Formative versus reflective measurement models:Two applications of
formative measurement ☆
Tim Coltman a,⁎,Timothy M.Devinney b,1,David F.Midgley c,2,Sunil Venaik d,3
a Centre for Business Services Science,University of Wollongong,Wollongong 2522,Australia b
Australian School of Business,The University of New South Wales,Sydney 2052,Australia c
Marketing Area,INSEAD,Boulevard de Constance,77305Fontainebleau,France d
UQ Business School,The University of Queensland,Brisbane 4072,Australia
Received 1May 2007;received in revid form 1November 2007;accepted 1January 2008
Abstract
This paper prents a framework that helps rearchers to design and validate both formative and reflective measurement models.The framework draws from the existing literature and includes both theoretical and empirical considerations.Two important examples,one from international business and
one from marketing,illustrate the u of the framework.Both examples concern constructs that are fundamental to theory-building in the disciplines,and constructs that most scholars measure reflectively.In contrast,applying the framework suggests that a formative measurement model may be more appropriate.The results reinforce the need for all rearchers to justify,both theoretically and empirically,their choice of measurement model.U of an incorrect measurement model undermines the content validity of constructs,misreprents the structural relationships between them,and ultimately lowers the ufulness of management theories for business rearchers and practitioners.The main contribution of this paper is to question the unthinking assumption of reflective measurement en in much of the business literature.
©2008Elvier Inc.All rights rerved.
Keywords:Formative;Reflective;International business;Integration-responsiveness;Marketing;Market orientation
1.Introduction
火的颜色Management scholars often identify structural relationships among latent,unobrved constructs by statistically relating covariation between the latent constructs and the obrved variables or indicator
s of the latent constructs (Borsboom et al.,2003,2004).This statistical covariation allows scholars to argue that if variation in an indicator X is associated with
variation in a latent construct Y ,then exogenous interventions that change Y can be detected in the indicator X.Most scholars assume that this relationship between construct and indicator is reflective.In other words,the change in X reflects the change in the latent construct Y .With reflective (or effect)measurement models,causality flows from the latent construct to the indicator.However,not all latent constructs are entities that are measurable with a battery of positively correlated items (Bollen and Lennox,1991;Edwards and Bagozzi,2000;Fornell,1982).A less common,but equally plausible approach is to combine a number of indicators to form a construct without any assumptions as to the patterns of intercorrelation between the items.A formative or causal index results (Blalock,1964;Diamantopoulos and Winklhofer,2001;Edwards and Bagozzi,2000)where causality flows in the opposite direction,from the indicator to the construct.Although the reflective view dominates the psychological and management sciences,the formative view is common in economics and sociology.
Available online at
Journal of Business Rearch 61(2008)1250–
1262
☆
The views here are solely tho of the authors,who appear in alphabetical order.The authors thank the anonymous reviewers and the editor of the special issue for their many helpful comments.
⁎Corresponding author.Tel.:+61242213912;fax:+61242214055.E-mail address:tcoltman@uow.edu.au (T.Coltman),
T.Devinney@agsm.edu.au (T.M.Devinney),david.midgley@inad.edu (D.F.Midgley),svenaik@business.uq.edu.au (S.Venaik).1
Tel.:+61293855671;fax:+61293855722.2
Tel.:+33160724977;fax:+33160745500.3
Tel.:+61733656841;fax:+61733656988.0148-2963/$-e front matter ©2008Elvier Inc.All rights rerved.doi:10.1016/j.jbusres.2008.01.013
The distinction between formative and reflective measures is important becau proper specification of a measurement model is necessary to assign meaningful relationships in the structural model(Anderson and Gerbing,1988).Theoretical work in construct validity(Blalock,1982;DeVellis,1991;Edwards and Bagozzi,2000)and structural equation modeling(Baumgartner and Homberg,1996;Chin and Todd,1995;Shook,Ketchen, Hult,and Kacmar,2004)enhances our understanding,however, considerable debate still exists regarding the procedures a working rearcher should follow to achieve construct validity (e.g.,Diamantopoulos,2005;Finn and Kayande,2005; Rossiter,2005).This paper is not to repeat or continue this debate.Rather,the authors take the middle ground,building on the work of both tho w
ho stress theoretical justifications for constructs and tho who argue for empirical validation as part of measure development.
汽修专修培训The paper prents an organizing framework for construct measurement that begins with theoretical justification to define the nature of the focal constructs,and then employs a ries of empirical tests to support the causal direction between constructs and their measures.The framework builds on the work of Jarvis et al.(2003)who provide a t of decision rules for deciding whether the measurement model should be formative or reflective.However,the framework here differs from Jarvis et al.'s decision rules in veral respects,most importantly in the specific procedures and the attention to measurement error.
The major contribution of this paper is to question the common assumption of a reflective measurement model en in much of the business literature.Applying the framework to two widely ud constructs in this literature is the vehicle for questioning this assumption.The two constructs are integration-responsiveness(from the discipline of international business) and market orientation(from the discipline of marketing).This choice of examples is important becau of:(1)the predomi-nance of the reflective modeling approach for the constructs, even though a formative model may be theoretically more appropriate,and(2)the criticality of the underlying phenomena to the
development of the two disciplines.
In the ca of the integration-responsiveness framework,the diver measures of each of the integration and responsiveness pressures are unlikely to intercorrelate highly as a reflective model requires.A priori,a formative approach to measurement would em worthy of consideration,yet most of the work in this area takes the reflective stance,often without any consideration of alternatives(Venaik et al.,2004).Similarly, most rearch on market orientation defines it as a one-dimensional construct which the rearcher measures through a multi-item reflective scale.Yet,the main scales that measure market orientation—MARKOR(Kohli and Jaworski,1990) and MORTN(Deshpande and Farley,1998)—are conceptua-lized as a t of activities that make up the attribute(e Narver and Slater,1990,p.21),implying a formative model.Further, the substantive inconsistencies in the market orientation literature(Langerak,2003)rai many questions about the dimensionality(Siguaw and Diamantopoulos,1995)and measurement(Narver et al.,2004)of the market orientation construct.The examples rve to illustrate a problem in the international business and marketing literature,which pay insufficient attention to measuring constructs.
The organization of the paper is as follows.Section2 prents the framework for designing and validating reflective and formative models using both theoretical and empirical considerations.Sectio
ns3and4then apply this framework to the two illustrative and important examples taken,respectively, from international business and marketing.The purpo here is to examine whether reflective or formative measurement models are more or less appropriate,not to debate the content validity of the measures that various scholars adopt.Finally, Section5provides discussion and conclusions.
2.An organizing framework for designing and validating reflective and formative models
In recent years,scholars have begun to challenge the blind adherence to Churchill's(1979)procedure with its strict emphasis on exploratory factor analysis(Spearman,1904), internal consistency(Cronbach,1951)and the domain sampling model(Nunnally and Bernstein,1994).In psychology, Borsboom et al.(2003,2004)u basic logic and measurement theory to argue that the choice of model is dependent upon the ontology the latent construct invokes.In marketing,Rossiter (2002)provides a general procedure for scale development which extends“accepted”practice by reemphasizing the importance of theoretical considerations.Borsboom and Rossiter both argue that scholars should focus only on theoretical considerations and resist the temptation to conduct empirical tests.
Alternatively,Diamantopoulos(2005)and Finn and Kayande(2005)argue that both theoretical and emp
irical criteria are necessary to design and validate measurement models.Empirical analys provide an important foundation for content validity,especially to detect errors and misspecifica-tions or wrongly conceived theories.For example,finding a negative relationship when theory and common n suggest a positive relationship should be a concern for rearchers.
This paper follows the stance of Diamantopoulos,and Finn and Kayande but takes a different perspective on empirical measurement and the role that measures play in the choice of a formative or reflective measurement model.To comprehen-sively capture the necessary theoretical and empirical aspects, the paper prents an organizing framework for designing and validating formative and reflective models(e Table1).As shown in the table,three theoretical considerations and three empirical considerations distinguish formative models from reflective ones.The following ctions briefly discuss each of the considerations.
2.1.Theoretical considerations
Three broad theoretical considerations are important in deciding whether the measurement model is formative or reflective.The considerations include:(1)the nature of the construct,(2)the direction of causality between the indicators
1251
T.Coltman et al./Journal of Business Rearch61(2008)1250–1262
and the latent construct,and(3)the characteristics of the indicators ud to measure the construct[numbering relates to the rows in Table1].
Consideration1:The nature of the construct.In a reflective model,the latent construct exists(in an absolute n) independent of the measures(Borsboom et al.,2004;Rossiter, 2002).Typical examples of reflective scenarios include measures of attitudes and personality.Practically all scales in business and related methodological texts on scale development (Bearden and Netmeyer,1999;Bruner et al.,2001;Netmeyer, Bearden,and Sharma,2003;Spector,1992)u a reflective approach to measurement.For example,examining papers in the Journal of International Business Studies and the Journal of Marketing for2006reveals that nearly95%of constructs with multiple items assume reflectivity without apparent considera-tion of an alternative formulation.
In contrast,in a formative model,the latent construct depends on a constructivist,operationalist or instrumentalist interpretation by the scholar(Borsboom et al.,2003).For example,the human development index(HDI)does not exist as an independent entity.Rather,it is a composite measure of
human development that includes:health,education and income(UNDP,2006).Any change in one or more of the components is likely to cau a change in a country's HDI score.In contrast to the reflective model,few examples of formative models are en in the business literature.
Consideration2:Direction of causality.The cond key theoretical consideration in deciding whether the measurement model is reflective or formative is the direction of causality between the construct and the indicators.As shown in Fig.1, reflective models assume that causality flows from the construct to the indicators.In the ca of formative models,the rever is the ca,causality flows from the indicators to the construct. Hence,in reflective models,a change in the construct caus a change in the indicators.In the ca of formative models,it is the other way around;a change in the indicators results in a change in the construct under study.Thus,the two models in Fig.1are different,both psychometrically and conceptually (Bollen and Lennox,1991).The difference in causal direction has profound implications both for measurement error(Dia-mantopoulos,2006)and model estimation;topics for Section 2.2.
Consideration3:Characteristics of indicators.Significant differences are prent in the characteristics of the indicators
Table1
A framework for asssing reflective and formative models:theoretical and empirical considerations
Considerations Reflective model Formative model Relevant literature Theoretical considerations
1.Nature of construct Latent construct exists Latent construct is formed Borsboom et al.(2003,2004)➢Latent construct exists independent of the
measures ud
➢Latent constructs is a combination of its
indicators
漆面修复2.Direction of causality between items and latent construct Causality from construct to items Causality from items to construct Bollen and Lennox(1991);
Edwards and Bagozzi(2000);
Rossiter(2002);Jarvis et al.
(2003)
➢Variation in the construct caus variation in the
item measures
➢Variation in the construct does not cau variation
in the item measures
➢Variation in item measures does not cau
variation in the construct
➢Variation in item measures caus variation in the
construct
3.Characteristics of items ud to measure the construct Items are manifested by the construct Items define the construct Rossiter(2002);Jarvis et al.
(2003)
➢Items share a common theme➢Items need not share a common theme
➢Items are interchangeable➢Items are not interchangeable
➢Adding or dropping an item does not change the
conceptual domain of the construct
➢Adding or dropping an item may change the
conceptual domain of the construct
Empirical considerations
4.Item intercorrelation Items should have high positive intercorrelations Items can have any pattern of intercorrelation but
should posss the same directional relationship
Cronbach(1951);Nunnally and
Bernstein(1994);Churchill
(1979);Diamantopoulos and
Siguaw(2006)
➢Empirical tests:asssing internal consistency
and reliability by Cronbach alpha,average variance
extracted,and factor ,from common
or confirmatory factor analysis)
➢Empirical test:no empirical asssment of indicator
reliability possible;various preliminary analys are
uful to check directionality between items and
construct
5.Item relationships with construct antecedents and conquences Items have similar sign and significance of
relationships with the antecedents/conquences as
the construct
Items may not have similar significance of
relationships with the antecedents/conquences as
the construct
沙滩蟹Bollen and Lennox(1991);
Diamantopoulos and Winklhofer
(2001);Diamantopoulos and
Siguaw(2006)
➢Empirical tests:establishing content validity by
theoretical considerations,asssing convergent
and discriminant validity empirically
➢Empirical tests:asssing nomological validity by
using a MIMIC model,and/or structural linkage with
another criterion variable
6.Measurement error and collinearity Identifying the error term in items is possible Identifying the error term is not possible if the
formative measurement model is estimated in isolation
Bollen and Ting(2000);
Diamantopoulos(2006)➢Empirical test:identifying and extracting
measurement error by common factor analysis
➢Empirical test:using the vanishing tetrad test to
determine if the formative items behave as predicted
➢Collinearity should be ruled out by standard
diagnostics such as the condition index
1252T.Coltman et al./Journal of Business Rearch61(2008)1250–1262
that measure the latent constructs under reflective and formative scenarios.In a reflective model,change in the latent variable must precede variation in the indicator(s).Thus,the indicators all share a common theme and are interchangeable.This interchangeability enables rearchers to measure the construct by sampling a few relevant indicators underlying the domain of the construct(Churchill,1979;Nunnally and Bernstein,1994). Inclusion or exclusion of one or more indicators from the domain does not materially alter the content validity of the construct.
However,the situation is different in the ca of formative models.Since the indicators define the construct,the domain of the construct is nsitive to the number and types of indicators the rearcher lects.Adding or removing an indicator can change the conceptual domain of the construct significantly. However,as Rossiter(2002)points out,this does not mean that rearchers need a census of indicators as Bollen and Lennox (1991)suggest.As long as the indicators conceptually reprent the domain of interest,they may be considered adequate from the standpoint of empirical prediction.
2.2.Empirical considerations
Paralleling the three theoretical considerations above,are three empirical considerations that inform understanding of the measurement model:(4)indicator intercorrelation,(5)indicator relationships with construct antecedents and conquences,and (6)measurement error and collinearity[numbering relates to the rows in Table1].
Consideration4:Indicator intercorrelation.In a reflective model,the underlying construct drives the indicators,which have positive and,desirably,high intercorrelations.In a formative model,the indicators do not necessarily share the same theme and hence have no preconceived pattern of intercorrelation.Indicators in a formative model can theoreti-cally posss no intercorrelation or high or low intercorrelation.
Regardless,rearchers should check that indicator inter-correlations are as they expect.Such checks are a necessary part of the various preliminary analys for questionnaire items which samples of respondents provide.The preliminary analys include checking for the prence of , using distances in factor spaces for reflective measurement models or regression influence diagnostics for formative models);checking that the dimensionality of the construct is consi
stent with a rearcher's ,using common factor models or principal components analysis);establishing that the correlations between items and constructs have the expected directionality and ,through bivariate correlations,factor or regression analysis);reliability statistics (only in the ca of the reflective measurement model);and, where veral constructs are part of a theoretical structure, showing that common method bias is not an ,by the abnce of one common factor).Some of the preliminary analys(and the diagnostics that go with them)shed uful light on issues of indicator intercorrelation and inferentially suggest whether the rearcher should prefer one measurement model or another.However,in themlves,they cannot either support or disconfirm theoretical expectations as to the nature of the measurement model.For that,rearchers require stronger tests.
Since reflective indicators have positive intercorrelations, rearchers can u statistics such as factor loading and communality,Cronbach alpha,average variance extracted and internal consistency to empirically asss the individual and composite reliabilities of their indicators(Trochim,2007). However,as the measures of reliability assume internal consistency—that is,high intercorrelations among the indica-tors in question—they are inappropriate for formative indica-tors,where no theoretical assumption is made about inter-item correlation.One of the key operational issues in the
u of formative indicators is that no simple,easy and universally accepted criteria exist for asssing their
均衡营养
玛咖的功效reliability. Fig.1.Reflective and formative measures.1253
T.Coltman et al./Journal of Business Rearch61(2008)1250–1262
Consideration5:Indicator relationships with construct antecedents and conquences.In the ca of r
eflective models, the indicators have a similar(positive/negative,significant/non-significant)relationship with the antecedents and conquences of the construct.The requirement for interrelated indicators is not the ca for formative indicators as they do not necessarily share a common theme and,therefore,do not have the same types of linkages with the antecedents and conquences of the construct.This lack of a common theme is a significant issue when using formative models,particularly as it has implications for the appropriate level of aggregation of formative indicators. While aggregating indicators to create a construct achieves the objective of model parsimony,it may come at a significant cost in terms of the loss of the rich,diver and unique information the individual indicators provide.Edwards(2001)makes a similar point for cond and higher order dimensions.
In the ca of formative measurement,Diamantopoulos and Winklhofer(2001)suggest three possible approaches.First,the rearcher can relate the indicators to some simple overall index variable,such as a summary or overall rating—this is the approach the cond example in this paper takes(market orientation).Second,the rearcher can apply a Multiple Indicators and Multiple Caus(MIMIC)model,where both formative and reflective indicators measure the construct.Third, the rearcher applies a structural model linking the formatively measured construct with another refl
ectively measured con-struct to which it relates theoretically.This approach establishes criterion and nomological validity,and is the approach the first example in this paper takes(integration-responsiveness pressures).
Consideration6:Measurement error and collinearity.A key difference between formative and reflective measurement models is the treatment of measurement error.As Fig.1shows,an important assumption underlying the reflective measurement model is that all the error terms(δi of Fig.1)associate with the obrved scores(x i)and,therefore,reprent measurement error in the latent variable.The formative measurement model does not assume such a correlational structure.For the formative ca the disturbance term(ζ)neither associates with the individual indicator,nor the t of indicators as a whole.This term therefore does not reprent measurement error(Diamantopoulos,2006).
In the ca of reflective measurement models,rearchers can identify and eliminate measurement error for each indicator using common factor analysis.This elimination occurs becau the factor score contains only that part of the indicator that is shared with other indicators,and excludes the errors in the underlying items(Spearman,1904).However,in the ca of formative models,the only way to overcome measurement error is to design it out of the study before collecting the data.Diama
ntopoulos(2006)suggests two possible ways to eliminate the error term:(1)capture all possible caus of the construct,and(2)specify the focal construct in such a way as to capture the full t of indicators. Both approaches legitimately exclude the error term(ζ=0).In the light of the above,it is clear that unlike the reflective model,no simple way exists to empirically asss the impact of measurement error in a formative model.
However,Bollen and Ting(2000)suggest that the tetrad test can provide some assistance in asssing measurement error.A tetrad refers to the difference between the products of two pairs of error covariances(Spearman and Holzinger,1924).The tetrad test involves examining the nested vanishing tetrads that a comparison of the two different measurement models implies. In the ca of a reflective model,the null hypothesis is that the t of non-overlapping tetrads vanishes.In simpler terms,when comparing the intercorrelations between pairs of errors,they should tend to zero.Referring back to Fig.1,the assumption underlying the reflective model is that the correlations between theδi are zero.The tetrad test confirms whether or not this is true.If not,the rearcher may wish to consider a formative measurement model.调到敌岛打特盗
The tetrad test is a confirmatory procedure and not for u as a stand-alone criterion for distinguishing formative from reflective models.Specifically,if the test rejects the hypothesis that the e
rrors are uncorrelated,it can be for one of two alternative reasons.One is that the construct is better measured formatively,not reflectively.The other is that reflective measurement is more appropriate but the error structure is contaminated.One possible source of contamination is common method error.Similarly,if the rearcher accepts the hypothesis that the errors are uncorrelated,this could still be a mistake.A possibility,although unlikely in practice,is that a formative model is correct but that the indicator error structures are uncorrelated.Thus,while rving as an important pointer,the results from the tetrad test do not provide definitive proof as to the correct measurement model.
Another measurement issue that rearchers need to check in formative models is collinearity.The prence of highly correlated indicators will make estimation of their weights in the formative model difficult and result in impreci values for the weights.Given a criterion variable,as above,an estimate of the impact of collinearity can be made by regressing the indicators on this variable and computing standard diagnostics such as the condition index.
回家的路作文
The next two ctions apply the three ts of theoretical criteria and three ts of empirical criteria to two key constructs in international business and marketing,integration-responsive-ness and marketing orientation.
3.Application one:measuring international
business pressures
The international business literature makes extensive u of the Integration Responsiveness(IR)framework of Prahalad and Doz(1987)to characterize the environmental pressures confronting firms as they expand worldwide.According to this framework,firms come under countervailing pressures to simultaneously coordinate the activities and strategies of their local business units to attain global competitive advantage (global integration)while adapting the activities and strate-gies to the unique circumstances of the countries in which they operate(local responsiveness).
Although this framework has been in u for two decades, the issue of relevance here is whether the formative or reflective
1254T.Coltman et al./Journal of Business Rearch61(2008)1250–1262
