocal. Our meta-analysis provided overall effect estimates for previously investigated correlates of ELM. 2 In general, we expected that correlations between ELM and member performance and attitudinal variables would be positive and strong (the exceptions being turnover processes and role conflict, in which cases the direction was expected to be negative). We expected ELM scale reliability to be generally acceptable and leader and member ELM ratings to be only moderately correlated.
We proposed and tested specific moderating variables. In particular, we hypothesized that the LAMB scale would demonstrate the best general predictive validity, although measures from the different perspectives (leader vs.. Member) might demon- 829 actions using the Dissertations Abstracts International (19751996) computer database. Although unpublished manuscripts were not actively solicited, some authors sent unpublished manuscripts when communicating about other published studies. These unpublished studies were also included in the analyses.
We identified a population of 164 studies using these procedures. The criteria for inclusion in the analyses were (a) dyadic exchange quality was measured in the study; (b) the reported results were sufficient to calculate an effect size for the relationship between exchange quality and a correlate, or between leader and member ELM perceptions; and (c) the relationship reported was also reported in at least five other studies. The last criterion was necessary to ensure that the number of studies for each meta-analysis was adequate for drawing generalize conclusions about the reposed relationships.
Although it has been suggested that a meta-analysis can be conducted with as few as two studies (Hunter, Schmidt, & Jackson, 1982), second- order sampling error poses a distinct threat to the validity of results when only a small number of studies are included (Hunter & Schmidt, 1990). There is no strict rule about the minimum number of studies to include in a meta-analysis, but we chose six as the cutoff for our study because of concerns about Type I and Type II errors in the moderator analyses (Jackets, Harris, & Orr, 1986; Specter & Levine, 1987).
On the basis of these arterial, we retained 79 studies for the analyses. 3 Six studies contained two samples; therefore, we worked with a total of 85 independent samples. The relationships analyzed are described in the following section. Strata divergent relationships due to variability in the constructs meaning inherent in these perspectives. Method Literature Search The studies included for analysis consisted of published articles, conference papers, doctoral dissertations, and unpublished manuscripts.
We identified relevant published articles primarily through computer-based searches of the Psychology 1984-1996), stuck (1974-1984), and ABA/INFORM (1975-1996) databases using the keywords leader-?member exchange, vertical dyad linkage, exchange quality, and dyadic leadership. We used the Social Sciences Citations Index to identify articles that referenced either of the seminal ELM articles (I. E. , Danseuses et al. , 1975; Green & Cushman, 1975).
The reference lists of identified empirical studies and theoretical reviews were also crosschecked for additional references. Finally, we conducted a manual search of Academy of Management Journal, Administrative Science Quarterly, Group and Organization Management, Human Relations, Journal of Applied Psychology, Organizational Behavior and Human Performance (later Organizational Behavior and Human Decision Processes), and Personnel Psychology from 1975 to 1996 to identify studies that measured ELM, although it may not have been the primary variable of interest.
We identified conference papers presented during the period from 1990 to 1994 by manually searching the programs for the annual Society for Industrial-organizational Psychology and Academy of Management meetings. We identified desserts- Constructs Included in the Meta-Analyses ELM. Although this term is commonly found in the literature to refer to the dyadic relationship between a leader and a member, construct personalization is far from consistent. We identified seven different versions of the ELM scales developed by Green and colleagues, two additional ELM measures (Laden & Measly, in press; Cherished et al. 1992), and several modified versions of these scales. The LAMB scale (Green, Novak, & Shoemakers, 1982) is by far the most frequently used ELM measure. LAMB differs from the earlier VOID measures in that it does not focus on the amount of negotiating latitude a leader allows a member but rather on he nature of their general working relationship. To be included in the analyses, the study had to measure dyadic exchange quality in a manner consistent with the general themes found in the LAMB scale.
Because of the variability From a conceptual standpoint, it makes sense to divide the literature on ELM correlates into two categories: antecedents (I. E. , those variables hypothesized to affect the development of leader-member relationships) and outcomes (I. E. , those variables hypothesized to result from ELM). Because most of the empirical research reviewed in this article is correlation and cross-sectional, we void discussing them in terms of causal inferences regarding the direction of these relationships. For purposes of the present analyses, we treat them all as correlates. A list of the reasons for study exclusion is available on request from Charlotte R. Serener. Found in scales explicitly labeled “ELM,” we also included a few other measures that were used to assess the ELM construct (e. G. , Leader Behavior Description Questionnaire; LBS.). We included the LBS. because it was used in the early empirical research before the specific ELM scales were developed (e. G. , Green, Danseuses, Minima, & Cushman, 973). However, we included studies using the LBS. only if the item referents were changed to reflect a dyadic perspective.
We did not include studies that used the Multiracial Leadership Questionnaire (Bass, 1985) because these relationships are not consistently dyadic in nature and have recently been reviewed by Lowe, Crock, and Submariner’s (1996). We evaluated studies on a case-by-case basis to determine if the constructs used could be classified as ELM. In cases of uncertainty, we contacted the author(s) to determine if they would classify their measure as ELM. Performance ratings. This construct was personalized as subjective, supervisory ratings of a member’s Job performance. We excluded global measures of competence (I. E. Not related to performance in a specific Job) from this analysis but analyzed them separately in the relationship between competence and Objective performance. Objective performance referred to all measures of performance that did not rely on a subjective rating, which typically meant indexes of the quantity or quality of work, such as the total number of cases processed or the total sales in dollars (e. G. , Green, Novak, & Shoemakers, 1982; Tanner & Strawberry, 1990). Satisfaction. We examined two constructs related to work satisfaction in the meta-analysis: overall Job satisfaction and satisfaction with the supervisor.
Overall satisfaction referred to how satisfied a member was with the Job in general or the organization, whereas satisfaction with the supervisor referred more explicitly to how satisfied a member was with his or her working relationship with a supervisor. Several different scales were used to measure satisfaction, including the Job Description Survey, Hopped Scale, and Minnesota Satisfaction Questionnaire. Satisfaction with supervisor was usually measured using one of the facet scales associated with these general measures of satisfaction. Organizational commitment.
Although there are many different types of commitment discussed in the literature, our analysis included only measures of members’ commitment to the employing organization. Measures of both attitudinal and calculated (or behavioral) commitment were included. With two exceptions, all studies in this analysis used a version of the commitment scales developed by Monday, Steers, and Porter (1979). Role perceptions. Many different aspects of role perceptions have been linked to ELM; however, there was an insufficient number of studies available to meta-analyze all role perceptions.
We conducted meta-analyses to assess the relationships between ELM and role conflict and role clarity. The analysis for role clarity also included studies that measured role ambiguity (reverse coded). The most common measure of role perceptions was a version of the scale developed by Orzo, House, and Larrikin(1970). Turnover processes. We examined the relationship between ELM and turnover in two separate analyses. One analysis was conducted using actual turnover data whereas the second analysis included member elf-reports of intentions to leave the organization. Member competence.
Measures of competence included both self- and supervisor ratings of general ability or expertise, cognitive ability tests, and experimental manipulations of competence. Because competence was conceptualized as a more global measure of ability, we did not use specific performance ratings in this analysis. Leader-member agreement. This analysis did not refer specifically to a correlate of ELM but rather to the agreement between leader and member regarding ELM quality. All studies that reported a correlation between leader ELM and member ELM were included in this analysis.
Coded Variables In addition to the correlations between ELM and the constructs described above, the following information was coded for each study: (a) date of publication, (b) publication form (I. E. , article, conference paper, dissertation, book chapter, or unpublished manuscript), (c) type of sample (lab versus field, type of field sample), (d) sample size, (e) demographic information regarding leaders and members, ( f) ELM measurement instrument, (g) ELM measurement perspective (I. E. , leader or member), (h) scale properties of ELM measure (I. E. Liability and range), and (I) measurement information for criterion variables (I. E. , measurement instrument and reliability). All studies were coded independently by two raters (Charlotte R. Serener and a psychology doctoral student). A total of 1 ,075 coding decisions were made, with 86% agreement between the two raters. All disagreements were resolved through discussion with David V. Day. Computation and Analysis of Effect Sizes We used correlation coefficients in the computation of all effect sizes. If a correlation coefficient could not be calculated from the reported statistics, the study was roped from further analyses.
In order to preserve the independence of samples, we included only one effect size from each sample in each meta-analysis. We computed estimates of overall effect sizes and homogeneity of effect sizes according to the formulas set forth by Hedges and Oilskin (1985). Before combining effect sizes, we converted RSI to DSL and corrected for sampling error. We then averaged the refs to obtain an estimate of the overall effect size, d+. When measurement reliability information was available for at least 75% of the effect sizes in a given analysis, we also corrected the responding sample-weighted correlations for measurement error.
Although the application of corrections is traditionally associated with the Hunter and Schmidt (1990) approach to meta-analysis, Hedges and Oilskin (1985) stated that this is an acceptable procedure provided that sufficient reliability information is available. In such instances, a closer estimate of the true population correlation can be attained by applying correction formulas to the study-level correlations, given that statistical artifacts such as measurement error can attenuate willing-study correlations (Huffiest, Arthur, & Bennett, 1993; Hunter & Schmidt, 1990).
In the present meta- analyses, we calculated summary effect sizes with both uncorrected and corrected correlations. In addition to the estimates of overall effect size, we also calculated a homogeneity statistic (Q) for each analysis. A significant Q statistic indicates that the overall mean effect size does not adequately describe all the effect sizes and further moderator analyses are warranted.
Because the formulas for homogeneity of effect sizes are inappropriate for disorientated effect sizes, we did not interpret the Q statistics associated with the correlations corrected for measurement unreliability. We used the corrected correlations only to estimate a mean weighted effect size. TTT is fairly common for overall effect sizes to be heterogeneous (Hedges, 1987). Although Hunter and Schmidt (1990) argued that nearly all heterogeneity can be attributed to statistical artifacts, heterogeneity may also result from other factors, such as between-study moderators and statistical outliers.
Some researchers have advocated the removal of statistical outliers in an attempt to achieve homogeneity of effect sizes, as well as a more reliable estimate the true population effect size (Hedges, 1987; Hedges & Oilskin, 1985; Huffiest & Earth 995). In the present meta-analyses, we conducted both categorical moderator analyses and statistical outlier analyses to examine potential sources of heterogeneity. On the basis of our review of the literature, we identified two a priori categorical moderators of overall effect sizes: (a) the ELM measurement instrument (I. . , LAMB all others), and (b) ELM perspective (I. E. , leader vs.. Member), which we used to examine the impact these two variables on the homogeneity of the overall effect sizes. These analyses analogous to a one-way analysis of variance that compares effect sizes by a specific lass variable determined on the basis of study characteristics. If a categorical moderator model completely fits the data, the between-class effect will be significant whereas the within-class effect will be insignificant (indicating homogeneity with classes; Johnson & Turbo, 1992).
In addition to the categorical moderator analyses, conducted outlier analyses to attain homogeneity of effect sizes. This was accomplished by ordering the effect sizes in each meta-analysis with regard to their deviations from the mean effect size and then sequentially eliminating the largest outlier until the overall Q tactics was insignificant (indicating homogeneity of effect sizes), or until 20% of the original studies were removed (see Hedges & Oilskin, 1985, up. 256-257).