Steven E. Markham - Conservancy - University of Minnesota

63

Leadership Convergence: An Application ofWithin and Between to Validity

Steven E. MarkhamVirginia Polytechnic Institute and State UniversityFred Dansereau, Jr. and Joseph A. AluttoState University of New York at Buffalo

MacDonald Dumas

University of the West Indies

Problems in drawing inferences about leadershipphenomena when multiple units of analysis (groupsand individuals) simultaneously exist in a data set areaddressed. Using a technique recommended by Dan-sereau and Dumas (1977), within-unit and between-unit sources of covariation are examined for data con-

taining matched superior-subordinate reports. In thisdata set matched superior-subordinate reports were notsignificantly correlated at the individual level. Whensupervisory group differences were held constant,however, the relationships between these matched re-ports were significantly greater than zero. This conver-gent validity within supervisory units suggests an ap-proach to validity which is not included in traditionaltheories of leadership.

In discussing data aggregation problems, 9 ~~ir~-

berly (1980, p. 367) asked, &dquo;How valid is it ...to define leadership style in terms of averaged sub-ordinate ratings?&dquo; The methodological issue of ag-gregation as it applies to the validity of superiorand subordinate descriptions of leadership is thefocus of this study. The problem raised by thisissue can be viewed as one variant of a more gen-eral unit-of-analysis problem in organizational re-search (see Dansereau & Dumas, 1977; Knapp,1977).A unit-of-analysis problem occurs when a data

set contains reports from or about individuals who

are located in naturally occurring groups. Groupaverages can be computed, correlated, and com-pared in such a data set. The inferences which aredrawn from these aggregated average scores canbe problematic, as discussed by Robinson ( 1950)in his fam®~s &dquo;ec®l~gical fallacy&dquo; article. Simplystated, correlations (or any statistic) based uponaggregated scores cannot be used to draw infer-ences about the behavior of the individuals repre-sented in the aggregation. Thus, the statistical unitof analysis should match the unit specified by atheory.

Nevertheless, a multiple-unit data set with in-dividuals identified as members of groups contains

untapped research potential. Dansereau and Dumas(1977), Knapp (1977), and Sirotnik (1980), amongothers, suggested that a variety of statistics andinferential procedures which do not contradict Ro-binson’s ecological fallacy argument can be de-rived from a data base which contains individualsembedded in their groups (or collectivities). TheCovariance Theorem from which Robinson’s (1950)argument originates illustrates this point. The theo-rem (cf. Przeworski & Teune, 1970) can be statedas follows:

Downloaded from the Digital Conservancy at the University of Minnesota, http://purl.umn.edu/93227. May be reproduced with no cost by students and faculty for academic use. Non-academic reproduction

requires payment of royalties through the Copyright Clearance Center, http://www.copyright.com/

64

where N = total number of individuals,B = group effect,J = total number of groups, and

W = within-group effect.In other words, as shown in Equation 1, the total

covariation between variable x and variable y with/V individuals embedded in J groups is equal to aweighted sum of the within-group and between-group covariances (Przeworski & Teune, 1970,p. 59). Thus, three types of correlations can bederived for this type of multiple-unit data base,which contains TV individuals and J groups. These

three correlations are (1) the total raw unadjustedcorrelation, which is based on the total N individ-uals ; (2) the between-unit correlation, which is basedon J units, with each unit represented by its averagescore; and (3) a within-unit correlation based onthe residual variation after removing between-unitdifferences.The Covariance Theorem has important impli-

cations for research in formal work organizationsbecause such organizations can be represented bya multiple-unit data base. An illustration from theleadership literature may help demonstrate this unit-of-analysis issue. Katerburg and Horn ( 19~ l ) ex-amined descriptions of leadership behaviors to de-termine whether between-supervisory group or

within-supervisory group variance was more im-portant. In this type of research, a common ap-proach is to ask subordinates their degree of agree-ment with statements such as ’ &dquo;My supervisor looksout for the personal welfare of the group mem-bers.&dquo; To the extent that all members of a super-

visory unit respond with similar answers, then theunit’ average score could be used to represent that

supervisor’s leadership style with some degree ofaccuracy. If some supervisors in an organizationhave high average scores on scales constructed fromquestions like the one above, and if other super-visors have low scores, a between-unit phenome-non could be described by using group averages torepresent each supervisor’s style. in this case, abetween-unit correlation, which forms the second

part of the Covariance Theorem, might prove use-ful.

There is, however, no guarantee that a super-visor treats all the members of his/her unit in a

similar manner. It is hypothetically possible thatsubordinate reports in the same group could varyso widely that the use of the group average torepresent the supervisor’s style would create 6 ‘a

fictitious average or middle range score which theleader never displays&dquo; (Schriesheim, House, & Kerr,1976, p. 361). For this case, a within-unit residualcorrelation, which is part of the third term of theCovariance Theorem, might prove more useful thana correlation based on unit averages.

Without empirical testing it is difficult to deter-mine if a within-unit correlation or a between-unitcorrelation is more descriptive of a data set. Thus,this paper uses a within and between analysis(WABA) technique (cf. Dansereau, Alutto, Mark-ham, & Dumas, 1982) that makes no a priori as-sumption about which source of variance for a mul-tiple-data base is more important. As an alternativeto a combined AI~1CO~IA/rrgultiple regression pro-cedure, the ~IA13A technique uses weighted unitaverages in calculating between-unit correlationsas well as correlations based on within-unit vari-

ation. In a single step, data are transformed priorto entry into a correlation program by partitioningeach respondent’ raw score on each measure intotwo components. These two components are a

weighted between-unit score, which is shared byall members of a supervisory unit, and a within-unit component, which is that individual ’ unique,relative position above or below the units’ average(see McNemar, 1955). Thus, supervisory units areequated with statistical cells. The following sub-stantive research question will be addressed in or-der to illustrate this procedure: To what extent arematched leadership reports from subordinates andsupervisors affected by the problem of between-unit and within-unit variation?

Method

In using any procedure which focuses upon su-pervisory groups as the unit of analysis and, there-fore, as statistical cells, an experimental designperspective may prove helpful. In a sense, reportsfrom individual respondents can be viewed as val-idated by the degree to which they are influencedby their group membership. If respondent reports



65

are not affected by work group membership, thenthere will be little variation based on these cells.In this procedure it is necessary to accurately defineeach supervisory unit and to locate each subordi-nate in his/her proper group.

The Setting

In this study, 107 managers were interviewed

concerning the leadership style of their superiors.The 107 managers formed the entire managerialhierarchy of the production department of a largemetals extraction firrn. (This subset of the entiresystem was taken to minimize the types of rolesand the number of management levels in the study.Three levels of management were represented.)

Four procedures were employed to insure thatwork groups were accurately defined. First, eachmanager’s reporting relationship was obtained fromorganizational records. Second, structured inter-views were scheduled so that managers who were

higher in the organization were interviewed beforetheir subordinates. These managers then listed the

people who worked for them. Third, their reportswere then checked against the information providedby the organization’s records. Finally, during theinterview, each respondent was asked to verify thatthe name of his/her superior was correctly listedon the questionnaire.

All 107 managers filled out questionnaires as

subordinates and described their relationship withtheir superiors. After completion of this portion ofthe interview, 30 of the 107 managers completeda second set of questions about each of their sub-ordinates. In other words, while 107 managers de-scribed their 30 supervisors, only 77 respondentscompleted just one set of instruments as subordi-nates. The remaining 30 respondents filled out twosets of interview questionnaires: one from a sub-ordinate perspective and one from a superior per-spective. In this way matched superior-subordinatereports were obtained describing 107 different dyadicreporting relationships. This multirater procedurefor assessing matched superior-subordinate reportshas been used in previous studies of leadership (seeDansereau, Graen, & Haga, 1975).

Measures

The results of analyzing three measures are re-ported here to illustrate the WABA technique andits inference-drawing procedure.

Leadership attention: Amount received. Lead-

ership was measured by the Leadership AttentionScale. This scale (cf. Dansereau et al. , 1975) con-tains nine items designed to assess various formsof personal consideration of the subordinate by thesupervisor. Sample items include the following:assurances of confidence from the superior, supportand attention for feelings, information about thecurrent and future states of the unit, feedback fromthe superior to the subordinate about job perfor-mance, and so forth. (For a complete listing, seeMarkham, 1978.) Subordinates responded to thesequestions in terms of their superior’s actions, e.g.,&dquo;How much attention from your supervisor are yougetting?&dquo; Response categories included almost none,a little, a moderate amount, quite a bit, and a greatdeal. The coefficient alpha for this scale was .92.The mean was 31.3~ with a standard deviation of6.38.

Leadership attention: Amount given. Each ofthe nine items above was also asked of those 30

managers who had other managers reporting to them.These managers described each of their subordi-

nates individually. Each question was rephrased soas to reflect a superior’s viewpoint. Thus, eachsuperior was asked to predict the amount of atten-tion and support that a particular subordinate per-ceived he/she was getting. This procedure resultedin reports about 107 subordinates. The coefficient

alpha for this scale wars .9. The mean was 29.11 Iwith a standard deviation of 5.63.

Subordinate satisfaction with superior. Be-

cause many studies of leadership have includedvarious measures of satisfaction (see Stogdill, 1974),this measure was also included to add to the scopeof the validity check. The use of a nonleadershipmeasure hclps rulc out the possibility that any re-sults are solely derived from response bias to sim-ilar questions. This scale is composed of six itemsdescribing the positive aspects of a supervisor fromthe perspective of a subordinate. It is a portion ofthe Role Orientation Index (Graen, Dansereau, &



66

~i~~rni9 1972). This index was adapted, in turn,from the Job Description Index. The coefficientalpha was .84. The mean was 8.32 with a standarddeviation of 2.59.

Analytical Procedure

The analytical procedure consisted of five steps.First, zero-order correlations based on individualswere computed. (This step corresponds to r~,~ inthe Covariance Theorem.) Second, group meanswere calculated with supervisory groups aligned ascells in a one-way ANOVA. Third, these groupaverages were used for the computation of weightedunit-average correlations. (This step correspondsto r.xy,~ in the Covariance Theorem.) Fourth, afterthe effects of the units were removed, the residualerror correlation matrix was computed. (This stepcorresponds to the last part of the last term of theCovariance Theorem, and it is labeled rxy,i4).) Fifth,after assessing levels of significance of these cor-relations, given the independence of the between-cell and within-cell scores, the between-unit cor-

relations were compared to the within-unit corre-lations to infer the appropriate unit of analysis. Aversion of the Z test was used for this last inferential

step.

Results

The results for the first two steps of this analy-tical procedure are shown in Table 1. This table

includes the measures’ means, standard deviations,and the RZ and for the one-way ANOVA that

was used to partial out supervisory group differ-ences.

The results of the third and fourth steps of theprocedure are shown in Table 2. This table includesthe raw individual correlations, the weightedbetween-unit correlations, and the residual within-unit correlations. For each of the relationships amongthe measures, three component correlations are dis-

played. The zero-order, total individual correla-tions are shown above the line for each relation-

ship. The degrees of freedom for any total correlationare based upon 107 individuals. For each relation-

ship, the two component correlations are shownbelow the line. The first correlation, shown on theleft, is based on weighted group average scores.The degrees of freedom for this correlation aredependent on the number of supervisory units, andin this case J = 30. The second shownon the right, is the residual, within-unit correlationafter the differences between supervisory units havebeen partialled out. The degrees of freedom for thiscorrelation are based on the residual term 1~1 - J.

Leadership Convergence

The total individual correlation between the su-

periors’ reports and the subordinates’ reports de-scribing the amount of leadership attention (Table2) ~aas ~ _ .15 (n.s). Given this result, a re-

searcher might conclude that there is no convergent

Table 1

WABA Computational Procedure for MatchedSuperior-Subordinate Correlated Reports

, ----- -

A _ - _ ~ -.

;’ _ . _ _ _ __ _ _ - _ _ _

* p < .05; ** p < .01.Note: The degrees of freedom for the zero-order, total individual

correlations are based on N = 107 ~ndwidu~ls9 those for the

between-unit correlations on J = 30 units, and those for theresidual within-unit correlations on N-J = 77. ©



67

validity. In traditional leadership research, how-ever, the component of variation of theoretical in-terest is based on supervisory units, not raw indi-vidual scores (Seeman, 1957; Sheridan &

Vredenburg, 1979). For superiors’ reports of lead-ership attention, a theoretical focus on groups seemsplausible because 68% of the variation in these

supervisory reports was attributable to differencesbetween the 30 supervisory groups. Furthermore,38% of the variance in subordinates’ reports oc-curred between groups for both the second and third

scale. These findings might result in an inferencethat the phenomenon being investigated is a between-unit phenomenon because a significant amount ofvariance occurs at that level of analysis.

Before such an inference can be drawn, how-

ever, there must be significant covariation betweensuperior and subordinate leadership descriptions atthe supervisory level of analysis. Because the cor-relation between the amount of leadership attentiongiven and the amount received based on unit av-erages is fb = . 12 (M. s.), the data do not supportsuch an inference. Contrary to traditional expec-tations, however, the within-unit correlation be-tween the two leadership reports was significantwith rw = .46. Thus, these data appear to describea situation in which there is convergence between

superiors and subordinates when describing indi-viduals who receive relatively higher amounts ofleadership attention when compared to just themembers of their own unit. Even though there wasa fair amount of variation between supervisory unitsfor both variables, significant covariation did notoccur based upon differences between groups.When the superiors’ reports of leadership atten-

tion given are compared with subordinates’ reportsof satisfaction with their supervisor, similar resultsare obtained. The total individual correlation wasr = .23. Removing the between-unit variation re-sulted in a within-unit correlation of ~ = .39. Thebetween-unit correlation was fb == .11(M.~.). ~.~~i~9the statistical imposition of supervisory units ap-peared to result in a nontraditional conclusion.

An Application of the Z Test

Because the WABA technique forces the si-multaneous consideration of two scores, a proce-

dure is needed to test the significance of the dif-ference between these two component correlations.Dansereau et al. (1982) have recommended the useof a standard Z test. Since the between- and within-

unit correlations are statistically independent~I~~;Ne~~r9 1955), it is possible to test the differ-ence between them using the standard f to Z for-mula for testing the difference between indepen-dent correlations (Guilford, 1965). When neithercomponent correlation can be considered error, the

equation can be written as

where Z’b is the transformed between-unit corre-lation and Z’ w is the transformed within-unit cor-relation. ~f ~’b or Z’w is considered error, a more

general form of this equation is

where Z’obs is either transformed between- or within-unit scores. In this case, these Z tests for the within-unit correlations have ~l -,~ - 3 degrees of free-dom and the between-unit correlations h~~e ~ - 3

degrees of freedom. Thus, the denominator in

Equation 3 is readily obtained.Table 3 shows the results of applying Equations

2 and 3 to the set of between- and within-unitcorrelations from Table 1. For each pair of cor-relations, three Z scores are shown. The Z scoresshown in the first and second columns of the tableare based upon Equation 3. The results in the firstcolumn assume that the between-unit correlation is

error. The second column’s results also use Equa-tion 3 and assume that the within-unit correlationis error. The last Z score shown in the last column,based on Equation 2, is a more conservative testand is used when there is no previous knowledgeor that either component correlation iserror.



68



69

Regardless of which assumption is made, all threeZ scores indicate that there are significant differ-ences between the component correlations for the

superiors’ and subordinates’ leadership attentionreports. Thus, the within-unit correlations arestronger than the between-unit correlations. A within-unit model is inferred.As shown in the first column of Table 3, the

difference between the within- and between-unitcorrelations of the superiors’ reports of leadershipattention and the subordinates’ reports of satisfac-tion is statistically significant only when an as-sumption is made that within-unit variation is noterror. As shown in the second and third columns,if this assumption is not made (and appropriatedegrees of freedom are used in the denominator ofthe Z tests), the differences between the two com-

ponent correlations are not statistically significant.The difference between the component corre-

lations for the subordinates’ descriptions of theamount of leadership attention they receive andtheir level of satisfaction follows the same patternas that of the correlations based on superiors’ lead-ership reports and subordinates’ satisfaction re-

ports. There is, however, a difference between thesetwo cases based on the results presented in Table2. In the case of the relationship between raters,the between-cell r was near zero. For the self-

reports, the between-cell r was greater than zero.Before considering the problem posed by these

self-reports, a summary of the implications of theZ test may be helpful. The Z test, a two-step pro-cedure, is used to discriminate between the be-tween-unit and within-unit correlations. First, com-ponent correlations must be tested for their levelof statistical significance. Second, the differencebetween the component correlations must also betested. When both steps have been accomplished,an inference is drawn by ruling out the variousalternatives. These alternatives can be listed as fol-lows :

1. A between-unit condition exists when r~ is sig-nificant, r, is null, and the difference betweenthem is significant;

2. A within-unit condition exists when rw is sig-nificant, ~°b is null, and the difference betweenthem is significant;

3. An ambiguous condition exists when both r,and rw are significant regardless of the differ-ence between them; and

4. The traditional null condition exists when nei-

ther r, nor r, is significant regardless of thedifference between them (see Dansereau et al.,1981, for a more detailed description).

The Problem with Self-ReportsThe correlations based only on subordinate self-

reports in Table 2 pose an interesting problem ofinterpretation because they fall in the third categorydescribed above. In this case, the subordinates’

reports of the amount of leadership attention re-ceived are correlated with their reports of satisfac-tion with their supervisor. The total individual cor-relation ®f ~ _ .63 was significant. Both componentcorrelations were also significant, thereby resultingin an ambiguous inference as indicated by the thirdcategory above. Although it would be tempting toconclude that both between-unit and within-unit

phenomena are occurring, there are at least threereasons not to draw this inference.

One possible explanation is that some supervi-sory units are homogeneous with respect to lead-ership attention, while others have a large amountof variation. If this were the case, then to makesense of the data one approach would be to definean appropriate structural boundary condition whichdivides the population of supervisory units into twocategories. In one hypothetical subsystem, data mightbe configured such that an rb = .46 and r, = .00.In a second subsystem, r, might be 00 and r, = .64.

In this example, by inadvertently combining thesehypothetical subsystems, an ambiguous conditioncould result. If this were true for the actual data

reported here, it would be inappropriate to inferthat both a between-unit phenomenon and a within-unit phenomenon are occurring in all of the su-pervisory groups.A second reason for assigning this result to an

ambiguous condition has to do with parsimony. Ifeach supervisory group is differentiated, and if thereis also differentiation between groups, then the useof supervisory groups as statistical cells has notcontributed any new knowledge. Thus, whatever



70

leadership processes are occurring are perhaps bet-ter understood as an individualized situation whichis independent of supervisory units. Rather thanconclude that both between-unit and within-unit

phenomena are occurring, it might be more par-simonious to suggest that neither is supported.The assumption underlying these two arguments

is that both component correlations of the subor-dinates’ self-reports in Table 2 are valid. This as-sumption also yields a third reason for interpretingthese results as indicating an ambiguous condition.Specifically, Eden and Leviatan (1975) have ar-gued that implicit theories of leadership may ac-count for some of the self-reported data in orga-nizational research. Thus, statistically significantcorrelations from self-reported data might not bevalid if they suffer from response bias and haveno correspondence to external referents.

The results in Table 2 for subordinate self-

reports could be interpreted in this manner. For theleadership-satisfaction relationship, both between-and within-cell correlations are significantly dif-ferent from zero. However, when the superiors’reports of leadership attention are compared withthe subordinates’ report of satisfaction, the between-unit correlation is no longer significant ( ~°b = .11,n.s.). Furthermore, the convergence correlation forvalidating the leadership reports seems to operateonly at the within-unit level of analysis. Thus, it

is possible that the between-unit correlation for thesubordinates’ self-reports, while statistically sig-nificant, may not be accurate because it cannot beconfirmed from the superiors’ perspective. In con-clusion, because any of these three possibilitiescannot be ruled out on the basis of the data pro-vided, these correlated self-reports will not be in-terpreted as representing both between- and within-unit phenomena.

Discussion

The research question posed earlier can now betentatively answered. The unit-of-analysis problemas represented by between-group and within-groupsources of variation has a major effect upon matchedsupervisor-subordinate leadership convergencecorrelations. If just the raw correlation based on

the total number of individuals were used, a nullinference would be drawn from the data (r = . .1 ~ 9~.s~.9 from Table 2). Despite the fact that a largeamount of variation is accounted for by differencesbetween units, these differences did not translateinto systematic covariation at the unit level for theseleadership descriptions (~ ~ &horbar;.12, n. s. ). The

imposition of supervisory groups as statistical cellsproduced a residual correlation (rv = .46) that was

significantly different from zero and significantlydifferent from its between-unit counterpart. Afterthe application of the Z test, an inference was drawnthat there appears to be valid convergence between

superiors and subordinates concerning the amountof leadership attention when a within-unit per-

spective is used. In traditional leadership research,this within-unit portion of variance is usually con-sidered error (Seeman, 1957).The results from applying the WABA procedure

also demonstrate that a leadership-satisfaction link-age might well fall into the same within-unit con-dition, depending on which a priori assumption ismade. However, the information derived from sub-ordinate self-reports in this study is ambiguous andcould not be interpreted. A distinct possibility ex-ists that either one of the component correlationsof the subordinate self-reports was not valid eventhough found to be statistically significant.The discovery of significant within-unit corre-

lations is contrary to the expectations of many stat-isticians (see Blalock, 1979). They have arguedthat the use of correlations based on group averageswill be larger than the correlation based on the totalnumber of individuals. The data presented heredemonstrate that this is not always true.The WABA procedure itself has direct impli-

cations for measurement in organizational re-

search. If a measure is believed to be a &dquo;group&dquo;construct, then empirical analyses must be viewedwith concern when they do not indicate both thatbetween-unit covariance for two variables is sig-nificant and that within-unit covariance is random.

Likewise, if a measure is believed to be a &dquo;dyadic&dquo; 9

construct, then the within-unit covariation should

be significant and the between-unit covariances shouldbe essentially random. This approach to the unit-of-analysis problem is not restricted to the lead-



71

ership area. Sirotnik (1980) has applied it to theissue of organizational climate in educational sys-tems. Theoretically, any issue involving variablessuch as satisfaction, participation, commitment, andabsenteeism is subject to these concerns if the database has multiple units of analysis that are clearlyidentified.The ~1A~A technique has some advantages over

the ANCOVA/regression technique. Both are formsof the General Linear Model and therefore resultin identical values for within-unit correlations.Between-unit correlations are similar except that~1A~~ makes an adjustment for the size of theunit. For example, in the calculation of the

between-unit correlations, WABA makes an ad-justment to the unit means by weighting each bythe size of the unit. (if unweighted unit averagesare entered into a standard least squares regressionmodel, no adjustment is made for the size of eachunit.) This is not a problem if all units are the samesize. In this data set, some units were larger thanothers, however, and in such a case, ordinary leastsquares is not the most efficient technique (Hannan& ~~rr~stei~, 1974, p. 378). When cells sizes arevery different, this problem becomes more pro-nounced. An additional comparative advantage ofWABA is that all results can be displayed in asingle table. More important, in WABA an attemptis made to clarify research assumptions about theinference-drawing process through the use of amodified Z test.

In general terms, WABA falls into a broad cat-egory of recently described techniques which areconcerned with separating individual from &dquo;con-

textual&dquo; or group effects. (Firebaugh, 1978; Han-nan & Burstein, 1974; James, Demaree, & Hater,1980; Kraemer, 1978; Li~c®ln ~ Zeitz, 1980). TheWABA technique is similar to the ANCOVA modelproposed by Lincoln and Zeitz (1980) with theexception that the WABA technique places equalemphasis on interpreting the within-unit effects asidefrom the between-unit effects. Because the Lincoln

and Zeitz (1980, p. 396) technique is primarilyconcerned with the between-unit regression lineand its decomposition into a predicted componentand a residual component for the purpose of testingthe between-unit effect, it can be categorized in

the macro end of the aggregation spectrum. On theother end of the spectrum, the James et al. (1980,p. 354) technique is oriented more toward the inter-pretation of individual effects by examining &dquo;thetotal possible variation in the person variable thatis associated with between-group differences.&dquo;The WABA technique attempts to establish a

middle ground in this micro-macro aggregation is-sue without a priori assumptions about the appro-priate unit of analysis. The ability to empiricallytest the assumptions concerning the appropriate unitof analysis of a construct or a relationship is a keyfeature of the WABA procedure. Thus, the ~1~~Atechnique can be viewed as an alternative to thetheoretical decision-making rule proposed by Fire-baugh (1978) for situations when aggregated dataare used to provide estimates of individual levelrelationship without committing the ecological fal-lacy described by Robinson (1950). Essentially,with WABA there is no need for a cross-level in-

ference from aggregate data to individual level the-

ory or vice versa because all units of analysis arepresent in the multiple data base, and are matchedand tested for the appropriate statistical unit.

References

Blalock, H. M. The presidential address: Measurementand conceptualization problems: The major obstacleto integrating theory and research, American Socio-logical Review, 1979,44, 881-894.

Dansereau, F., Alutto, J., Markham, S., & Dumas, M.Multiplexed supervision and leadership: An applica-tion of within and between analysis. In J. Hunt, U.Sekaran, & C. Schriesheim (Eds.), Leadership: Be-yond establishment views. Carbondale IL: SouthernIllinois University Press, 1982.

Dansereau, F., & Dumas, M. Pratfalls and pitfalls indrawing inferences about leader behavior in organi-zations. In J. Hunt and L. L. Larson (Eds.), Lead-ership: The cutting edge. Carbondale IL: SouthernIllinois University Press, 1977.

Dansereau, F., Graen, G., & Haga, W. A vertical dyadlinkage approach to leadership within formal organi-zations : A longitudinal investigation of the role mak-ing process. Organizational Behavior and HumanPerformance, 1975, 13, 46-78.

Eden, D., & Leviatan, U. Implicit leadership theory asa determinant of factor structure underlying supervi-sory behavior scales. Journal of Applied Psychology,1975, 60, 736-741.



72

Firebaugh, G. A rule for inferring individual-level re-lationships from aggregate data. American Sociolog-ical Review, 1978, 43, 557-572.

Graen, G., Dansereau, F., & Minami, T. Dysfunctionalleadership styles. Organizational Behavior and Hu-man Performance. 1972, 7, 216-236.

Guilford, J. P. Fundamental statistics in psychology andeducation. New York: McGraw-Hill, 1965.

Hannan, M. T., & Burstein, L. Estimation from groupedobservations. American Sociological Review, 1974,39, 374-392.

James, L. R., DeMaree, R. G., & Hater, J. J. A sta-tistical rationale for relating situational variables andindividual differences. Organization Behavior andHuman Performance, 1980, 25, 354-364.

Katerberg, R., & Horn, P. The effects of within-groupand between-groups variation in leadership. Journalof Applied Psychology, 1981, 66, 218-223.

Kimberly, J. R. Data aggregation in organization re-search : The temporal dimension. Organization Stud-ies, 1980, 1, 367-377.

Knapp, T. R. The unit-of-analysis problem in applica-tions of simple correlation analysis in educational re-search. Journal of Educational Statistics, 1977, 2,171-186.

Kraemer, H. C. Individual and ecological correlation ina general context. Behavior Science, 1978, 23, 67-72.

Lincoln, J., & Zeitz, G. Organizational properties fromaggregate data: Separating individual and structuraleffects. American Sociological Review, 1980, 45, 391-408.

McNemar, Q. Psychological statistics. New York: Wiley,1955.

Markham, S. Leadership and motivation: An empiricalexamination of exchange and its outcomes. Unpub-lished doctoral dissertation, State University of NewYork at Buffalo, 1978.

Przeworski, A., & Teune, H. The logic of comparativesocial inquiry. New York: Wiley & Sons, 1970.

Robinson, W. S. Ecological correlations and the behav-ior of individuals. American Sociological Review, 1950,15, 351-357.

Schriesheim, C. A., House, R., & Kerr, S. Leader in-itiating structure: A reconciliation of discrepant re-search results and some empirical tests. OrganizationBehavior and Human Performance, 1976, 15, 297-321.

Seeman, M. A comparison of general and specific leaderbehavior descriptions. In R. M. Stogdill & A. E. Coons

(Eds.), Leader behavior: Its descriptions and mea-surement. Columbus: Ohio State University, Bureauof Business Research, 1957.

Sheridan, J. E., & Vredenburg, D. J. A structural modelof leadership influence in a hospital organization.Academy of Management Journal, 1979, 22, 6-21.

Sirotnik, K. A. Psychometric implications of the unit-of-analysis problem (with examples from the mea-surement of organizational climate). Journal of Ed-ucational Measurement, 1980, 17, 245-282.

Stogdill, R. M. Handbook of leadership: A survey oftheory and research. New York: The Free Press, 1974.

Acknowledgments

The authors acknowledge the assistance and support ofthe Virginia Tech Productivity Research Center and theState University of New Y®rlz c~t Buffalo’s ResearchFoundation.

Author’s Address

Send request for reprints or further information to StevenE. Markham, Pamplin Hall/Management, Virginia Poly-technic Institute and State University, Blacksburg VA24061, U.S.A.



Steven E. Markham - Conservancy - University of Minnesota

Documents

Transcript of Steven E. Markham - Conservancy - University of Minnesota